Content uploaded by Hachik Yazadzhiyan
Author content
All content in this area was uploaded by Hachik Yazadzhiyan on Dec 26, 2023
Content may be subject to copyright.
5
№ 4 (5), 2023
ISSN: 2682 – 9517 (print) ISSN: 2683 – 0930 (online)
JOURNAL OF INFORMATICS AND INNOVATIVE TECHNOLOGIES (JIIT)
What are LLM hallucinations? Causes, ethical
concerns and prevention
Hachik Yazadzhiyan
Institute of Informatics and Innovative Technolgies,
Plovdiv, Bulgaria,
hachik.yazadjian@iiit.bg; ORCID
0009-0004-1900-3639
Abstract. In this paper, hallucinations of large
language models (LLM) are discussed. The
hallucinations of LLM make researchers worry
about progress in this field because if they cannot
control the outcome of the models, then they cannot
build critical systems to serve humanity.
Hallucinations are one of the major ethical problems
of LLMsand can have harmful consequences as
users without sufficient domain knowledge begin to
rely too heavily on these increasingly persuasive
language models.
Keywords: AI Hallucinations, LLM,
I.
I
NTRODUCTION
Large Language Models (LLM) are artificial
intelligence systems capable of analyzing and
generating human text. But they have a problem - LLMs
hallucinate, i.e. they make things up. The hallucinations
of LLMs make researchers worry about progress in the
field because if they can't control the outcome of the
models, then they can't build critical systems to serve
humanity. In general, LLMs use huge amounts of
training data and complex learning algorithms to
generate realistic results. In some cases, context
learning is used to train these models using just a few
examples. LLMs are becoming increasingly popular in a
variety of application areas ranging from machine
translation, sentiment analysis, virtual AI assistance,
image annotation, natural language processing, and
more.
Despite the cutting-edge nature of LLMs, they are
still prone to biases, errors, and hallucinations. Jan
Lekun, current Chief Scientist of Artificial Intelligence at
Meta, recently mentioned a central defect in LLMs that
causes hallucinations: "Big language models have no
idea of the underlying reality that the language
describes. These systems generate text that sounds
good, grammatically and semantically, but they don't
really have any purpose beyond just satisfying
statistical consistency with the prompt.
II.
H
ALLUCINATIONS IN
LLM
Hallucinations refer to the pattern generating
results that are syntactically and semantically correct,
but unrelated to reality and based on false
assumptions. Hallucinations are one of the major
ethical problems of LLMsand can have harmful
consequences as users without sufficient domain
knowledge begin to rely too heavily on these
increasingly persuasive language models.
Some degree of hallucination is inevitable with all
autoregressive LLMs. For example, a model might
attribute a fake quote to a celebrity that was never said.
They may assert something about a particular topic that
is factually incorrect or cite non-existent sources in
scholarly articles, thus spreading misinformation.
However, making AI models hallucinate does not
always have adverse effects. For example, a suggests a
new study scientists discover "novel proteins with an
unlimited set of properties" by hallucinating LLMs.
III.
W
HAT CAUSES
LLM
HALLUCINATIONS
?
LLMs can hallucinate due to a variety of factors ranging
from rekeying errors in encoding and decoding to
training bias.
A.
Readjust
Overfitting is a problem where the AI model fits the
training data too well. However, it cannot fully
represent the entire set of input data it may encounter,
i.e., it fails to generalize its predictive power to new,
unseen data. Overfitting can result in the model
producing hallucinated content.
6
№ 4 (5), 2023
ISSN: 2682 – 9517 (print) ISSN: 2683 – 0930 (online)
JOURNAL OF INFORMATICS AND INNOVATIVE TECHNOLOGIES (JIIT)
B.
Encoding and decoding errors
If there are errors in the encoding and decoding of text
and its subsequent representations, this can also lead
to the generation of meaningless and erroneous model
results.
C.
Addiction to learning
Another factor is the presence of certain outliers in the
training data, which can cause the model to produce
results that represent these outliers rather than the
true nature of the data. This is similar to the lack of
diversity in the training data, which limits the model's
ability to generalize to new data.
The complex structure of LLM makes it quite
challenging for AI researchers and practitioners to
identify, interpret, and correct for these underlying
causes of hallucinations.
IV.
E
THICAL PROBLEMS OF
LLM
HALLUCINATIONS
LLMs can perpetuate and reinforce harmful biases
through hallucinations and can in turn negatively affect
users and have detrimental social consequences. Some
of these most important ethical concerns are listed
below:
A.
Discriminatory and toxic content
As LLM training data is often full of sociocultural
stereotypes due to inherent biases and lack of diversity.
Thus LLM can, produce and reinforce these harmful
ideas against disadvantaged groups in society.
They can generate this discriminatory and hate-
mongering content based on race, gender, religion,
ethnicity, etc.
B.
Privacy issues
LLMs are taught in a massive learning campus that
often includes people's personal information. There
have been cases where such models have violated
people's privacy. They can leak specific information
such as social security numbers, home addresses, cell
phone numbers and medical records.
C.
Misinformation and disinformation
Language models can produce human-like content that
seems accurate but is actually false and not supported
by empirical evidence. This may be accidental, leading
to misinformation, or there may be malicious intent
behind it to deliberately spread misinformation. If left
unchecked, it can create adverse socio-cultural,
economic and political trends.
V.
P
REVENTING
LLM
HALLUCINATIONS
Researchers and practitioners are taking different
approaches to address the problem of hallucinations in
LLM. These include improving the variety of training
data, eliminating inherent biases, using better
regulation techniques, and using adversarial training
and reinforcement learning, among others:
- Developing better regulation techniques is at the
heart of dealing with hallucinations. They help prevent
overload and other problems that cause hallucinations.
- Increasing data can reduce the frequency of
hallucinations, as seen in a study. Data augmentation
involves expanding the training set by adding an
arbitrary token anywhere in the sentence. It doubles
the size of the training set and leads to a reduction in
hallucination frequency.
- OpenAI and Google's DeepMind developed a
technique called reinforcement learning with human
feedback (RLHF) to address the hallucination problem
of ChatGPT. It involves a human evaluator that
frequently reviews model responses and selects the
most relevant prompts for the user. This feedback is
then used to correct the model's behavior. Ilya
Sutskever, OpenAI's chief scientist, recently mentioned
that this approach could potentially resolve
hallucinations in ChatGPT: "I'm pretty hopeful that by
just improving that subsequent reinforcement training
from the human feedback step, we can teach him not
to hallucinate."
- Identifying hallucinated content to use as an example
for future training is also a method used to address
hallucinations. A new technique in this regard detects
hallucinations at the token level and predicts whether
7
№ 4 (5), 2023
ISSN: 2682 – 9517 (print) ISSN: 2683 – 0930 (online)
JOURNAL OF INFORMATICS AND INNOVATIVE TECHNOLOGIES (JIIT)
each token in the output is hallucinated. It also includes
a method for learning unsupervised hallucination
detectors.
Detection of hallucinations at symbol level
Simply put, LLM hallucinations are a growing
anxiety. And despite efforts, much work still needs to
be done to address the problem. The complexity of
these patterns means that it is generally challenging to
correctly identify and correct the inherent causes of
hallucinations.
However, with continued research and
development, mitigating hallucinations in LLMs and
reducing their ethical implications is possible.
If you would like to learn more about LLM and the
preventative techniques being developed to correct
LLM hallucinations, see unite.ai to expand your
knowledge.