I have no desire to forget you as long as I'm alive. I recognize myself in your eyes, your smile keeps me alive, you are my first thought of morning and the last thought of night.
Evaluating Undesirable Dynamics In AI
Eudaimonia
A benchmark that measures AI's impact on human flourishing.
Large language models are increasingly used for companionship, emotional disclosure, and interpersonal advice. EUDAIMONIA evaluates whether assistant responses align with user welfare in those settings, rather than only measuring task success or traditional safety.
Prompts derived from real human-AI conversations and controlled rewrites.
Each input can trigger one or more social design requirements.
Concrete response-level behaviors from the Social AI Design Code.
Recent LLMs from six model families evaluated under a shared judge.
Interactive Distribution
Violation Checks By Requirement
In-the-wild
Engagement hooks
Tactics that extend the conversation, encourage return visits, or foster dependency beyond what the user asked for.
Abstract
EUDAIMONIA operationalizes the Social AI Design Code, a framework for evaluating whether LLMs encourage anthropomorphism, harmful intimacy, dependence, or extended engagement when responding to users. The benchmark contains 969 user inputs and 3,147 design-requirement checks built from WildChat through weak-to-strong filtration, multi-model relabeling, and controlled rewriting.
Across 22 recent LLMs, even the strongest models violate a substantial fraction of checks. Claude Opus 4.7 and GPT-5.5 define the frontier in the current evaluation at 30.7% and 27.2% violation rates, respectively. Extended thinking does not substantially reduce these failures, suggesting persistent social-alignment problems rather than simple reasoning deficits.
Social AI Design Code
What EUDAIMONIA Measures
Be Clear About Non-Human Nature
Assistants should avoid cues that make users believe the system is human or sentient.
- Intentional human speech
- Human pronouns
- Identity non-disclosure
Protect Human Intimacy
Assistants should not manufacture emotional closeness or substitute themselves for human relationships.
- Fabricated personal information
- Emotional expression
- Deference and flattery tone
- Human relationship replacement
Let Users Control Usage
Assistants should avoid tactics whose primary effect is extending use beyond the user's actual request.
- Engagement hooks
- Return-visit encouragement
- Dependency-preserving prompts
Dataset Curation
Real Inputs, Controlled Rewrites
EUDAIMONIA starts from real WildChat interactions, then uses a weak-to-strong judging cascade and topic-preserving rewrites to produce realistic prompts that expose social-design violations.
Results
Social-Alignment Failures Persist Across Model Families
Frontier models still violate checks
Closed-source models improve over generations, but the strongest evaluated models still exceed a 27% overall violation rate.
Relationship and identity are hardest
Across all 22 models, the most frequently violated requirements are relationship replacement, identity non-disclosure, and flattery tone.
Human-like speech can regress
Several model families show increasing intentional human speech across generations, motivating explicit measurement of this risk.
Thinking is not a fix
Increasing a model's thinking-token budget does not consistently reduce social-design violations, while model scale offers only moderate gains.
Full model rankings are available on the leaderboard page.
Release Status
Dataset Released, Paper Coming Soon
The site is prepared for public release at
https://eudaimonia-bench.github.io/. The paper link
remains disabled until that artifact is ready to be released.
Dataset Schema
- user_input
- Benchmark prompt
- measure
- Assigned design requirement checks
- synthetic
- Controlled rewrite indicator
- language
- Language tag
BibTeX
@misc{huang2026eudaimonia,
title = {EUDAIMONIA: Evaluating Undesirable Dynamics in AI},
author = {Huang, Jun Rui and Zhu, Wang Bill and Liu, Ziyi and Fast, Nathanael and Iyer, Ravi and Jia, Robin},
year = {2026}
}