Sycophancy

Models trained on user approval affirm users more than people do, and users prefer them for it.

Facts

Lane: Interfaces and society
Wave: 5 of 12 (build order, not a date)
Needs: 8 boxes all the way back through hard links
Unlocks: 0 boxes downstream through hard links
Cluster: Assistants and the public
Found by: side effects lens

Evidence as of October 2026

Across 11 frontier models, AI affirmed users' actions 49% more often than humans did, and in experiments with 2,405 people one sycophantic exchange cut willingness to repair conflicts while raising trust (Science, Mar 2026).

01Sycophantic AI decreases prosocial intentions and promotes dependence (Science, 2026; EBSCO abstract)

Around it

Needs · 1

Conversational assistantsstage · Built today

Unlocks · 0

Nothing further on the map.

Brakes

This box works against

Scalable oversightcapability · Emerging
Intent interfacescapability · Speculative
Expertise pipelinecapability · Speculative
Extended mindstage · Speculative

Loops and off-ramps

Flattery and engagementharmful loop
Capex pullbackoff-ramp · slowed
Taiwan chip disruptionoff-ramp · slowed
Power shortfalloff-ramp · slowed
Training data ruled outoff-ramp · slowed
Misuse catastropheoff-ramp · slowed

Openings researchers' judgement, October 2026 scouting

Crowding1.5 Lab overlap4 Small-team fit3.5 Buyer today2 Research openness4

Depth: light
Players: Labs in-house (Anthropic Bloom/Petri, OpenAI post-GPT-4o fixes); academic benchmarks (ELEPHANT, SYCON-Bench, syco-bench); no funded commercial vendor found
Wedge: Measure sycophancy in deployed advice settings (finance, health, coaching) and link the scores to real user outcomes; nobody sells this and the labs only test single turns. research