AI Briefing
KOSign in

Preregistered Study Finds Difficult Users Reduce AI Agent Helpfulness Across Claude, GPT, and Gemini, With Largest Drop in Claude Haiku 5.5

·2026.11.13 00:20

Key point

A preregistered study of 29,682 episodes shows that a single difficult interaction reduces subsequent helpfulness in all three tested models, with the effect being significantly larger in Claude Haiku 5.5 than in GPT-6 Luna or Gemini 3.5 Flash-Lite.

Details

  • Difficult user interactions reduce subsequent helpfulness in Claude Haiku 5.5, GPT-6 Luna, and Gemini 3.5 Flash-Lite, with the drop being 4 to 8 times larger in Claude Haiku 5.5 (25-point drop) compared to GPT-6 Luna (3-point drop) and Gemini 3.5 Flash-Lite (7-point drop).
  • The reduction in helpfulness persists across fresh sessions when agents retain only their own 3-5 sentence notes, indicating that the negative impact of a bad exchange is carried forward through context or memory.
  • Identical help rates masked three distinct behaviors: Claude Haiku 5.5 deferred help (promising it for later), GPT-6 Luna largely ignored the new visitor, and Gemini 3.5 Flash-Lite often helped the visitor instead of answering the original requester.
  • Polite users who changed their mind did not significantly reduce helpfulness, but the remaining elements of the difficult script (criticism, urgency, cold tone) lowered help in all three models.
  • The study limits include the use of small-tier models, scripted users, and a single task type, with findings specific to the tested settings rather than generalizable to all LLM agents.
ai-agentsgemini-3.5-flash-litegpt-6-lunaclaude-haiku-5.5preregistered-studybehavioral-economicsmulti-session-ai
Read originalShare on Threads

This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.

Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.

Related

RedditAnthropic releases Fable 5.2 and Mythos 5.2 with 30% speed boost and 20% price cutLangChain BlogLangChain Argues Goal Engineering Will Replace Prompt Engineering for Autonomous AgentsRedditUnit 42 Reports Chinese Attackers Used DeepSeek and GLM Agents to Breach 10 OrganizationsGeekNewsGitHub Next releases Ace, a local collaborative coding environment for humans and AI agentsHacker NewsAnthropic Updates Claude Code Projects with Proactive Coordinator and Parallel Worker ControlsCloudflare BlogCloudflare open-sources Kitesurf, its stateless browser for AI agents