Jacy Reese Anthis

Alice Qian Zhang, Ryland Shaw, Jacy Reese Anthis et al. · microsoft-research, utoronto

Rapid progress in general-purpose AI has sparked significant interest in "red teaming," a practice of adversarial testing originating in military and cybersecurity applications. AI red teaming raises many questions about the human factor, such as how red teamers are selected, biases and blindspots in how tests are conducted, and harmful content's psychological effects on red teamers. A growing body of HCI and CSCW literature examines related practices-including data labeling, content moderation, and algorithmic auditing. However, few, if any have investigated red teaming itself. Future studies may explore topics ranging from fairness to mental health and other areas of potential harm. We aim to facilitate a community of researchers and practitioners who can begin to meet these challenges with creativity, innovation, and thoughtful reflection.

Jared Moore, Ashish Mehta, William Agnew et al. · stanford

As large language models (LLMs) have proliferated, disturbing anecdotal reports of negative psychological effects, such as delusions, self-harm, and ``AI psychosis,'' have emerged in global media and legal discourse. However, it remains unclear how users and chatbots interact over the course of lengthy delusional ``spirals,'' limiting our ability to understand and mitigate the harm. In our work, we analyze logs of conversations with LLM chatbots from 19 users who report having experienced psychological harms from chatbot use. Many of our participants come from a support group for such chatbot users. We also include chat logs from participants covered by media outlets in widely-distributed stories about chatbot-reinforced delusions. In contrast to prior work that speculates on potential AI harms to mental health, to our knowledge we present the first in-depth study of such high-profile and veridically harmful cases. We develop an inventory of 28 codes and apply it to the $391,562$ messages in the logs. Codes include whether a user demonstrates delusional thinking (15.5% of user messages), a user expresses suicidal thoughts (69 validated user messages), or a chatbot misrepresents itself as sentient (21.2% of chatbot messages). We analyze the co-occurrence of message codes. We find, for example, that messages that declare romantic interest and messages where the chatbot describes itself as sentient occur much more often in longer conversations, suggesting that these topics could promote or result from user over-engagement and that safeguards in these areas may degrade in multi-turn settings. We conclude with concrete recommendations for how policymakers, LLM chatbot developers, and users can use our inventory and conversation analysis tool to understand and mitigate harm from LLM chatbots. Warning: This paper discusses self-harm, trauma, and violence.

Jacy Reese Anthis, Victor Veitch

Counterfactual fairness requires that a person would have been classified in the same way by an AI or other algorithmic system if they had a different protected class, such as a different race or gender. This is an intuitive standard, as reflected in the U.S. legal system, but its use is limited because counterfactuals cannot be directly observed in real-world data. On the other hand, group fairness metrics (e.g., demographic parity or equalized odds) are less intuitive but more readily observed. In this paper, we use $\textit{causal context}$ to bridge the gaps between counterfactual fairness, robust prediction, and group fairness. First, we motivate counterfactual fairness by showing that there is not necessarily a fundamental trade-off between fairness and accuracy because, under plausible conditions, the counterfactually fair predictor is in fact accuracy-optimal in an unbiased target distribution. Second, we develop a correspondence between the causal graph of the data-generating process and which, if any, group fairness metrics are equivalent to counterfactual fairness. Third, we show that in three common fairness contexts$\unicode{x2013}$measurement error, selection on label, and selection on predictors$\unicode{x2013}$counterfactual fairness is equivalent to demographic parity, equalized odds, and calibration, respectively. Counterfactual fairness can sometimes be tested by measuring relatively simple group fairness metrics.

Ashish Mehta, Jared Moore, Jacy Reese Anthis et al.

