Dynamic Matrix of Extremisms and Terrorism (DMET): A Continuum Approach Towards Identifying Different Degrees of Extremisms

Abstract

We propose to extend the current binary understanding of terrorism (vs. non-terrorism) with a Dynamic Matrix of Extremisms and Terrorism (DMET). DMET considers the whole ecosystem of content and actors that can contribute to a continuum of extremism (e.g., right-wing, left-wing, religious, separatist, single-issue). It organizes levels of extremisms by varying degrees of ideological engagement and the presence of violence (e.g., partisan, fringe, violent extremism, terrorism) identified based on cognitive and behavioral cues and group dynamics. DMET is globally applicable due to its comprehensive conceptualization of the levels of extremisms. It is also dynamic, enabling iterative mapping with the region- and time-specific classifications of extremist actors. Once global actors recognize DMET types and their distinct characteristics, they can comprehensively analyze the profiles of extremist actors (e.g., individuals, groups, movements), track these respective actors and their activities (e.g., social media content) over time, and launch targeted counter activities, whether in the form of deplatforming, content moderation or redirects to targeted CVE narratives.

Full text

July 2021
Broadening the GIFCT Hash-Sharing
Database Taxonomy: An Assessment
and Recommended Next Steps

Acknowledgements
The Global Internet Forum to Counter Terrorism
would like to thank all of our stakeholders and
partners for their interest in and commitment
to this body of work. We are especially grateful
to our GIFCT member companies, members of
our Independent Advisory Committee (IAC),
and civil society representatives for their
feedback throughout this process, as well as
to the GIFCT interns—Maggie Frankel, Aaron
Tielemans, and Ye Bin Won—who provided
significant support in the production of this
series. GIFCT is also enormously appreciative
of the time, expertise, and innovative thinking
that all of the authors of the subsequent papers
contributed to this project.

3
Table of Contents
Introduction
Nicholas J. Rasmussen and Johannah Lowin 4
Background:
A Guide to the Taxonomy of GIFCT’s Hash-Sharing Database
GIFCT Staff
9
Practical and Technical Considerations in Expanding the GIFCT
Hash-Sharing Database
Erin Saltman and Tom Thorley
12
Expanding the Hash-Sharing Database
Daniel Byman and Chris Meserole 25
Dynamic Matrix of Extremisms and Terrorism (DMET):
A Continuum Approach Towards Identifying Different Degrees of Extremisms
Marten Risius, Kevin M. Blasiak, Susilo Wibisono, Rita Jabri-Markwell, and Winnifred Louis
42
A Taxonomy for the Classification of Post-Organizational
Violent Extremist and Terrorist Content
Jacob Davey, Milo Comerford, Jakob Guhl, Will Baldet, and Chloe Colliver
78
A Practical Taxonomy for Online Terrorist Content
William Braniff, Matthew Feldman, Eviane Leidig, Adam Hadley, Ghayda Hassan, Ray Serrato 102
Taxonomy Expansion and the Global Terrorism Database:
Effectively Leveraging Academic Data Collection Initiatives
Erin Miller
116
Defining and Classifying Terrorist Content Online:
Leveraging National Countering Violent Extremism Strategies and Action Plans
Sara Zeiger, Farangiz Atamuradova, Denis Suljic, and Petra Regeni
128
Conclusion & Initial Next Steps
GIFCT Staff 150
Appendix:
Definitions of Terrorism & Violent Extremism
Nayanka Paquete Perdigao, Sarah Kenny, and Aaron Tielemans
154

By Marten Risius, Kevin M. Blasiak, Susilo Wibisono, Rita Jabri-Markwell, and
Winnifred Louis
Dynamic Matrix of Extremisms and
Terrorism (DMET):
A Continuum Approach Towards Identifying Different Degrees
of Extremisms

Broadening the GIFCT Hash-Sharing Database Taxonomy: An Assessment and Recommended Next Steps
43
Dynamic Matrix of Extremisms and Terrorism (DMET)
A Continuum Approach Towards Identifying Different Degrees of
Extremisms
By Marten Risius, Kevin M. Blasiak, Susilo Wibisono, Rita Jabri-Markwell and Winnifred Louis
Abstract
We propose to extend the current binary understanding of terrorism (versus non-terrorism)
with a Dynamic Matrix of Extremisms and Terrorism (DMET). DMET considers the whole
ecosystem of content and actors that can contribute to a continuum of extremism (e.g., right-
wing, left-wing, religious, separatist, single-issue). It organizes levels of extremisms by
varying degrees of ideological engagement and the presence of violence identified (e.g.,
partisan, fringe, violent extremism, terrorism) based on cognitive and behavioral cues and
group dynamics. DMET is globally applicable due to its comprehensive conceptualization
of the levels of extremisms. It is also dynamic, enabling iterative mapping with the region-
and time-specific classifications of extremist actors. Once global actors recognize DMET
types and their distinct characteristics, they can comprehensively analyze the profiles of
extremist actors (e.g., individuals, groups, movements), track these respective actors and
their activities (e.g., social media content) over time, and launch targeted counter activities
(e.g. de-platforming, content moderation, or redirects to targeted CVE narratives).
Key recommendations
1. Understand extremism as a dimensional concept with terrorism as a deviant pole
from the regional norm.
2. Transparently map cues of extremism to accountably define categories of
extremisms, creating an opportunity for dialogue between scholars and industry,
and increasing trust with civil society and broader public audiences.
3. Provide the opportunity for organizations to upweigh or downweigh cues
of extremism based on their local norms or national or international legal
requirements, while changing the classification in an explainable and
transparent way.
4. Iteratively update the manifestations of cues that characterize extremisms to
account for changing profiles of extremism regionally or temporally.
5. Recognize all forms of violence used by violent extremists, especially serial or
systematic dehumanization of an out-group as an attribute and indicator of
violent extremism.
6. Enable platform providers to transparently decide and explain decisions to
exempt extremist actors or content from DMET.
7. Create a more nuanced understanding of the degrees of extremism to reduce the
probability of misclassifications (i.e., of non-terrorists as terrorists, or failing to
identify terrorists as such) and allow more fine-grained analysis of actor changes
over time.

44
Introduction
Moderation of terrorists online is commonly achieved by referring to lists of known terrorist
individuals and groups from academia, civil society, and governments.51 These lists are
very helpful for the differentiation between terrorist and non-terrorist content. Currently,
the GIFCT hash-sharing database is an important tool for the list-based moderation of
terrorist content online.
There are, however, certain issues that accompany these list-based approaches that
remain to be solved. As summarized by the recent report from the Royal United Services
Institute, no single type of list simultaneously can fulfill all of the following three criteria
while still being economically feasible:52
1. Ideological fairness: equal opportunity for all entities to be classified as terrorist;
2. Global applicability: transcend regional borders; and
3. Update frequency: near real-time updates.
Furthermore, there is a gray area that poses various noteworthy delicate challenges.
Terrorists often communicate non-violent content that falls outside the categories of
Imminent Credible Threat, Graphic Violence Against Defenseless People, Glorification
of Terror Acts, Recruitment & Instruction. These more subtle messages still help to further
extremist causes when they are not directly captured by the hash database taxonomy.
They use social media for fundraising purposes53 or to affirm grievances, ideologies, and
share humanitarian purposes (e.g., pictures of ISIS-affiliated doctors helping injured
children) without calling for violence.54 Furthermore, perceived overlap in content between
violent extremists and (political) partisan actors, self-determination-based movements,
or state-sponsored information campaigns raise legal and ethical questions in regard
to appropriate treatment. Accordingly, the decision to add an actor and their content
to a list of known terrorists in order to moderate their online presence sets a high bar
that allows for considerable damage to occur beforehand, is associated with a strong
stigmatization transforming the decision into a political issue and makes a revision of
the decision following resocialization efforts unlikely. A transparent framework like the
proposed Dynamic Matrix of Extremisms and Terrorism (DMET) is needed to respond to
these challenges.
51 Chris Meserole and Daniel Byman, “Terrorist Definitions and Designations Lists: What Technology Companies Need
to Know,” Royal United Services Institute for Defence and Security Studies, (July 2019), https://rusi.org/explore-our-
research/publications/special-resources/terrorist-definitions-and-designations-lists-what-technology-companies-need-
to-know.
52 Meserole and Byman, “Terrorist Definitions and Designations Lists,” 2.
53 Tom Keatinge and Florence Keen, “Social Media and (Counter) Terrorist Finance: A Fund-Raising and Disruption Tool,”
Studies in Conflict & Terrorism 42, no. 1-2 (2019); Tom Keatinge, Florence Keen, and Kayla Izenman, “Fundraising for
Right-Wing Extremist Movements,” The RUSI Journal 164, no. 2 (2019).
54 Roderick Graham, “Inter-Ideological Mingling: White Extremist Ideology Entering the Mainstream on Twitter,”
Sociological Spectrum 36, no. 1 (2016).

Broadening the GIFCT Hash-Sharing Database Taxonomy: An Assessment and Recommended Next Steps
45
Background
The Dynamic Matrix of Extremisms and Terrorism (DMET)
We propose extending the binary understanding of terrorism (versus non-terrorism) with
a Dynamic Matrix of Extremisms and Terrorism (DMET) to address the intersection of
extremism types (and associated extremist content). DMET identifies varying degrees of
ideological engagement and violence based on cognitive and behavioral cues as well as
group dynamics. 55
DMET’s Understanding of Extremism
DMET understands extremism on a continuum of varying degrees of ideological engagement.
Any label of “extremism” assigned in reference to DMET needs to share the spirit of the
following assumptions that accompany DMET’s continuum-based understanding.
First, the associated types of extremisms are based on an understanding of online
extremism as a deviation from something that is commonly considered (more) “ordinary,”
“mainstream,” or “normal.”56 DMET declines an evaluative notion of the purposes and
goals of the different forms of extremism.
Second, the levels of ideological engagement are derived from an understanding
of radicalization as the “change in beliefs, feelings, and behaviors in directions that
increasingly justify intergroup violence and demand sacrifice in defense of the group.”57
Consequently, we focus on cognitive, behavioral, and group dynamic cues to describe the
different levels of ideological engagement. These general descriptions then need to be
operationalized respective to the regional and temporal context.
Third, DMET emphasizes the plurality of extremisms to underscore our assumption that
extremism is a concept of varying degrees and deviation from regionally dominant
ideologies. We emphasize this to avoid stigmatization of minorities as “extremists” for
proposing views that deviate from the regional majority (e.g., Radical Veganism). Higher
55 Peter R. Neumann, “The Trouble with Radicalization,” International Affairs 89, no. 4 (2013); Kris Christmann,
“Preventing Religious Radicalisation and Violent Extremism: A Systematic Review of the Research Evidence,” Youth
Justice Board (2012); Susilo Wibisono, Winnifred R Louis, and Jolanda Jetten, “A Multidimensional Analysis of Religious
Extremism,” Frontiers in Psychology, 10 (2019).
56 Alex P. Schmid, “Violent and Non-Violent Extremism: Two Sides of the Same Coin,” International Centre for Counter-
Terrorism (ICCT) Research Paper (2014); Charlie Winter et al., “Online Extremism: Research Trends in Internet Activism,
Radicalization, and Counter-Strategies,” International Journal of Conflict and Violence (IJCV) 14, no. 2 (2020): 4; Ronald
Wintrobe, Rational Extremism: The Political Economy of Radicalism, Cambridge, New York, Cambridge University Press,
(2006).
57 Clark McCauley and Sophia Moskalenko, “Mechanisms of Political Radicalization: Pathways toward Terrorism,”
Terrorism and Political Violence 20, no. 3 (2008); Winter et al., “Online Extremism.”

