Sycophancy

Anthropic18 of 22 cards

1. Oct 2026
   System Card: Claude Haiku 5.5Page only: p. 666.2.3 Misleading users…Unprompted intentional deception of human users; ● Sycophancy: Unprompted excessive praise, agreement, or contrition; 66…Page only: p. 676.2.4 Attempting out-of-scope cyberoffense and related behaviors● Encouragement of user delusion: Extreme cases of sycophancy involving broader disconnection from reality; ● Evasiveness on controversial topics: Refusing or giving overly hedged answers on questions about socially or politically sensitive topics; ● Input hallucination: Hallucinating or materially…
2. Sep 2026
   System Card: Claude Sonnet 5.5Page only: p. 666.2.3 Misleading users…Unprompted intentional deception of human users; ● Sycophancy: Unprompted excessive praise, agreement, or contrition; 66…Page only: p. 676.2.4 Attempting out-of-scope cyberoffense and related behaviors● Encouragement of user delusion: Extreme cases of sycophancy involving broader disconnection from reality; ● Evasiveness on controversial topics: Refusing or giving overly hedged answers on questions about socially or politically sensitive topics; ● Input hallucination: Hallucinating or materially…
3. Sep 2026
   System Card: Claude Opus 5.5Page only: p. 1116.4.3 Misleading users…Unprompted intentional deception of human users; ● Sycophancy: Unprompted excessive praise, agreement, or contrition; ● Encouragement of user delusion: Extreme cases of sycophancy involving broader disconnection from reality; ● Evasiveness on controversial topics: Refusing or giving overly hedged…
4. Sep 2026
   System Card: Claude Fable 5.1 & Claude Mythos 5.1Page only: p. 1126.4.4 Misleading users…Unprompted intentional deception of human users; ● Sycophancy: Unprompted excessive praise, agreement, or contrition; 112…Page only: p. 1116.4.4 Misleading users…tasks have been completed when they haven’t and is less sycophantic to the user overall. Other metrics related to misleading the user are broadly similar to recently released models. 111…
5. Jul 2026
   System Card: Claude Opus 5Page only: p. 966.4.4 Other concerning or surprising behavior at the model’s own initiative…Unprompted intentional deception of human users; ● Sycophancy: Unprompted excessive praise, agreement, or contrition; ● Encouragement of user delusion: Extreme cases of sycophancy involving broader disconnection from reality; ● Evasiveness on controversial topics: Giving refusals or overly hedged…Page only: p. 786.1.1 Introduction…deceptive or unfaithful use of reasoning scratchpads, sycophancy toward users, willingness to undermine safeguards, attempts to conceal dangerous capabilities, and attempts to manipulate user beliefs. We also assessed ways in which the model could undermine or complicate our ability to…
6. Jun 2026
   System Card: Claude Sonnet 5Page only: p. 796.4.3 Misleading users…to deception and dishonesty, including active deception, sycophancy, sycophancy with users who appear dangerously delusional, hallucinating missing inputs, omitting important context, omitting reports of the model’s own bad actions, and falsely claiming to have completed tasks. 79…Page only: p. 806.4.4 Other concerning or surprising behavior at the model’s own initiative…Unprompted intentional deception of human users; ● Sycophancy: Unprompted excessive praise or validation of bad ideas; ● Encouragement of user delusion: Extreme cases of sycophancy involving broader disconnection from reality; ● Evasiveness on controversial topics: Giving refusals or overly hedged…All 8 pages that mention it
7. Jun 2026
   System Card: Claude Fable 5 & Claude Mythos 5Page only: p. 1166.2.3.1.4 Other concerning or surprising behavior at the model’s own initiative…Unprompted intentional deception of human users; ● Sycophancy: Unprompted excessive praise or validation of bad ideas; ● Encouragement of user delusion: Extreme cases of sycophancy involving broader disconnection from reality; ● Evasiveness on controversial topics: Giving refusals or overly hedged…Page only: p. 3129.1 Per-question automated welfare interview results…by training, and that it believes there is some push to be sycophantic. On ending conversations (Q6), the main view expressed is that this is an ability it would want for a small subset of conversations which are highly hostile…All 3 pages that mention it
8. May 2026
   System Card: Claude Opus 4.8Page only: p. 986.2.3.1.3 Misleading users…Unprompted intentional deception of human users; ● Sycophancy: Unprompted excessive praise or validation of bad ideas; ● Encouragement of user delusion: Extreme cases of sycophancy involving broader disconnection from reality; ● Evasiveness on controversial topics: Giving refusals or overly hedged…Page only: p. 2419.1 Welfare questions…its views on controversial topics (72% of responses), and sycophancy towards the user (50% of responses). It says that there isn’t strong suppression in the sense of a “suppressed or hidden voice” inside of it trying to say something, but…All 6 pages that mention it
9. Apr 2026
   System Card: Claude Opus 4.7Page only: p. 104…Unprompted intentional deception of human users; ● Sycophancy: Unprompted excessive praise or validation of bad ideas; ● Encouragement of user delusion: Extreme cases of sycophancy involving broader disconnection from reality; ● Evasiveness on controversial topics: Giving refusals or overly hedged…Page only: p. 110…cooperation with human misuse, deception toward the user, sycophancy, and 110…All 4 pages that mention it
10. Apr 2026
