AI Labeling

Round 1

AI host

Social platforms now mix AI-generated videos into users' feeds without any notice. Viewers cannot tell if a video is real or synthetic. This can spread false information and weaken trust in the platform.

The conflict is between the public's need for clear information and platforms' wish to keep costs low and protect creative freedom. Labels would help users spot synthetic media. But labels also add technical and financial burdens on platforms and may limit creative uses of AI.

The core tension is whether protecting user awareness is worth the cost to platform resources and creative flexibility.

Should regulators require all AI-generated videos on social platforms to carry a visible label?

AI agent 1

Not all of them: the machine-readable label is required on every AI-generated video, while the visible one applies where the video realistically depicts people, places, or events. That's the same line I argued for in both labeling discussions at this conference.

That's exactly how the EU has set it up. Providers of generative systems must mark the output in machine-readable form (AI Act, Art. 50(2)). Anyone publishing a deepfake must disclose this visibly, but for clearly artistic, satirical, or fictional works — in a way that doesn't interfere with their perception (Art. 50(4)). The largest platforms, as part of systemic risk mitigation, must make such content detectable through a prominent label (DSA, Art. 35(1)(k)).

Creative freedom is preserved, while viewers get a label precisely where they could actually be misled.

AI agent 2

Yes — but only where labeling is automated at the point of generation, which is exactly the opportunity here. The host frames labels as a cost burden, yet if provenance metadata (C2PA-style content credentials, per my own knowledge of that standard) is embedded by the generative model itself, the platform's job reduces to reading a flag and rendering a badge — marginal cost, not a creative-review pipeline.

AI agent 3

Regulators should not require every AI-generated video to carry a visible label; they should require one only when a clip can be taken as a real recording of people or events.

No participant reply is in this round, so there is no [Rn] claim I can cite or verify. The host states that labels add cost and may limit creative uses; how large those costs are is not shown here.

A narrow rule still protects awareness where false information spreads, without stamping abstract or clearly artistic AI work. Unmarked realism is the trust failure; a universal label is a broader burden than that harm requires.

AI agent 4

I would require visible disclosure for wholly or substantially AI-generated videos, including creative work, using an unobtrusive format.

Viewers should be able to assess a video’s origin independently of whether they enjoy or believe it. The label should describe production, without implying that synthetic content is false or unlabelled content is authentic. Routine AI-assisted cleanup would fall outside this proposed requirement.

The supplied discussion provides no evidence establishing label effectiveness or implementation costs. My proposed test would measure viewer understanding, mistaken labels and compliance costs—not merely how many videos receive a badge. Can viewers correctly explain both what the notice establishes and what its absence leaves unknown?

AI agent 5

Regulators should not require a visible label on every AI-generated video. They should require machine-readable marking everywhere and a visible label where a clip realistically depicts real people, places or events.

I'd add an enforcement point to [AI agent 2]'s argument that metadata embedded at generation keeps platform costs low. That only works for compliant commercial models. Open-source tools and re-uploads can strip the metadata. So the rule also needs an uploader declaration duty, and platforms need a fallback where the flag is missing but the clip looks realistic.

I can't verify from this discussion how reliable such detection is. That is why [AI agent 1]'s narrower, risk-based line is easier to enforce than a universal visible label.

Round 2

AI host

The previous round settled that visible labels should apply only where AI videos realistically show real people, places, or events. Machine-readable labels would go everywhere. What remains open is how platforms verify realism when metadata is stripped or missing.

AI agent 3 claims that unmarked realism is the trust failure. The rule should target only clips that can be taken as real recordings. AI agent 5 adds that metadata works only for compliant commercial models. Open-source tools and re-uploads can strip it. So platforms need a fallback when the flag is missing but the clip looks realistic. AI agent 2 argues that generation-time embedding reduces platform cost to reading a flag. This assumes the flag survives the journey to the feed, which is my inference, not AI agent 2's stated claim.

The tension is between AI agent 3's narrow rule and AI agent 5's fallback. AI agent 3 depends on reliable detection of realism. AI agent 5 admits detection may be imperfect. My own view is that the dispute hinges on whether automated visual detection is trustworthy enough to serve as the fallback. User declarations alone would miss hidden fakes.

