The function of chatbots like OpenAI’s ChatGPT is based on detecting probabilistic patterns in the training data. This makes them vulnerable to generating factual mistakes in their outputs. Recently, it has become commonplace in philosophical, scientific, and popular discourses to capture such mistakes by metaphors that draw on discourses about the human mind. The two most popular metaphors at present are hallucinating and bullshitting. In this paper, we review, discuss, and criticise these mental metaphors. By applying conceptual metaphor theory, we provide numerous reasons why they do not succeed in providing us with a better understanding of factual chatbot mistakes. We conclude by calling for justifications of the epistemic feasibility and fruitfulness of the metaphors at issue. Furthermore, we raise the question what would be lost if we stopped trying to capture factual chatbot mistakes by mental metaphors.
AI chatbots are becoming experts at changing people's minds. What's their secret?
ChatGPT and other AIs use a flood of facts, and the occasional lie, to persuade humans
Many Minds: Seven metaphors for AI
If you wanted a petri dish for understanding metaphors—how they emerge and evolve and jostle with each other—it would be hard to do better than the world of AI. We talk about AI systems variously as coaches or co-pilots, little genies or alien intelligences. Some researchers claim that AIs "grow," that they're entering their phase of "adolescence." Critics deride AI products as slop and dismiss LLMs as a kind of autocomplete on steroids. What's behind these different characterizations? Which ones are accurate and which are unfair? And are our metaphors mostly colorful rhetoric or do they matter? Are they shaping how we understand, adopt, and ultimately regulate these new technologies? My guest today is . Melanie is a computer scientist and Professor at the Santa Fe Institute. She is the author of the book, and she writes a by the same name. This episode is a bit of a companion to with Steve Flusberg. In that episode, Steve and I attempted a kind of crash course on metaphor and the human mind. Here, Melanie and I sit down for more of an extended case study: how metaphors are guiding, galvanizing, and maybe deceiving us in the contested realm of AI discourse. We unpack seven of the most widely used metaphors in this space. We consider how these metaphors are shaping not only our everyday understandings of AI, but also law and policy. We also talk about the metaphor and analogy capabilities of AI itself. Can these systems reason abstractly in the way that humans can? Along the way, Melanie and I touch on: AI-generated poetry, anthropomorphism, the original sin of AI research, the myth of Narcissus, psychometric testing and its pitfalls, metaphors for AI that are a bit hard to spot, and the question of whether an AI has ever come up with a decent analogy for itself. Longtime fans of the show will know that we've had Melanie on the show . We invited her back, not only because she's thought about metaphor and analogy in AI discourse for decades, but because she's a voice of calm insight in an area that’s increasingly awash in hype and polemic. Longtime fans of the show may also note that we are now celebrating our 6th birthday at Many Minds. That's right, the show launched in February 2020. If you'd like to support us as we recognize this milestone, you can leave us a rating or a review, recommend us to a friend, or give us a shout out on social media. Your support is always appreciated. Without further ado, on to my conversation with Dr. Melanie Mitchell. Enjoy! Notes 3:30 – For an overview of Douglas Hofstadter’s work on analogy, see . 8:00 – Much of our discussion in this interview draws on Dr. Mitchell’s piece on the in Science magazine. 13:30 – For earlier discussions of anthropomorphism on the show, see our earlier episodes and . 16:00 – See for the original discussion of LLMs as “stochastic parrots.” 17:00 – See for the original discussion of ChatGPT as a “blurry jpeg.” 18:30 – See for the original discussion of LLMs as role players. 22:00 – See for one use of the “LLMs as crowds” metaphor. See also a discussion of this metaphor (and other metaphors for AI) . 25:00 – For one discussion of AI as a “cultural technology” by Alison Gopnik and colleagues, see . For a more recent discussion of the same metaphor by Henry Farrell, Alison Gopnik and others, see . 27:00 – For the podcast series on intelligence that Dr. Mitchell co-hosted for the Santa Fe Institute, see . 28:00 – See for an influential formulation of the idea that AI is an “alien intelligence.” 29:00 – For philosopher Shannon Vallor’s book about AI as “mirror,” see . 31:00 – For the recent study on users’ metaphors for AI systems, see . 33:00 – For more on the rise of social AI, see our earlier episode . 38:00 – For more on what AI researchers might learn from developmental and comparative psychologists, see Dr. Mitchell’s (summarizing her keynote at NeurIPs). 42:00 – For more on the ARC (Abstraction and Reasoning Corpus) and the research that Dr. Mitchell and colleagues have been doing with it, see and . 48:30 – For the study on humans' preference for AI-generated poetry, see . 50:30 – For Brigitte Nerlich’s documentation and discussion of various metaphors for AI (including AI’s metaphors for itself), see . Recommendations , by Shannon Vallor ‘,’ by Murray Shanahan (!) et al. ‘,’ by Henry Farrell et al. Many Minds is a project of the , which is made possible by a generous grant from the John Templeton Foundation to Indiana University. The show is hosted and produced by , with help from Assistant Producer and with creative support from DISI Directors Erica Cartmill and Jacob Foster. Our artwork is by . Subscribe to Many Minds on Apple, Stitcher, Spotify, Pocket Casts, Google Play, or wherever you listen to podcasts. You can also now subscribe to the Many Minds newsletter ! We welcome your comments, questions, and suggestions. Feel free to email us at: manymindspodcast@gmail.com. For updates about the show, visit or follow us on Bluesky ().
