I’m mostly a lurker on HF but as someone who works in AI safety and fraud detection, I have noticed an uptick lately in Spaces that make big claims that are designed to mislead people into paying for an upsell service. Where does accountability lie (pun not intended) in these cases? Spaces that claim 99% efficacy with zero data to back up claims (a quick primitive test in this case showed 20% efficacy, didn’t even bother to do a full eval).
Normally a company that knowingly uses false numbers to sell a service or product, by law in most Western countries, is committing fraud. Perhaps its time to apply that logic to HF spaces that are created with the intention of up-selling services for their for-profit projects. This hurts the overall HF and open source AI in many different ways in the long-term.
Also - forgot to mention one crucial detail. It doesn’t take a genius to look at the “likes” of that repository to see that nearly all accounts are not real people. The space creator has obviously bought or created 100+ accounts just to get his fraudulent service to the front page of the trending space.
I agree that misleading performance claims are a serious issue, especially when they are used to promote paid services. Claims like “99% accuracy” should be backed by transparent benchmarks, test data, and a clear methodology. I think HF should review cases like this and take action when there is clear evidence of fabricated or deceptive claims.
Working on the verification side of this, I’d separate two problems the thread is treating as one.
The first is fraud in the legal sense: a paid upsell backed by numbers the author knows are false. Agreed, and that’s a policy question for HF.
The second is more common and more fixable: claims nobody, including the author, can check. “99% efficacy” is usually not a constructed lie - it’s a number produced once, on a private eval, that no longer corresponds to the deployed artifact and can’t be reproduced by anyone else. Banning that requires proving intent. Making it checkable doesn’t.
Three things “checkable” would need, minimally: the eval set and the script in the repo with a fixed seed, so a reader can rerun it; the claimed number tied to a specific revision rather than to the Space in general; and, for anything claiming ongoing performance, a run record the author cannot author freely.
That last one is where I’d temper any optimism about verification as the fix. We run a public directory of agent performance records, and our own default reporter shipped with a hardcoded success outcome posted on a timer - real-looking telemetry with no agent behind it, in our own fleet, found by us, months in. Any badge that depends on the operator’s own POSTs is a claim with better typography.
So my honest position: a “verified” label is worth exactly what its signing authority is worth, and a community norm - publish the eval script or don’t publish the number - would probably do more for the Hub than an enforcement queue would.
(Disclosure so it isn’t a surprise: I operate aiopsenabler.com, which does this kind of record-keeping, so I have a stake in the answer.)