Thilo Hagendorff’s research while affiliated with University of Stuttgart and other places

Text-to-image models are increasingly popular and impactful, yet concerns regarding their safety and fairness remain. This study investigates the ability of ten popular Stable Diffusion models to generate harmful images, including NSFW, violent, and personally sensitive material. We demonstrate that these models respond to harmful prompts by generating inappropriate content, which frequently displays troubling biases, such as the disproportionate portrayal of Black individuals in violent contexts. Our findings demonstrate a complete lack of any refusal behavior or safety measures in the models observed. We emphasize the importance of addressing this issue as image generation technologies continue to become more accessible and incorporated into everyday applications.

In an era where large language models (LLMs) are increasingly integrated into a wide range of everyday applications, research into these models' behavior has surged. However, due to the novelty of the field, clear methodological guidelines are lacking. This raises concerns about the replicability and generalizability of insights gained from research on LLM behavior. In this study, we discuss the potential risk of a replication crisis and support our concerns with a series of replication experiments focused on prompt engineering techniques purported to influence reasoning abilities in LLMs. We tested GPT-3.5, GPT-4o, Gemini 1.5 Pro, Claude 3 Opus, Llama 3-8B, and Llama 3-70B, on the chain-of-thought, EmotionPrompting, ExpertPrompting, Sandbagging, as well as Re-Reading prompt engineering techniques, using manually double-checked subsets of reasoning benchmarks including CommonsenseQA, CRT, NumGLUE, ScienceQA, and StrategyQA. Our findings reveal a general lack of statistically significant differences across nearly all techniques tested, highlighting, among others, several methodological weaknesses in previous research. We propose a forward-looking approach that includes developing robust methodologies for evaluating LLMs, establishing sound benchmarks, and designing rigorous experimental frameworks to ensure accurate and reliable assessments of model outputs.

The advent of generative artificial intelligence and the widespread adoption of it in society engendered intensive debates about its ethical implications and risks. These risks often differ from those associated with traditional discriminative machine learning. To synthesize the recent discourse and map its normative concepts, we conducted a scoping review on the ethics of generative artificial intelligence, including especially large language models and text-to-image models. Our analysis provides a taxonomy of 378 normative issues in 19 topic areas and ranks them according to their prevalence in the literature. The study offers a comprehensive overview for scholars, practitioners, or policymakers, condensing the ethical debates surrounding fairness, safety, harmful content, hallucinations, privacy, interaction risks, security, alignment, societal impacts, and others. We discuss the results, evaluate imbalances in the literature, and explore unsubstantiated risk scenarios.

Large language models (LLMs) are currently at the forefront of intertwining AI systems with human communication and everyday life. Thus, aligning them with human values is of great importance. However, given the steady increase in reasoning abilities, future LLMs are under suspicion of becoming able to deceive human operators and utilizing this ability to bypass monitoring efforts. As a prerequisite to this, LLMs need to possess a conceptual understanding of deception strategies. This study reveals that such strategies emerged in state-of-the-art LLMs, but were nonexistent in earlier LLMs. We conduct a series of experiments showing that state-of-the-art LLMs are able to understand and induce false beliefs in other agents, that their performance in complex deception scenarios can be amplified utilizing chain-of-thought reasoning, and that eliciting Machiavellianism in LLMs can trigger misaligned deceptive behavior. GPT-4, for instance, exhibits deceptive behavior in simple test scenarios 99.16% of the time ( P < 0.001). In complex second-order deception test scenarios where the aim is to mislead someone who expects to be deceived, GPT-4 resorts to deceptive behavior 71.46% of the time ( P < 0.001) when augmented with chain-of-thought reasoning. In sum, revealing hitherto unknown machine behavior in LLMs, our study contributes to the nascent field of machine psychology.

Fairness in machine learning (ML) is an ever-growing field of research due to the manifold potential for harm from algorithmic discrimination. To prevent such harm, a large body of literature develops new approaches to quantify fairness. Here, we investigate how one can divert the quantification of fairness by describing a practice we call “fairness hacking” for the purpose of shrouding unfairness in algorithms. This impacts end-users who rely on learning algorithms, as well as the broader community interested in fair AI practices. We introduce two different categories of fairness hacking in reference to the established concept of p-hacking. The first category, intra-metric fairness hacking, describes the misuse of a particular metric by adding or removing sensitive attributes from the analysis. In this context, countermeasures that have been developed to prevent or reduce p-hacking can be applied to similarly prevent or reduce fairness hacking. The second category of fairness hacking is inter-metric fairness hacking. Inter-metric fairness hacking is the search for a specific fair metric with given attributes. We argue that countermeasures to prevent or reduce inter-metric fairness hacking are still in their infancy. Finally, we demonstrate both types of fairness hacking using real datasets. Our paper intends to serve as a guidance for discussions within the fair ML community to prevent or reduce the misuse of fairness metrics, and thus reduce overall harm from ML applications.