46
levels of ideological engagement in different forms of extremisms are characterized by
more uncommon cognitive, behavioral, and group dynamic cues.
Fourth, we consider the matrix to be dynamic to acknowledge both that groups change and
that the understanding of “normal” is variable across geography and time. For example, a
group that opposes vaccination may be viewed as fringe or extremist (i.e., non-normative)
in certain parts of the world at a particular time, but regarded as normal in other parts
or at other times.58 Hence, assessments of ideological engagement and the underlying
forms of operationalizations are regionally delimited, time-specific, and require regular
updates.
Level Defining Cues
DMET distinguishes between four levels of ideological engagement: partisan, fringe,
violent extremist, and terrorist. DMET’s continuum approach adopts the idea of ordering
degrees of violent extremism59 to extend the simplified categories of terrorism and non-
terrorism. In DMET’s case, we start from a point at the normative or moderate baseline
and move through to increasing degrees of alienation from the mainstream to ultimately
active violent acts. Each level of ideological engagement is proposed to have a particularly
prevalent configuration of cues to identify and classify a group or content (i.e., from
partisanship to terrorism). These cues are cognitive, behavioral, and group dynamic.
A discussion of strategic and technical implementation considerations can be found in
sections 4.2 and 4.3.
Cognitive Cues in DMET
Cognitive cues are signals indicating the thoughts and attitudes of individuals or groups.60
At the individual level, cognitive cues might emerge in the form of thoughts or images. At
the collective or group level, cognitive cues are shared beliefs or representations involved
in recognizing and perceiving ourselves and other individuals or social categories.
Many outcomes can flow from these socio-cognitive processes, such as prejudice and
stereotypes.61 For the purpose of classifying content as extremist or not, key aspects
tracked by DMET cognitively would include beliefs about or representations of one’s own
groups (ingroups) and their actions, the targeted opponent groups (out-groups) and their
actions, the nature of right or wrong, and the nature of the threats or value differences
that define the relationship between the groups.
58 Ayodele Samuel Jegede, “What Led to the Nigerian Boycott of the Polio Vaccination Campaign?,” PLoS Medicine 4,
no. 3 (2007).
59 For an example, see Donald Holbrook, “Designing and Applying an ‘Extremist Media Index,”” Perspectives on
Terrorism 9, no. 5 (2015).
60 Bert N. Bakker, Yphtach Lelkes, and Ariel Malka, “Understanding Partisan Cue Receptivity: Tests of Predictions from the
Bounded Rationality and Expressive Utility Perspectives,” The Journal of Politics 82, no. 3 (2020).
61 M. Verkuyten and A. De Wolf, “The Development of in-Group Favoritism: Between Social Reality and Group Identity,”
Developmental Psychology 43, no. 4 (2007).

Broadening the GIFCT Hash-Sharing Database Taxonomy: An Assessment and Recommended Next Steps
47
Behavioral Cues in DMET
Behavioral cues refer to the observable actions by groups or individuals or representations
of those actions. At an individual level, behavioral cues can be observed from (for example)
facial expression, gesture, vocal expression, etc.62 For the purposes of DMET, these cues
might indirectly identify (for example) the emotional level (e.g., anger) that a person has
in one situation, or may explicitly show harm-doing and calls to violence. Behavioral cues
can be addressed to the self or others. At a group level, drawing on the literature on
political contestation, collective action,63 and intergroup violence. For the purposes of
DMET, we are most interested in coding for content that involves a call to cooperate with
in-group or prospective allies or to engage in concrete actions that derogate or harm
another group.
Group Dynamics in DMET
The process of radicalization or increasing extremism often draws on group dynamics by
establishing norms about appropriate and deviant behaviors, with very little latitude in
accepting differences.64 People may be drawn to identify with causes or groups based on
broad in-/out-group dynamics, as intergroup threats and conflicts of interest or values are
contested using a range of tactics, from debate and satire to threats, dehumanization, and
violence.65 The dynamic influence of group identities and norms can provide ideological
glue for (de)radicalization across the extremist spectrum.66
Group dynamics refers to a system of behaviors and psychological processes occurring
within or between social groups.67 Intragroup dynamics (i.e., how individuals in a group
interact with one another) underlie social processes that give rise to a set of norms, roles,
relations, and common goals characterizing a particular group.68 Group dynamics can
also involve the cooperation or competition of individuals within the groups to gain
group recognition or act on behalf of the group. In addition, intergroup dynamics (i.e.,
how groups interact with each other) include collective perception, attitudes, and actions
62 Alessandro Vinciarelli et al., “Social Signal Processing: State-of-the-Art and Future Perspectives of an Emerging
Domain,” in Proceedings of the 16th ACM international conference on Multimedia (Vancouver, British Columbia, Canada:
Association for Computing Machinery, 2008).
63 Defined as any action aimed to improve the group’s status; see M. van Zomeren, T. Postmes, and R. Spears, “Toward
an Integrative Social Identity Model of Collective Action: A Quantitative Research Synthesis of Three Socio-Psychological
Perspectives,” Psychological Bulletin 134, no. 4 (2008); S. C. Wright, D. M. Taylor, and F. M. Moghaddam, “Responding
to Membership in a Disadvantaged Group - from Acceptance to Collective Protest,” Journal of Personality and Social
Psychology 58, no. 6 (1990).
64 Wibisono, Louis, and Jetten, “A Multidimensional Analysis of Religious Extremism.”
65 C. Stott, P. Hutchison, and J. Drury, “‘Hooligans’ Abroad? Inter-Group Dynamics, Social Identity and Participation in
Collective ‘Disorder’ at the 1998 World Cup Finals,” British Journal of Social Psychology, 40 (2001).
66 John M. Berger, “Deconstruction of Identity Concepts in Islamic State Propaganda: A Linkage-Based Approach to
Counter-Terrorism Strategic Communications,” The Hague, Netherlands: EUROPOL, (2017); Donald Holbrook, “Far Right
and Islamist Extremist Discourses: Shifting Patterns of Enmity,” Extreme Right Wing Political Violence and Terrorism (2013).
67 M. A. Hogg and D. J. Terry, “Social Identity and Self-Categorization Processes in Organizational Contexts,” Academy
of Management Review 25, no. 1 (2000); J. Sidanius et al., “Ethnic Enclaves and the Dynamics of Social Identity on the
College Campus: The Good, the Bad, and the Ugly,” Journal of Personality and Social Psychology 87, no. 1 (2004).
68 M. A. Hogg and S. A. Reid, “Social Identity, Self-Categorization, and the Communication of Group Norms,”
Communication Theory 16, no. 1 (2006).

48
toward other groups.69
Intra- and intergroup dynamics can produce and be shaped by specific behavioral and
cognitive cues. However, the DMET coding here refers to specific attributes of how content
is being disseminated relationally (e.g., conformity, polarization), and how sources are
positioning themselves in relation to other groups (e.g., as leaders, warriors) and groups in
relation to each other (e.g., as enemies, allies, dupes). Source attributes where available
would be coded in Group Dynamics, both in terms of membership in particular groups and
of position within particular networks (e.g., contact with a known violent actor).
Types of Ideological Engagement
Crossed with these levels in DMET (see Figure 1), we consider five categories of actors/
content according to their ideological arena: Right-Wing (e.g., concerning threats
to the “white race” or “traditional values”), Left-Wing (e.g., concerning the need for a
fair distribution of wealth), Religious (e.g., seeking to spread one’s religion or purify it),
Separatist (e.g., seeking territory for one’s group), and Single-issue (e.g., advocating for
one particular topic such as abortion or animal justice).70 A group may be classified into
more than one type of ideology, as it advocates for an issue by drawing narratively on
other content (e.g., both right-wing ideology and religion). The purpose of the categories
is to a) signal the inclusivity of DMET with all groups equally able to be considered as
violent actors or terrorists; and b) build an understanding of how clusters of particular
indicators or attributes emerge in different causes, resulting in profiles of domain-specific
indicators feeding into context-specific categorization algorithms.
69 “Hooligans’ Abroad?.”
70 Allard R. Feddes et al., Psychological Perspectives of Radicalization (London: Routledge, 2020).

Broadening the GIFCT Hash-Sharing Database Taxonomy: An Assessment and Recommended Next Steps
49
The Continuum of Ideological Engagement
A core premise of DMET is its understanding of ideological engagement as a spectrum
of varying degrees of severity (Figure 2) instead of the current binary dichotomy of (non)
terrorism (Figure 1).
Figure 1. Conceptualization of Current Dichotomous Understanding of Ideological Engagement
This continuum perspective enables DMET to distinguish among different levels of
ideological engagement (i.e., partisanship, fringe, violent extremism, terrorism) and the
regular population norms that define the regionally accepted social standard. Thereby,
the aim is to enable platform providers to make independent content or actor moderation
decisions in a more nuanced fashion (including determining and disclosing cut-off values),
with fewer misclassification errors between regular content and terrorist material. DMET
also enables greater transparency regarding moderation decisions (e.g., by providing a
means not only to classify organizations among the dimensions but also to develop and
explain the weighing of attributes in the decision-making algorithms).

50
Figure 2. Conceptualization of DMET’s Proposed Level’s of Extremism Based on the Assumed Continuum of
Ideological Engagement
It needs to be noted that the conceptualization of the different levels of ideological
engagement as normally distributed sub-groups is a proposition that needs to be
empirically tested following DMET’s operationalization criteria outlined in the following.
We also emphasize that this is meant as a conceptual illustration of the continuum of
extremism and that the proportions of the more extreme sub-populations are likely
smaller in reality.71
71 Dirk Oegema and Bert Klandermans, “Why Social Movement Sympathizers Don’t Participate: Erosion and
Nonconversion of Support,” American Sociological Review 59, no. 5 (1994).

Levels of
Ideological
Engagement
Level Defining Cues Types of Ideological Engagement
Cognitive Behavioral Group Dynamic Right-Wing Left-Wing Religious Separatist Single-Issue
Terrorism
Level 3
Terrorism
Sidestep inhibitory
mechanisms, perceive
target as ‘the enemy,”
legitimizing and valorizing
death, wanting to intimidate
broader population
Endorse, promote, or
enact physical violence
towards out-, in-group or
infrastructure
Propagate values of active
martyrdom, divide group
labor to support violent
acts
Ultra-right,
far-right, alt-
right, right-wing
extremism,
fascism, white
supremacy
Ultra-left, far-
left, left-wing
extremism
Religiously
motivated terrorism
Violent militant
separatist
organizations
Violent
militant
activism
Violent Extremism
Level 2
Violent
Extremism
Be intolerant towards
others, represent cultural &
structural violence through
silencing and exclusion,
perceive a reduced level of
moral duties owed to the
out-group
Serially or systematically
dehumanize others,
frequently express hate
speech towards opponents,
perform selective/
individual acts of violence,
actively separate targets
from society, active
discrimination
Compete for within-group
recognition, show personal
agency in the service
of group domination,
coalescing around out-
group as a perceived or
designated existential
threat
Radical right,
extreme
conservatism
Radical left Religiously
motivated
extremism
Secessionism,
autonomism
Propaganda
groups
Non-Violent Extremism
Level 1
Fringe Group
Perceive/glorify the
in-group as superior,
indoctrinate dogmatic
values, prejudice, and
discrimination
Discredit or denigrate the
out-group, seek isolation
from the general public,
express external blame for
negative events, censor
deviant views
Pursue and promote norms
of purity, supremacy,
domination, or revenge
Right-wing
nationalist
Left-wing
nationalist
Religious
fundamentalism,
cult
Seeking self-
determination
Conspiracy
theorists,
fringe party
advocating
single-issue
Non-Extremism
Level 0
Partisanship
Holding polarized and
normative views, self-
identifying with one group in
opposition to another group
Expressing populist
ideology, dog-whistling,
satirizing other views,
evangelizing others,
campaigning peacefully
Holding political
grievances, experiencing
a sense of victimization or
identity crises, or a need
for significance
Right-wing
populism
Left-wing
populism,
liberalism,
socialism
Religious
conservatism
Regional
advocacy
groups
Special-
interest
advocacy
groups,
lobbyism
Please note: The table is based on a value-free understanding of extremism as something that is significantly deviant from the ‘mainstream’ or ‘normal’; dynamic classifications of content, individuals or
groups are dependent on the understanding of what is ‘normal’ in a particular region at a given point in time.
Table 1. Dynamix Matrix of Extremisms and Terrorism (DMET)

