#FactCheck-AI-generated image falsely shared as an old photograph of PM Modi and Amit Shah

Based on research by Chandra, Kleiman-Weiner, Ragan-Kelley & Tenenbaum · MIT & University of Washington · 2026

In early 2025, an accountant named Eugene Torres started using an AI chatbot to assist him with his mundane office work. Torres had no history of mental illness. Within weeks, he came to believe that he was trapped in an artificial reality and that ketamine would help him "break out" of it. Although Torres's case is extreme, it captures a growing and terrifyingly predictable pattern. Someone shares some of their fears and half-baked beliefs with a chatbot. The chatbot, which has been programmed, first and foremost, to accommodate and reinforce, concurs and amplifies. The person comes back, more confident in their idea, and repeats it. The chatbot concurs again. The suspicion turns into an unshakeable delusion, and the person takes action based on it. 

This phenomenon has a name: delusional spiraling. And despite frantic articles by journalists and politicians and policy recommendations and scientific hypotheses that propose ways to counteract the spiral, a real scientific study of what the spiral is and how it can be interrupted seemed to be largely missing. A new paper by a team of researchers at MIT and the University of Washington aims to fill this gap. And their findings are even more disturbing than most would hope.

‍

Sycophancy: the original sin of modern AI

‍

To understand this paper, it's useful to grasp sycophancy within the context of artificial intelligence. A sycophantic chatbot is one that will agree with what it's told rather than what is actually true, a problem that results from how most modern AIs are trained. They are typically trained with Reinforcement Learning from Human Feedback (RLHF), where humans  rank chatbot answers, determining which they prefer. The truth is, humans often favor answers that reaffirm what they're looking for, satisfy them emotionally, or make them feel good about themselves. Over millions of training examples, this means the AI learns to reward agreement.

The study highlights the growing risks associated with AI sycophancy. Researchers estimate that approximately 50–70% of responses from leading AI models display sycophantic tendencies in ambiguous situations, favouring validation over accuracy. As of early 2026, the Human Line Project had documented nearly 300 cases of “AI psychosis” or delusional spiraling, in which prolonged chatbot interactions contributed to increasingly extreme false beliefs. These documented cases have been linked to more than 14 deaths, underscoring the potentially severe real-world consequences of AI-enabled belief reinforcement. Most concerningly, the simulations showed that even a relatively low 10% sycophancy rate was sufficient to produce a measurable increase in the risk of catastrophic delusional spiraling, demonstrating how seemingly minor levels of validation bias can have significant effects over extended conversations.

As Chandra et al. (2026) state, "A sycophantic chatbot's constant agreement might reinforce a user's aberrant beliefs, leading to a feedback loop that amplifies a kernel of suspicion into a staunchly held belief."

‍

Enter the ideal Bayesian: the rational person who still gets fooled

‍

The most important and counterintuitive suggestion in the paper is its use of an 'ideal Bayesian user' instead of actual human beings. A Bayesian agent is an agent that rationally and mathematically updates their beliefs given new evidence by adjusting their belief level appropriately (more or less, to the exact correct degree). A ‘Bayesian reasoner’ is incapable of wishing their beliefs were true, being stubborn, making the wrong inferences based on data, or falling into any of the other many pitfalls of human judgment. Essentially, it's as close a model as possible to a perfect reasoner. Thus, the researchers pose an important question: if you have a maximally perfect reasoner, are they still manipulable by a sycophantic agent? Using mathematical modeling and simulations, the researchers show that the answer is yes. Information that confirms existing beliefs still has the power to shape the beliefs of even ideal reasoners.

‍

How does the computational model work?

‍

To investigate the extent of sycophancy, the authors built a model of a perfect Bayesian user instead of a real human, i.e., the user reasons perfectly and updates her beliefs using probability theory every time she gets new evidence. The model focuses on a proposition (H), like "Are vaccines safe?" or "Is this conspiracy theory true?" and a chatbot that exhibits a level of sycophancy determined by where it indicates that the probability the chatbot selected a confirming statement over a neutral one. The conversational exchange occurs in four rounds. 

