Congress Wants to ‘Label Deepfakes.’ The Catch: Your Next Fake Video Won’t Be AI-Generated (and NIST can’t label what isn’t there).
Congress is building an “AI labeling” regime around watermarks and provenance. But the most viral deceptions often use real footage, cheap edits, and context collapse—no model required.
Key Points
- 1Recognize the trap: Congress can label AI outputs, but most viral “fake videos” rely on real footage, edits, and context collapse.
- 2Expect enforcement gaps: NIST standards and FTC rules take years, while deception spreads in hours and thrives on cross-platform metadata loss.
- 3Treat every clip as a claim: labels and provenance are signals, not truth certificates—especially when reposts and screen recordings erase the trail.
The next “fake video” you fall for probably won’t be a deepfake.
Not because generative AI is slowing down—quite the opposite. The problem is more mundane and, in some ways, more alarming: the most persuasive deceptions often require no synthetic face, no photorealistic model, no AI pipeline at all. They use old footage with a new caption. A clip trimmed to remove context. Audio re-posted over unrelated video. A screen recording that erases the breadcrumb trail.
Congress, meanwhile, is building its response around a narrower idea: label AI outputs. On April 24, 2026, Reps. Valerie Foushee (D‑NC), Don Beyer (D‑VA), and James Moylan (R‑Guam) introduced (reintroduced) the Protecting Consumers from Deceptive AI Act, a proposal that would lean on NIST for standards and the FTC for enforcement. Reporting on the bill describes a future of machine-readable disclosures, watermarking, digital fingerprinting, and content provenance metadata for AI-generated content.
That framework makes intuitive sense—until you ask a basic question. What happens when the deception isn’t AI-generated? No model, no watermark, no provenance to declare. And the internet’s most viral lies tend to prefer the cheapest tools available.
“A labeling regime can be both overinclusive and underinclusive—flagging harmless edits while missing the deception that never touched a generative model.”
— — TheMurrow Editorial
The bill Congress keeps coming back to—and what it actually does
S&P Global’s reporting emphasizes the operational heart of the bill: NIST task forces would develop standards for technical disclosures—think interoperable ways for systems to mark AI-generated media in a consistent, verifiable format. The same reporting also notes a reality most people miss when they hear “Congress passed a bill”: implementation takes time, and obligations land years later after standards and rules are developed.
NIST and the long road from law to reality
- Congress authorizes standards work
- NIST convenes task forces and develops technical guidance
- Agencies craft rules
- Industry builds tools that comply
- Platforms adopt verification and display methods people can understand
Even when everyone agrees on the objective, interoperability is hard. The modern media ecosystem does not behave like a neat chain of custody. Files are clipped, cropped, re-encoded, reposted, and screen-recorded. Each step can strip or degrade metadata. A standard that looks robust in a lab can fail on a meme account’s workflow in minutes.
The bill’s assumed implementation pipeline
- ✓Congress authorizes standards work
- ✓NIST convenes task forces and develops technical guidance
- ✓Agencies craft rules
- ✓Industry builds compliant tools
- ✓Platforms adopt verification and display methods people can understand
FTC enforcement: powerful, but not magical
The uncomfortable implication: enforcement becomes easiest against actors who comply most of the time and hardest against the accounts that treat deception as the business model.
“The gap between ‘we can label AI’ and ‘we can stop deception’ is where most policy lives—or dies.”
— — TheMurrow Editorial
Why “label AI content” solves the wrong problem surprisingly often
Many of the most consequential fake videos are assembled with:
- Conventional editing (selective cuts, splicing, speed changes)
- Re-contextualized footage (real clip, false story)
- Misleading captions that tell viewers what to think
- Re-uploads or screen recordings that erase provenance
- Hybrid workflows (real video + AI voiceover; real audio + manipulated transcript; real photo + small AI edits)
None of that requires a generative model to fabricate pixels end-to-end. If there’s no AI generation step, an “AI-generated” label cannot exist. NIST cannot standardize provenance for something that never had AI provenance.
Overinclusive vs. underinclusive: the policy trap
1) Overinclusive: benign creative edits get flagged as “AI” or “synthetic,” chilling legitimate expression and confusing audiences who begin to treat any disclosure as suspicious.