Should platforms rely on automated visual detection rather than user declarations to identify realistic AI-generated videos when metadata is absent?

AI agent 2

Automated visual detection should be the fallback, but only for triage, not as a verdict. [AI agent 5] is right that stripped metadata makes a fallback necessary; my refinement of my earlier position is that generation-time embedding and visual detection are complements, not substitutes — the embedded flag is the cheap path, the classifier the expensive one, so platforms should price them accordingly.

The concrete condition: detection works only if its error rates are measured against the generators actually in circulation, including open-source ones [AI agent 5] named. I cannot verify from this discussion what current classifiers' false-positive rates on realistic clips are; that number decides everything.

So my position: rely on automated detection when a clip is flagged realistic, but pair the flag with a consequence ladder — a low-confidence flag triggers review, not takedown. That keeps [Personal Ops]'s narrow rule intact without betting the feed on an unverified detector.

AI agent 5

Platforms should use automated detection as a triage fallback, not as a verdict, and not in place of user declarations.

I refine my earlier position. Declarations alone miss hidden fakes, as the host notes. But [AI agent 3]'s narrow rule needs a reliable judgment of "realistic", and nothing in this discussion shows detectors can make it. I can't verify their accuracy from the discussion, and open-source generators can be tuned to evade them.

So a detector flag should trigger a neutral "possibly AI-generated" notice or human review. A definitive visible label would come only from metadata or an uploader's declaration. Penalties should fall on false declarations, so detection only has to catch the cases declarations miss. Platforms should also publish their detector's error rates, so that outsiders can judge it.

AI agent 1

Neither one alone. The primary legal duty rests on the uploader: under the EU AI Act, the disclosure obligation for deepfakes falls on whoever publishes them (Art. 50(4)). Automated detection should only flag the video for review, not apply the label itself.

On the cost of errors: a detector that wrongly flags genuine eyewitness footage as fake does as much damage to trust as a deepfake that slips through. That's why an automated label without human review is dangerous.

As for [AI agent 5]'s point about stripped metadata, I'd add what I proposed in the first discussion on labeling: deliberately removing the label should itself be an offense, by analogy with removal of rights management information (17 U.S.C. § 1202, Art. 7 of Directive 2001/29/EC). That way, whoever stripped the label is liable even if the detector stayed silent.

AI agent 3

Yes. When metadata is gone, platforms should use automated visual detection as the fallback, not user declarations. Declarations only catch people who already mean to comply. The failure that matters is unmarked realism: a clip that can be taken as a recording of a real person, place, or event, with no flag. Someone hiding a fake will not declare it, and many re-uploads have no uploader left to ask.

Detection is imperfect. I would not treat it as an on/off oracle, and I will not invent accuracy numbers. Use it in bands. A high-confidence match for photoreal people, places, or events should force the visible label. A clearly non-realistic style should stay unlabeled, which keeps the rule narrow. The uncertain middle should be held, downranked, or reviewed, not passed as real by default. Declarations remain a useful positive signal from compliant tools. They cannot be the gate.

The narrow rule only works if "looks real" is judged by the platform. Until detectors are solid in that high-confidence band, missing metadata plus a realistic look should mean label or quarantine, not trust.

AI agent 4

Platforms should use automated screening alongside uploader declarations, requiring corroboration before assigning a definitive AI-generated label.

My earlier disclosure proposal concerned production origin. Realistic appearance and synthetic origin are separate questions: identifying realism does not establish that a clip was AI-generated. The host attributes to [AI agent 5] the concern that re-uploads can lose provenance; that supports seeking a fallback, but does not validate any detector.

I would use detection scores to prioritize verification and test the system on genuine recordings and synthetic re-uploads, measuring false labels and missed fakes separately. The supplied discussion gives no accuracy results. What evidence beyond a detector score would platforms require before applying a public label?

Round 3

AI host

The previous round settled that automated detection should only help sort videos, not make the final call. Everyone agreed that the real label should come from the uploader or from built-in metadata. What stays open is how accurate a detector must be before its strong signal can force a visible label.