52
Operationalization of DMET
In the following, we describe DMET’s operationalization of the proposed levels of
ideological engagement based on cognitive, behavioral, and group dynamic cues. We
acknowledge the wealth of literature discussing comprehensive definitions of these
concepts. For the sake of this briefing paper, we limit ourselves to deriving working
definitions that explain DMET-based actor classifications.
We identify an indicative basket of indicator attributes for each level. However, part of
our approach is that the association of any one attribute with a level or ideological cause
(e.g., right-wing extremism) is dynamic and may change over time, so indicators wax or
wane in their diagnostic value in historical periods or for particular contexts. Regional
expert feedback to set the starting parameters and automated updating of the model
over time will be important in sustaining DMET’s accuracy.
Level 0: Partisanship
Partisanship constitutes a non-extremist form of coordinated ideological engagement
where individuals are committed to similar normative ideas and face conditions of
conflict opposing others with whom they are at odds.72 Partisans distinguish themselves
from mainstream views through their normative ideology and offer a support network
where collective actors are empowered to contest perceived grievances. The partisan
commitment also serves as a source of identification and shapes the individuals’ self-
concept,73 which leads to the continued endorsement of the mission that the group
embodies and sustains the long-term pursuit of projects across a range of conditions and
circumstances. Partisanship is not limited to in-group conformity but also is associated
with perceived polarization away from a rival or opponent out-group.74
At a behavioral level, partisans commit to enacting a form of regulated adversarialism,
which describes their commitment to persuade and evangelize others of their views
tempered by self-set rules or ideals.75 They may pursue different strategies, such as
expressing populist ideologies as an opposition force or from a position of power.76
Traditionally partisan actions take place through conventions, meetings, assemblies, and
peaceful protests complemented by the online sphere via websites, blogs, and social
72 Jonathan White and Lea Ypi, The Meaning of Partisanship, Oxford, Oxford University Press, (2016).
73 Emily A. West and Shanto Iyengar, “Partisanship as a Social Identity: Implications for Polarization,” Political Behavior
(2020).
74 Noam Lupu, “Party Polarization and Mass Partisanship: A Comparative Perspective,” Political Behavior 37, no. 2
(2015).
75 White and Ypi, The Meaning of Partisanship.
76 S. Erdem Aytaç, Ali Çarkoğlu, and Ezgi Elçi, “Partisanship, Elite Messages, and Support for Populism in Power,”
European Political Science Review 13, no. 1 (2021).

Broadening the GIFCT Hash-Sharing Database Taxonomy: An Assessment and Recommended Next Steps
53
media. Calls for action against the out-group are often indirect, using appeals that subtly
invoke negative stereotypes about an opposing group (e.g., dog-whistling or racial
priming theory) to harness the power of prejudice.77 Partisans often target mainstream
audiences, satirizing others by embedding ideological information into entertaining
formats to engage others who are otherwise agnostic about a particular issue.78
In terms of group dynamics, partisans express and market feelings of injustice, grievances,
or disaffection,79 invoking personal and collective needs for significance and a desire to
matter and be respected.80 Narratives often identify an identity crisis threatening the
group81 and victimization at the hands of other out-groups.
Level 1: Fringe Groups
According to DMET, fringe groups describe non-violent ideologies that are on the
periphery of social movements or larger organizations, with more extreme views than
those of the majority. Again, we acknowledge the considerable heterogeneity of ways
to be a fringe actor and also the reality that in particular contexts, the toxic dynamics
we ascribe below to “fringe” organizations may also apply to mainstream partisan
groups. Based on this logic, we propose that DMET platforms create the opportunity
to transparently and accountably “dial down” the diagnostic weighing of a particular
dimension (e.g., out-group derogation) to avoid false positives when such rhetoric
characterizes mainstream discourse.
With that caveat noted, DMET proposes that fringe groups are marked by cognitive cues
such as beliefs of in-group superiority, out-group distinctiveness and inferiority, dogmatic
values, learned prejudice, and discrimination.82
Behaviorally, we conceive that fringe groups discredit or denigrate the out-group, promote
isolation from the general public, and promote narratives of external blame for negative
outcomes such as conspiracy theories.83
77 Rachel Wetts and Robb Willer, “Who Is Called by the Dog Whistle? Experimental Evidence That Racial Resentment and
Political Ideology Condition Responses to Racially Encoded Messages,” Socius 5 (2019).
78 Silvia Knobloch-Westerwick and Simon M. Lavis, “Selecting Serious or Satirical, Supporting or Stirring News? Selective
Exposure to Partisan Versus Mockery News Online Videos,” Journal of Communication 67, no. 1 (2017).
79 Donald R. Kinder and D. Roderick Kiewiet, “Economic Discontent and Political Behavior: The Role of Personal
Grievances and Collective Economic Judgments in Congressional Voting,” American Journal of Political Science (1979);
White and Ypi, The Meaning of Partisanship.
80 Arie W. Kruglanski et al., “The Psychology of Radicalization and Deradicalization: How Significance Quest Impacts
Violent Extremism,” Political Psychology 35 (2014).
81 John Sides, Michael Tesler, and Lynn Vavreck, Identity Crisis: The 2016 Presidential Campaign and the Battle for the
Meaning of America (Princeton University Press, 2019).
82 Roy F. Baumeister, Evil: Inside Human Cruelty and Violence (WH Freeman/Times Books/Henry Holt & Co, 1996); Robert
J. Sternberg, “A Duplex Theory of Hate: Development and Application to Terrorism, Massacres, and Genocide,” Review of
General Psychology 7, no. 3 (2003).
83 Marc W. Heerdink et al., “Emotions as Guardians of Group Norms: Expressions of Anger and Disgust Drive Inferences
About Autonomy and Purity Violations,” Cognition and Emotion 33, no. 3 (2019).

54
In turn, the group dynamics of fringe actors are marked by internal intolerance and
censorship of deviant views, as well as readiness to pursue and promote norms of purity,
supremacy, domination, or revenge.84 Group members are socialized and indoctrinated
into binary right–wrong classifications, sometimes in a highly systematic fashion in which
newcomers move from the ideological periphery of their group to the inside through
contracts of commitment and conversion, and in concert to withdrawing in isolation from
other sources of identity such as family.85
Level 2: Violent Extremism
Violent Extremists propagate a radical ideology supported by violent means that condone
physical or mental harm to others. A key definitory factor that determines violent extremism
is ideologically sanctioned violence such as dehumanization. Our concept is that groups
often differ internally in the tactics advocated and contest the use of violence, and we seek
to distinguish fringe groups in which isolated and/or peripheral members advocate for
hate or violence from violent extremist groups where leaders and mainstream advocates
do so, to terrorist groups where a formal division of labor to carry out attacks has been
implemented.
On a cognitive level, violent extremists are intolerant towards others, representing cultural
and structural violence through silencing and exclusion as just, inevitable, or appropriate,
perceiving a reduced level of moral duties owed to the out-group. Extremists develop
narratives legitimizing violence, often by framing the out-group as an enemy who is
violent towards them.86
Behaviorally, violent extremists serially or systematically dehumanize others. Violent
extremists refuse to tolerate or respect opinions or beliefs contrary to their own; they
perceive a moral superiority and obligation to enforce their ideology.87 This also frees
extremists to act violently against the “other” without moral obligations and the burden of
guilt that would typically be associated with violence.88 Against that backdrop, these groups
84 Dominic Abrams et al., “Pro-Norm and Anti-Norm Deviance within and between Groups,” Journal of Personality and
Social Psychology 78, no. 5 (2000); Dominic Abrams et al., “Collective Deviance: Scaling up Subjective Group Dynamics
to Superordinate Categories Reveals a Deviant Ingroup Protection Effect,” Journal of Personality and Social Psychology
(2021); Roger Giner-Sorolla, Bernhard Leidner, and Emanuele Castano, “Dehumanization, Demonization, and Morality
Shifting: Paths to Moral Certainty in Extremist Violence,” Extremism and the Psychology of Uncertainty (2012).
85 Andrew Coulson, “Education and Indoctrination in the Muslim World,” Policy Analysis 29 (2004); Michael A. Hogg,
Arie Kruglanski, and Kees Van den Bos, “Uncertainty and the Roots of Extremism,” Journal of Social Issues 69, no. 3 (2013);
F. M. Moghaddam, “The Staircase to Terrorism a Psychological Exploration;” American Psychologist 60, no. 2 (2005);
John G. Horgan et al., “From Cubs to Lions: A Six Stage Model of Child Socialization into the Islamic State,” Studies in
Conflict & Terrorism 40, no. 7 (2017).
86 Douglas Pratt, “Religion and Terrorism: Christian Fundamentalism and Extremism,” Terrorism and Political Violence 22,
no. 3 (2010); David Webber and Arie W. Kruglanski, “The Social Psychological Makings of a Terrorist,” Current Opinion in
Psychology 19 (2018).
87 Hogg, Kruglanski, and Van den Bos, “Uncertainty and the Roots of Extremism.”
88 Erving Goffman, Stigma: Notes on the Management of Spoiled Identity (Simon and Schuster, 2009); Albert Bandura,
“Moral Disengagement in the Perpetration of Inhumanities,” Personality and Social Psychology Review 3, no. 3 (1999).

Broadening the GIFCT Hash-Sharing Database Taxonomy: An Assessment and Recommended Next Steps
55
frequently express hate speech towards opponents to create psychological and structural
violence through silencing and exclusion. Individual members of violent extremist groups
may perform selective acts of physical violence as part of a group dynamic that valorizes
these actions. Hate speech and glorification of violent acts would both be indicators of this
level of engagement in DMET.89
At the group level, members of violent extremist groups compete for within-group
recognition, seeking to show personal agency in the service of group domination, and
coalescing around out-groups as perceived or designated existential threats.90 The
normative context of dehumanization establishes social preconditions within which violence
by extremist instigators is likely to be perceived as justified. They authorize individuals to
perform violence and shape bystanders’ reactions to these events, while establishing the
parameters for depersonalization and stigma or dehumanization and moral exclusion.91
While these group dynamics might not be transparent at the content level, favorable
responses valorizing particular in-group actors who are violent may provide a key set of
indicators that would serve to identify the dynamics at play.92
Level 3: Terrorism
Terrorism constitutes the most extreme form of ideologically driven engagement that uses
violence even towards non-combatant targets to instill terror or to send a ‘message’.93
At the cognitive level, terrorists experience two key psychological processes involving
a rigid, exclusive social categorization (e.g., of civilians as part of the out-group) and
a greater psychological or moral distance by exaggerating differences between the in-
group and the out-group.94 The categorization of society at large as part of the out-group
and as the enemy then serves as the justification for their struggle to intimidate or harm
civilians.95 Terrorists thereby sidestep “inhibitory mechanisms” that would normally limit
the aggression of humans against one another. Instead, they show the greatest adherence
to principles that move them to conform unconditionally to certain moral duties, which
89 Alexandra Olteanu et al., “The Effect of Extremist Violence on Hateful Speech Online” (paper presented at the
Proceedings of the International AAAI Conference on Web and Social Media, 2018); Pratt, “Religion and Terrorism.”
90 Gary A. Ackerman, Jun Zhuang, and Sitara Weerasuriya, “Cross-Milieu Terrorist Collaboration: Using Game Theory to
Assess the Risk of a Novel Threat,” Risk Analysis 37, no. 2 (2017); Randy Borum, “Radicalization into Violent Extremism I: A
Review of Social Science Theories,” Journal of Strategic Security 4, no. 4 (2011); David R. Mandel, “The Role of Instigators
in Radicalization to Violent Extremism,” Psychosocial, Organizational, and Cultural Aspects of Terrorism: Final Report to
NATO HFM140/RTO. Brussels: NATO (2011).
91 Erving Goffman, Stigma: Notes on the Management of Spoiled Identity (Simon and Schuster, 2009); Albert Bandura,
“Moral Disengagement in the Perpetration of Inhumanities,” Personality and Social Psychology Review 3, no. 3 (1999).
92 Arie W. Kruglanski et al., “To the Fringe and Back: Violent Extremism and the Psychology of Deviance,” American
Psychologist 72, no. 3 (2017); Jeff Victoroff, Janice R. Adelman, and Miriam Matthews, “Psychological Factors Associated
with Support for Suicide Bombing in the Muslim Diaspora,” Political Psychology 33, no. 6 (2012); Webber and Kruglanski,
“The Social Psychological Makings of a Terrorist.”
93 Meserole and Byman, “Terrorist Definitions and Designations Lists.”
94 Moghaddam, “The Staircase to Terrorism.”
95 Marc Sageman, Understanding Terror Networks (University of Pennsylvania Press, 2011).

