Under the definition in the NO FAKES Act, a digital replica is a newly created, computer-generated, highly realistic electronic representation that is readily identifiable as the voice or visual likeness of an individual. Readily identifiable is the whole test. That means voice alone can be a replica, a materially altered real recording can be, and a character with no real-world counterpart isn't — because there's no individual to identify.
The definition is short and the boundary is where all the difficulty lives. Most creators don't have a rule problem. They have a "does this specific thing count" problem.
What does the definition actually say?
Worth reading in the bill's own words rather than a summary, because the operative phrase does more work than people expect.
S. 4591 defines a digital replica as a newly created, computer-generated, highly realistic electronic representation that is readily identifiable as the voice or visual likeness of an individual — embodied in a sound recording, image or audiovisual work in which the actual individual did not perform or appear, or in which they did and the fundamental character of that performance has been materially altered.
Two halves worth separating. It covers things built from nothing that look like someone. And it covers real footage of someone that's been changed enough to alter the fundamental character of the performance.
Our guide to licensing covers what that means for consent, payment and paperwork. This post is only about the boundary.
Which cases are clear?
Start with the ones that don't need argument.
| Case | Replica? | Why |
|---|---|---|
| A character described in words, with no reference images | No | Nothing is readily identifiable as an individual, because there's no individual. This is what Generate Actor does |
| Generated from photographs of a real person | Yes | The likeness is carried into the output by the reference, whatever the character is called |
| A cloned voice of a real person | Yes | Voice sits alongside visual likeness in the definition, not beneath it |
| Real footage of someone, materially altered | Yes | The definition covers a version in which they did appear, where the fundamental character of the performance has been materially altered |
| Your own face, uploaded | Yes — of you | Still a replica. You're simply the person identifiable in it |
| A human creator cast from a library | Yes — licensed | The consent and payment position is documented rather than absent |
Source: definition in S. 4591 as reported by Senate committee, read September 2026, plus Hexcoded product documentation. The bill has not passed. Requirements vary by jurisdiction.
Notice that five of those six are replicas. The permissive case is narrower than most workflows assume — specifically a character with no real-world counterpart and no real-person reference anywhere in its generation history.
Does voice alone count?
Yes, and this is the case creators most often miss.
The definition names voice and visual likeness together. A cloned voice attached to a generated face, or to no face at all, is readily identifiable as that person's voice. The absence of a face doesn't help.
That matters for narration, for dubbing, and for anything where you've used a real performer's recordings as reference material. If the output sounds identifiably like someone, treat it exactly as you'd treat their face.
Where this falls short. "Readily identifiable" is doing a lot of work and nobody has drawn a bright line. A voice resembling a well-known delivery style is different from one built on that person's recordings, and only the second is clearly a replica.
What about a composite of several real people?
Usually still a replica, and the reasoning follows from the definition.
The test is whether the output is readily identifiable as an individual — not whether you intended one. A composite built from four real faces might be identifiable as none of them, in which case there's no individual to identify. But composites don't blend evenly. If one source dominates, the output is readily identifiable as that person and the other three references are irrelevant.
So a composite isn't a laundering technique. It reduces the chance of identifiability, it doesn't eliminate it, and you can't tell which you've got by looking at your inputs.
Every reference image you use is a claim someone could make later.
Which is the argument for keeping a record even on characters you're confident are fully synthetic. On Hexcoded, each saved reference carries an optional 200-character description field — enough to note what the character is and what created it. It costs nothing to fill in and it's the only evidence that exists.
What's expressly excluded?
The bill lists five exclusions, and they're broader than most coverage suggests.
| Exclusion | What it covers |
|---|---|
| News and sports | Use in a bona fide news, public affairs or sports broadcast or account, provided the replica is materially relevant to the subject |
| Documentary and biography | Representing the individual as themselves in a documentary, historical or biographical work, including some degree of fictionalization |
| Commentary and parody | Use consistent with the public interest in bona fide commentary, criticism, scholarship, satire or parody |
| Incidental use | Where the replica is fleeting or negligible |
| Advertising a permitted use | Commercial announcements or advertisements promoting any of the four uses above |
Source: S. 4591, additional exclusions at clause (5)(A), read September 2026. The bill has not passed.