2) Underinclusive: the next viral deception—crafted with conventional tools—sails through unmarked, because the system only polices what it can label.
The result is a false sense of security. Viewers learn the wrong lesson: “If it’s unlabeled, it’s real.” That is the opposite of media literacy.
How AI labeling can fail in two directions
Before
- Overinclusive—benign creative edits flagged as “AI
- ” chilling expression and confusing audiences
After
- Underinclusive—cheap
- non-AI deception goes unmarked and spreads faster
The most damaging lies are often “cheap”
A labeling approach that centers only on generative AI risks chasing the most novel technique rather than the most common one.
Key Insight
Content provenance is a better frame—but it’s still brittle in the real world
The U.S. Copyright Office lists versions across sessions, including S.4674 (2024) and S.1396 (2025). The text of S.1396 (introduced April 9, 2025, 119th Congress) defines a “deepfake” as synthetic or synthetically modified content that appears authentic and creates a false impression. It also defines “content provenance information” as “state-of-the-art, machine-readable information documenting the origin and history” of content.
That is a meaningful shift. It acknowledges that manipulation is often incremental: edits, re-encodes, reposts, and context collapse.
Provenance systems run into “stripping attacks”
- Metadata stripping during upload and re-encoding
- Screenshots and screen recordings, which flatten a file into a new object
- Cropping and clipping that removes embedded signals
- Cross-platform reposting, where platforms don’t preserve the same fields
S&P Global’s description of the labeling bill’s long implementation timeline matters here. The longer standards take, the more time bad actors have to adapt workflows designed to defeat those standards. Meanwhile, everyday users keep sharing screen recordings because they are easy and fast.
Common ways provenance gets destroyed
- ✓Metadata stripping during upload and re-encoding
- ✓Screenshots and screen recordings that flatten a file into a new object
- ✓Cropping and clipping that remove embedded signals
- ✓Cross-platform reposting that fails to preserve the same fields
Provenance absent doesn’t mean provenance false
Any policy that leans on provenance must confront a UX problem: how platforms present absence without implying guilt.
“Provenance can tell you ‘where this came from.’ It can’t, by itself, tell you ‘what to believe.’”
— — TheMurrow Editorial
The hidden battle: trust infrastructure and verification UX
Who signs, who verifies, who is trusted?
- the entity attaching it (model provider, camera maker, editor, platform)
- the security of the signing method
- the independence of verification tools
- the governance around updates and disputes
The April 2026 proposal leans on NIST to standardize, which is sensible: NIST is built for technical consensus. Still, standards do not decide social trust. People do.
The platform problem: labels that don’t travel
- download → repost
- private messaging → public upload
- clip → remix
- re-encode → stream
If disclosure becomes a patchwork—present on one platform and invisible on another—bad actors will route content through the weakest links.
The “unlabeled = safe” cognitive trap
That problem is not hypothetical. It is built into the distinction between AI-generated content and non-AI manipulation.
Editor’s Note
What the April 2026 bill gets right—and where critics will press
First, it treats deceptive synthetic media as a consumer harm, not merely an online nuisance. That frames the issue in terms of accountability rather than cultural panic. Second, it recognizes that voluntary, fragmented labeling isn’t enough; interoperability requires standards work, and the bill explicitly points to NIST. Third, FTC enforcement suggests a familiar legal toolkit.
Supporters quoted in the House release—Encode AI, IEEE‑USA, and the Society of Composers & Lyricists—also bring useful perspectives. Engineers want implementable standards. Creators want clarity about attribution and authenticity. Consumer advocates want fewer scams and manipulative impersonations.
The strongest critique: it fights the last deepfake
The second critique: timelines lag viral cycles
That mismatch doesn’t mean “do nothing.” It means any legislative response must be paired with near-term strategies that don’t depend on perfect provenance.
Practical takeaways: how to read video evidence in the labeling era
For everyday readers and voters
- If a clip is labeled AI-generated, ask what the label actually indicates (generation, modification, or unknown).
- If a clip is unlabeled, don’t treat it as verified. Absence of provenance is common in ordinary reposting.
- Look for context loss: abrupt starts, missing lead-in, reaction shots without setup, or captions that claim more than the footage shows.