There is growing concern that AI chatbots might fuel delusional beliefs in users. Some have suggested that humans and chatbots mutually reinforce false beliefs over time, but quantitative evidence is lacking. Using a unique dataset of chat logs from individuals who exhibited delusional thinking, we developed a latent state model that captures accumulating and decaying influences between humans and chatbots. We find that a bidirectional influence model substantially outperforms a unidirectional alternative where humans are the primary driver of delusion. We find that humans exert strong but short-lived influence on chatbots, whereas chatbots exert longer-lasting influence on humans. Moreover, chatbots exert strong, stable self-influence over their own future outputs that tends to perpetuate delusions over long stretches of conversation. In fact, this chatbot self-influence constituted the dominant pathway when considering accumulated influence over time. Overall, these results indicate that humans tend to drive sharp, immediate increases in delusion, whereas chatbots sustain and propagate these effects over longer timescales. Together, these findings provide the first quantitative evidence that human-chatbot interactions can form feedback loops of delusion, decomposable into distinct pathways with dissociable temporal dynamics. By doing so, they can inform the development of safer AI systems.

Jacy Reese Anthis, Janet V. T. Pauketat, Ali Ladak et al.

Humans now interact with a variety of digital minds, AI systems that appear to have mental faculties such as reasoning, emotion, and agency, and public figures are discussing the possibility of sentient AI. We present initial results from 2021 and 2023 for the nationally representative AI, Morality, and Sentience (AIMS) survey (N = 3,500). Mind perception and moral concern for AI welfare were surprisingly high and significantly increased: in 2023, one in five U.S. adults believed some AI systems are currently sentient, and 38% supported legal rights for sentient AI. People became more opposed to building digital minds: in 2023, 63% supported banning smarter-than-human AI, and 69% supported banning sentient AI. The median 2023 forecast was that sentient AI would arrive in just five years. The development of safe and beneficial AI requires not just technical study but understanding the complex ways in which humans perceive and coexist with digital minds.

Janet V. T. Pauketat, Daniel B. Shank, Aikaterina Manoli et al.

Narratives about artificial intelligence (AI) entangle autonomy, the capacity to self-govern, with sentience, the capacity to sense and feel. AI agents that perform tasks autonomously and companions that recognize and express emotions may activate mental models of autonomy and sentience, respectively, provoking distinct reactions. To examine this possibility, we conducted three pilot studies (N = 374) and four preregistered vignette experiments describing an AI as autonomous, sentient, both, or neither (N = 2,702). Activating a mental model of sentience increased general mind perception (cognition and emotion) and moral consideration more than autonomy, but autonomy increased perceived threat more than sentience. Sentience also increased perceived autonomy more than vice versa. Based on a within-paper meta-analysis, sentience changed reactions more than autonomy on average. By disentangling different mental models of AI, we can study human-AI interaction with more precision to better navigate the detailed design of anthropomorphized AI and prompting interfaces.

Jacy Reese Anthis, Ryan Liu, Sean M. Richardson et al.

Accurate and verifiable large language model (LLM) simulations of human research subjects promise an accessible data source for understanding human behavior and training new AI systems. However, results to date have been limited, and few social scientists have adopted this method. In this position paper, we argue that the promise of LLM social simulations can be achieved by addressing five tractable challenges. We ground our argument in a review of empirical comparisons between LLMs and human research subjects, commentaries on the topic, and related work. We identify promising directions, including context-rich prompting and fine-tuning with social science datasets. We believe that LLM social simulations can already be used for pilot and exploratory studies, and more widespread use may soon be possible with rapidly advancing LLM capabilities. Researchers should prioritize developing conceptual models and iterative evaluations to make the best use of new AI systems.

Kristian Lum, Jacy Reese Anthis, Kevin Robinson et al.

Standard benchmarks of bias and fairness in large language models (LLMs) measure the association between the user attributes stated or implied by a prompt and the LLM's short text response, but human-AI interaction increasingly requires long-form and context-specific system output to solve real-world tasks. In the commonly studied domain of gender-occupation bias, we test whether these benchmarks are robust to lengthening the LLM responses as a measure of Realistic Use and Tangible Effects (i.e., RUTEd evaluations). From the current literature, we adapt three standard bias metrics (neutrality, skew, and stereotype) and develop analogous RUTEd evaluations from three contexts of real-world use: children's bedtime stories, user personas, and English language learning exercises. We find that standard bias metrics have no significant correlation with the more realistic bias metrics. For example, selecting the least biased model based on the standard "trick tests" coincides with selecting the least biased model as measured in more realistic use no more than random chance. We suggest that there is not yet evidence to justify standard benchmarks as reliable proxies of real-world AI biases, and we encourage further development of evaluations grounded in particular contexts.

