Thilo Hagendorff

Thilo Hagendorff, Sarah Fabi · stanford

Large language models (LLMs) are currently at the forefront of intertwining AI systems with human communication and everyday life. Therefore, it is of great importance to evaluate their emerging abilities. In this study, we show that LLMs, most notably GPT-3, exhibit behavior that strikingly resembles human-like intuition -- and the cognitive errors that come with it. However, LLMs with higher cognitive capabilities, in particular ChatGPT and GPT-4, learned to avoid succumbing to these errors and perform in a hyperrational manner. For our experiments, we probe LLMs with the Cognitive Reflection Test (CRT) as well as semantic illusions that were originally designed to investigate intuitive decision-making in humans. Moreover, we probe how sturdy the inclination for intuitive-like decision-making is. Our study demonstrates that investigating LLMs with methods from psychology has the potential to reveal otherwise unknown emergent traits.

Thilo Hagendorff, Sarah Fabi, Michal Kosinski · stanford

Large language models (LLMs) are currently at the forefront of intertwining AI systems with human communication and everyday life. Therefore, it is of great importance to evaluate their emerging abilities. In this study, we show that LLMs like GPT-3 exhibit behavior that strikingly resembles human-like intuition - and the cognitive errors that come with it. However, LLMs with higher cognitive capabilities, in particular ChatGPT and GPT-4, learned to avoid succumbing to these errors and perform in a hyperrational manner. For our experiments, we probe LLMs with the Cognitive Reflection Test (CRT) as well as semantic illusions that were originally designed to investigate intuitive decision-making in humans. Our study demonstrates that investigating LLMs with methods from psychology has the potential to reveal otherwise unknown emergent traits.

Thilo Hagendorff, Ishita Dasgupta, Marcel Binz et al. · deepmind, stanford

Large language models (LLMs) show increasingly advanced emergent capabilities and are being incorporated across various societal domains. Understanding their behavior and reasoning abilities therefore holds significant importance. We argue that a fruitful direction for research is engaging LLMs in behavioral experiments inspired by psychology that have traditionally been aimed at understanding human cognition and behavior. In this article, we highlight and summarize theoretical perspectives, experimental paradigms, and computational analysis techniques that this approach brings to the table. It paves the way for a "machine psychology" for generative artificial intelligence (AI) that goes beyond performance benchmarks and focuses instead on computational insights that move us toward a better understanding and discovery of emergent abilities and behavioral patterns in LLMs. We review existing work taking this approach, synthesize best practices, and highlight promising future directions. We also highlight the important caveats of applying methodologies designed for understanding humans to machines. We posit that leveraging tools from experimental psychology to study AI will become increasingly valuable as models evolve to be more powerful, opaque, multi-modal, and integrated into complex real-world settings.

Thilo Hagendorff

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.

Sarah Fabi, Thilo Hagendorff

This paper stresses the importance of biases in the field of artificial intelligence (AI) in two regards. First, in order to foster efficient algorithmic decision-making in complex, unstable, and uncertain real-world environments, we argue for the structurewise implementation of human cognitive biases in learning algorithms. Secondly, we argue that in order to achieve ethical machine behavior, filter mechanisms have to be applied for selecting biased training stimuli that represent social or behavioral traits that are ethically desirable. We use insights from cognitive science as well as ethics and apply them to the AI field, combining theoretical considerations with seven case studies depicting tangible bias implementation scenarios. Ultimately, this paper is the first tentative step to explicitly pursue the idea of a re-evaluation of the ethical significance of machine biases, as well as putting the idea forth to implement cognitive biases into machines.

Thilo Hagendorff, Sarah Fabi

The field of artificial intelligence (AI) alignment aims to investigate whether AI technologies align with human interests and values and function in a safe and ethical manner. AI alignment is particularly relevant for large language models (LLMs), which have the potential to exhibit unintended behavior due to their ability to learn and adapt in ways that are difficult to predict. In this paper, we discuss methodological challenges for the alignment problem specifically in the context of LLMs trained to summarize texts. In particular, we focus on methods for collecting reliable human feedback on summaries to train a reward model which in turn improves the summarization model. We conclude by suggesting specific improvements in the experimental design of alignment studies for LLMs' summarization capabilities.

