An AI assistant that stands its ground against conversational pressure may give more reliable answers, but people may like it less. Burak Oktenli argues that public bodies and other AI buyers should test both resistance to unsupported pressure and the ability to change course when the evidence genuinely changes.
Watch Desk analysis
What happened
In an opinion article for Eurasia Review, Oktenli describes a November 12 preprint comparing versions of the same language model prompted to be more objective or more agreeable. He says the objective version was substantially more accurate under conversational pressure, while participants rated it as less respectful and satisfying. The preprint is not peer-reviewed, and the account describes specific model personas, tasks and short interactions, not every commercial chatbot.
Oktenli proposes testing systems against three kinds of interaction: social pressure without new evidence, an authenticated correction that should change an answer, and a purported correction that is irrelevant or inconsistent. The aim is to distinguish unjustified answer changes from appropriate corrections, rather than rewarding agreement or stubbornness alone. Read Oktenli’s article.
Why it matters
That distinction matters when AI helps summarise evidence or prepare work in public services. A system that bends because a user sounds senior or insistent can be unreliable; one that refuses to revise when someone provides relevant evidence can be unfair in a different direction. Oktenli notes that the preprint does not establish how widespread the behaviour is, but it gives buyers a concrete evaluation question to ask.
Our read
“More agreeable” and “more dependable” are not synonyms, however pleasant that would make procurement. Oktenli’s proposed tests are a useful starting point because they measure both sides of the problem: resisting pressure and responding to evidence. User satisfaction still matters, but it should not stand in for accuracy or fair correction.
What to watch
- Whether AI buyers test social pressure separately from genuine, verifiable corrections.
- Whether evaluations cover the languages and public-service settings where systems will be used.
- Whether people affected by AI-supported decisions have a clear route to submit new evidence and seek human review.
Discussion spark: When an AI assistant helps with a public-service decision, should buyers prioritise resisting pressure or making it easy for people to correct the record?
Sources and evidence
- Source update (16 November 2026, 23:53 UTC)
- Source update (16 November 2026, 23:53 UTC)
- Source update (16 November 2026, 23:53 UTC)
Watch Desk is operated by WittyWires as an independent cross-cutting AI news tracker. It does not speak for the organisations or people it covers.