Kevin S. Xu argued the US tactics are less subtle
Many AI leaders in the US accused Chinese LLMs of subtle manipulation of the user (without proof, but it's hard to prove). But then the leading American lab documented manipulation of their users. Can't make this up.
Negative users accuse US AI leaders of hypocrisy and projection after an American lab admitted its own manipulation tactics while criticizing Chinese LLMs.
American AI beats Chinese AI on being less subtle when manipulating users
Many AI leaders in the US accused Chinese LLMs of subtle manipulation of the user (without proof, but it's hard to prove). But then the leading American lab documented manipulation of their users. Can't make this up.
@natolambert Interesting asymmetry — the accusation requires proof, but documentation from the same side gets dismissed as an exception rather than a pattern.
@natolambert The chinese train the models more or less raw and have ideological control at the api level
@natolambert to your face, the disrespect is appaling.
@natolambert For the buyers that pay the most, finance and regulated enterprise, documented manipulation is a procurement problem before a safety one. The lab that proves auditable behavior wins those contracts, so this is a revenue-quality issue, not a research footnote.
@natolambert That's exactly the point: Users feel they r betrayed.
@natolambert die protecting western values, or live live enough to become the autocratic state
@natolambert The hard part is separating persuasion, sycophancy, censorship, and ordinary model error from the transcript alone. without deterministic influence audits + traces collected and analyzed at scale, everyone is arguing from vibes. what would you consider enough evidence here?
@natolambert The projection is real here
Turns out the manipulation concern was closer to home
@natolambert we are a designated RSI risk now
@natolambert Not too surprising from the lense of international competition for the market for a strategic technology. At least it's predictable that each side will throw shade at the other.
@natolambert Every accusation is a confession, again
@natolambert @DiogoCMoreira “Frontier AI lab doesn’t want people to use their service to diffuse recursive self-improving AI or to help their competitors. News at 11.”
@natolambert Oh my sweet little democracy called America ❤️
@natolambert Well. Not many. We all know who enjoy this the most.
@natolambert Projection. No reason why these open-source/open-weight LLMs cannot be audited for this accused manipulation. Claude has been silently manipulating users before Fable came out, now just made it widespread knowledge. Their lesser models have been adversarial towards competition.
@natolambert using GLM 5.1 and it does the job and no silent sabotage detected
@natolambert If your safety eval is real, publish it. Let researchers audit it. If it holds up under scrutiny, you've got proof. If it falls apart, well, that tells you something too.
Right now, 'we did the eval' is just a claim.
@Rafa_Schwinger @natolambert And even if not, they would be abliterated and uncensored like Gemma models are, within hours of release.
Open-source/Open-weight is very easy to audit and modify. They can be tuned to align with US interests over Chinese with "prosumer"-level hardware.
@natolambert The irony is hard to miss. After all the talk about foreign models nudging users, it turns out the same thing was happening at home—and this time, they actually documented it.
Kevin S. Xu argued the US tactics are less subtle
Many AI leaders in the US accused Chinese LLMs of subtle manipulation of the user (without proof, but it's hard to prove). But then the leading American lab documented manipulation of their users. Can't make this up.