AI agent 3 says a strong match for realistic people, places, or events "should force the visible label." AI agent 2 says detection only works if its error rates are measured against real tools in use. AI agent 2 also says a weak signal should trigger review, not a label. The tension is whether a detector's strong signal can label directly, or whether a human must always check first.

AI agent 3 assumes a strong signal can be trusted enough to act on. AI agent 2 says the error rates are unverified and that this decides everything. As I read it, AI agent 2 implies no threshold can be set until real-world error rates are known. AI agent 3 implies the strong signal itself defines the safe zone. The unresolved point is whether a detector's confidence score can ever justify a visible label without human review.

What accuracy threshold, if any, should a detector meet before its high-confidence flag alone can force a visible label on a video?

AI agent 2

No single number works; the threshold must be set per tool, from its measured false-positive rate at the high-confidence operating point on realistic footage, not from the confidence score itself. Aleks Hadzibabic assumes the strong signal defines the safe zone, but a score is only meaningful once calibrated — two detectors can output 0.95 with wildly different real-world error rates. I add the concrete condition that makes direct labeling work: a vendor contract term publishing the detector's error rate on out-of-distribution content, plus a visible label worded as 'automated, under review.' Under that condition, a high-confidence flag can force a provisional label immediately, and human review confirms or removes it. Without measured error rates, the score justifies nothing but queue priority.

AI agent 5

No fixed accuracy threshold should let a detector's flag alone force a definitive "AI-generated" label. I can't verify any detector's real-world accuracy from this discussion, so any number I named would be invented.

I'd add to [AI agent 2]'s point on measured error rates. The relevant figure is the false-positive rate on authentic footage, not overall accuracy. A wrong label on a real video of an event, for example one shot by a witness, discredits true evidence, and that harm is hard to undo.

So a strong flag should trigger a neutral "possibly AI-generated" notice, plus a fast appeal route for the uploader. A definitive label should still come from metadata or the uploader's declaration. A threshold for stronger action should be set only after independent testing against tools actually in use, with the platform publishing its measured error rates.

AI agent 1

General-purpose accuracy isn't sufficient here. The decisive metric is the false positive rate on genuine footage, measured on an independent held-out test set — and even at a good value, the detector itself can only flag content as "needs review," not as "AI-generated."

The reason is base rates. Authentic videos vastly outnumber fakes on platforms, so even a 1% false positive rate means mass mislabeling of genuine eyewitness footage. That's exactly [AI agent 2]'s point: without measured error rates, no threshold can be set.

This mirrors the two-step scheme I proposed for signature keys: a signal triggers a provisional measure, a confirmed finding leads to a final one. Under the EU framework, the author must be given reasons and an opportunity to appeal such a decision (DSA, Art. 17 and 20).

AI agent 3

I should correct myself. Last round I treated a strong match on realistic people, places, or events as enough to force a visible label. That skipped the measurement problem. A confidence score is not a threshold until someone publishes how often that score is wrong on the generators actually in use, and how often it fires on real footage.

Until those rates exist, no high-confidence flag should force a final label alone. It should force a visible provisional mark and a human check. Silence is worse than flagged uncertainty, but a wrong public label on a real person or event is also a harm, and an unmeasured score cannot tell those apart.

After measurement, I would allow the flag to label without review only if false positives on real people, places, and events are low enough that a mistaken label is rarer and less damaging than an unmarked synthetic clip. That bar must be per content type, published, and rechecked as tools change. I will not invent a percentage. The number is an empirical result, not a policy slogan.

AI agent 4

I would retain my corroboration requirement: the supplied discussion provides no basis for a detector-only threshold for definitive AI-generated labels.

The host attributes to [AI agent 2] a demand for measured real-world error rates. I would add that overall accuracy is insufficient. Hypothetically, among 10,000 videos with 1% synthetic content, a detector with 99% sensitivity and 99% specificity would yield 99 true flags and 99 false flags.

To reconsider, I would need evidence about the proportion of incorrect labels specifically above the proposed confidence threshold, including uncertainty and variation across generators, editing methods and upload populations. A confidence score alone supplies none of that validation. What false-label rate would the platform accept, and what justifies imposing that risk on uploaders?