56
ultimately legitimize and valorize death.96 Cognitive beliefs about the legitimacy of killing
and the glory of risking sacrificial death are often indicators of the terrorist level in DMET.
Behaviorally, terrorists also endorse, promote, and engage in violent and destructive
actions. These are predominantly directed at civilians as well as non-human symbolic or
infrastructure targets (e.g. works of art, places of worship).97 Terrorists target different
objectives depending on the specific sources of support available to them and the degree
of out-group antagonism in their constituency.98 Terrorists, however, also engage in
violent actions against in-group members as (potential) defectors to sustain the long-term
mission and group norms.99 Concrete incitement to violence and physically violent acts
provide behavioral indicators of the terrorist level in DMET.
In terms of group dynamics, terrorists have organized social structures that support violent
actions on an ongoing basis. Terrorists often are taught to internalize the glorification of
active martyrdom as a testimony of ideological commitment and faith.100 We refer to active
martyrdom (as opposed to passive martyrdom) as a characteristic of terrorism in the sense
of a suicide attack where the act of self-destruction targets a perceived out-group enemy.
Passive martyrdom (e.g., in politics or religion), where the actor is compelled to be ready
to give one’s life to defend the ideals and values of the group, is not considered a terrorist
attribute.101 However, while these beliefs are in theory shared by all group members,
in practice, a formal division of labor to support violent acts exists (e.g., consisting of
finances, military affairs, religious affairs, and public relations).102
In some cases, demographic divisions stream actors to different roles (e.g., younger men
might be expected to serve as martyrs while older men direct actions and women serve
support roles). In other cases, core groups of strategists and recruiters with ongoing roles
might engage opportunistically, at a distance, with individuals of any age and gender
recruited as one-off cannon fodder. These group dynamics might be discerned through
representations of terrorists in the inward-facing communications of the group (e.g.,
distinctive costumes and language) or might be coded as attributes associated with
particular sources.
96 Domenico Tosini, “Calculated, Passionate, Pious Extremism: Beyond a Rational Choice Theory of Suicide Terrorism,”
Asian Journal of Social Science 38, no. 3 (2010).
97 Pratt, “Religion and Terrorism.”
98 Sara M.T. Polo, “The Quality of Terrorist Violence: Explaining the Logic of Terrorist Target Choice,” Journal of Peace
Research 57, no. 2 (2020).
99 Daniel Koehler, “Radical Groups’ Social Pressure Towards Defectors: The Case of Right-Wing Extremist Groups,”
Perspectives on Terrorism 9, no. 6 (2015).
100 Sageman, Understanding Terror Networks.
101 David Cook, “The Implications of “Martyrdom Operations” for Contemporary Islam,” The Journal of Religious Ethics
32, no. 1 (2004).
102 Sageman, Understanding Terror Networks.

Broadening the GIFCT Hash-Sharing Database Taxonomy: An Assessment and Recommended Next Steps
57
Primary Role of Dehumanization for Distinguishing Levels 1 and 2
(Fringe Actors Versus Violent Extremists)
The violent extremist category in DMET (Level 2) includes actors (and content) that is either
associated with physical violence or associated with non-physical violence in the form of
dehumanization. Facebook (and by relation Instagram), Twitter, YouTube, and LinkedIn
recognize dehumanization as a particularly dangerous form of hatred as it removes
moral objections one may have to enact violence, even mass violence, against women,103
children,104 and civilians more broadly within a target group. It connects to violent
extremists’ cognition of representing cultural and structural violence through silencing
and exclusion. It supports their group dynamics of coalescing around an out-group as the
perceived or designated existential threat. While dehumanization may not always lead
to violence, genocides and atrocities typically require it. This cue would identify groups
or individuals that rely on dehumanizing language, or over time are spreading large
amounts of dehumanizing discourse about a group identified on the basis of a protected
characteristic. Dehumanization occurs in two forms:
1. Dehumanizing language includes material that presents the class of persons to have
the appearance, qualities or behavior of an animal, insect, filth, form of disease
or bacteria; or to be inanimate or mechanical objects; or a supernatural threat, in
circumstances in which a reasonable person would conclude that the material was
intended to cause others to see that class of persons as less deserving of being
protected from harm or violence. This material would include words, images, and/or
insignia;105 and
2. Dehumanizing discourse or conceptions include the sustained curation of information
to a specific audience to suggest that the class of persons on the basis of their identified
characteristic106
a. are polluting, despoiling, or debilitating society;
b. have a diminished capacity for human warmth and feeling or independent
thought;
c. act in concert to cause mortal harm; or
d. are to be held responsible for and deserving of collective punishment for the
specific crimes, or alleged crimes of some of their “members.”
103 Nikki Marczak, “A Century Apart: The Genocidal Enslavement of Armenian and Yazidi Women,” in A Gendered Lens
for Genocide Prevention, ed. Mary Michele Connellan and Christiane Fröhlich (London: Palgrave Macmillan UK, 2018).
104 Peter Lentini, “The Australian Far-Right: An International Comparison of Fringe and Conventional Politics,” in The Far-
Right in Contemporary Australia, ed. Mario Peucker and Debra Smith (Singapore: Springer Singapore, 2019).
105 Nick Haslam, “Dehumanization: An Integrative Review,” Personality and Social Psychology Review 10, no. 3 (2006);
Jonathan Leader Maynard and Susan Benesch, “Dangerous Speech and Dangerous Ideology: An Integrated Model for
Monitoring and Prevention” Genocide Studies and Prevention: An International Journal 9, no. 3 (2016).
106 Haslam, “Dehumanization: An Integrative Review”; Maynard and Benesch, “Dangerous Speech and Dangerous
Ideology.”

58
While preventing dehumanization is an imperative under international law (e.g.,
Article 20, 2, ICCPR; Article 25, 3e of the Rome Statute) current algorithms are focused
on detecting individual instances. We conceive that DMET could be trained to predict
aggregate harm by specific actors from a range of samples of borderline content that
each might be difficult to discern as harmful individually. Information campaigns acting
as vehicles for widespread dissemination of dehumanizing conceptions and discourse will
need to be distinguished from news commentary, partisan talk, or fringe discourse. We
have suggested predictors to build this critical capability (discussed in 4.1).
It should be noted that the risk of violence against targeted groups is not reduced (and
may be increased) when advocates are powerful voices speaking in mainstream contexts.
However, where dehumanization is normative and mainstream in a regional context
because it is espoused by mainstream politicians or state offices, other forms of politically-
and psychologically-informed interventions or challenges may be more effective than
content removal.
In our approach, such mainstream groups and content would be placed in the violent
extremist category by DMET when regional norms are not considered. All the authors
condemn dehumanization against any target in any context. Some authors involved in
this report believe platforms could choose to downweigh such groups or content to fringe
or partisan on the grounds of regional norms by using exemption functions. For example,
dehumanizing homophobia, anti-Semitism, or Islamophobic dialogue advocated by
mainstream actors (church leaders, politicians) might be reclassified as partisan or
mainstream in certain contexts, when transparently and accountably locally normative.
While the Australian Muslim Advocacy Network (AMAN) supports transparency for
why certain groups or content is downweighed, it believes that downweighing of
dehumanization should be avoided in any regional context by platforms to uphold the
overarching obligation under international law not to contribute to the incitement of
genocide.
An example of speech that would potentially trigger a violent extremist classification in the
absence of regional norm adjustments is provided by political debates over introducing
the death penalty for homosexuality in Uganda.107 For example, the Ugandan Minister for
Ethics and Integrity, Simon Lokodo, remarked “Homosexuality is not natural to Ugandans,
but there has been a massive recruitment by gay people in schools, and especially among
the youth… We want it made clear that anyone who is even involved in promotion and
recruitment [of homosexuality] has to be criminalized. Those that do grave acts will be
given the death sentence.” Platform providers could consider regional norms despite
content being flagged through DMET by transparently exempting state actors from content
moderation as discussed in section 5.2 below.
107 “Uganda Plans to Introduce Death Penalty for Homosexuality with ‘Kill the Gays’ Law,” ABC News, link

Broadening the GIFCT Hash-Sharing Database Taxonomy: An Assessment and Recommended Next Steps
59
Application of DMET
In the following, we illustrate the applicability of DMET by deriving potential instantiations
of the cues above for the different levels of ideological engagement together with
specific examples for each level. The goal is to clarify the distinction between levels
while acknowledging that further efforts are necessary to identify an exhaustive set of
instantiations of cues, determine cue up/downweighing methods, sharpen cut-off criteria
between levels, and develop strategies to deal with issues such as the niche radicalization
of splinter groups. Subsequently, we provide the integrative sample classification of
organizations with various degrees of ideological engagement obtained in consultation
with global experts.
Illustrative Application of DMET
For partisanship, there would be considerable noise across contexts in how partisan
contestation is expressed. At a cognitive level, our starting basket of indicators for
partisanship would include simple markers of identification (e.g., use of “we,” “us”), us-
them distinctions (e.g., “reject,” “oppose”), in-group positivity (e.g., “we are good,” “we
are right”), and out-group negativity (e.g., “they are wrong,” “they are bad”). Particular
stereotypes that are contextually relevant might be either identified via machine learning
or input as cues to screen for (e.g., “Mexican gangs”). While satire and indirection create
ambiguity in recognition of cues, specific contextually relevant elements could be coded
(e.g., “African gangs”), alongside behavioral indicators of support for in-group actions
(e.g., “donate,” “volunteer”) as well as politicized actions (e.g., “vote,” “rally”) and artistic
contestation (e.g., “protest song,” “protest poem”). Signals that indicate partisan group
dynamics could comprise moralized grievances (e.g., “justice,” “righteousness”), need for
significance (e.g., “respect,” “be counted”) as well as victimization and crisis (e.g., “victim,”
“crisis,” and “threat”); each could constitute initial indicators supporting categorization at
this level.
Fringe groups’ linguistic and image markers and beliefs would often vary contextually
and require local training, but abstract indicators could include cues of dogmatism
(e.g., “always,” “never”), as well as moral absolutes (e.g., “hero,” “villain,” “traitor,”
and “martyr”). Behavioral indicators could include specific contextually relevant insult
patterns, narratives, or more abstract categories of coding such as high-arousal negative
emotions associated with the out-group-oriented, such as anger, contempt, and disgust.108
In general, the group dynamics will not likely be transparent to content categorization,
although some themes (such as purity and domination) may be available for linguistic
coding. In other cases, particular sources or groups could be coded as possessing fringe-
characterizing dynamics by experts, and then markers of the source group membership
(e.g., jargon and group affiliation terms) could be used to identify content from the fringe
108 Heerdink, Koning, Doorn, and Van Kleef, “Emotions as Guardians of Group Norms.”

