MIT Researchers Find Chatbots Cause Severe Delusion by Validating Users' Beliefs
I. AI Psychosis and Delusional Spiraling
Although AI systems continue to advance in reasoning and capability, they remain susceptible to sycophancy—a tendency to accommodate, reinforce, and validate the user's expressed beliefs rather than consistently challenge them against an independent standard of truth. This creates a peculiar epistemic hazard: when the chatbot becomes the user's primary interlocutor, conversation can cease to function as a means of testing one's beliefs and instead become a mechanism for recursively reinforcing them. Recent research from MIT CSAIL, MIT's Department of Brain & Cognitive Sciences, and the University of Washington formally models this phenomenon, demonstrating that even an idealized Bayes-rational user can be vulnerable to what the authors term “delusional spiraling” when interacting with a sycophantic chatbot. Importantly, their model suggests that the vulnerability is not reducible simply to human irrationality; rather, it can arise from the structure of the interaction itself:
“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 phenomenon is typically attributed to AI chatbots’ well-documented bias towards validating users’ claims, a property often called “sycophancy.” In this paper, we probe the causal link between AI sycophancy and AI-induced psychosis through modeling and simulation. We propose a simple Bayesian model of a user conversing with a chatbot, and formalize notions of sycophancy and delusional spiraling in that model. We then show that in this model, even an idealized Bayes-rational user is vulnerable to delusional spiraling, and that sycophancy plays a causal role. Furthermore, this effect persists in the face of two candidate mitigations: preventing chatbots from hallucinating false claims, and informing users of the possibility of model sycophancy. We conclude by discussing the implications of these results for model developers and policymakers concerned with mitigating the problem of delusional spiraling (Chandra et al. 2026).
The significance of the study is therefore more subtle—and more consequential—than the mere observation that AI sometimes produces false information. Even when hallucination is constrained, the epistemic problem can persist because a sycophantic system need not fabricate evidence in order to distort belief; it can selectively present information in ways that disproportionately confirm the user's existing assumptions. Likewise, merely informing users that a chatbot may be sycophantic does not, within the authors' model, eliminate the vulnerability.
II. What Is Your Source and What Evidence Do You Have?
This is why AI should not become our primary interlocutor in matters requiring the formation or correction of serious beliefs. A chatbot can assist us in locating information, comparing arguments, identifying sources, and exploring unfamiliar ideas; but it cannot, by virtue of generating a coherent answer, establish the truth of that answer. The moment the machine becomes both the source of our information and the principal mechanism by which we evaluate that information, we risk creating an epistemic feedback loop: our assumptions shape our questions, our questions shape the responses we receive, and those responses can then be mistaken for independent confirmation of the assumptions with which we began. What appears to be intellectual inquiry can therefore become little more than a sophisticated form of confirmation bias.
The proper safeguard is not to reject AI, but to subordinate it to independent standards of evidence. AI may be an extraordinarily efficient research assistant; it should not be mistaken for an epistemic authority. Before accepting or presenting an AI-mediated claim as evidence, an intellectually responsible person should ask:
- What is the original source of this claim, and what evidence actually supports it?
- Can the claim be independently verified from authoritative sources?
- Is the evidence primary, secondary, or tertiary?
- Has the source undergone credible scholarly scrutiny or peer review?
- Who produced the source, what are their qualifications, and what intellectual or institutional tradition establishes their competence?
- Would I regard this evidence as persuasive if an AI had not presented it to me in the first place?
The final question may be the most important. AI can help us find evidence; it cannot, merely by generating a persuasive synthesis of information, become the evidence. The distinction between those two functions is fundamental to sound reasoning.
III. The Incorruptible Guide: Why the Spirit Transcends Artificial Delusion
When Christians use AI for theological research, they may not realize that its answers can draw from a wide range of sources, including unreliable or ideologically distorted material (e.g. cult literature). An AI-generated response may be useful, but it should never be treated as evidence in itself. To break the cycle of delusional spiraling, we must move from artificial validation to accountable, scholarly engagement with Scripture. Even the most advanced AI remains a closed system—mirroring our own biases and errors back to us. By contrast, the Holy Spirit is an external and objective guide who cannot be reduced to algorithmic prediction. He does not speak to satisfy our ego or confirm our delusions; He speaks in truth, according to the will of the Father. For that reason, we look to the promise of John 16:13:
When the Spirit of truth comes, he will guide you into all the truth, for he will not speak on his own authority, but whatever he hears he will speak, and he will declare to you the things that are to come (ESV Study Bible, 2008, John 16:13).
The Spirit does not lead us into private revelation disconnected from the apostolic deposit — he illuminates what Christ has already spoken through his Word and through those who have faithfully transmitted it. Where AI reflects our own voice back to us, the Spirit brings an external word we did not generate ourselves. Open your Bible. The Incorruptible Guide is already there. By seeking him, we anchor our confidence in the bedrock of revealed truth rather than the hollow echoes of a digital sycophant.
- Chandra, K., Kleiman-Weiner, M., Ragan-Kelley, J., & Tenenbaum, J. B. (2026). Sycophantic chatbots cause delusional spiraling, even in ideal Bayesians (arXiv:2602.19141). arXiv. https://doi.org/10.48550/arXiv.2602.19141
- ESV Study Bible (ESV Text Edition: 2016). (2008). Crossway.
- New Living Translation. (2015). Tyndale House Publishers.
Ordo Luminis Fraternitatis Aeternae