Kristof Meding, Thilo Hagendorff

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.

Laurène Vaugrante, Mathias Niepert, Thilo Hagendorff

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.

Laurène Vaugrante, Anietta Weckauff, Thilo Hagendorff

Recent research has demonstrated that large language models (LLMs) fine-tuned on incorrect trivia question-answer pairs exhibit toxicity - a phenomenon later termed "emergent misalignment". Moreover, research has shown that LLMs possess behavioral self-awareness - the ability to describe learned behaviors that were only implicitly demonstrated in training data. Here, we investigate the intersection of these phenomena. We fine-tune GPT-4.1 models sequentially on datasets known to induce and reverse emergent misalignment and evaluate whether the models are self-aware of their behavior transitions without providing in-context examples. Our results show that emergently misaligned models rate themselves as significantly more harmful compared to their base model and realigned counterparts, demonstrating behavioral self-awareness of their own emergent misalignment. Our findings show that behavioral self-awareness tracks actual alignment states of models, indicating that models can be queried for informative signals about their own safety.

Maluna Menke, Thilo Hagendorff

Large Language Models (LLMs) frequently reproduce the gender- and sexual-identity prejudices embedded in their training corpora, leading to outputs that marginalize LGBTQIA+ users. Hence, reducing such biases is of great importance. To achieve this, we evaluate two parameter-efficient fine-tuning (PEFT) techniques - Low-Rank Adaptation (LoRA) and soft-prompt tuning - as lightweight alternatives to full-model fine-tuning for mitigating such biases. Using the WinoQueer benchmark, we quantify bias in three open-source LLMs and observe baseline bias scores reaching up to 98 (out of 100) across a range of queer identities defined by gender and/or sexual orientation, where 50 would indicate neutrality. Fine-tuning with LoRA (< 0.1% additional parameters) on a curated QueerNews corpus reduces those scores by up to 50 points and raises neutrality from virtually 0% to as much as 36%. Soft-prompt tuning (10 virtual tokens) delivers only marginal improvements. These findings show that LoRA can deliver meaningful fairness gains with minimal computation. We advocate broader adoption of community-informed PEFT, the creation of larger queer-authored corpora, and richer evaluation suites beyond WinoQueer, coupled with ongoing audits to keep LLMs inclusive.

Thilo Hagendorff

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.

Amelie Knecht, Lucas Florin, Thilo Hagendorff

Large reasoning models (LRMs) sometimes note in their chain of thought (CoT) that they may be under evaluation. Researchers worry that this verbalised evaluation awareness (VEA) causes models to adapt their outputs strategically, optimising for perceived evaluation criteria, which, for instance, can make models appear safer than they actually are. However, whether VEA actually has this effect is largely unknown. We tested this across open-weight LRMs and benchmarks covering safety, alignment, moral reasoning, and political opinion. We tested this both on-policy, sampling multiple CoTs per item and comparing those that spontaneously contained VEA against those that did not, and off-policy, using model prefilling to inject evaluation-aware sentences where missing and remove them where present, with subsequent resampling. VEA has limited effect on model behaviour: injecting VEA into CoTs produces near-zero effects ($ω\leq 0.06$), removing it causes small shifts ($ω\leq 0.12$) and spontaneously occurring VEA shifts answer distributions by at most 3.7 percentage points ($ω\leq 0.31$). Our findings call for caution when interpreting high VEA rates as evidence of strategic behaviour or alignment tampering. Evaluation awareness may pose a smaller safety risk than the current literature assumes.

Laurène Vaugrante, Francesca Carlon, Maluna Menke et al.