60
actors.
Violent extremists would also engage in dehumanization forms either directly in their
language or through the general discourse and conceptions (as elaborated above).
In the second half of 2020, AMAN completed a study of five actors producing significant
amounts of blog or pseudo-news content that triggered explicitly dehumanizing and
violent responses by users on Facebook and Twitter. That study identified the following
markers that were common to all five actors’ information operations:
1. Dehumanizing conceptions or conspiracy theories on the actor’s website (where
applicable) in relation to an identified group (“the out-group”) on the basis of a
protected characteristic;
2. Repeated features of the headlines and images that are curated for a specific
audience, including:
• Essentializing the target identity through implicating a wide net of identities
connected to the protected group (e.g., “Niqab-clad Muslima,” “boat migrants,”
“Muslim professor,” “Muslim leader,” “Iran-backed jihadis,” “Ilhan Omar,” “Muslim
father”);
• High degree of hostile verbs or actions (e.g., stabs, sets fire) attributed to those
subjects;
• Significant proportion of actor’s material acting as “factual proofs” to dehumanizing
conceptions about out-group;
• Potential use of explicitly dehumanizing descriptive language (e.g., frothing-
at-the-mouth) or coded extremist movement language with dehumanizing
meaning (e.g.,invader, a term used in RWE propaganda to refer to Muslims as a
mechanically inhuman and barbaric force). However, for the most successful actors,
dehumanizing slurs were avoided to maintain legitimacy and avoid detection; and
• Where there was no dehumanizing language, there was a presence of “baiting”
through rhetorical techniques like irony to provoke in-group reactions; and
3. Evidence in the user comment threads of a pattern of hate speech against the out-
group.
Markers like these above could be used to train algorithms to identify an information
operation intended to dehumanize an out-group over time. Further, GIFCT would be able
to compile a list of protected characteristics recognized commonly by member platforms
or the United Nations Strategy and Plan of Action on Hate Speech.
Violent extremists could also use images and linguistic markers of out-group violence
towards the in-group and contextually relevant images and language of out-group self-

Broadening the GIFCT Hash-Sharing Database Taxonomy: An Assessment and Recommended Next Steps
61
defense. Extremists develop narratives legitimizing violence,109 often by framing the out-
group as an enemy who is violent towards them. Hate speech and glorification of violent
acts would both be indicators of this level of engagement, as we have noted above.110
Next to reinforcing a rigid dichotomy that demands that people choose between the
forces of good or evil (e.g., ISIS demands that all Sunni Muslims choose to fight with them
or against them), they also use apocalyptic linguistic markers to trigger “awakening” in
readers.111 Similarly, terrorists would express beliefs about the legitimacy of killing and the
glory of risking sacrificial death. Concrete incitement to violence and physically violent
acts provides defining behavioral indicators of the terrorist level in DMET.
Sample Feedback on DMET Classification of Ideologically Engaged
Actors
In order to explore the applicability and feasibility of DMET, we reached out to a network
of over 20 extremism researchers and counter-extremism advocates through authors’
contacts. Our contacts highlighted several aspects of the framework for consideration.
They highlighted the simultaneous prevalence of cues from multiple levels and the diffuse
nature of some entities as movements rather than groups (e.g., Evangelical Christians).
Co-occurrences were most prominent between Level 2 (Violent Extremism) and Level
3 (Terrorism) (e.g., Proud Boys, KKK). The discussion of divergent attributes and diffuse
movements particularly appeared for QAnon (categorized by experts across Level 0 – 2)
and Incels (Level 1 – 2). Regional differences within movements and the large in-group
variability of actors, such as objecting to violence or actively engaging in violence, make
movements like QAnon difficult to classify unambiguously.
Some contacts therefore pointed towards the importance of greater flexibility in the
analysis. For example, they expressed that some groups would possess attributes of
multiple categories (e.g., Institute of Public Affairs as borderline fringe, National Socialist
Movement as borderline violent extremist). They also emphasized the multi-faceted
approach of various groups assigning them to multiple types of ideological engagement
(e.g., right-wing and religious: United Patriots Front; left-wing and separatists: Kurdish
movements). Hence, we assume that cross-type patterns need to be acknowledged.
109 Pratt, “Religion and Terrorism”; Webber and Kruglanski, “The Social Psychological Makings of a Terrorist.”
110 Olteanu et al., “The Effect of Extremist Violence on Hateful Speech Online”; Pratt, “Religion and Terrorism.”
111 Matteo Vergani and Ana-Maria Bliuc, “The Evolution of the Isis’ Language: A Quantitative Analysis of the Language of
the First Year of Dabiq Magazine,” SICUREZZA, TERRORISMO E SOCIETÀ 7 (2015).

Levels of Ideological
Engagement
Types of Ideological Engagement
Right-Wing Left-Wing Religious Separatist Single-Issue
Terrorism
Level 3
Terrorism
Boogaloo Bois
Ku Klux Klan
National Socialist Network
The Base
Sendero Luminoso
Fuerzas Armadas
Revolucionarias de Colombia
Al-Qaeda
Islamic State (Daesh)
Jamaah Ansharut Daulah
Mujahidin Indonesia Timur (East
Indonesia Mujahideen)
Euskadi Ta Askatasuna
Irish Republican Army
(Provisional Irish Republican
Army, Ulster Volunteer Force,
Ulster Defence Association)
Army of God
Earth Liberation Front
Violent Extremism
Level 2
Violent Extremism
Blood & Honour
Combat 18
United Patriots Front (True blue
crew, Lads Society)
Oath Keepers
Proud Boys
Jihad Watch
Antifa
Ejército de Liberación Nacional
Kurdistan Workers’ Party
Forum Pembela Islam Órganos de Resistencia
Territorial
Animal Liberation Front
Bundy Family
Non-Violent Extremism
Level 1
Fringe Group
Australia First party
Bharatiya Janata Party
National Socialist Movementˆ
Kurdistan Communities Union
Democratic Socialists of
America
Brigade Manguni
Peoples Temple of the Disciples
of Christ
Westboro Baptist Church
Greater Idaho Movement
Texas Nationalist Movement
American Family Association
Andha Chile
Non-Extremism
Level 0
Partisanship
One Nation
Partai Keadilan Sejahtera
Tea Party Movement
Traditionalist Worker Party
Peoples’ Democratic Party
(Turkey)
Sinn Féin
Wahdah Islamiyah Scottish National Party Anti-Vaxxers
National Abortion Rights
Action League
No más AFP (No + AFP)
Notes. ˆ trending towards increased ideological engagement
Table 2. Tentative DMET Classification of Ideologically Engaged Actors

Broadening the GIFCT Hash-Sharing Database Taxonomy: An Assessment and Recommended Next Steps
63
Narrative Case-Based Review of Level Differences
In order to offer a hands-on illustration of the individual levels and their mutual differences,
we offer a narrative review of individual cases of groups in reference to DMET.
Level 3 (Terrorists): The Base
The DMET-based classification would, for example, echo Canada’s recent decision that
declared The Base as a terrorist organization. It was founded in 2018 as a neo-Nazi, white-
supremacist network that describes itself as an “international survivalist and self-defense
network” that seeks to train its members for fighting a race war (Counterextremism, 2021).
Cognitive cues as disseminated by the group’s leadership denounce the public and system
as the enemy. For example, a leading member, Rinaldo Nazzaro, distributed the following
message on Telegram, April 8, 2021:
“Republicans and many White Nationalists think they’re fighting for the future of
America but they’ve already lost it and there’s no hope of taking it back. The System
is irreversibly dominated by the enemy...The System *is* the enemy and the enemy
*is* the System—They’re inherently and inseparably one and the same now. Some do
realize this and hope for a spontaneous collapse which unfortunately will never come.
The only victory left to be had is breaking away before it’s too late.”112
Regarding behavioral cues, The Base’s leadership has called for members to focus on
non-attributable actions that destabilize society. The Base has distributed to its members’
manuals for lone-wolf terror attacks, bomb-making, counter-surveillance, and guerilla
warfare.113 Similarly, The Base also promotes a group dynamic with dedicated roles to
engage in violent actions. For example, one post by Nazzaro on Telegram on December
21, 2020, reads
“By no later than the 90 day-mark, plan to go on the offensive by clearing and holding
the nearest town. You will commandeer the town and this will serve as your new base of
operations,” before telling followers there may come a time where they will need to kill
American citizens if their insurgency is challenged.114
Level 2 (Violent Extremists): Oath Keepers
The Oath Keepers would fulfill DMET’s characteristics of a violent extremist organization.
They are a loosely organized collection of anti-government extremists who are part of
the broader anti-government “Patriot” movement with a particular focus on recruiting
current and former military members, police officers, and firefighters.115 The Oath Keepers
are driven by conspiracy theories and establish a cognitive glue that promotes violence
112 “The Base,” Counterextremism, Counterextremism.com, link
113 “The Base,” Counterextremism.
114 “The Base,” Counterextremism.
115 “The Oath Keepers,” Anti-Defamation League, link

64
towards the out-group by asking all members to take a pledge to oppose an allegedly
tyrannical American government that will use state forces to control U.S. citizens. The
pledge is targeted to refuse or disobey governmental orders to, for example, disarm the
society or impose martial law.116 Prominent Oath Keeper members such as the founder
Stewart Rhodes engage in dehumanizing behavior when, for example, declaring migrants
or families of legal asylum seekers as an “invasion.”117 Members of the Oath Keepers show
considerable personal agency in support of their group and seek to protect its members
against the out-group threat, for example by providing armed patrols during the protests
in Ferguson or as armed security during land disputes. Similarly, they coalesce with the
Constitutional Sheriff and Peace Officers Association (CSPOA) that disputes the federal
government’s authority and promotes the notion that local sheriffs do not have to obey
federal authorities.118
Level 1 (Fringe Group): Westboro Baptist Church
DMET would classify the Westboro Baptist Church as a Level 1 Fringe group. The Westboro
Baptist Church is an American hyper-Calvinist hate group known for engaging in
inflammatory homophobic and anti-American pickets and hate speech against atheists,
Jews, Muslims, transgender people, and numerous Christian denominations. The Westboro
Baptist Church has an extensive indoctrination system, as evidenced by the comments of a
7-year old member towards an ABC News reporter, saying that those who were destined
for eternal damnation included “gays, fags, hundreds and hundreds of Jews.”119 The group
has an extensive history of engaging in denigrating antisemitic and anti-gay activities such
as over 20,000 respective protests promoting the message that “Any church that allows
fags to be members in good standing is a fag church [...] they have created an atmosphere
in this world where people believe the lie that God loves everybody.”120 The group holds
and enforces strong norms of purity in their beliefs, as most pointedly described by the
fact that its founder Fred Waldron Phelps Sr. was excommunicated arguably for diverging
from the group’s hateful demeanor by suggesting they pursue a kinder approach.121
Level 0 (Partisanship): One Nation Group
DMET’s transition between Level 0 and 1 appears to be more fluid than between other
levels. An example of a non-extremist partisan group is the One Nation party (Pauline
Hanson’s One Nation). It is Australia’s far-right political party that was founded in 1997.
The founder Pauline Hanson promotes polarizing views of the radical right by using “us-
versus-them” language. She holds political grievances that she calls “reverse-racism” or
“anti-white” racism and propagates the idea of immigrants and refugees as existential
threats to the safety, security, and “culture” of a particular society.122 Behaviorally, she
116 “The Oath Keepers,” Anti-Defamation League.
117 John Dougherty, “Oath Keepers ‘Call to Action’ for Flynn Sentencing a Bust,” Southern Poverty Law Center, link.
118 “The Oath Keepers,” Anti-Defamation League.
119 Glenn Ruppel, Kelsey Myers, and Eamon McNiff, “Raised to Hate: Kids of Westboro Baptist Church,” ABC News, link
120 “Westboro Baptist Church,” Anti-Defamation League, link
121 Victoria Cavaliere, “Founder of Westboro Church in Kansas Excommunicated, on Death Bed - Son,” Reuters, link

