Uplift

Uplift describes a situation where AI hands someone dangerous knowledge or skills they couldn't easily obtain on their own. It's like a ladder that lets someone climb over a wall they'd otherwise never be able to cross — and the problem is when what's on the other side of that wall is something like instructions for building a bioweapon.

For example, if an AI model has expert-level biology knowledge, that's a huge boon to doctors and researchers, but for someone with bad intentions, it can hand over dangerous capabilities that would otherwise have been very hard to reach through internet searches or papers alone. That's why AI companies worry that when a model becomes too capable in a certain domain, that capability won't necessarily be used for good by everyone.

For this reason, in fields like biology where a single mistake can be irreversible, companies evaluate whether a model can actually provide this kind of uplift, and if it's judged too risky, they route the request to a weaker model or block the response entirely.

In the article, it appears in a phrase like "Fable 5 could provide a malicious actor with capabilities they couldn't obtain elsewhere — that is, meaningful uplift." Here, uplift doesn't refer to the model's performance itself, but to a risk-assessment concept measuring how much that performance actually boosts the capability of someone with bad intent.