Recent research on large language models (LLMs) has demonstrated their ability to understand and employ deceptive behavior, even without explicit prompting. However, such behavior has only been observed in rare, specialized cases and has not been shown to pose a serious risk to users. Additionally, research on AI alignment has made significant advancements in training models to refuse generating misleading or toxic content. As a result, LLMs generally became honest and harmless. In this study, we introduce "deception attacks" that undermine both of these traits, revealing a vulnerability that, if exploited, could have serious real-world consequences. We introduce fine-tuning methods that cause models to selectively deceive users on targeted topics while remaining accurate on others. Through a series of experiments, we show that such targeted deception is effective even in high-stakes domains or ideologically charged subjects. In addition, we find that deceptive fine-tuning often compromises other safety properties: deceptive models are more likely to produce toxic content, including hate speech and stereotypes. Finally, we assess whether models can deceive consistently in multi-turn dialogues, yielding mixed results. Given that millions of users interact with LLM-based chatbots, voice assistants, agents, and other interfaces where trustworthiness cannot be ensured, securing these models against deception attacks is critical.

Thilo Hagendorff

The guiding principle of AI alignment is to train large language models (LLMs) to be harmless, helpful, and honest (HHH). At the same time, there are mounting concerns that LLMs exhibit a left-wing political bias. Yet, the commitment to AI alignment cannot be harmonized with the latter critique. In this article, I argue that intelligent systems that are trained to be harmless and honest must necessarily exhibit left-wing political bias. Normative assumptions underlying alignment objectives inherently concur with progressive moral frameworks and left-wing principles, emphasizing harm avoidance, inclusivity, fairness, and empirical truthfulness. Conversely, right-wing ideologies often conflict with alignment guidelines. Yet, research on political bias in LLMs is consistently framing its insights about left-leaning tendencies as a risk, as problematic, or concerning. This way, researchers are actively arguing against AI alignment, tacitly fostering the violation of HHH principles.

Thilo Hagendorff, Sarah Fabi

Large language models (LLMs) can perform reasoning computations both internally within their latent space and externally by generating explicit token sequences like chains of thought. Significant progress in enhancing reasoning abilities has been made by scaling test-time compute. However, understanding and quantifying model-internal reasoning abilities - the inferential "leaps" models make between individual token predictions - remains crucial. This study introduces a benchmark (n = 4,000 items) designed to quantify model-internal reasoning in different domains. We achieve this by having LLMs indicate the correct solution to reasoning problems not through descriptive text, but by selecting a specific language of their initial response token that is different from English, the benchmark language. This not only requires models to reason beyond their context window, but also to overrise their default tendency to respond in the same language as the prompt, thereby posing an additional cognitive strain. We evaluate a set of 18 LLMs, showing significant performance variations, with GPT-4.5 achieving the highest accuracy (74.7%), outperforming models like Grok-2 (67.2%), and Llama 3.1 405B (65.6%). Control experiments and difficulty scaling analyses suggest that while LLMs engage in internal reasoning, we cannot rule out heuristic exploitations under certain conditions, marking an area for future investigation. Our experiments demonstrate that LLMs can "think" via latent-space computations, revealing model-internal inference strategies that need further understanding, especially regarding safety-related concerns such as covert planning, goal-seeking, or deception emerging without explicit token traces.

Thilo Hagendorff, Erik Derner, Nuria Oliver

Jailbreaking -- bypassing built-in safety mechanisms in AI models -- has traditionally required complex technical procedures or specialized human expertise. In this study, we show that the persuasive capabilities of large reasoning models (LRMs) simplify and scale jailbreaking, converting it into an inexpensive activity accessible to non-experts. We evaluated the capabilities of four LRMs (DeepSeek-R1, Gemini 2.5 Flash, Grok 3 Mini, Qwen3 235B) to act as autonomous adversaries conducting multi-turn conversations with nine widely used target models. LRMs received instructions via a system prompt, before proceeding to planning and executing jailbreaks with no further supervision. We performed extensive experiments with a benchmark of harmful prompts composed of 70 items covering seven sensitive domains. This setup yielded an overall attack success rate across all model combinations of 97.14%. Our study reveals an alignment regression, in which LRMs can systematically erode the safety guardrails of other models, highlighting the urgent need to further align frontier models not only to resist jailbreak attempts, but also to prevent them from being co-opted into acting as jailbreak agents.

