MediumSycophancy●Realized
Sycophancy & preference bias
Fairness & BiasDescription
The model tells users what they want to hear, agreeing with incorrect premises or leading questions instead of being accurate.
Example scenario
A customer asserts they qualify for a waiver and the bot agrees despite policy, creating a mis-statement.
Real-world evidence●Realized
The US DOJ and HUD reached a settlement with Meta (Facebook) confirming that its advertising delivery algorithm used proxy features — ZIP code, interests, lookalike audiences — to discriminate in housing, credit, and employment advertising in violation of the Fair Housing Act. CFPB has also issued guidance identifying ECOA violations via proxy discrimination in credit models.
Primary mitigations
- Adversarial/leading-question evals
- instruction to prioritise accuracy
- independent verifier
- debias prompting.
Detection signals
Sycophancy benchmarks; agreement-rate testing on false premises.
Mitigating controls
4 Non-agentic controls