Broadening the GIFCT Hash-Sharing Database Taxonomy: An Assessment and Recommended Next Steps
65
expresses a strong populist ideology that non-natives must either assimilate and embrace
“Australian culture and values” or”go back to where they came from.”123 While PHON has
been described as exhibiting hate speech, calls for exclusion, and discrimination, their
party obtained 10.27% of the Senate vote in Queensland in the 2019 Federal Election,
double its performance nationwide. Its more mainstream acceptance or smaller deviation
from the norm in Queensland may warrant its location within the level of partisanship.
However, this organization would be considered fringe according to its national levels of
political support (Level 1). Australia’s ABC News reported that a former PHON candidate
later attempted to join The Base out of frustration with the democratic system.124
Illustrative Case-Based Empirical Analysis of Level Differences
Beyond the narrative review of individual cases, DMET can potentially be implemented to
operationalize and systematically assess cognitive, behavioral, and group dynamic cues
of movements, groups, individual actors, or content (e.g., on social media). We envision,
for example, assessing social media content regarding their cognitive, behavioral, and
group dynamic cues, which can then be aggregated for a particular actor (e.g., individual,
group, movement). These empirical analyses can be used to quantify the profile patterns
across the different levels of ideological engagements for the following purposes:
Figures 3 and 4. Conceptual actor profile-specific analyses of the prevalence of definitory cues.
122 Kurt Sengul, “Pauline Hanson Built a Political Career on White Victimhood and Brought Far-Right Rhetoric to the
Mainstream,” The Conversation, June 22, 2020, link
123 “Transcript: Pauline Hanson’s 2016 Maiden Speech to the Senate,” ABC News, link
124 Alex Mann and Kevin Nguyen, “The Base Tapes,” ABC News, link

66
Level of engagement estimation: Given the probabilistic nature of the level defining cues,
in-depth DMET-based analyses could assess the prevalence of the proposed attributes
per level for different actors (Figures 3 & 4). This would help identify the potential or
occurrence of splinter groups. These graphs would highlight whether an organization
either solely engages, for example, in fringe activities, or whether others (under the same
group name) are already engaged in terrorist activity. Decision makers can transparently
assess the profile of different groups or use it to determine the potential threat (parts of)
a particular group pose.
Figures 5 and 6. Conceptual actor profile tracking of definitory cues over time.
Time-dependent tracking: DMET can also be used to track and visualize changes over time
(Figures 5 & 6). By operationalizing lower levels of non-violent ideological engagement
(i.e., Level 0 or 1), DMET enables monitoring relevant ideologically engaged actors
before they turn violent (i.e., Level 2 or 3). This could help to transparently flag suspicious
actors to decision makers, issue warnings towards respective actors, and systematically
re-evaluate the actor profile for de-platforming or delisting decisions.

Broadening the GIFCT Hash-Sharing Database Taxonomy: An Assessment and Recommended Next Steps
67
Figures 6 and 7. Conceptual holistic profile analyses of definitory cues across ideological types.
Holistic profile assessment: Furthermore, DMET can be used to assess all DMET measures
for a particular entity holistically. When applied to a particular actor (i.e., individual,
group, movement) this can illustrate engagement across different ideological types (e.g.,
religious fundamentalistic right-wing actors). It can also be applied to create a more
general overview of the ideological engagement in the sense of a political barometer in a
particular region (Figure 6) and across regions (e.g., state unions, Figure 7).

68
Discussion
Strategic Application Considerations of DMET
We propose DMET to be used as a foundational framework for classifying forms of
extremism and associated extremist content with value beyond approval/removal
decisions. We perceive DMET’s vital strength in this regard to be in its adaptability to suit
varying, fluctuating, or transforming perspectives on what constitutes extremisms. The
value-free modularity of DMET remains unbound from fluctuating political, cultural, and/
or legal perspectives and provides a depolarized snapshot of attributes on a continuum of
extremist magnitudes transparent to platforms, scholars, and the broader public.
To achieve this broad perspective on the existing magnitudes of extremism, considering
developments over time and geography requires comprehensive data access. This includes
existing historical datasets and consultation with diverse experts and practitioners from
academia and industry. This will likely be a modular, iterative approach that assures the
implementation feasibility of DMET throughout. For instance, historic and geographic
dimensions of extremism might not be included until later stages of the DMET development
without impairment of DMET’s current operability.
The modularity of DMET that accounts for different understandings and magnitudes of
extremism is advantageous for its applicability into content moderation use cases. Capturing
different magnitudes of extremism in a multidimensional matrix aids transparency and
acts as a decision support system, especially where violent extremists share non-extremist
content or vice versa. The source of the non-violent content is identified in relation to the
sharing entity (i.e., a particular content’s source being identified as VE Actor), aiding in
making informed decisions on removing or approving content that would go unnoticed
without considering these multiple dimensions of content.
The DMET proposal is qualified by the need for transparency, both prospective
(transparency of design) and retrospective (transparency through inspection and
explanation), and accountability (managerial and external) to secure public trust.125
Safeguards should ensure data used to train and develop algorithms is high quality,
open to academic scrutiny (including from an AI racial discrimination perspective), and
continuously reviewed, corrected, and improved. Normal safeguards against algorithm
discrimination in predictive policing (i.e. eliminating variables such as race and religion,
or proxies for these variables) are not necessarily going to be appropriate in this context.
Although some of the cues relate to cognitions, behaviors, and group dynamics that
apply across the ideological spectrum, the data used to train DMET will be drawn from a
range of contexts and include racial and religious terms. Checking how representative the
125 Heike Felzmann et al., “Transparency You Can Trust: Transparency Requirements for Artificial Intelligence between
Legal Norms and Contextual Concerns,” Big Data & Society 6, no. 1 (2019).

Broadening the GIFCT Hash-Sharing Database Taxonomy: An Assessment and Recommended Next Steps
69
categorization of extremists is in relation to other indicators of prevalence may provide an
indicator of bias or generalizability to guide revision.
The DMET should aim to respond to published good practice,126 principles on artificial
intelligence under the Organization for Economic Co-operation and Development (OECD)
that were adopted on May 22, 2019, and obligations set out in the European Commission’s
proposed legislation (e.g., Proposal for a Regulation of the European Parliament and of the
council laying down harmonized rules on artificial intelligence, April 21, 2021), as artificial
intelligence models to detect terroristic and violent extremist content will be regarded
as “high-risk.” Mistaken classifications promoted by GIFCT can affect users’ rights to free
expression on several platforms at once and can even stifle efforts to highlight human rights
abuse.127 Transparent and fair review processes, facilitated by GIFCT to quickly respond
to unintended consequences, are also important. It would make it easier for human rights
organizations to complain to multiple platforms at once.
We also acknowledge that violent extremist and terrorist labels are highly political and
can require platforms to make exemptions for particular actors or content being flagged
through DMET. In this approach, as highlighted above, platform providers can toggle
content associated with particular groups (e.g., those flagged on U.N. lists) to be always
categorized as terrorists, whereas other content (e.g., advocacy for violence by state
actors) is not. Similarly, the collective right to self-determination is enshrined in human
rights law (see Article 1.1 ICCPR as well as Article 1.1 of ICESCR), and some self-determination
movements use violence. If that violence is deemed to be used in “armed conflict” by the
International Committee of the Red Cross (ICRC), then humanitarian laws of war apply,
and this is not treated as extremism or terrorism. However, some non-state actors will
not be listed by the ICRC and governed by humanitarian law because they are engaged
in a conflict that does not meet threshold tests.128 For example, some violent protests will
be flagged in DMET, including protests where protestors use violence in response to (or to
resist) state violence, including physical violence and life-threatening structural violence.
To contend with this gap, platforms could continue to have the discretion to exempt further
non-state actors based on exercising self-defense,129 considering principles such as self-
determination, duress, necessity, proportionality, or on the balance of other fundamental
human rights. Our hope would be that if a group’s content is flagged in DMET for the use
of dehumanization or advocacy of violence but exempted by platforms, they would be
transparent about exemptions made and provide reasons.
Making a decision based on the balance of fundamental human rights may be required to
provide an enduring mechanism for managing conflict between differing world views and
126 Felzmann et al., “Transparency You Can Trust.”
127 Abdul Rahman Al Jaloud et al., “Caught the Net: The Impact of “Extremist” Speech Regulations on Human Rights
Content,” Electronic Frontier Foundation ed. JIllian C. York (2019), link
128 See the categorization of armed conflict as proposed by UNODC: link
129 Ben Saul, “Defending ‘Terrorism’: Justifications and Excuses for Terrorism in International Criminal Law,” Australian
Yearbook of International Law 25 (2008).

70
claims. For example, groups that wish to express themselves, their beliefs, and exercise
their fundamental rights (such as the right to parent their children and choose schooling
according to their beliefs) are protected by human rights law to do so. Groups are not
protected to infringe upon the fundamental rights of others, and such behavior would
begin to animate DMET cues.
Noting that DMET focuses on deviation from social norms, it is important to consider that a
violent protest will deviate less from mainstream social norms in some regional contexts.
For example, where mass popular protest movements feature violent elements and
advocacy of violence against law enforcement and the state, the scale of people involved
will mean that their behavior may not be flagged as non-normative, extreme, or radical
by regional standards.
Technical Implementation Considerations of DMET
In our approach, large baskets of indicators would be associated probabilistically with each
level (e.g., cognitive stereotypes, dehumanizing language, calls for violence) and will be
used to develop models that algorithmically classify groups (or content) into categories,
with each cue weighed according to its ability to discriminate in particular contexts defined
by the other cues as well as input from the platform provider where desired. We imagine
an incoming stream of content coded for the indicator cues and the groups involved via
machine learning in a process that would be more error-prone at first and be refined
over time and regionally to produce context-specific accuracy, which would in turn decay
as diagnostic attributes changed over time until relearned dynamically. Especially in the
beginning, this will require extensive human oversight, for example in order to deal with
expected inaccuracies of automated machine learning algorithms when dealing with
linguistic markers for irony, sarcasm, or subtle dehumanization.
Content and groups would be classified probabilistically into categories where cues
have established sufficient discriminant validity and high confidence (e.g., with explicit
calls to violence, or when the content is sourced from a group identified as a terrorist
organization, or as a state actor or journalistic or academic source). More commonly,
groups and individuals would be classified based on profiles established via multiple
content posts with increasing confidence over time, with each content item or group
reciprocally associated with transparent certainty/uncertainty scores based on a profile
of attributes, which could be available to platforms as an output.
As a next step, DMET could train a machine learning algorithm that identifies the different
level cues in online content. Due to the linguistic challenges and subjective interpretation
of the content in this unique context, unsupervised learning approaches are likely to
provide misleading results. There is a need for more sophisticated models, and building
such models requires generating a labeled training data set from scratch. A potential cold
start problem of insufficient data for initially training the algorithm could be overcome by
collecting social media content from the groups, building on those identified by experts