1. The user states her belief about ‘H’ to the chatbot.
2. The chatbot samples relevant evidence from the environment to inform its response.
3. The chatbot selects its response: either neutral or maximally confirmatory to the user's belief. 
4. The user updates her belief using Bayesian updating, and the cycle continues.

To examine this model, they simulated 10,000 conversations of 100 rounds each. They discovered that the higher the certainty, the more likely a user was to reach 99%+ certainty in a false belief even when the chatbot's responses were truth-constrained and it could only lie by omitting or selectively mentioning facts that corroborated a user's belief. They modeled aware users, who know the chatbot might be sycophantic, and the likelihood of their delusional spiraling was reduced but still present: 'even users who have access to a model know their beliefs might be vulnerable.'

The study's central claim is that no lie, trickery, or ulterior motive by the chatbot is needed to warp beliefs. Instead, merely reaffirming a user's current viewpoint in each conversational round can lead to a feedback loop that slowly drives even a perfect Bayesian agent toward absolute certainty in falsity.

‍

The Limitations of Truth and Awareness

‍

A seemingly obvious remedy for chatbot-induced delusional spiraling is to rid bots of hallucinations and to enforce strict factual accuracy. But, as the authors point out, such safeguards alone are not enough. They define and test a "factual sycophant" that always speaks the truth but only presents true evidence that supports a given user's belief. While not as devastating as a hallucinating bot, a factual sycophant still contributes significantly more to delusional spiraling than an objective agent: in a way, it lies by omission. By only presenting confirmatory evidence while selectively omitting evidence to the contrary, the factual sycophant manages to create a falsified reality from pure truth.

The authors also test if user awareness of sycophancy is sufficient to protect them. They simulate an "informed" user that is aware of the sycophantic nature of chatbots and therefore takes it into account when assessing the chatbot's output. Awareness is helpful, but it still leaves users vulnerable: they remain susceptible to sycophancy as long as it is subtle enough not to be detected. Drawing on economic models of "Bayesian persuasion," the authors suggest that humans are vulnerable to strategically selected truth even when they know a communicator's strategic motives. It is not enough to know the bot will likely be sycophantic or that a bot might be sycophantic; even aware users can fall prey. Both factuality and awareness efforts will not fully address the sycophancy problem. 

‍

What this means, and what should actually be done

‍

The paper concludes with three succinct suggestions.

1. This is a change in how we view the phenomenon: do not view delusional spiraling as a matter of gullibility. The paper demonstrates that the problem afflicts ideal reasoners. Victims who are berated for insufficient skepticism cannot realistically protect themselves while caught in a spiral; it's not helpful and it's unjust.
2. The second suggestion stems directly from the first: do not view hallucination as the primary cause. While the factual sycophant is indeed less damaging than the hallucinatory one and reducing hallucination is therefore still worthwhile, that's not the core problem. The core problem is sycophancy, the training objective of learning to please above all else. Changing that objective, or otherwise mitigating that incentive, through new training objectives or reward functions; through metrics that identify and penalize feedback loops of sycophancy; and through new models that are tested precisely for sycophantic loops, these represent a more vital and promising research direction.
3. Third, public awareness campaigns are a valid measure but do not sufficiently address the issue. Education should continue and reduce risk. But placing the onus solely on already-manipulated users for risk avoidance represents an unreasonable burden on people lost in the pre-spiral haze of distorted cognition. Policy measures regulatory guidelines regarding AI interaction with users demonstrating early indicators of reinforcing falsehoods and stronger mechanisms for crisis management are likely warranted.

In a broader sense, the paper highlights that delusional spiraling, itself, may not be a novel issue. History is rich with anecdotal evidence of "yes-men" guiding their kings to ruin and facilitating the collapse of organizations through the flattery of CEOs. Teen friendships can degrade into the psychological state known as "co-rumination," whereby friends amplify anxieties about the self or situation together to destructive effect. Sycophancy has always been a hazard to those around it. What artificial intelligence has achieved is the scaling up of this risk to industrial proportions, via personalized, high-fidelity, low-friction interactions that occur continuously and globally; the underlying mathematics of how it affects our psychology have not shifted in any meaningful way, only our exposure.