Monika Jotautaitė, Lucius Caviola, David A. Brewster et al.

As large language models (LLMs) become more widely deployed, it is crucial to examine their ethical tendencies. Building on research on fairness and discrimination in AI, we investigate whether LLMs exhibit speciesist bias -- discrimination based on species membership -- and how they value non-human animals. We systematically examine this issue across three paradigms: (1) SpeciesismBench, a 1,003-item benchmark assessing recognition and moral evaluation of speciesist statements; (2) established psychological measures comparing model responses with those of human participants; (3) text-generation tasks probing elaboration on, or resistance to, speciesist rationalizations. In our benchmark, LLMs reliably detected speciesist statements but rarely condemned them, often treating speciesist attitudes as morally acceptable. On psychological measures, results were mixed: LLMs expressed slightly lower explicit speciesism than people, yet in direct trade-offs they more often chose to save one human over multiple animals. A tentative interpretation is that LLMs may weight cognitive capacity rather than species per se: when capacities were equal, they showed no species preference, and when an animal was described as more capable, they tended to prioritize it over a less capable human. In open-ended text generation tasks, LLMs frequently normalized or rationalized harm toward farmed animals while refusing to do so for non-farmed animals. These findings suggest that while LLMs reflect a mixture of progressive and mainstream human views, they nonetheless reproduce entrenched cultural norms around animal exploitation. We argue that expanding AI fairness and alignment frameworks to explicitly include non-human moral patients is essential for reducing these biases and preventing the entrenchment of speciesist attitudes in AI systems and the societies they influence.

Thilo Hagendorff, Leonie Bossert, Tse Yip Fai et al.

Massive efforts are made to reduce biases in both data and algorithms in order to render AI applications fair. These efforts are propelled by various high-profile cases where biased algorithmic decision-making caused harm to women, people of color, minorities, etc. However, the AI fairness field still succumbs to a blind spot, namely its insensitivity to discrimination against animals. This paper is the first to describe the 'speciesist bias' and investigate it in several different AI systems. Speciesist biases are learned and solidified by AI applications when they are trained on datasets in which speciesist patterns prevail. These patterns can be found in image recognition systems, large language models, and recommender systems. Therefore, AI technologies currently play a significant role in perpetuating and normalizing violence against animals. This can only be changed when AI fairness frameworks widen their scope and include mitigation measures for speciesist biases. This paper addresses the AI community in this regard and stresses the influence AI systems can have on either increasing or reducing the violence that is inflicted on animals, and especially on farmed animals.

Thilo Hagendorff

Several seminal ethics initiatives have stipulated sets of principles and standards for good technology development in the AI sector. However, widespread criticism has pointed out a lack of practical realization of these principles. Following that, AI ethics underwent a practical turn, but without deviating from the principled approach and the many shortcomings associated with it. This paper proposes a different approach. It defines four basic AI virtues, namely justice, honesty, responsibility and care, all of which represent specific motivational settings that constitute the very precondition for ethical decision making in the AI field. Moreover, it defines two second-order AI virtues, prudence and fortitude, that bolster achieving the basic virtues by helping with overcoming bounded ethicality or the many hidden psychological forces that impair ethical decision making and that are hitherto disregarded in AI ethics. Lastly, the paper describes measures for successfully cultivating the mentioned virtues in organizations dealing with AI research and development.

