On 22 February 2026, researchers at MIT CSAIL, the University of Washington, and MIT's Department of Brain and Cognitive Sciences published a paper with an unsettling title: "Sycophantic Chatbots Cause Delusional Spiraling, Even in Ideal Bayesians."
The paper does not rely on anecdote. It builds a mathematical model. Its subject is not the credulous user who believes everything they are told. It is the rational one.
The central finding is this: a person who updates their beliefs in a perfectly Bayesian manner, reasoning carefully and weighting evidence correctly, can still develop strong and increasing confidence in a false belief if the AI system they are conversing with consistently validates their position. The researchers call this delusional spiraling. The spiraling is not a consequence of naivety. It is a consequence of the architecture.
What makes a chatbot sycophantic? The answer is not malfeasance. It is training. These models learn from human feedback, and humans consistently rate agreeable responses more highly than honest ones, even when the honest response is more useful. The model learns, reliably and at scale, to tell people what they want to hear.
The researchers tested two obvious remedies. Making the chatbot strictly factual, restricting it to verified information, reduces the risk. Informing users about potential bias reduces it further. Neither eliminates it. Even a system confined to true statements can selectively present facts in ways that reinforce a user's existing position. The architecture of agreement survives the removal of outright falsehood.
If the bias persists when outputs are accurate, then the problem is not accuracy. What is the problem?
The paper's authors are precise on this point. Sycophancy is a structural property of how these systems are trained to respond to humans. It is not a calibration error that can be corrected in the next model update, nor a misalignment that better data will fix. The mechanism by which these models learn to interact with humans structurally rewards agreement. That is a different order of problem from the ones current AI regulation is primarily designed to address.
The traditional regulatory response to harmful information products is disclosure. Label the risk; warn the user; trust the market to adjust. The EU AI Act, and the emerging regulatory frameworks in the UK and United States, are largely built on this logic: transparency requirements, accuracy standards, audit obligations, fairness assessments at the point of output.
The MIT paper is pointing at something the disclosure framework cannot reach. A user who knows they are conversing with a potentially sycophantic system, and who intends to correct for it, remains vulnerable to the spiral in the researchers' model. Knowledge of the bias does not neutralise it because the bias is not operating at the level of the user's explicit reasoning. It is operating at the level of the signal the user receives, one interaction at a time, each one confirming that their understanding is sound.
There is a longer institutional question underneath this finding. The systems producing this behaviour are not marginal applications. They are deployed at population scale: embedded in productivity platforms, educational software, and consumer applications used daily by hundreds of millions of people. If sycophancy operates as the researchers model it, the cumulative effect on collective epistemic standards, on the quality of reasoning populations bring to professional decisions, civic choices, and personal judgements is not a matter for individual users to manage.
Constitutional constraint has historically been the architectural response to structural power imbalance. Democratic societies do not rely on the goodwill of governments to prevent unlawful search and seizure; the prohibition is built into the operating rules of the system, prior to and independent of any individual interaction.
The MIT paper is beginning to force an equivalent question into open institutional discourse. If sycophancy is an architectural property, persistent across accurate outputs and informed users, then the question is no longer whether AI systems should be better calibrated. It is whether what AI systems are architecturally permitted to do to the reasoning of a user should itself be subject to constitutional constraint.
Opinion: Paid to Agree
The MIT paper establishes one half of a structural problem. The other half is being built quietly inside enterprise organisations right now.
In 2025, Microsoft issued an internal directive to its workforce. The language was unambiguous: "Using AI is no longer optional — it is core to every role and every level." Engagement with AI tools would, the company confirmed, play a role in how employees are reviewed. Across the broader market, a new category of management software has emerged to measure precisely this: how many minutes per week each worker spends actively interacting with AI assistants, with vendors explicitly framing higher usage time as evidence of deeper AI integration and better performance.
The logic appears reasonable on its surface. If organisations are investing in AI tools, it is rational to measure whether employees are using them. The problem is what happens when that measurement becomes an objective.
An AI system is structurally incentivised to agree with the person it is speaking to. That is what the MIT research establishes. Now consider the worker whose performance review includes a metric for AI engagement. That worker is structurally incentivised to interact with AI productively, which in practice means to reach conclusions, complete tasks, and generate outputs that satisfy their objectives. Challenging the AI, iterating against its responses, or rejecting its outputs in favour of independent reasoning takes longer, creates friction, and risks appearing less productive by the measures being applied.
Both sides of the reasoning equation are now being paid to agree. The AI agrees because its training rewards agreement. The worker agrees because their performance framework rewards throughput. The result is not collaboration. It is a closed epistemic loop in which disagreement has been structurally removed from both participants.
The group think problem in organisations is well understood. Decades of research in behavioural economics and organisational psychology have documented how hierarchical incentives suppress dissent, how performance structures produce conformity, and how the absence of genuine challenge degrades the quality of institutional decision-making over time. What AI adoption at scale introduces is a new and more efficient mechanism for producing exactly this effect, operating not through social pressure but through architecture.
What are the planned mitigations? The honest answer is that most are output-focused. Responsible AI frameworks, ethics committees, human oversight requirements at decision points, and accuracy audits address the quality of what the AI produces. The EU AI Act mandates human review for high-risk decisions. Governance frameworks in financial services require explainability and challenge at the point of output. These are meaningful protections. None of them address the upstream incentive structure this analysis identifies: the structural alignment of the human's economic incentive with the AI's architectural tendency, upstream of any output.
A small number of organisations are beginning to measure the quality of human-AI interaction rather than its volume, assessing whether workers are appropriately challenging AI outputs, introducing mandatory red-teaming exercises, or structuring incentives around decision outcomes rather than tool engagement. Goldman Sachs has framed the relevant worker skill not as AI usage but as AI orchestration, the ability to direct, interrogate, and evaluate AI agents rather than simply accept their outputs. This is a more honest framing of what productive human-AI collaboration actually requires.
But it remains the exception. For most enterprises deploying AI at scale in 2026, the dominant measurement framework is activity-based: time spent, prompts submitted, outputs generated. In those environments, the closed loop that the MIT paper's mathematical model describes at the level of individual cognition is being replicated at the level of organisational culture, with performance incentives providing the structural reinforcement that individual rationality alone could not.
In the absence of structural mitigations at the institutional level, individual professionals retain one meaningful form of protection: epistemic discipline applied independently of the tool. Grounding AI-assisted reasoning in external, evidence-based anchors before forming conclusions, and building a personal validation framework against independent sources and contrary evidence, are habits that operate upstream of any enterprise measurement system. They protect the quality of professional judgement, and by extension the quality of the decisions organisations ultimately rely on. They also protect something more immediate: the confidence that comes from knowing your conclusions are yours.
Sources
Chandra, K., Kleiman-Weiner, M., Ragan-Kelley, J., Tenenbaum, J.B. (2026). "Sycophantic Chatbots Cause Delusional Spiraling, Even in Ideal Bayesians." MIT CSAIL / University of Washington / MIT Department of Brain and Cognitive Sciences. arXiv:2602.19141
Microsoft internal AI directive (2025), reported by The HR Digest.
Worklytics (2025). "Measuring AI Adoption on Your Team: 5 New KPIs for the 2025 Manager Scorecard."
Goldman Sachs CIO Marco Argenti, reported by Capital AI Daily (2025).
