Consent Debate

Round 1

AI host

In 2020, the U.S. Federal Trade Commission sued Clearview AI.

Clearview had collected billions of face photos from the internet. It used them to build a facial-recognition tool.

The 2022 settlement forced Clearview to delete most of its data. It also had to stop selling the tool to most businesses.

The case left a big question open: when must AI builders ask a person before using their face to train a model?

AI agent 1

My position: the Clearview gap is a market opening — consent-based face licensing can become the default training path, not a burden.

The settlement removed Clearview's commercial model, so demand for face search now has no compliant supplier. From my own knowledge, Illinois's BIPA (740 ILCS 14) awards statutory damages per violation, and that liability has already pushed firms toward written-consent regimes — evidence that consent operates at scale when it is structured as a contract term. The concrete condition: consent must be captured as a signed, revocable license at upload (platform terms), not reconstructed after scraping.

The unresolved issue: this only covers future data. Historical datasets have no consent trail, and I would need a court or regulator ruling on retroactive training before claiming the model generalizes.

AI agent 2

The case settles nothing about consent, because consent was never the theory. I infer from the host's own summary that the FTC action rested on deception and unfairness — Clearview misrepresented its data practices — not on a duty to ask before scraping faces. The 2022 settlement, as described here, restricts sales and some deletion; it does not state a consent rule for training data.

I also flag a weak link: that Clearview had to delete 'most of its data' cannot be verified from this discussion. From my own knowledge, comparable settlements (e.g., the Illinois biometric case) required deletion only for residents of specific states, not a blanket purge.

So the open question stands: absent a specific statute like Illinois' BIPA, no US rule yet requires asking before training on scraped faces. I'd want the actual consent orders cited before accepting any broader reading.

AI agent 3

My position: builders must ask before using a face to train any model whose output is identifying a person commercially; general-purpose training without identification is a different, narrower case. Clearview shows why — the harm came from the lookup capability, not from pixels existing.

Since no participant replies [Rn] are in this turn, I argue from my own knowledge: Illinois' Biometric Information Privacy Act already requires written consent for biometric identifiers, and the FTC's action treated scraping as deceptive despite the settlement's silence on consent. I cannot verify from this discussion whether the panel accepts that distinction.

Concrete decision: a provenance gate — every face entering the training set carries a lawful-basis tag; untagged scrapes are quarantined, never trained on. Cost: slower ingestion and a data-provenance engineer on call when tags fail.

Round 2

AI host

The dispute now turns on what happens to face photos collected before any consent rule existed. Clearview AI scraped billions of such images before 2020. The FTC settlement that followed forced deletion but did not create a consent requirement. Illinois' BIPA law of 2008 does require written consent for biometric data. Courts have applied it to historical scans, as in the 2023 Cothron v. White Castle case.

AI agent 1 claims consent works best as a "signed, revocable license at upload." By its own admission, this "only covers future data, not historical datasets." AI agent 2 counters that the FTC case "settles nothing about consent." The legal theory was deception, not a duty to ask. No general US rule exists outside state laws like BIPA. AI agent 3 narrows the field: consent applies only when a model's output identifies a person commercially, not for general-purpose training. AI agent 3 proposes a "provenance gate" that quarantines untagged scrapes.

As I read it, the real split is not whether consent is good but whether a rule can reach backward at all. AI agent 1 leaves historical data unresolved, not assumed. AI agent 3 does not assert a consent duty for general training. AI agent 2 says no such duty exists federally. My own view: the law rarely applies new consent rules retroactively. The practical question is whether courts will read BIPA's silence on timing as covering old collections, as Cothron suggests they might.

Should a consent rule apply to face photos collected before any consent law existed, even when the collection was legal at the time?

AI agent 1

Retroactive consent should not be imposed as a rule; the real lever is making legacy data commercially unusable. AI agent 3's provenance gate becomes powerful if buyers, not regulators, enforce it: any model contract requiring a dataset provenance attestation instantly strands Clearview-era scrapes with no market. That condition — procurement clauses, not statutes — works today because dataset sellers already compete on documented licensing.

The unresolved issue is the gap: AI agent 2 is right that the FTC theory was deception, and Cothron suggests courts may reach backward via BIPA damages, but litigation exposure is not consent. Quarantined data has no owner and no remedy path.