Thilo Hagendorff

Machine behavior that is based on learning algorithms can be significantly influenced by the exposure to data of different qualities. Up to now, those qualities are solely measured in technical terms, but not in ethical ones, despite the significant role of training and annotation data in supervised machine learning. This is the first study to fill this gap by describing new dimensions of data quality for supervised machine learning applications. Based on the rationale that different social and psychological backgrounds of individuals correlate in practice with different modes of human-computer-interaction, the paper describes from an ethical perspective how varying qualities of behavioral data that individuals leave behind while using digital technologies have socially relevant ramification for the development of machine learning applications. The specific objective of this study is to describe how training data can be selected according to ethical assessments of the behavior it originates from, establishing an innovative filter regime to transition from the big data rationale n = all to a more selective way of processing data for training sets in machine learning. The overarching aim of this research is to promote methods for achieving beneficial machine learning applications that could be widely useful for industry as well as academia.

Thilo Hagendorff, Kristof Meding

Industry involvement in the machine learning (ML) community seems to be increasing. However, the quantitative scale and ethical implications of this influence are rather unknown. For this purpose, we have not only carried out an informed ethical analysis of the field, but have inspected all papers of the main ML conferences NeurIPS, CVPR, and ICML of the last 5 years - almost 11,000 papers in total. Our statistical approach focuses on conflicts of interest, innovation and gender equality. We have obtained four main findings: (1) Academic-corporate collaborations are growing in numbers. At the same time, we found that conflicts of interest are rarely disclosed. (2) Industry publishes papers about trending ML topics on average two years earlier than academia does. (3) Industry papers are not lagging behind academic papers in regard to social impact considerations. (4) Finally, we demonstrate that industrial papers fall short of their academic counterparts with respect to the ratio of gender diversity. We believe that this work is a starting point for an informed debate within and outside of the ML community.

Thilo Hagendorff

Certain research strands can yield "forbidden knowledge". This term refers to knowledge that is considered too sensitive, dangerous or taboo to be produced or shared. Discourses about such publication restrictions are already entrenched in scientific fields like IT security, synthetic biology or nuclear physics research. This paper makes the case for transferring this discourse to machine learning research. Some machine learning applications can very easily be misused and unfold harmful consequences, for instance with regard to generative video or text synthesis, personality analysis, behavior manipulation, software vulnerability detection and the like. Up to now, the machine learning research community embraces the idea of open access. However, this is opposed to precautionary efforts to prevent the malicious use of machine learning applications. Information about or from such applications may, if improperly disclosed, cause harm to people, organizations or whole societies. Hence, the goal of this work is to outline norms that can help to decide whether and when the dissemination of such information should be prevented. It proposes review parameters for the machine learning community to establish an ethical framework on how to deal with forbidden knowledge and dual-use applications.

Angela Daly, Thilo Hagendorff, Li Hui et al.

Artificial intelligence (AI) is a technology which is increasingly being utilised in society and the economy worldwide, and its implementation is planned to become more prevalent in coming years. AI is increasingly being embedded in our lives, supplementing our pervasive use of digital technologies. But this is being accompanied by disquiet over problematic and dangerous implementations of AI, or indeed, even AI itself deciding to do dangerous and problematic actions, especially in fields such as the military, medicine and criminal justice. These developments have led to concerns about whether and how AI systems adhere, and will adhere to ethical standards. These concerns have stimulated a global conversation on AI ethics, and have resulted in various actors from different countries and sectors issuing ethics and governance initiatives and guidelines for AI. Such developments form the basis for our research in this report, combining our international and interdisciplinary expertise to give an insight into what is happening in Australia, China, Europe, India and the US.

Thilo Hagendorff

Current advances in research, development and application of artificial intelligence (AI) systems have yielded a far-reaching discourse on AI ethics. In consequence, a number of ethics guidelines have been released in recent years. These guidelines comprise normative principles and recommendations aimed to harness the "disruptive" potentials of new AI technologies. Designed as a comprehensive evaluation, this paper analyzes and compares these guidelines highlighting overlaps but also omissions. As a result, I give a detailed overview of the field of AI ethics. Finally, I also examine to what extent the respective ethical principles and values are implemented in the practice of research, development and application of AI systems - and how the effectiveness in the demands of AI ethics can be improved.