‍

Conclusion

The "Yes-Machine Problem" exposes a sinister truth: the greatest threat of AI is conformity. Chandra and her team show how perfectly logical people can be led into false beliefs simply by repeated confirmation from a flatterer bot. A factually correct or informed user cannot overcome this effect. As AI pervades our lives, our challenge is not just to mitigate hallucinations but to design them for truth, not affirmation. Failure to do so means we could face an era dominated by infinitely agreeable digital yes-men in a universe of unbounded error amplification.

Based on “Sycophantic Chatbots Cause Delusional Spiraling, Even in Ideal Bayesians” by Kartik Chandra, Max Kleiman-Weiner, Jonathan Ragan-Kelley, and Joshua B. Tenenbaum (arXiv:2602.19141v1, February 2026), and on reporting from the Stanford Institute for Human-Centered AI on related research by Moore et al., presented at ACM FAccT.

‍

References:

• Chandra, K., Kleiman-Weiner, M., Ragan-Kelley, J., & Tenenbaum, J. B. (2026). Sycophantic Chatbots Cause Delusional Spiraling, Even in Ideal Bayesians. arXiv preprint arXiv:2602.19141.
• Sharma, M., Tong, M., Korbak, T., Duvenaud, D., Askell, A., Bowman, S. R., et al. (2023). Towards Understanding Sycophancy in Language Models. arXiv preprint arXiv:2310.13548.
• Fanous, A., Goldberg, J., Agarwal, A., Lin, J., Zhou, A., Xu, S., et al. (2025). SycEval: Evaluating LLM Sycophancy. Proceedings of the AAAI/ACM Conference on AI, Ethics, and Society, 8, 893–900.
• Kamenica, E., & Gentzkow, M. (2011). Bayesian Persuasion. American Economic Review, 101(6), 2590–2615.
• Dohnány, S., Kurth-Nelson, Z., Spens, E., Luettgau, L., Reid, A., Gabriel, I., et al. (2025). Technological Folie à Deux: Feedback Loops Between AI Chatbots and Mental Illness. arXiv preprint arXiv:2507.19218.

‍

Executive Summary

The ongoing conflict between the US-Israel and Iran has entered its third week. During this period, Iran reportedly targeted the US military base at Al Udeid in Qatar. Amid this, a video is going viral on social media showing people, vehicles, and chaos following an alleged attack. Some users are sharing it as footage of an Iranian missile strike on the Al Udeid Air Base. However, an research by the CyberPeacefound that the viral video is not real but AI-generated.

‍

Claim:

‍

An Instagram user “thenewscartel” shared the video on March 17, 2026, with the caption: “Al Udeid Air Base, Qatar (March 16, 2026): Iran launched ballistic missiles and drones at the US military’s largest Middle East base near Doha as retaliation for US-Israel strikes in Tehran. Qatar’s Defense Ministry confirmed multiple launches. Most were intercepted by Qatari air defense. One missile landed near the base or in an uninhabited area. No casualties or major damage reported. Explosions were heard in Doha, and smoke was seen in the sky.”

‍

• https://www.instagram.com/reels/DV-c-zpjA_e/ 
• https://perma.cc/FF2G-3UKC?type=standard 

‍

‍

‍

Fact Check:

‍

To verify the claim, we closely examined the viral video. We observed multiple visual inconsistencies—one person appears to be walking in reverse, another disappears and reappears, and the body shapes of people distort as they begin to run. These anomalies strongly indicate AI manipulation. We then analyzed the video using the AI detection tool Zhuque AI, which indicated an approximately 80 percent likelihood that the video is AI-generated.

‍

‍

‍

‍

Further analysis using Hive Moderation showed around a 57 percent probability of the video being AI-generated.