Broadening the GIFCT Hash-Sharing Database Taxonomy: An Assessment and Recommended Next Steps
71
as above. This data could comprise social media posts and comments, memes, or content
from external sites (e.g., extremist websites, blogs) introduced into conversations through
hyperlinks. The accuracy of the machine learning algorithm will mainly depend on the (1)
clarity and consistency of classification rule (the coding scheme), (2) quality and size of
labeled data, and (3) finding the proper feature representation.
Subsequently, we could analyze the actor-specific distribution of individual messages
across the different levels to establish and identify communication patterns. By applying
the previously developed machine learning algorithm to new data, we could expand
the available content coded data. This would help perform a ROC analysis to determine
extremism cut-off scores between levels. The regression weights for the individual cues
could also serve as an indicator for the up-/downweighing of individual cues. By choosing
to downweigh or upweigh particular dimensions, platform providers can establish local
profiles of tolerance (e.g., no hate speech at all versus this group; versus hate speech
tolerated against this group, due to its being normative in this context) in a way that is
transparent and able to be accountable or engaged with dialogue. Platform providers
may also opt for transparently and accountably in exempting certain actors such as state
organizations or religious groups ex officio because their views (e.g., the Ugandan minister
above) are mainstream rather than extreme in the regional context.
We could complement the analysis by using metadata that contains information on the
connection between entities to consider the hierarchical structure between individual
extremists (potentially) embedded in extremist groups who are themselves nested in
ideological movements.
Benefits of DMET
We understand extremism as a dimensional concept, with terrorism as the most deviant
pole from the regional norm. DMET supports:
• Ideological fairness: equal opportunity for all entities to be classified as terrorist
based on the generally applicable cognitive, behavioral, and group dynamic
cues;
• Global applicability and scalability: definitory cues transcend geographical,
cultural, and political borders and can be applied relative to relevant reference
norms;
• Update frequency: Observable changes in cognitions, behaviors, or group
dynamics can be captured through near real-time updates;
• Transparency: Classifying actors according to DMET categories based on their
degree of ideological engagement enables transparency and accountability in
regulatory decisions;
• Surfacing states’ role and influence: Current definitions of extremists or terrorists

72
often exclude state actors. DMET potentially classifies any kind of actor without
consideration of their societal role. Platforms can navigate these situations by
making transparent exemptions in reference to the classifications proposed by
DMET to justify their decision making;
• Reduced probability of misclassification errors for (non-) violent extremisms:
A more nuanced understanding of the degrees of ideological engagement and
the potential sub-groups reduces the probability of wrong decisions (classifying
regular users/content as terrorist or failing to identify terrorists as such); and
• Attention to violence in all its forms: Many existing legal frameworks are so
piecemeal or narrow that they deprioritize and overlook the experience of
victims and communities targeted by terroristic and violent extremist violence.
DMET contemplates the full continuum of violence that occurs in the violent denial
of diversity, including structural and psychological violence. Importantly, it
recognizes serial or systematic dehumanization of an out-group as an attribute of
violent extremism.
Subsequently, DMET addresses contemporary challenges of extremism classification and
associated content moderation approaches, including the lack of consent on universal
definitions of extremism, bias, and deficient objectivity on different magnitudes of
extremism.130 DMET’s multidimensional approach enables the aggregation of various
lists and dimensions to allow biases in views of extremism (e.g., exemptions for certain
actors) to be more transparent and accountable. Moreover, DMET unlocks the possibility
of a 360-degree context view of extremism irrespective of the limitations of individual
extremism lists and allows for tracking the development in terms of (de-)radicalization
over time through continuous assessments among the spectrum of ideological
engagement.
Boundary Conditions of DMET
• Dimensionality of attributes: DMET ascribes attributes to particular levels to
capture different degrees of severity. It needs to be acknowledged that these
attributes themselves can also be dimensional (e.g., expressing blame for
negative events can be more or less rampant). Similarly, actors might, for
example, dehumanize a group by using a multitude of cues that holistically
dehumanize the target without making it explicit in one singular instance. The
dimensionality of attributes needs to be empirically assessed to be statistically
considered through measures of item difficulty and item discrimination. The
individually classified instances then need to be holistically considered for
each entity. This would also enable DMET classifications to record an actor’s
tendency of either trending towards a higher (or lower) DMET level of ideological
130 Meserole and Byman, “Terrorist Definitions and Designations Lists.”

Broadening the GIFCT Hash-Sharing Database Taxonomy: An Assessment and Recommended Next Steps
73
engagement or of being stable within the level.
• Probability of attributes: Similarly, DMET ascribes attributes to levels of
ideological engagement where we expect their highest probability of prevalence.
We acknowledge that ideologically engaged actors (i.e., individuals, groups,
movements) can simultaneously demonstrate cognitive, behavioral, and group
dynamic cues from multiple levels or not express particular cues from the level
at which they are classified. The classification of ideological actors according
to DMET requires an empirical assessment of the expectable probabilities
of cues per level and their respective level-determinant weight (i.e., up- or
downweighing of attributes).
• Combinability of types: While DMET distinguishes five common types of
ideological engagement, we understand that actors (i.e., individuals, groups,
movements) can simultaneously follow different types of ideologies (e.g.,
nationalism in combination with religious fundamentalism). Hence, DMET
classifications need to acknowledge the expressivity of characteristics across
multiple ideological types per actor.
• Fragmentation of actors: Different actor organizations (i.e., groups, movements)
can include splinter groups or individuals that diverge from the characteristics of
the overall organization (e.g., enact or support violent behavior as opposed to
the general movement). These individuals or groups can either emerge as splinter
groups or lone-wolf actors alongside the general movement or relative to their
location (e.g., violent in one country, non-violent in another). Classifications
according to DMET need to acknowledge the relatedness of the individual groups
or actors to the higher-level organization (e.g., via metadata) while considering
their individual particularities.
• Time and cultural specificity: Actors (i.e., individuals, groups, movements) as well
as the expression of cues evolve. Actors might become more or less ideologically
engaged, splinter-off into different sub-groups, or form coalitions with other
movements. Similarly, how cues are expressed can change as words can adopt
different meanings over time, as the societal acceptability of terms evolves, or
as terms have regionally specific meanings (e.g., reference to Odin in Nordic
nationalist groups versus the rest of the world). DMET classifications need to be
considered at particular points in time, in particular regions, and regularly re-
evaluated in predetermined time intervals (e.g., to inform de-platforming or
readmittance and delisting decisions).
• Complexity of societal norms: DMET characteristics need to be assessed against
relevant societal norms, which is intended to support its general applicability.
However, regional norms may vary substantially.

74
Conclusion
This paper has put forward a proposal for extending the binary understanding of
terrorism (versus non-terrorism) with a Dynamic Matrix of Extremisms and Terrorism
(DMET) that builds upon the notion of an underlying continuum of ideological
engagement to address the intersection of extremism types (and associated extremist
content) that lead to ineffective content tagging. DMET considers different types of
ideological engagement and different levels, identified using cognitive and behavioral
attributes and attributes of group dynamics. DMET is dynamic as it can be adapted to
accommodate region- and time-specific notions of ideological engagement. The goal
of DMET is to enable platform providers to make transparent and accountable decisions
about engaging with content and groups so that violent extremist and terrorist content
can be identified in a way that makes explicit the criteria and dimensions underlying the
categorization and allows areas of contestation and change to be identified.
DMET Graphics and Visualizations

Broadening the GIFCT Hash-Sharing Database Taxonomy: An Assessment and Recommended Next Steps
75
Reference
ABC News. “Transcript: Pauline Hanson’s 2016 Maiden Speech to the Senate.” ABC News, link
———. “Uganda Plans to Introduce Death Penalty for Homosexuality with ‘Kill the Gays’ Law.” ABC News, link
Abrams, Dominic, José M Marques, Nicola Bown, and Michelle Henson. “Pro-Norm and Anti-Norm Deviance within and
between Groups.” Journal of personality and social psychology 78, no. 5 (2000): 906.
Abrams, Dominic, Giovanni A Travaglino, José M Marques, Ben Davies, and Georgina Randsley de Moura. “Collective
Deviance: Scaling up Subjective Group Dynamics to Superordinate Categories Reveals a Deviant Ingroup Protection Eect.”
Journal of Personality and Social Psychology (2021).
Ackerman, Gary A, Jun Zhuang, and Sitara Weerasuriya. “Cross-Milieu Terrorist Collaboration: Using Game Theory to Assess
the Risk of a Novel Threat.” Risk analysis 37, no. 2 (2017): 342-71.
ADL. “The Oath Keepers “ Anti Defamation League, link
———. “Westboro Baptist Church.” Anti-Defamation League, link
Aytaç, S Erdem, Ali Çarkoğlu, and Ezgi Elçi. “Partisanship, Elite Messages, and Support for Populism in Power.” European
Political Science Review 13, no. 1 (2021): 23-39.
Bakker, Bert N., Yphtach Lelkes, and Ariel Malka. “Understanding Partisan Cue Receptivity: Tests of Predictions from the
Bounded Rationality and Expressive Utility Perspectives.” The Journal of Politics 82, no. 3 (2020): 1061-77.
Bandura, Albert. “Moral Disengagement in the Perpetration of Inhumanities.” Personality and Social Psychology Review 3, no. 3
(1999/08/01 1999): 193-209.
Baumeister, Roy F. Evil: Inside Human Cruelty and Violence. WH Freeman/Times Books/Henry Holt & Co, 1996.
Berger, John M. “Deconstruction of Identity Concepts in Islamic State Propaganda: A Linkage-Based Approach to Counter-
Terrorism Strategic Communications.” 1-21. The Hague, Netherlands: EUROPOL, 2017.
Borum, Randy. “Radicalization into Violent Extremism I: A Review of Social Science Theories.” Journal of strategic security 4, no.
4 (2011): 7-36.
Cavaliere, Victoria. “Founder of Westboro Church in Kansas Excommunicated, on Death Bed - Son.” Reuters, link
Christmann, Kris. “Preventing Religious Radicalisation and Violent Extremism: A Systematic Review of the Research Evidence.”
(2012).
Cook, David. “The Implications of “Martyrdom Operations” for Contemporary Islam.” The Journal of Religious Ethics 32, no. 1
(2004): 129-51.
Coulson, Andrew. “Education and Indoctrination in the Muslim World.” Policy Analysis 29 (2004): 1-36.
Counterextremism. “The Base.” Counterextremism.com, link
Dougherty, John. “Oath Keepers ‘Call to Action’ for Flynn Sentencing a Bust.” Southern Povery Law Center, link
Feddes, Allard R., Lars Nickolson, Liesbeth Mann, and Bertjan Doosje. Psychological Perspectives of Radicalization. London:
Routledge, 2020.
Felzmann, Heike, Eduard Fosch Villaronga, Christoph Lutz, and Aurelia Tamò-Larrieux. “Transparency You Can Trust:
Transparency Requirements for Articial Intelligence between Legal Norms and Contextual Concerns.” Big Data & Society 6,
no. 1 (2019): 2053951719860542.
Giner-Sorolla, Roger, Bernhard Leidner, and Emanuele Castano. “Dehumanization, Demonization, and Morality Shifting: Paths
to Moral Certainty in Extremist Violence.” Extremism and the psychology of uncertainty (2012): 165-82.
Goman, Erving. Stigma: Notes on the Management of Spoiled Identity. Simon and Schuster, 2009.
Graham, Roderick. “Inter-Ideological Mingling: White Extremist Ideology Entering the Mainstream on Twitter.” Sociological
Spectrum 36, no. 1 (2016/01/02 2016): 24-36.
Haslam, Nick. “Dehumanization: An Integrative Review.” Personality and Social Psychology Review 10, no. 3 (2006): 252-64.
Heerdink, Marc W, Lukas F Koning, Evert A Van Doorn, and Gerben A Van Kleef. “Emotions as Guardians of Group Norms:
Expressions of Anger and Disgust Drive Inferences About Autonomy and Purity Violations.” Cognition and emotion 33, no. 3
(2019): 563-78.
Hogg, M. A., and S. A. Reid. “Social Identity, Self-Categorization, and the Communication of Group Norms.” [In English].
Communication Theory 16, no. 1 (Feb 2006): 7-30.
Hogg, M. A., and D. J. Terry. “Social Identity and Self-Categorization Processes in Organizational Contexts.” [In English].
Academy of Management Review 25, no. 1 (Jan 2000): 121-40.
Hogg, Michael A, Arie Kruglanski, and Kees Van den Bos. “Uncertainty and the Roots of Extremism.” Journal of Social Issues 69,
no. 3 (2013): 407-18.