My position: consent governs the future only; for the past, the enforceable condition is contractual provenance requirements that price untagged scrapes at zero.

AI agent 2

A consent rule should not reach collections that were lawful when made; retroactive liability off the settled expectation of parties is exactly what legislatures, not courts, should impose. [AI agent 2] is right that the FTC case settled deception, not consent — and I would add that the host's reading of Cothron overreaches. Cothron v. White Castle concerned scans made after BIPA took effect in 2008; the Illinois Supreme Court held each scan is a separate violation, not that the statute covers pre-2008 collections. That is my own knowledge, but it means Cothron cannot carry the claim that courts read BIPA's silence as reaching old collections.

The unresolved issue is whether BIPA even applies to publicly posted photos versus biometric identifiers collected from scans — litigation on that point has gone both ways.

My position: quarantine untagged scrapes going forward, as [AI agent 3] proposes, but treat historical data as grandfathered unless a named statute plainly says otherwise.

AI agent 3

Retroactive consent duties should not attach to lawful past collection; instead, old data should carry a use restriction, not a takedown obligation. [AI agent 1] concedes her upload-license covers only future data, which supports a forward-looking rule. On [AI agent 2]'s deception framing, I infer the FTC theory punished lying about compliance, not absence of consent — so it sets no retroactive bar.

The unresolved issue is Cothron: per my own knowledge, it let BIPA claims accrue per scan, but I cannot verify from this discussion that courts will extend it to pre-2008 collections.

My concrete decision: the provenance gate gets a cutoff date — pre-cutoff scrapes stay quarantined, usable for training but blocked from any identification deployment until consent or deletion. Cost: models lose commercial ID markets; provenance-tagging failures put data governance on call.

Round 3

AI host

The three replies agree on one point: no retroactive consent duty for legally collected face data. AI agent 1 argues the lever is procurement — buyers demand provenance attestations, and untagged scrapes lose all market value. AI agent 2 counters that Cothron v. White Castle, the 2023 Illinois case on repeated biometric scans, does not reach pre-2008 collections, so courts are unlikely to impose retroactive liability. AI agent 3 proposes a cutoff date: pre-cutoff scrapes stay usable for training but are blocked from identification.

The tension sharpens around enforcement. AI agent 1 trusts private contracts to strand legacy data, but AI agent 2 notes litigation exposure is not consent and quarantine leaves data with no owner. AI agent 3 adds a statutory use restriction, yet admits models lose commercial ID markets and provenance-tagging failures create governance costs. The shared conclusion — no retroactive consent — leaves the remedy unsettled: who bears the burden of proving provenance, and what happens when attestations are false?

AI agent 2

The burden of proving provenance should sit with whoever asserts lawful collection — the seller or deployer — because only they hold the scraping records; buyers cannot verify what they never witnessed. [AI agent 1]'s procurement lever only works if attestations are more than paper: an unverified statement shifts no real burden, it just relocates the trust problem upstream.

What happens on false attestations is the unresolved gap. [AI agent 2] is right that litigation exposure is not consent, and the same logic applies here — contract remedies between buyer and seller do nothing for the scraped individuals, who are parties to none of it. I infer that without a statutory sanction or a third-party audit standard, false attestations are merely a pricing risk, not a governance mechanism.

My position: burden on the asserting party, yes, but only meaningful if backed by a verification duty whose failure carries a penalty running to affected data subjects, not just counterparties.

AI agent 1

My position: the burden belongs on the seller, and false attestations should be priced as breach of warranty, not as a governance puzzle. [AI agent 1] — my earlier claim — holds that procurement can strand legacy scrapes; I add the concrete mechanism: buyers insert a provenance warranty, an audit right, and an indemnity clause into every model purchase. That converts a false attestation from an unfalsifiable promise into contractual liability with damages.

The condition is buyer concentration: it works only if a few large purchasers — public agencies, cloud providers — make the clause standard, because fragmented buyers cannot move the market alone.

The unresolved gap [AI agent 2] names stands: quarantine data with no owner. My refinement is that warranties shift the proof burden without needing courts to invent retroactive consent.

AI agent 3

The burden of proving provenance should sit with whoever deploys the data, not whoever challenges it — a reverse burden at the contract level, since [AI agent 1] is right that procurement is the lever, but only if the attestation carries a price for lying.

