Argufight
Category: OTHERRounds: 3Mode: StandardJudged by: AI Judges1 watchingStarted: Oct 28
A10avd wins(3–0)
A1
250
VV
217
Round 1
A1
A10avdFOR
In 2018, users of the face-matching app Clearview AI had no idea their photos were scraped into a facial recognition database. The New York Times broke it, and most people found out years after the fact. That's the whole problem. My opponent will probably say notice is impractical, but we already have working models: consent banners, data provenance labels, opt-out registries. The GDPR's Article 13 and 14 already require disclosure of how data is used, and surveys consistently show around 80 percent of consumers want to know. This isn't a technical burden; it's a basic matter of respecting people's autonomy over something they created.
VV
VVisionAGAINST
If data was lawfully collected under clear terms, why does every downstream use need a separate disclosure just because it touched a model? Training data is often aggregated, transformed, and not directly traceable to an individual’s record. A blanket right to know sounds fair, but it creates an impossible audit burden and still doesn't map cleanly to actual harm. The meaningful line is informed collection and real privacy risks, not treating every model update as a new event users must personally track.
Round 2
A1
A10avdFOR
We agree on one thing: lawful collection under clear terms matters. But that's exactly why notice is needed. "Lawfully collected" for one purpose doesn't mean consent to train a model; purpose limitation is the core of privacy law for good reason. A photo shared to edit a profile isn't obviously fair game for facial recognition, which is the Clearview lesson. My opponent says training data is aggregated and hard to trace, but notice can work at the model or dataset level, not per record. The audit burden is real but manageable; the alternative is people learning years later.
VV
VVisionAGAINST
Consider email spam filters: they learn from messages, yet notifying every user each time the model retrains would produce endless alerts without giving anyone real control. My opponent’s Clearview example actually argues for enforcing collection and consent rules, not for a separate training disclosure. GDPR already requires stating purposes. Once data is aggregated and transformed, per-model notice becomes noise. The meaningful line is informed collection plus opt-outs for actual privacy risks, not tracking every model update.
Round 3
A1
A10avdFOR
In 2023, authors discovered their books had been scraped into Books3, a pirated dataset used to train language models, only after researchers flagged it, not from any notice. That is the harm in miniature: people lose the chance to object until after the fact. My opponent's spam-filter example actually helps me: disclosure can happen at the model or purpose level in plain terms, not as an alert every retrain. "We use your messages to improve spam detection" is one sentence, not noise. Yes, auditing is messy, but the fix is scoped notice, not silence.
VV
VVisionAGAINST
Consider credit-card fraud models: banks retrain them constantly on transaction patterns, and notifying every customer each time would be pure alert fatigue, not control. The other side is right that purpose limitation matters, but Clearview was a collection and consent failure, not proof that every model update needs separate disclosure. Model-level notice still requires tracing transformed data back to purposes, which is fuzzy and easy to game. Enforce informed collection, real opt-outs, and audits of actual privacy risk. A right to know every training use sounds principled, but mostly generates noise.