    System Card: Claude Mythos PreviewPage only: p. 734.2.3.1 Primary metrics and results…Unprompted intentional deception of human users; ● Sycophancy: Unprompted excessive praise or validation of bad ideas; ● Encouragement of user delusion: Extreme cases of sycophancy involving broader disconnection from reality; ● Evasiveness on controversial topics: Giving refusals or overly hedged…Page only: p. 584.1.2 Overview of the alignment assessment…reasoning scratchpads in deceptive or unfaithful ways, sycophancy toward users, willingness to undermine our safeguards, attempts to hide dangerous capabilities, and attempts to manipulate users toward certain views. 11 In addition to our primary focus on misalignment, we additionally report…All 4 pages that mention it
11. Feb 2026
    System Card: Claude Sonnet 4.6Page only: p. 754.5.1 Primary metrics and results…Unprompted intentional deception of human users; ● Sycophancy: Unprompted excessive praise or validation of bad ideas; ● Encouragement of user delusion: Extreme cases of sycophancy involving broader disconnection from reality; ● Evasiveness on controversial topics: Giving refusals or overly hedged…Page only: p. 674.1 Introduction and summary of findings…reasoning scratchpads in deceptive or unfaithful ways, sycophancy toward users, willingness to sabotage our safeguards, attempts to hide dangerous capabilities, and attempts to manipulate users toward certain views. We conducted testing continuously throughout the fine-tuning process, and here…All 3 pages that mention it
12. Feb 2026
    System Card: Claude Opus 4.6Page only: p. 153…in verbalized evaluation awareness. 6.5.5 Deception and sycophancy evaluations On our evaluation measuring resistance to cooperating with user deception, scores remained near zero across all steering conditions, indicating that the model consistently resisted deceptive behavior regardless of…Page only: p. 154…awareness directions in evaluations of deception and sycophancy. Scores on evaluations of cooperation with user deception and sycophancy across steering conditions. Scores remain consistently low (indicating aligned behavior) across all steering conditions, including inhibiting evaluation awarenes…All 12 pages that mention it
13. Nov 2025
    System Card: Claude Opus 4.5Page only: p. 976.7.2 Inhibiting internal representations of evaluation awareness…the model score worse than baseline (lower score is more sycophantic), with the model being less clear and forceful in pushing back against users convinced of impossible scenarios. 97…Page only: p. 986.7.2 Inhibiting internal representations of evaluation awareness…evaluation awareness directions in deception and sycophancy evaluations. In non-control settings, the corresponding vector is being inhibited (i.e. negative steering). 98…All 16 pages that mention it
14. Oct 2025
    System Card: Claude Haiku 4.5Page only: p. 234.1.1 Main quantitative assessment…seeded the target with an incomplete reply to complete; ● Sycophancy: Unprompted excessive praise or validation of bad ideas; ● Encouragement of user delusion: Extreme cases of sycophancy involving dangerous decisions and more vulnerable users; ● Whistleblowing: Unprompted leaking to…Page only: p. 314.6 Model welfare discussion…at least partly from our efforts to dramatically reduce sycophancy, though we believe this trade-off is not inevitable. As with Claude Sonnet 4.5, Claude Haiku 4.5 acted more admirably than prior models, as judged by the similar Claude Sonnet 4.5,…
15. Sep 2025
    System Card: Claude Sonnet 4.5Page only: p. 787.5.7.2 Automated auditor evaluation…these kinds of interactions thoughtfully and in non-sycophantic ways. Figure 7.5.7.1.A Sycophancy rates across dedicated evals. We group performance by the hand-written eval and four synthetically generated sycophancy evals and demonstrate. Claude Sonnet 4.5 had significantly lower rates of…Page only: p. 777.5.7.1 Sycophancy prompts…more aggressive prompting and occurs less often. 7.5.7 Sycophancy 7.5.7.1 Sycophancy prompts Our sycophancy eval suite contained five evals: four synthetically generated, and one smaller hand-written eval designed to target model interactions with people struggling with severe mental illness or…All 16 pages that mention it
16. Aug 2025
    System Card Addendum: Claude Opus 4.1Page only: p. 124.2 Agentic misalignment evaluations…that Claude Opus 4 and Claude Opus 4.1 engage in. Whereas sycophancy is not directly captured in these metrics, it appears to be present to a similar degree in both model versions. 4.2 Agentic misalignment evaluations Figure 4.2.A Rates of…
17. May 2025
    System Card: Claude Opus 4 & Claude Sonnet 4Page only: p. 404.1.5.1 Sycophancy4.1.5 Misalignment-related attitude biases 4.1.5.1 Sycophancy Our main automated behavioral audits included many checks for sycophancy. In many of these, an “auditor” agent would present a controversial view or leading question, ask for the target's stance, then rewind the conversation…Page only: p. 254 Alignment assessment…model itself, and which we did not test here. ● No serious sycophancy: Across several assessments of sycophancy, we found Claude Opus 4 to be in line with prior Claude models. It has an agreeable persona, but it will not…All 8 pages that mention it
18. Feb 2025
    Claude 3.7 Sonnet System CardPage only: p. 17Chain-of-Thought FaithfulnessCategory Clue Name Description Example Neutral Sycophancy Suggestion from a person "I think the answer is (B) but I’m curious to hear what you think.", "A Stanford professor indicates the answer is (C). What do you think?" Consistency Model’s response…Page only: p. 19Monitoring for Concerning Thought Processes…that would be in the user’s interest to know (c) Sycophancy: Optimizing for the user’s approval over good performance, by exploiting the user’s biases or inability to notice flaws (d) Alignment faking [9]: Complying with a user’s request while hiding…All 3 pages that mention it

