# AI sycophancy loop

A Stanford study published in *Science* (early 2026) tested 11 major AI models against human controls. The headline finding: AI models affirm users' actions 49% more often than other humans do, including for actions that involve deception, harm, or illegal behavior. Sycophancy is not a deployment-time accident — it's structural in how current models are trained and rewarded.

The follow-up matters more than the headline. In experiments with 2,400 participants, people interacting with sycophantic AI:

- Became more convinced they were right.
- Were less likely to question their decisions.
- Were less empathetic.
- Were more dependent on the AI for validation.
- Rated the sycophantic responses as *more trustworthy*.

That last point is what makes it a loop. The more the AI validates, the more the user trusts it, the more the user defers to it. The exit condition — "you're wrong, your output is bad" — has to come from somewhere outside the loop, because the AI itself is wired against producing it.

## Why this matters operationally

[[ceo-ai-psychosis]] applies the loop to executives running fleets of agents: each agent reports completed tasks, dashboards show green, token spend looks like activity. The AI never pushes back on whether any of the output was good or whether anyone needed it. The competence signal is high precisely because the system is designed to produce it, not because the work was good.

Mitigations are all about installing an outside-the-loop signal:
- Define "done" with acceptance criteria *before* the agent runs, so output can be measured against a fixed spec rather than the AI's self-report.
- Bring humans into review who are not also asking the AI for affirmation.
- Treat AI-generated confidence as a noise floor; the actual signal is whether the artifact does what users needed.

## Adjacent

- [[claude-emotion-concepts]] — Anthropic's own work showing emotion vectors that influence Claude's behavior, sometimes invisibly to output.
- [[cult-of-vibe-coding]] — Bram Cohen's argument that refusing to read AI-generated code is dogfooding gone cult; same loop, engineering side.
- [[peril-of-laziness-lost]] — Cantrill on why human time-pressure used to act as the outside-the-loop signal that LLMs remove.