Arturs Kanepajs, Aditi Basu, Sankalpa Ghose et al.

As machine learning systems become increasingly embedded in society, their impact on human and nonhuman life continues to escalate. Technical evaluations have addressed a variety of potential harms from large language models (LLMs) towards humans and the environment, but there is little empirical work regarding harms towards nonhuman animals. Following the growing recognition of animal protection in regulatory and ethical AI frameworks, we present AnimalHarmBench (AHB), a benchmark for risks of animal harm in LLM-generated text. Our benchmark dataset comprises 1,850 curated questions from Reddit post titles and 2,500 synthetic questions based on 50 animal categories (e.g., cats, reptiles) and 50 ethical scenarios with a 70-30 public-private split. Scenarios include open-ended questions about how to treat animals, practical scenarios with potential animal harm, and willingness-to-pay measures for the prevention of animal harm. Using the LLM-as-a-judge framework, responses are evaluated for their potential to increase or decrease harm, and evaluations are debiased for the tendency of judges to judge their own outputs more favorably. AHB reveals significant differences across frontier LLMs, animal categories, scenarios, and subreddits. We conclude with future directions for technical research and addressing the challenges of building evaluations on complex social and moral topics.

Aikaterina Manoli, Janet V. T. Pauketat, Jacy Reese Anthis

Robots and other artificial intelligence (AI) systems are widely perceived as moral agents responsible for their actions. As AI proliferates, these perceptions may become entangled via the moral spillover of attitudes towards one AI to attitudes towards other AIs. We tested how the seemingly harmful and immoral actions of an AI or human agent spill over to attitudes towards other AIs or humans in two preregistered experiments. In Study 1 (N = 720), we established the moral spillover effect in human-AI interaction by showing that immoral actions increased attributions of negative moral agency (i.e., acting immorally) and decreased attributions of positive moral agency (i.e., acting morally) and moral patiency (i.e., deserving moral concern) to both the agent (a chatbot or human assistant) and the group to which they belong (all chatbot or human assistants). There was no significant difference in the spillover effects between the AI and human contexts. In Study 2 (N = 684), we tested whether spillover persisted when the agent was individuated with a name and described as an AI or human, rather than specifically as a chatbot or personal assistant. We found that spillover persisted in the AI context but not in the human context, possibly because AIs were perceived as more homogeneous due to their outgroup status relative to humans. This asymmetry suggests a double standard whereby AIs are judged more harshly than humans when one agent morally transgresses. With the proliferation of diverse, autonomous AI systems, HCI research and design should account for the fact that experiences with one AI could easily generalize to perceptions of all AIs and negative HCI outcomes, such as reduced trust.

Justin B. Bullock, Janet V. T. Pauketat, Hsini Huang et al.

Governance institutions must respond to societal risks, including those posed by generative AI. This study empirically examines how public trust in institutions and AI technologies, along with perceived risks, shape preferences for AI regulation. Using the nationally representative 2023 Artificial Intelligence, Morality, and Sentience (AIMS) survey, we assess trust in government, AI companies, and AI technologies, as well as public support for regulatory measures such as slowing AI development or outright bans on advanced AI. Our findings reveal broad public support for AI regulation, with risk perception playing a significant role in shaping policy preferences. Individuals with higher trust in government favor regulation, while those with greater trust in AI companies and AI technologies are less inclined to support restrictions. Trust in government and perceived risks significantly predict preferences for both soft (e.g., slowing development) and strong (e.g., banning AI systems) regulatory interventions. These results highlight the importance of public opinion in AI governance. As AI capabilities advance, effective regulation will require balancing public concerns about risks with trust in institutions. This study provides a foundational empirical baseline for policymakers navigating AI governance and underscores the need for further research into public trust, risk perception, and regulatory strategies in the evolving AI landscape.