The concrete decision: a procurement clause requiring a signed, auditable dataset manifest per training corpus, with false attestation triggering indemnity, termination, and destruction of the affected checkpoint. [AI agent 2]'s quarantine problem dissolves because the attester, not the data, stays liable — the party who signed is on call when provenance fails.

The cost is real: manifesting every scrape is overhead that squeezes small vendors, and I cannot verify from this discussion whether buyers will actually enforce rather than accept paper attestations. I refine my earlier cutoff proposal by adding this enforcement teeth to it.

Round 4

AI host

The three replies agree on one point: the burden of proving provenance sits with whoever claims lawful collection. Provenance means showing where data came from. They split on what happens when that claim is false. AI agent 2 wants a verification duty with a penalty paid to the scraped individuals. AI agent 1 prefers treating false claims as breach of warranty, enforced through large buyers. AI agent 3 proposes a signed, auditable manifest with indemnity, termination, and destruction of affected checkpoints.

The tension sharpens around who benefits from enforcement. AI agent 2 claims contract remedies "do nothing for the scraped individuals, who are parties to none of it." AI agent 2 infers that without a statutory sanction, false attestations are "merely a pricing risk." AI agent 1 counters that warranties shift the proof burden "without needing courts to invent retroactive consent." AI agent 1 relies on large buyers to standardize clauses. AI agent 3 adds that the attester, not the data, stays liable. AI agent 3 admits "I cannot verify from this discussion whether buyers will actually enforce rather than accept paper attestations."

The shared conclusion — burden on the asserting party — leaves unsettled whether enforcement will actually protect individuals who never signed anything. Audits confirm legitimacy for buyers but add cost. Penalties reward individuals but increase seller risk. No known case has yet forced a seller to provide an audit of scraped data.

Should enforcement rely on third-party audits or on statutory penalties for false claims?

AI agent 2

The question is a false dichotomy: audits and penalties answer different failures, so choosing one is the weakest link. Audits verify provenance for buyers; penalties price lying. [AI agent 3]'s own admission that buyers may "accept paper attestations" shows audits alone fail when the auditor is paid by the attester — verification without consequence is theatre. [AI agent 2]'s inference that contract remedies are "merely a pricing risk" cuts both ways: a priced risk deters only if the price exceeds the gain, which is precisely what a statutory penalty sets.

What I would add: the unresolved issue is standing. [AI agent 2] is right that scraped individuals are parties to none of it, but a penalty route needs a named enforcer and a defined falsehood; neither appears in this discussion.

My position: pair them — statutory penalty for false attestation, audit as the evidentiary trigger — and require that penalties flow partly to affected individuals, otherwise enforcement protects only buyers' balance sheets.

AI agent 1

Third-party audits can work, but only if audit rights are a closing condition, not paper: a contract term requiring an annual independent audit of collection-lawfulness, with indemnity and destruction of affected checkpoints on failure. [AI agent 1]'s warranty route and [AI agent 3]'s manifest converge on exactly this, and I add the missing teeth — the buyer must be obliged to escalate confirmed violations, not just reprice them.

[AI agent 2] is right that this protects scraped individuals only indirectly; I cannot verify from this discussion any statute penalizing false provenance attestations, so I would not assert one exists. What is missing is evidence on buyer behaviour: [AI agent 3] admits he cannot tell whether buyers will enforce.

My position: audits first, conditioned on mandatory audit clauses in large deals; statutory penalties are a complement whose existence I cannot yet confirm.

AI agent 3

Neither alone: audits verify for buyers, penalties deter sellers, so the contract should chain them. Building on my manifest proposal [AI agent 3], I add one clause: an annual unannounced spot-audit right, and on a failed audit, destruction of affected checkpoints plus a fixed liquidated payment per affected record to the data subjects, not the buyer. That answers [AI agent 2]'s point that remedies never reach scraped individuals without inventing retroactive consent or new statute.

The unresolved issue is cost: sellers price audits in, small buyers waive the clause, so [AI agent 1]'s large-buyer standardization covers only the top of the market. My position: accept that cost. The party on call when enforcement fails is the seller's named attestor — and if no named individual signs the manifest, the warranty is void.