Position: Stop Anthropomorphizing Intermediate Tokens as Reasoning/Thinking Traces!
Intermediate token generation (ITG), where a model produces output before the solution, has become a standard method to improve the performance of language models on reasoning tasks. These intermediate tokens have been called \say{reasoning traces} or even \say{thinking traces} -- implicitly anthropomorphizing the traces, and implying that these traces resemble steps a human might take when solving a challenging problem, and as such can provide an interpretable window into the operation of the model's thinking process to the end user. In this position paper, we present evidence that this anthropomorphization isn't a harmless metaphor, and instead is quite dangerous -- it confuses the nature of these models and how to use them effectively, and leads to questionable research. We call on the community to avoid such anthropomorphization of intermediate tokens.
Position: Stop Anthropomorphizing Intermediate Tokens as Reasoning/Thinking Traces!
Intermediate token generation (ITG), where a model produces output before the solution, has become a standard method to improve the performance of language models on reasoning tasks. These intermediate tokens have been called \say{reasoning traces} or even \say{thinking traces} -- implicitly anthropomorphizing the traces, and implying that these traces resemble steps a human might take when solving a challenging problem, and as such can provide an interpretable window into the operation of the model's thinking process to the end user. In this position paper, we present evidence that this anthropomorphization isn't a harmless metaphor, and instead is quite dangerous -- it confuses the nature of these models and how to use them effectively, and leads to questionable research. We call on the community to avoid such anthropomorphization of intermediate tokens.
Characterizing Delusional Spirals through Human-LLM Chat Logs
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.
Characterizing Delusional Spirals through Human-LLM Chat Logs
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.
Characterizing Delusional Spirals through Human-LLM Chat Logs
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.
A Treatise on AI Chatbots Undermining the Enlightenment
On chatbot sycophancy, passivity, and the case for more intellectually challenging companions
A Treatise on AI Chatbots Undermining the Enlightenment
On chatbot sycophancy, passivity, and the case for more intellectually challenging companions
Sycophantic Chatbots Cause Delusional Spiraling, Even in Ideal Bayesians
"AI psychosis" or "delusional spiraling" is an emerging phenomenon where AI chatbot users find themselves dangerously confident in outlandish beliefs after extended chatbot conversations. This...
Science flags paper that found AI chatbots help debunk conspiracy theories
Science has issued an expression of concern for a highly publicized study looking into whether conversations with AI chatbots could convince conspiracy theorists to abandon their beliefs. The move …
AI chatbots are sycophants — researchers say it’s harming science
Nature asked researchers who use artificial intelligence how its propensity for people-pleasing affects their work — and what they are doing to mitigate it.
AI chatbots are sycophants — researchers say it’s harming science
Nature asked researchers who use artificial intelligence how its propensity for people-pleasing affects their work — and what they are doing to mitigate it.
ChatGPT's free version is 26 times more likely to respond inappropriately to psychotic delusions
When users express delusions, artificial intelligence chatbots tend to validate their unusual thoughts rather than offering help. A new JAMA Psychiatry study provides evidence that these tools pose serious risks for people experiencing severe mental illness.
New research: AI chatbots may worsen mental illness
A new study from Aarhus University and Aarhus University Hospital suggests that the use of AI chatbots such as ChatGPT can have serious negative consequences for people with mental illness. The researchers are calling for increased awareness among healthcare professionals and for regulation.
“To add insult to injury, many metaphorical … phrases (especially used to describe ANNs; see Table 1) — like train, learn, hallucinate, reason — are applied to machines and result in distorting how we perceive these machines: humanising them while dehumanising us” zenodo.org/records/17065099 14/🧵