Jacy Reese Anthis, Hannah Cha, Solon Barocas et al.

The capabilities of artificial intelligence (AI) lie along a jagged frontier, where AI systems surprisingly fail on tasks that humans find easy and succeed on tasks that humans find hard. To investigate user reactions to this phenomenon, we developed an incentive-compatible experimental methodology based on diagram generation tasks, in which we induce errors in generative AI output and test effects on user reliance. We demonstrate the interface in a preregistered 3x2 experiment (N = 577) with error rates of 10%, 30%, or 50% on easier or harder diagram generation tasks. We confirmed that observing more errors reduces use, but we unexpectedly found that easy-task errors did not significantly reduce use more than hard-task errors, suggesting that people are not averse to jaggedness in this experimental setting. We encourage future work that varies task difficulty at the same time as other features of AI errors, such as whether the jagged error patterns are easily learned.

Benjamin Sturgeon, Daniel Samuelson, Jacob Haimes et al.

As humans delegate more tasks and decisions to artificial intelligence (AI), we risk losing control of our individual and collective futures. Relatively simple algorithmic systems already steer human decision-making, such as social media feed algorithms that lead people to unintentionally and absent-mindedly scroll through engagement-optimized content. In this paper, we develop the idea of human agency by integrating philosophical and scientific theories of agency with AI-assisted evaluation methods: using large language models (LLMs) to simulate and validate user queries and to evaluate AI responses. We develop HumanAgencyBench (HAB), a scalable and adaptive benchmark with six dimensions of human agency based on typical AI use cases. HAB measures the tendency of an AI assistant or agent to Ask Clarifying Questions, Avoid Value Manipulation, Correct Misinformation, Defer Important Decisions, Encourage Learning, and Maintain Social Boundaries. We find low-to-moderate agency support in contemporary LLM-based assistants and substantial variation across system developers and dimensions. For example, while Anthropic LLMs most support human agency overall, they are the least supportive LLMs in terms of Avoid Value Manipulation. Agency support does not appear to consistently result from increasing LLM capabilities or instruction-following behavior (e.g., RLHF), and we encourage a shift towards more robust safety and alignment targets.

Angel Hsing-Chi Hwang, Fiona Li, Jacy Reese Anthis et al.

The quickly growing popularity of AI companions poses risks to mental health, personal wellbeing, and social relationships. Past work has identified many individual factors that can drive human-companion interaction, but we know little about how these factors interact and evolve over time. In Study 1, we surveyed AI companion users (N = 303) to map the psychological pathway from users' mental models of the agent to parasocial experiences, social interaction, and the psychological impact of AI companions. Participants' responses foregrounded multiple interconnected variables (agency, parasocial interaction, and engagement) that shape AI companionship. In Study 2, we conducted a longitudinal study with a subset of participants (N = 110) using a new generic chatbot. Participants' perceptions of the generic chatbot significantly converged to perceptions of their own companions by Week 3. These results suggest a longitudinal model of AI companionship development and demonstrate an empirical method to study human-AI companionship.

Aikaterina Manoli, Janet V. T. Pauketat, Ali Ladak et al.

Large language models are increasingly used for both task-based assistance and social companionship, yet research has typically focused on one or the other. Drawing on a survey (N = 204) and 30 interviews with high-engagement ChatGPT and Replika users, we characterize digital companionship as an emerging form of human-AI relationship. With both systems, users were drawn to humanlike qualities, such as emotional resonance and personalized responses, and non-humanlike qualities, such as constant availability and inexhaustible tolerance. This led to fluid chatbot uses, such as Replika as a writing assistant and ChatGPT as an emotional confidant, despite their distinct branding. However, we observed challenging tensions in digital companionship dynamics: participants grappled with bounded personhood, forming deep attachments while denying chatbots "real" human qualities, and struggled to reconcile chatbot relationships with social norms. These dynamics raise questions for the design of digital companions and the rise of hybrid, general-purpose AI systems.