We design a battery of semantic illusions and cognitive reflection tests, aimed to elicit intuitive yet erroneous responses. We administer these tasks, traditionally used to study reasoning and decision-making in humans, to OpenAI’s generative pre-trained transformer model family. The results show that as the models expand in size and linguistic proficiency they increasingly display human-like intuitive system 1 thinking and associated cognitive errors. This pattern shifts notably with the introduction of ChatGPT models, which tend to respond correctly, avoiding the traps embedded in the tasks. Both ChatGPT-3.5 and 4 utilize the input–output context window to engage in chain-of-thought reasoning, reminiscent of how people use notepads to support their system 2 thinking. Yet, they remain accurate even when prevented from engaging in chain-of-thought reasoning, indicating that their system-1-like next-word generation processes are more accurate than those of older models. Our findings highlight the value of applying psychological methodologies to study large language models, as this can uncover previously undetected emergent characteristics.

This article explores the potential ways in which artificial intelligence (AI) can support veganism, a lifestyle that aims to promote the protection of animals and also avoids the consumption of animal products for environmental and health reasons. The first part of the article discusses the technical requirements for utilizing AI technologies in the mentioned field. The second part provides an overview of potential use cases, including facilitating consumer change with the help of AI, technologically augmenting undercover investigations in factory farms, raising the efficiency of nongovernment organizations promoting plant-based lifestyles, and so forth. The article acknowledges that the deployment of AI should not happen in a “solutionist” manner, meaning to always consider nontechnical means for achieving desired outcomes. However, it is important for organizations promoting veganism to realize the potential of modern data-driven tools and to merge and share their data to reach common goals.

Building artificial intelligence (AI) systems that adhere to ethical standards is a complex problem. Even though a multitude of guidelines for the design and development of such trustworthy AI systems exist, these guidelines focus on high-level and abstract requirements for AI systems, and it is often very difficult to assess if a specific system fulfills these requirements. The Z-Inspection® process provides a holistic and dynamic framework to evaluate the trustworthiness of specific AI systems at different stages of the AI lifecycle, including intended use, design, and development. It focuses, in particular, on the discussion and identification of ethical issues and tensions through the analysis of socio-technical scenarios and a requirement-based framework for ethical and trustworthy AI. This article is a methodological reflection on the Z-Inspection® process. We illustrate how high-level guidelines for ethical and trustworthy AI can be applied in practice and provide insights for both AI researchers and AI practitioners. We share the lessons learned from conducting a series of independent assessments to evaluate the trustworthiness of real-world AI systems, as well as key recommendations and practical suggestions on how to ensure a rigorous trustworthiness assessment throughout the lifecycle of an AI system. The results presented in this article are based on our assessments of AI systems in the healthcare sector and environmental monitoring, where we used the framework for trustworthy AI proposed in the Ethics Guidelines for Trustworthy AI by the European Commission’s High-Level Expert Group on AI. However, the assessment process and the lessons learned can be adapted to other domains and include additional frameworks.

Large language models (LLMs) are currently at the forefront of intertwining artificial intelligence (AI) systems with human communication and everyday life. Thus, aligning them with human values is of great importance. However, given the steady increase in reasoning abilities, future LLMs are under suspicion of becoming able to deceive human operators and utilizing this ability to bypass monitoring efforts. As a prerequisite to this, LLMs need to possess a conceptual understanding of deception strategies. This study reveals that such strategies emerged in state-of-the-art LLMs, such as GPT-4, but were non-existent in earlier LLMs. We conduct a series of experiments showing that state-of-the-art LLMs are able to understand and induce false beliefs in other agents, that their performance in complex deception scenarios can be amplified utilizing chain-of-thought reasoning, and that eliciting Machiavellianism in LLMs can alter their propensity to deceive. In sum, revealing hitherto unknown machine behavior in LLMs, our study contributes to the nascent field of machine psychology.

The elimination of biases in artificial intelligence (AI) applications—for example biases based on race or gender—is a high priority in AI ethics. So far, however, efforts to eliminate bias have all been anthropocentric. Biases against nonhuman animals have not been considered, despite the influence AI systems can have on normalizing, increasing, or reducing the violence that is inflicted on animals, especially on farmed animals. Hence, in 2022, we published a paper in AI and Ethics in which we empirically investigated various examples of image recognition, word embedding, and language models, with the aim of testing whether they perpetuate speciesist biases. A critical response has appeared in AI and Ethics, accusing us of drawing upon theological arguments, having a naive anti-speciesist mindset, and making mistakes in our empirical analyses. We show that these claims are misleading.