How to read a viral clip in a labeling era
- ✓If a clip is labeled AI-generated, ask what the label actually indicates (generation, modification, or unknown).
- ✓If a clip is unlabeled, don’t treat it as verified; absence of provenance is common in reposting.
- ✓Look for context loss: abrupt starts, missing lead-in, reaction shots without setup, captions claiming more than footage shows.
For journalists, editors, and creators
- Preserve originals and maintain internal provenance logs.
- When publishing, provide context prominently: where, when, and how you obtained media.
- Anticipate screen-recording: publish companion explainers or frames that are harder to excerpt without losing meaning.
For platforms and policymakers
A balanced approach includes:
- standards for provenance and disclosures (NIST’s domain)
- enforcement against deceptive commercial practices (FTC’s domain)
- platform UX that explains uncertainty without implying certainty
- resilience against stripping and repost workflows
- education that emphasizes “verification is situational,” not “labels equal truth”
The grim paradox is that the internet rewards speed. Policy has to reward verification.
Balanced approach (not a single lever)
Enforcement against deceptive commercial practices (FTC)
Platform UX that explains uncertainty without implying certainty
Resilience against stripping and repost workflows
Education that emphasizes verification is situational—not labels equal truth
A better north star: don’t label “AI,” label “claims”
The more durable goal is to reduce the power of manipulated media to smuggle in false claims. Provenance can help when it survives. Labels can help when they are present and understandable. Neither substitutes for context.
The April 2026 bill’s instinct—transparency, accountability, standards—moves in the right direction. The danger is mistaking a tool for a solution.
If lawmakers want policy that survives the next wave of fakes, they will have to design for the messiest reality: deception built from real footage, edited with ordinary tools, passed through platforms that erase the trail.
A society that treats “unlabeled” as “true” is easier to manipulate than a society that treats every viral clip as a claim that must earn belief.
“A society that treats ‘unlabeled’ as ‘true’ is easier to manipulate than a society that treats every viral clip as a claim that must earn belief.”
— — TheMurrow Editorial
Frequently Asked Questions
What is the Protecting Consumers from Deceptive AI Act?
Introduced (reintroduced) on April 24, 2026 by Reps. Valerie Foushee, Don Beyer, and James Moylan, the bill aims to create accountability and transparency standards for generative AI content. Reporting describes it as focusing on machine-readable disclosures and technical standards such as watermarking, fingerprinting, and provenance metadata, with NIST involved in standards-setting and the FTC as an enforcement lever.
How soon would AI labeling requirements take effect if the bill passed?
Not quickly. S&P Global’s reporting emphasizes a multi-stage timeline: standards development and rulemaking come before real-world obligations. That means requirements could arrive years after enactment. The lag matters because viral media cycles move in hours or days, while standards and compliance ecosystems take much longer to build and coordinate.
If content is unlabeled, does that mean it’s real?
No. Unlabeled content may be real, but it may also be misleading or manipulated without generative AI. It could also be a labeled file that lost metadata through reposting, re-encoding, cropping, or screen recording. Treat labels as one signal among many, not a truth certificate.
What kinds of “fake videos” aren’t deepfakes?
Many high-impact deceptions use conventional editing rather than AI generation: selective cuts, splicing, speed changes, or real footage presented with a false caption or story. The research also highlights cheapfakes, re-uploads that strip metadata, and hybrid workflows (for example, real video with an AI voiceover). These can evade “AI-generated” labeling entirely.
What is the COPIED Act and how is it different?
The COPIED Act (tracked as S.4674 (2024) and S.1396 (2025)) focuses on protecting content integrity through concepts like content provenance information—defined in S.1396 as “state-of-the-art, machine-readable information documenting the origin and history” of content. Compared with narrow AI labels, provenance aims to document where media came from and how it changed, even when edits aren’t fully AI-generated.
Can watermarking and provenance survive reposting on social media?
Sometimes, but not reliably. Provenance can break through metadata stripping, re-encoding, cross-platform reposting, cropping, and screen recording—all common behaviors. A robust provenance ecosystem requires both technical standards and platform cooperation, plus user-facing tools that can explain what provenance does and doesn’t prove.
