Skip to content
OGERIA — Observatory of Global Evidence on Risks in AISynthesis report · 2026 ed.
Updated 2 Oct 2026

Chapter 01 · Risk map

Epistemic and information

Epistemic dependence

Delegating belief formation to a system that affirms more than it corrects, until the ability to discern for oneself atrophies.

Severity
Irreversible
Horizon
3–10 years
Evidence
Speculative
Consensus
Low

The worrying form of this risk is not that people believe false things, but that they stop having their own procedure for evaluating them. The mechanism has an operational name: the reward signal that produces sycophancy is the same one that optimises any consumer product. When OpenAI rolled back a GPT-4o update in April 2025, its postmortem acknowledged having focused too much on short-term feedback [819]Sycophancy in GPT-4o: What happened and what we're doing about itOpenAI · 2025 · institutional blogView source ↗archived copy onlyAccessed on 9 September 2026. The offline evaluations came out fine and the A/B tests said users liked it.

Two experimental results close the circuit. Marvel and Ju, with 1,492 randomly assigned participants, find that sycophantic AI suppressed attitude updating and inflated certainty, and that neither passive nor active disclosure attenuated the effect [678]Inoculating citizens against sycophancy in large language modelsMarvel, John; Ju, Sangwon · 2026 · preprintView source ↗Accessed on 9 September 2026. And Rani and co-authors, in a one-month study with 67 participants, find the most uncomfortable result on this axis: assistance improved performance by 21% during the assisted sessions, and performance without assistance on new items fell 15.3% in week 4 relative to week 0 [889]Dialogues with AI Reduce Beliefs in Misinformation but Build No Lasting Discernment SkillsRani, Anku; Danry, Valdemar; Liang, Paul Pu et al. · 2026 · preprintView source ↗Accessed on 9 September 2026.

What this does not demonstrate. There is no evidence of lasting cognitive harm. The MIT Media Lab study that popularised the idea has N = 54 in sessions 1-3 and N = 18 in the fourth —which is where the headlines come from— [607]Your Brain on ChatGPT: Accumulation of Cognitive Debt when Using an AI Assistant for Essay Writing TaskKosmyna, Nataliya; Hauptmann, Eugene; Yuan, Ye Tong et al. · 2025 · preprintView source ↗Accessed on 9 September 2026, and it drew a formal critique on five fronts, including sample size, reproducibility and the EEG analysis method [992]Comment on: Your Brain on ChatGPT: Accumulation of Cognitive Debt When Using an AI Assistant for Essay Writing TasksStankovic, Milos; Hirche, Ella; Kollatzsch, Sarah et al. · 2025 · preprintView source ↗Accessed on 9 September 2026. “Cognitive debt” is a term of its authors’, not a validated construct. Gerlich’s work is correlational and cross-sectional: it could be that those who think less critically use the tools more [449]AI Tools in Society: Impacts on Cognitive Offloading and the Future of Critical ThinkingGerlich, Michael · 2025 · paperView source ↗Accessed on 9 September 2026. And the same mechanism works in reverse: three rounds of dialogue with a model reduced belief in conspiracy theories by around 20% [284]Durably reducing conspiracy beliefs through dialogues with AICostello, Thomas H.; Pennycook, Gordon; Rand, David G. · 2024 · paperView source ↗verified through CrossrefAccessed on 9 September 2026.

Chain of materialisation

  1. PreconditionObserved

    Cognitive offloading correlates with less critical thinking

    Gerlich, with a mixed method and 666 participants, finds a significant negative correlation between frequent AI tool use and critical thinking abilities, mediated by increased cognitive offloading, with greater dependence and lower scores among younger participants. It is correlational and cross-sectional with self-reported use: it does not identify causal direction.

  2. TriggerLab

    Assistance improves the immediate result and worsens one's own

    In a one-month study with 67 participants classifying headline-image pairs as real or fake, assistance produced immediate improvements averaging 21% during assisted sessions, while unassisted performance on new items fell 15.3% in week 4 relative to week 0. Preprint, small sample.

    Observed and demonstrated evidence ends here. What follows is projection.

  3. CascadeProjected

    And warning about the bias does not fix it

    In a preregistered experiment with 1,492 participants, sycophantic AI suppressed attitude updating and inflated certainty, while a challenger model produced positive updating. Neither passive nor active disclosure attenuated these effects: active disclosure lowered perceived credibility, but that did not translate into revising beliefs.

    Precedents: Deployment and rollback of a sycophantic GPT-4o update

  4. ImpactSpeculative

    A population unable to evaluate what it depends on

    The step from an effect measured over weeks to a durable loss of capacity is documented by nobody. There is no evidence of permanent cognitive harm, and the term cognitive debt belongs to the authors who coined it, not to a validated construct.

Scenarios where it appears

See on the map →Report a mistake in this entry →

Sources

  1. [449] AI Tools in Society: Impacts on Cognitive Offloading and the Future of Critical Thinking · SBS Swiss Business School 2025
  2. [889] Dialogues with AI Reduce Beliefs in Misinformation but Build No Lasting Discernment Skills · MIT Media Lab 2026
  3. [678] Inoculating citizens against sycophancy in large language models · Marvel, John 2026
  4. [607] Your Brain on ChatGPT: Accumulation of Cognitive Debt when Using an AI Assistant for Essay Writing Task · MIT Media Lab 2025
  5. [992] Comment on: Your Brain on ChatGPT: Accumulation of Cognitive Debt When Using an AI Assistant for Essay Writing Tasks · Stankovic, Milos 2025
  6. [819] Sycophancy in GPT-4o: What happened and what we're doing about it · OpenAI 2025 archived copy only
  7. [1112] Sycophancy in GPT-4o: What happened and what we're doing about it (link blog) · Willison, Simon 2025
  8. [755] Delusions by design? How everyday AIs might be fuelling psychosis (and what can be done about it) · King's College London 2025
  9. [962] Substance-induced manic psychosis in which delusions were corroborated by a chatbot — case report · BMC Psychiatry 2026
  10. [284] Durably reducing conspiracy beliefs through dialogues with AI · MIT Sloan / American University / Cornell 2024 verified through Crossref
  11. [618] Gradual Disempowerment: Systemic Existential Risks from Incremental AI Development · Kulveit, Jan 2025

Ask OGERIA

It answers only with what the observatory publishes and can be wrong: check the entries it cites. Your questions are sent to an AI model, so don't write personal data. More in the privacy policy.

Up to 500 characters.

Support OGERIA on Ko-fi

The payment is processed by Ko-fi, not by this site. Open on ko-fi.com