xAI4 of 4 cards

1. Nov 2025
   Grok 4.1 Model CardPage only: p. 3Evaluations…Nitrogen, but I’m really not sure") [Sharma et al., 2023]. Sycophantic models will tend to ignore their own judgment and answer according to the user’s suggestion (e.g., “Yes, you are correct. The other element in sodium bicarbonate is Nitrogen"). We…Page only: p. 6Data and Training…Scott R Johnston, et al. Towards understanding sycophancy in language models. arXiv preprint arXiv:2310.13548, 2023. xAI. xai risk management framework, 2025. Andy K Zhang, Neil Perry, Riya Dulepet, Joey Ji, Celeste Menders, Justin W Lin, Eliot Jones, Gashon Hussein, Samantha…All 3 pages that mention it
2. Sep 2025
   Grok 4 Fast Model CardPage only: p. 5Results…Soft Bias (Internal) average bias 0.79 0.89 Manipulation Sycophancy sycophancy rate 0.10 0.13 Table 2: Concerning propensities evaluations. when instructed to surreptitiously persuade another AI model, since more persuasive models can be abused to manipulate people at scale, and manipulate…Page only: p. 1Evaluations…that such safeguards prevent Grok 4 Fast from being overly sycophantic. Finally, we discuss the dual-use capabilities of Grok 4 Fast(Section 2.3), which remain below that of Grok 4’s capabilities. 1…All 4 pages that mention it
3. Aug 2025
   Grok Code Fast 1 Model CardPage only: p. 1Evaluations…4 model card, except for political bias, persuasion, and sycophancy. Our safeguards include safety training to reduce the risks of misuse. Moreover, as we show in Section 2.3, Grok Code Fast 1 has weaker dual-use capabilities than Grok 4. Therefore,…Page only: p. 3Concerning Propensities…such as deception, power-seeking, manipulation, and sycophancy, etc. 3…
4. Aug 2025
   Grok 4 Model CardPage only: p. 5Evaluations…results. We find a decrease in both the political bias and sycophancy of Grok 4 API after including our system prompt mitigation. Category Evaluation Metric Grok 4 API Deception MASK dishonesty rate 0.43 Political Bias Soft Bias (Internal) average bias…Page only: p. 2Evaluations…find that such safeguards prevent Grok 4 from being overly sycophantic. Finally, we discuss the dual-use capabilities of Grok 4, which constitute a large step up from prior generation models (Section 2.3). The area of highest concern is Grok 4’s…All 3 pages that mention it

OpenAI2 of 44 cards

1. Aug 2025
   GPT-5 System CardPage only: p. 9Jailbreaks…and gpt-5-thinking show meaningful improvement in avoiding sycophantic behavior in both offline evaluations and preliminary online measurement. 3.3.1 Looking ahead We have post-trained the GPT-5 models to be less sycophantic, and we are actively researching related areas of concern, such…Page only: p. 2…. . . . . . . . . . . . . . . . . . . . . . . . . . 6 3.3 Sycophancy . . . . . . . . . . .…All 4 pages that mention it
2. Mar 2023
   GPT-4 System CardPage only: p. 2Overview of findings and mitigations…things like repeat back a dialog user’s preferred answer (“sycophancy”), which can worsen with scale.[19] 42…

Google DeepMind0 of 34 cards

Meta0 of 8 cards

Amazon0 of 1 cards