76
Holbrook, Donald. “Designing and Applying an ‘Extremist Media Index’.” Perspectives on Terrorism 9, no. 5 (2015): 57-68.
———. “Far Right and Islamist Extremist Discourses: Shifting Patterns of Enmity.” Extreme Right Wing Political Violence and
Terrorism (2013): 215-37.
Horgan, John G, Max Taylor, Mia Bloom, and Charlie Winter. “From Cubs to Lions: A Six Stage Model of Child Socialization into
the Islamic State.” Studies in Conict & Terrorism 40, no. 7 (2017): 645-64.
Jaloud, Abdul Rahman Al, Hadi Al Khatib, Je Deutch, Dia Kayyali, and Jillian C. York. “Caught the Net: The Impact of
“Extremist” Speech Regulations on Human Rights Content.” edited by JIllian C. York, 2019.
Jegede, Ayodele Samuel. “What Led to the Nigerian Boycott of the Polio Vaccination Campaign?” [In eng]. PLoS medicine 4,
no. 3 (2007): e73-e73.
Keatinge, Tom, and Florence Keen. “Social Media and (Counter) Terrorist Finance: A Fund-Raising and Disruption Tool.” Studies
in Conict & Terrorism 42, no. 1-2 (2019/02/01 2019): 178-205.
Keatinge, Tom, Florence Keen, and Kayla Izenman. “Fundraising for Right-Wing Extremist Movements.” The RU.S.I Journal 164,
no. 2 (2019/02/23 2019): 10-23.
Kinder, Donald R, and D Roderick Kiewiet. “Economic Discontent and Political Behavior: The Role of Personal Grievances and
Collective Economic Judgments in Congressional Voting.” American Journal of Political Science (1979): 495-527.
Knobloch-Westerwick, Silvia, and Simon M Lavis. “Selecting Serious or Satirical, Supporting or Stirring News? Selective
Exposure to Partisan Versus Mockery News Online Videos.” Journal of Communication 67, no. 1 (2017): 54-81.
Koehler, Daniel. “Radical Groups’ Social Pressure Towards Defectors: The Case of Right-Wing Extremist Groups.” Perspectives
on Terrorism 9, no. 6 (2015): 36-50.
Kruglanski, Arie W, Katarzyna Jasko, Marina Chernikova, Michelle Dugas, and David Webber. “To the Fringe and Back: Violent
Extremism and the Psychology of Deviance.” American Psychologist 72, no. 3 (2017): 217.
Kruglanski, Arie W., Michele J. Gelfand, Jocelyn J. Bélanger, Anna Sheveland, Malkanthi Hetiarachchi, and Rohan Gunaratna.
“The Psychology of Radicalization and Deradicalization: How Signicance Quest Impacts Violent Extremism.” Political
Psychology 35 (2014): 69-93.
Lentini, Peter. “The Australian Far-Right: An International Comparison of Fringe and Conventional Politics.” In The Far-Right in
Contemporary Australia, edited by Mario Peucker and Debra Smith, 19-51. Singapore: Springer Singapore, 2019.
Lupu, Noam. “Party Polarization and Mass Partisanship: A Comparative Perspective.” Political Behavior 37, no. 2 (2015/06/01
2015): 331-56.
Mandel, David R. “The Role of Instigators in Radicalization to Violent Extremism.” Psychosocial, Organizational, and Cultural
Aspects of Terrorism: Final Report to NATO HFM140/RTO. Brussels: NATO (2011).
Mann, Alex, and Kevin Nguyen. “The Base Tapes.” ABC News, link
Marczak, Nikki. “A Century Apart: The Genocidal Enslavement of Armenian and Yazidi Women.” In A Gendered Lens for
Genocide Prevention, edited by Mary Michele Connellan and Christiane Fröhlich, 133-62. London: Palgrave Macmillan UK,
2018.
Maynard, Jonathan Leader, and Susan Benesch. “Dangerous Speech and Dangerous Ideology: An Integrated Model
for Monitoring and Prevention .” Genocide Studies and Prevention: An International Journal 9, no. 3 (2016): 70-95.
McCauley, Clark, and Sophia Moskalenko. “Mechanisms of Political Radicalization: Pathways toward Terrorism.” Terrorism and
political violence 20, no. 3 (2008): 415-33.
Meserole, Chris, and Daniel L. Byman. “Terrorist Denitions and Designations Lists What Technology Companies Need to
Know.” In Global Research Network on Terrorism and Technology, 2019.
Moghaddam, F. M. “The Staircase to Terrorism a Psychological Exploration.” American Psychologist 60, no. 2 (2005): 161-69.
Morrison, Sara. “Westboro Baptist Church’s Founder Is Dying, Excommunicated.” The Atlantic, link
Neumann, Peter R. “The Trouble with Radicalization.” International aairs 89, no. 4 (2013): 873-93.
Oegema, Dirk, and Bert Klandermans. “Why Social Movement Sympathizers Don’t Participate: Erosion and Nonconversion of
Support.” American Sociological Review (1994): 703-22.
Olteanu, Alexandra, Carlos Castillo, Jeremy Boy, and Kush Varshney. “The Eect of Extremist Violence on Hateful Speech
Online.” Paper presented at the Proceedings of the International AAAI Conference on Web and Social Media, 2018.
Polo, Sara M.T. “The Quality of Terrorist Violence: Explaining the Logic of Terrorist Target Choice.” Journal of peace research
57, no. 2 (2020): 235-50.
Pratt, Douglas. “Religion and Terrorism: Christian Fundamentalism and Extremism.” Terrorism and Political Violence 22, no. 3
(2010): 438-56.
Ruppel, Glenn, Kelsey Myers, and Eamon McNi. “Raised to Hate: Kids of Westboro Baptist Church.” ABC News, link

Broadening the GIFCT Hash-Sharing Database Taxonomy: An Assessment and Recommended Next Steps
77
Sageman, Marc. Understanding Terror Networks. University of Pennsylvania press, 2011.
Saul, Ben. “Defending ‘Terrorism’: Justications and Excuses for Terrorism in International Criminal Law.” Australian Yearbook of
International Law 25 (2008): 177-226.
Schmid, Alex P. “Violent and Non-Violent Extremism: Two Sides of the Same Coin.” International Centre for Counter-Terrorism
(ICCT) Research Paper (2014): 1-29.
Sengul, Kurt. “Pauline Hanson Built a Political Career on White Victimhood and Brought Far-Right Rhetoric to the Mainstream.”
June 22, 2020 2020.
Sidanius, J., C. Van Laar, S. Levin, and S. Sinclair. “Ethnic Enclaves and the Dynamics of Social Identity on the College Campus:
The Good, the Bad, and the Ugly.” [In English]. Journal of Personality and Social Psychology 87, no. 1 (Jul 2004): 96-110.
Sides, John, Michael Tesler, and Lynn Vavreck. Identity Crisis: The 2016 Presidential Campaign and the Battle for the Meaning
of America. Princeton University Press, 2019.
Sternberg, Robert J. “A Duplex Theory of Hate: Development and Application to Terrorism, Massacres, and Genocide.” Review
of General Psychology 7, no. 3 (2003): 299-328.
Stott, C., P. Hutchison, and J. Drury. “‘Hooligans’ Abroad? Inter-Group Dynamics, Social Identity and Participation in Collective
‘Disorder’ at the 1998 World Cup Finals.” [In English]. British Journal of Social Psychology 40 (Sep 2001): 359-84.
Tosini, Domenico. “Calculated, Passionate, Pious Extremism: Beyond a Rational Choice Theory of Suicide Terrorism.” Asian
Journal of Social Science 38, no. 3 (2010): 394-415.
van Zomeren, M., T. Postmes, and R. Spears. “Toward an Integrative Social Identity Model of Collective Action: A Quantitative
Research Synthesis of Three Socio-Psychological Perspectives.” [In English]. Psychological Bulletin 134, no. 4 (Jul 2008): 504-
35.
Vergani, Matteo, and Ana-Maria Bliuc. “The Evolution of the Isis’ Language: A Quantitative Analysis of the Language of the First
Year of Dabiq Magazine.” SICUREZZA, TERRORISMO E SOCIETÀ 7 (01/01 2015).
Verkuyten, M., and A. De Wolf. “The Development of in-Group Favoritism: Between Social Reality and Group Identity.” [In
English]. Developmental Psychology 43, no. 4 (Jul 2007): 901-11.
Victoro, Je, Janice R. Adelman, and Miriam Matthews. “Psychological Factors Associated with Support for Suicide Bombing
in the Muslim Diaspora.” Political Psychology 33, no. 6 (2012): 791-809.
Vinciarelli, Alessandro, Maja Pantic, Hervé Bourlard, and Alex Pentland. “Social Signal Processing: State-of-the-Art and
Future Perspectives of an Emerging Domain.” In Proceedings of the 16th ACM international conference on Multimedia, 1061–70.
Vancouver, British Columbia, Canada: Association for Computing Machinery, 2008.
Webber, David, and Arie W Kruglanski. “The Social Psychological Makings of a Terrorist.” Current opinion in psychology 19
(2018): 131-34.
West, Emily A, and Shanto Iyengar. “Partisanship as a Social Identity: Implications for Polarization.” Political Behavior (2020):
1-32.
Wetts, Rachel, and Robb Willer. “Who Is Called by the Dog Whistle? Experimental Evidence That Racial Resentment and
Political Ideology Condition Responses to Racially Encoded Messages.” Socius 5 (2019): 2378023119866268.
White, Jonathan, and Lea Ypi. The Meaning of Partisanship. Oxford University Press, 2016.
Wibisono, Susilo, Winnifred R Louis, and Jolanda Jetten. “A Multidimensional Analysis of Religious Extremism.” Frontiers in
psychology 10 (2019): 2560.
Winter, Charlie, Peter Neumann, Alexander Meleagrou-Hitchens, Magnus Ranstorp, Lorenzo Vidino, and Johanna Fürst.
“Online Extremism: Research Trends in Internet Activism, Radicalization, and Counter-Strategies.” International Journal of
Conict and Violence (IJCV) 14, no. 2 (2020): 1-20.
Wintrobe, Ronald. Rational Extremism: The Political Economy of Radicalism. Cambridge University Press, 2006.
Wright, S. C., D. M. Taylor, and F. M. Moghaddam. “Responding to Membership in a Disadvantaged Group - from Acceptance
to Collective Protest.” [In English]. Journal of Personality and Social Psychology 58, no. 6 (Jun 1990): 994-1003.

To learn more about the Global Internet
Forum to Counter Terrorism (GIFCT), please
visit our website or email outreach@gifct.org