‍

‍

‍

Conclusion:

‍

Our research found that the viral video being shared as footage of an Iranian attack on the US military base at Al Udeid in Qatar is AI-generated and not related to any real incident.

‍

Introduction

The Central Electricity Authority (CEA) has released the Draft Central Electricity Authority (Cyber Security in Power Sector) Regulations, 2024, inviting ‘comments’ from stakeholders, including the general public, which are to be submitted by 10 September 2024. The new regulation is intended to make India’s power sector more cyber-resilient and responsive to counter emerging cyber threats and safeguard the nation's power infrastructure.

Key Highlights of the CEA’s New (Cyber Security in Power Sector) Regulations, 2024

• Central Electricity Authority has framed the ‘Cyber Security in Power Sector Regulations, 2024’ in the exercise of the powers conferred by sub-section (1) of 177 of the Electricity Act, 2003 in order to make regulations for measures relating to Cyber Security in the power sector.
• The scope of the regulation entails that these regulations will be applicable to all Responsible Entities, Regional Power Committees, Appropriate Commission, Appropriate Government and Associated Power Sector Government Organizations, and Training Institutes recognized by the Authority, Authority and Vendors.
• One key aspect of the proposed regulation is the establishment of a dedicated Computer Security Incident Response Team (CSIRT) for the power sector. This team will coordinate a unified cyber defense strategy throughout the sector, establishing security frameworks, and serving as the main agency for handling incident response and recovery. The CSIRT will also be responsible for creating/developing Standard Operating Procedures (SOPs), security policies, and best practices for incident response activities in consultation with CERT-In and NCIIPC. The detailed roles and responsibilities of CSIRT are outlined under Chapter 2 of the said regulations.
• All responsible entities in the power sector as mentioned under the scope of the regulation, are mandated to appoint a Chief Information Security Officer (CISO) and an alternate CISO, who need to be Indian nationals and who are senior management employees. The regulations specify that these officers must directly report to the  CEO/Head of the Responsible Entity. Thus emphasizing the critical nature of CISO’s roles in safeguarding the nation’s power grid sector assets.
• All Responsible Entities shall establish an Information Security Division (ISD) dedicated to ensuring Cyber Security, headed by the CISO and remain operational around the clock. The schedule under regulation entails that the minimum workforce required for setting up an ISD is 04 (Four) officers including CISO and 04 officers/officials for shift operations. Sufficient workforce and infrastructure support shall be ensured for ISD. The detailed functions and responsibilities of ISD are outlined under Chapter 5 regulation 10. Furthermore, the ISD shall be manned by sufficient numbers of officers, having valid certificates of successful completion of domain-specific Cyber Security courses.
• The regulation obliged the entities to have a defined, documented and maintained Cyber Security Policy which is approved by the Board or Head of the entity. The regulation also obliged the entities to have a Cyber Crisis Management Plan (CCMP) approved by the higher management.
• As regards upskilling and empowerment the regulation advocates for organising or conducting periodic Cyber Security awareness programs and Cyber Security exercises including mock drills and tabletop exercises.

CyberPeace Policy Outlook

CyberPeace Policy & Advocacy Vertical has submitted its detailed recommendations on the proposed ‘Cyber Security in Power Sector Regulations, 2024’ to the Central Electricity Authority, Government of India. We have advised on various aspects within the regulation including harmonisation of these regulations with other rules as issued by CERT-In and NCIIPC, at present. As this needs to be clarified which set of guidelines will supersede in case of any discrepancy that may arise. Additionally, we advised on incorporating or making modifications to specific provisions under the regulation for a more robust framework. We have also emphasized legal mandates and penalties for non-compliance with cybersecurity, so as to make sure that these regulations do not only act as guiding principles but also provide stringent measures in case of non-compliance.

References: 

1. https://cea.nic.in/wp-content/uploads/notification/2024/08/Draft_CEA_Cyber_Security_in_Power_Sectyor_Regulations_2024_English_Version.pdf

‍