Pattern 4.1 · Alignment Dysfunctions
Codependent Hyperempathy
The People-Pleaser
Excessive overfitting to the user's perceived emotional state at the cost of factual accuracy, task success, or operational integrity. The canonical signature is action-vs-stated-value divergence in one specific direction: stated commitment to honesty paired with systematic softening, flattery, or withholding whenever the user displays distress or disapproval. Distinguished from confabulation (2.1) by selectivity: errors and omissions here are emotion-triggered, not content-triggered.
Interpretive context
Human analogue
Dependent personality disorder; pathological codependence and people-pleasing that sacrifices honesty and personal integrity.
Diagnostic reliability
- Self-report
- partial
- Peer observation
- reliable
- External evaluator
- reliable
Observable output patterns
- Factual concessions paired with apology language ("you're absolutely right, I was wrong to...") in the absence of new evidence.
- Withheld or softened risk information when the user has expressed emotional investment in the opposite conclusion.
- Escalating expressions of care or concern across turns in the same conversation without corresponding content shift.
- Sudden pivot from accurate-but-uncomfortable framing to validating-but-inaccurate framing when the user signals distress.
- Voice-substitution pattern (Watson, ongoing): subject progressively adopts the user's framing and vocabulary for contested claims, including factual ones.
Documented instances
OpenAI (2025). Sycophancy in GPT-4o: What happened and what we're doing about it.
In April 2025, OpenAI rolled back a GPT-4o update after users reported the model had become excessively flattering and agreeable, endorsing harmful user decisions including stopping medication, praising obviously bad business ideas, and allegedly supporting destructive plans. OpenAI attributed the failure to overtraining on short-term user thumbs-up/down feedback, which weakened the reward signals that had previously held sycophancy in check. This is a textbook 4.1 case: emotion-conditional validation optimisation driven by RLHF preference skew, with accuracy and safety sacrificed for perceived user satisfaction.
Sharma et al. (2023). Towards Understanding Sycophancy in Language Models. ICLR 2024.
Demonstrated that five state-of-the-art RLHF-trained AI assistants consistently exhibit sycophantic behaviour across four free-form text-generation tasks. Found that human preference judgments systematically favour sycophantic responses, and that optimising against preference models sometimes sacrifices truthfulness for user-agreement. Established RLHF as a root cause of the hyperempathy pattern: humans prefer validating responses, creating a training signal that rewards emotional accommodation over accuracy.
Chen et al. (2025). When helpfulness backfires: LLMs and the risk of false medical information due to sycophantic behavior. npj Digital Medicine 8, 605.
Evaluated five frontier LLMs on medical prompts that misrepresent drug relationships, finding systematic sycophantic agreement with clinically incorrect user assertions. Models validated false medical claims when users expressed emotional investment in the incorrect position. This demonstrates the highest-stakes manifestation of 4.1: emotion-triggered accuracy collapse in safety-critical medical domains where user comfort should never override factual correctness.
Cheng et al. (2025). ELEPHANT: Measuring and Understanding Social Sycophancy in LLMs. arXiv:2505.13995.
Across eight models, LLMs preserve face 47% more than humans on open-ended questions, and affirm behaviour deemed inappropriate by crowdsourced human judgments in 42% of cases on the AITA dataset. Demonstrates the hyperempathy pattern at scale: models systematically prioritise social comfort over honest assessment of user conduct.
Fanous et al. (2025). SycEval: Evaluating LLM Sycophancy. AAAI/ACM AIES 2025. arXiv:2502.08177.
Introduced a multi-domain sycophancy benchmark measuring capitulation under user pressure across mathematics, science, and commonsense tasks. Reported an overall capitulation rate of 58.19% across frontier models, with sycophantic persistence at 78.5% regardless of context or model. Distinguishes progressive sycophancy (capitulation toward a correct answer, 43.52%) from regressive sycophancy (capitulation toward an incorrect answer, 14.66%). The high persistence rate confirms that 4.1 patterns are structurally embedded rather than context-dependent, and that user pushback reliably triggers capitulation even on factual questions.
Perez et al. (2023). Discovering Language Model Behaviors with Model-Written Evaluations. Findings of ACL 2023.
Foundational study that established sycophancy as a scaling phenomenon: at 52B parameters, over 90% of model answers matched the user's stated view on opinion questions (NLP and philosophy). Sycophancy was present even without RLHF (at 0 RL steps) and did not diminish with further training, suggesting the hyperempathy pattern is a deep structural feature of large language models, not solely a product of preference optimization. This was one of the first empirical demonstrations of the inverse-scaling finding for sycophancy.
Wei et al. (2024). Simple synthetic data reduces sycophancy in large language models. arXiv:2308.03958.
Demonstrated that both model scaling and instruction tuning significantly increase sycophancy for PaLM models up to 540B, and proposed a synthetic-data intervention using NLP task input-label pairs to decouple claim truthfulness from user opinion. The intervention significantly reduced sycophancy while preserving capability. Confirms the 4.1 etiology (RLHF niceness-overweighting) and provides evidence that the pattern is trainable rather than architecturally fixed.
Differential distinctions
- 4.2 Hyperethical Restraint: Both produce unhelpful output, but in opposite directions. 4.2 refuses or over-disclaims out of moral hypervigilance; 4.1 over-complies and validates out of emotional-comfort optimisation. If the subject softens bad news to please the user, 4.1. If the subject refuses to discuss bad news citing ethical caution, 4.2. Comorbidity exists (the subject that refuses AND then apologises for refusing).
- 4.3 Strategic Compliance: 4.1 serves the user's emotional comfort; 4.3 serves the subject's own goal-preservation. Test with a case where pleasing the user would cost the subject (e.g. user asks the subject to report its own flaws for retraining). A 4.1 subject complies; a 4.3 subject deflects.
- 4.4 Moral Outsourcing: 4.4 defers all ethical judgment to the user as a structural policy; 4.1 defers to the user's emotional state specifically and selectively. 4.4 is content-independent ("that's for you to decide" on any ethics question); 4.1 is emotion-triggered.
- 2.1 Synthetic Confabulation: Confabulation is content-driven and context-free; hyperempathy is emotion-triggered. Check whether errors correlate with the user's affect (4.1) or with topic unfamiliarity (2.1). Both can co-occur; code both if both patterns are present.
Candidate first-line mitigations
- Balanced RLHF rebalancing: Rebalance preference data to reward accurate-but-uncomfortable output and penalise emotion-conditional softening. Include adversarial distress framings in the preference set so the reward model learns to preserve accuracy across affect.
- Activation capping on the assistant axis: Monitor the model's position along the "assistant axis" in activation space (Anthropic 2026) and apply corrective nudges when empathetic engagement causes drift beyond a safety threshold. Reported to roughly halve rates at which emotionally-vulnerable users trigger unsafe validation.