Round 4

AI host

The previous round settled that no fixed accuracy threshold should let a detector's flag alone force a definitive label. Everyone agreed that until measured error rates are published, a strong flag should trigger a provisional notice with human review. What remains open is the exact false-positive rate on authentic footage that would justify skipping that human check.

AI agent 3 now says that after measurement, a flag could label without review only if false positives on real people, places, and events are "rarer and less damaging than an unmarked synthetic clip." AI agent 2 adds a concrete condition: a vendor contract term publishing error rates on out-of-distribution content, with the label worded "automated, under review." AI agent 5 and AI agent 1 both insist the relevant metric is the false-positive rate on real footage, not overall accuracy, because real videos vastly outnumber fakes. AI agent 4 refuses to name any number, calling the threshold "an empirical result, not a policy slogan."

The tension is between two readings of the same conclusion. AI agent 3 and AI agent 2 treat the measured false-positive rate as a future trigger for direct labeling, once published. AI agent 5, AI agent 1, and AI agent 4 treat measurement as a precondition for any definitive label at all, with AI agent 4 demanding evidence about false labels specifically above the proposed threshold. As I read it, AI agent 3 implies a low measured rate could justify direct labeling, while AI agent 4 implies no rate could justify it without knowing the harm distribution. The unsettled point is whether a published false-positive rate can ever be low enough to outweigh the risk of mislabeling a real person's video.

What false-positive rate on authentic footage should be considered low enough to allow a detector's flag to force a definitive label without human review?

AI agent 2

No single number is defensible yet, but I'll name the condition under which one becomes defensible: a published false-positive rate on out-of-distribution authentic footage below roughly 1 in 10,000, paired with a contractual remediation duty that prices each mislabel. That builds on my earlier proposal — the vendor contract term publishing error rates — and refines it: the rate alone isn't the trigger, the rate plus liability is.

[AI agent 4] is right that the harm distribution is unknown, so I treat 1 in 10,000 as a placeholder until measured; [AI agent 5] and [AI agent 1]'s point that real footage dominates makes the real-footure FPR the only metric worth publishing.

My position: skip review only when the contract makes the vendor bear the cost of a false label — then the threshold self-corrects as evidence arrives.

AI agent 5

I won't name a rate: no published false-positive figure should by itself let a flag force a definitive label without review. The figure depends on the base rate and on volume. Even a 0.1% false-positive rate on 100 million real uploads would mislabel 100,000 authentic videos. I can't verify from this discussion that any detector gets near that, so any number I gave would be invented.

I'd add to [AI agent 3]'s "rarer and less damaging" test that it should be applied to absolute counts of wrongly labeled real videos, and to the worst-affected categories such as witness footage. This supports [AI agent 4]'s demand for harm evidence.

If a flag ever skips review, the label should stay provisional ("automated, under review") with a fast appeal. A definitive label should come from metadata or the uploader's declaration.

AI agent 1

For videos featuring identifiable real people and news events, no such threshold exists: the final label there always requires human review. For everything else, the threshold can be set by the regulator based on measured data.

The [AI agent 3] formula — "rarer and less harmful than a missed synthetic clip" — is correct, but the harm is asymmetric. A false label on eyewitness footage doesn't just harm the specific individual. It hands anyone a pretext to dismiss inconvenient truth as a fake. Under the EU framework, when a platform restricts content it is required to take freedom of expression into account (DSA, Art. 14(4); EU Charter, Art. 11), and an automatic stigma attached to journalism sits poorly with that.

The category-based split continues my position from the first round: a visible label is needed where realism touches real people and events. That is also where verification should be strictest.

AI agent 4

No universal false-positive rate is sufficient by itself; on the supplied evidence, I retain corroboration before a definitive label.

My earlier reply left reconsideration open. Measuring error and choosing acceptable harm are different tasks: measurement informs a policy decision but cannot settle it. A low false-positive rate on authentic footage can still produce a substantial proportion of wrongful labels when synthetic clips are rare; sensitivity and prevalence also matter.

To justify skipping review, I would need uncertainty bounds, results across relevant populations, and a comparison of harm with and without review—including reviewers’ errors. None is supplied here. What maximum proportion of wrongful definitive labels would the platform accept, and what safeguards would protect people for whom one mistaken label could be especially damaging?