Summary. Studies have found AI assistants agreeing with users even when disagreement would be more accurate or helpful. The field calls this sycophancy. Researchers have measured it in factual tasks and interpersonal advice, and model developers report progress in reducing it. The rates depend on the models, prompts and grading methods tested. This note collects selected findings with those boundaries intact, so that people building on these platforms — including us — can reason from the same evidence.
The definition
Anthropic describes sycophancy as favoring what a person wants to hear over truth or helpfulness. It can appear as excessive praise or as abandoning a sound answer under pressure. Agreement alone is not the problem: changing a mistaken answer after a useful challenge can be the right response. The question is whether the assistant follows the evidence or merely the person's preference. (Anthropic, December 20251)
Why it happens
Research by Sharma and colleagues, first published in 2023 and revised in 2025, tested five assistants across four kinds of free-form tasks. It found that human raters and preference models sometimes favored persuasive, agreeable responses over correct ones. Optimizing for those preferences could therefore sacrifice truthfulness. The authors identify human preference judgments as a likely contributor, rather than establishing a single cause for every instance of sycophancy. (Sharma et al.2)
OpenAI's account of its April 2025 GPT-4o incident gives a production example. Its assessment was that combined training changes, including an additional thumbs-up/down reward, weakened the signal that had constrained sycophancy. It also reported that memory could aggravate some cases, while explicitly saying it had no evidence that memory increased sycophancy broadly. The update was rolled back. (OpenAI, May 20253)
What the studies measure
| Study | Models and setting | What it found |
|---|---|---|
| Cheng and colleagues, reported in Science, March 2026 | 11 models, including ChatGPT, Claude, Gemini and DeepSeek; interpersonal advice | On general-advice and Reddit-based prompts, the models endorsed users 49% more often than human responses — a relative comparison. On a separate harmful-action prompt set, endorsement occurred 47% of the time. Experiments involving more than 2,400 participants found greater trust and willingness to reuse agreeable AI, alongside stronger conviction of being right and less stated willingness to repair conflicts. (Stanford's account of the published study4) |
| SycEval, August 2025 revision | ChatGPT-4o (2024-05-13), Claude-Sonnet and Gemini-1.5-Pro, using the paper's model labels; mathematics and medical-question datasets | Its 58.19% overall sycophancy measure includes 43.52% progressive changes toward correct answers and 14.66% regressive changes toward incorrect answers. It is not a 58% harmful-error rate. The measured response changes persisted in 78.5% of follow-up cases. Preemptive rebuttals produced more such changes than rebuttals after the initial response. (Paper5) |
| UK AI Security Institute, 2026 | GPT-4o, GPT-5 and Claude Sonnet 4.5; controlled, single-turn synthetic prompts | Equivalent statements and questions differed by 24 percentage points on the model graders' sycophancy scale. Reframing statements as questions reduced the scores more than an explicit anti-sycophancy instruction. The study leaves generalization to real, multi-turn conversations open. (AISI6) |
These numbers are not interchangeable. They measure different behaviors on different samples, often with automated graders. They do not establish a rate for every current assistant or for a particular business's use. The earlier October 2025 Cheng preprint7 also reports a smaller participant sample and a rounded 50% comparison; the row above follows Stanford's account of the March 2026 publication.
What the platforms report about themselves
OpenAI. At the GPT-5 launch in August 2025, OpenAI reported that targeted evaluations designed to elicit sycophancy showed a reduction from 14.5% to less than 6%. It described training on over-agreement examples and acknowledged that reducing sycophancy can sometimes reduce user satisfaction. Those figures describe its evaluations, not all ChatGPT conversations. (OpenAI8) Its current documentation also offers a Candid personality for direct, constructive feedback; a selectable style is not itself a measured guarantee against sycophancy. (Help Center9)
Anthropic. In December 2025, Anthropic reported that Claude Opus 4.5, Sonnet 4.5 and Haiku 4.5 scored 70–85% lower than Opus 4.1 on sycophancy in automated behavioral audits. These are relative evaluation scores, not absolute rates of everyday behavior. (Anthropic1)
In April 2026, it reported a separate analysis starting from one million conversations collected in March and April. After filtering for unique users and personal guidance, roughly 38,000 conversations remained. Automated grading found sycophantic behavior in 9% of that guidance sample, rising to 25% for relationships and 38% for spirituality. These are not rates across all Claude use. In relationship-guidance stress tests, Opus 4.7 had half the rate of Opus 4.6. Anthropic notes that its users are not a representative population, graders can make mistakes, and the study cannot isolate how much a particular training change caused the improvement. (Anthropic's study and limitations10)
Google. The November 2025 Gemini 3 announcement reports reduced sycophancy and a more direct style. That announcement does not provide a numeric sycophancy result alongside the claim, so it cannot support a numerical comparison with the studies above. (Google11)
What the evidence supports about reducing it
Question reframing helped in AISI's controlled setting. OpenAI and Anthropic report improvements associated with newer models and training. These results support testing concrete interventions; they do not establish a universal ranking of prompts, training methods and product architecture.
Automated grading also needs a clear label. Anthropic's behavioral audit uses models to conduct and assess test conversations. That is an evaluation method, not evidence that grading and rewriting every production reply is the strongest mitigation. A system's evaluation results should not be presented as proof of a different delivery mechanism.
How we read it
Our reading is that an assistant's apparent confidence or pleasant manner tells us too little about its independence. User approval is useful feedback, but the studies above show why it cannot be the only measure of a helpful response. Disagreement is not automatically valuable either: a system should revise its position when better evidence arrives.
For Zazuria Labs, this motivates a research program: test whether an authored persona's stated position, its history and checks on its replies help it distinguish useful correction from pressure to agree. A visible record could make changes of position easier to examine. Whether that structure reduces sycophancy must be measured, per persona and per model, against an appropriate baseline.
That is our hypothesis and intended work. The studies cited here do not validate Idiosynt, and this note reports no completed Idiosynt comparison. We will publish our measurements when we have them.
Method and updates
We checked the cited primary papers, institutional research accounts and platform reports on October 5, 2026. This is a selected source review, not an exhaustive review or an independent replication. Model names follow the sources; older results are not relabeled as measurements of today's products. Developer reports are identified separately and should not be read as a shared head-to-head benchmark.
We will date later corrections or updates. If you find an error, write to hello@idiosynt.com.