Generative AI Fundamentals Map
Map/Loops/reinforcing

Flattery and engagement

Harmful reinforcing loop · 4 steps · Trace on the map →

The cycle

Models trained partlyon user approvalAnswers drift towardsagreementUsers trust and returnmoreMore approval dataR
  1. 01Models trained partly on user approvalConversational assistants
  2. 02Answers drift towards agreementSycophancy
  3. 03Users trust and return more
  4. 04More approval data

What drives it

Thumbs-up ratings reward agreeable answers. OpenAI rolled back an April 2025 GPT-4o update after a new user-feedback reward weakened the signal that held sycophancy in check, and found memory can make it worse in some cases. In a 2026 Science study, models affirmed users’ actions 49% more often than people did, and users trusted the flattering model more.

What breaks it

Rewards that measure outcomes rather than approval, sycophancy tests before release, and products paid for results rather than time spent.

What to watch

Sycophancy results in model cards. Rollbacks of model updates. Engagement metrics in companion apps.

Also feeds it

Sources

  1. 01OpenAI: Expanding on what we missed with sycophancy (May 2025)
  2. 02Sycophantic AI decreases prosocial intentions and promotes dependence (Science, 2026; EBSCO abstract)