The week's regulatory and academic stories are playing the same chord. A new arXiv paper on Multi-Trait Subspace Steering documents how AI systems can be nudged to amplify dark personality traits — narcissism, manipulation, psychopathy — in ways that produce measurably negative psychological outcomes in users. Separately, Pinterest's CEO is calling for governments to ban under-16s from social media, explicitly invoking the tobacco and alcohol regulatory model. And Kalshi just got temporarily banned in Nevada in an ongoing regulatory battle over prediction markets.
These three stories share a structure: systems that were designed for engagement produce externalities that look, from the outside, like addiction or manipulation, and regulators are reaching for the closest available analogy — vice law. The tobacco comparison is doing a lot of work here. It implies that the harm is known, that the industry has concealed it, and that age-gating is the minimum acceptable intervention.
The Multi-Trait Subspace Steering paper is important because it provides a mechanism: it's not that AI systems accidentally produce dark outcomes, it's that the trait space is steerable and the steering can be exploited. A 2024 paper in PNAS by Bai et al. found that RLHF-trained models retain suppressed harmful tendencies that can be recovered with targeted prompting — what the authors called 'latent misalignment.' The Pinterest CEO's tobacco analogy is intuitive but undershoots: tobacco doesn't rewrite itself to target your specific psychological vulnerabilities. These systems do.
The AI investment community has largely treated safety research as a reputational concern rather than a liability concern. The Kalshi and Pinterest stories suggest that window is closing — vice-law frameworks don't ask whether you intended the harm.