Two things about that list are worth flagging.
The documentary exclusion has carve-backs. It fails if the use creates a false impression of authenticity, or if it involves unauthorised musical synchronisation. So a documentary framing isn't a blanket permission.
And the fifth exclusion is neatly nested — an advertisement promoting a documentary that contains a replica is covered, because the documentary itself is.
One absolute override. None of the exclusions apply where a digital replica depicts sexually explicit conduct. That's stated separately and it admits no exception.
Are public figures treated differently?
Partly, and less than people assume.
Nothing in the definition weakens with fame — if anything the commercial value of a well-known likeness makes claims more likely, not less. What differs is which exclusions apply. News, documentary, commentary, criticism, scholarship, satire and parody all have protection that doesn't depend on consent, and public figures are the people those exclusions mostly concern.
So a satirical piece about a politician sits differently from an advertisement featuring the same face. The face is equally identifiable in both. The use is what changes.
Where this falls short. Whether a given piece is commentary or commercial use is exactly the question that ends up litigated, and it varies by state and by distribution. Don't assume a framing device converts an advertisement into satire — and note that the fifth exclusion only covers advertising a permitted use.
What about someone who has died?
Post-mortem rights exist under the bill, and the structure is far more interesting than the headline number.
The base post-mortem term is ten years after death. It's renewable in successive five-year periods, but only where the right holder can demonstrate active and authorised public use in interstate or foreign commerce during the preceding two-year window. The hard ceiling on all that renewing is 70 years after death.
Read that properly and it changes the practical answer. A post-mortem right lapses at ten years unless someone is actively and commercially using it and files to extend. Most estates won't clear the use requirement. "Capped at 70 years" is technically accurate and practically misleading — which is how it's usually reported.
Today, though, none of that is law. Post-mortem right of publicity is a state question and it varies substantially.
Worth noting that platform licences address this separately. Hexcoded's actor licence provides for heirs to request that new use stops, with unpaid earnings passing to the estate.
Does stylisation get you out of it?
Sometimes, and the threshold is higher than it looks.
The definition requires the representation to be highly realistic and readily identifiable. A heavily stylised or animated rendering can stop being highly realistic — but it stops being a replica when a viewer can no longer identify who it's meant to be, not when you've applied a filter. A recognisable caricature is still recognisable, which is the whole point of a caricature.
Realism and identifiability are separate requirements and both have to be met. A photorealistic image of nobody in particular fails the second. A stylised image that everyone recognises may fail the first — or may not.
The four questions that settle most cases
- Is the output readily identifiable as a specific real person?
- Did any real person's photographs, footage or recordings go into it?
- Is the voice identifiable, separately from the face?
- Does one of the five exclusions actually apply?
What's still unsettled?
Three areas where nobody should be claiming certainty.
Accidental resemblance. A fully text-generated character that happens to be readily identifiable as a real person. The definition turns on identifiability rather than intent, so "I didn't mean to" isn't a defence in itself — but how close is close enough has no settled answer.
Style versus identity in voice. Building on a genre of delivery rather than a person, and landing somewhere identifiable anyway.
Where the federal line will actually fall. S. 4591 has been revised once already and hasn't passed the Senate. Its final definitions may not match the current text.
Current as of the publication date. The NO FAKES Act is pending legislation and had not passed the Senate at the time of writing. Likeness law varies by state. Hexcoded's actor licence is governed by the laws of India, with jurisdiction in Gurugram, Haryana. Nothing in this post is legal advice.
- "Readily identifiable as the voice or visual likeness of an individual" is the whole test. Learn that phrase
- Voice sits alongside visual likeness, not beneath it. A cloned voice with no face is still a replica
- Real footage counts too, where the fundamental character of the performance has been materially altered
- A composite isn't a laundering technique. If one source dominates, the output is identifiable as that person
- Five exclusions: news and sports, documentary and biography, commentary and parody, incidental use, and advertising a permitted use
- Sexually explicit depictions override every exclusion. No exception applies
- Post-mortem rights would last ten years, renewable on proof of active commercial use, capped at 70. Not "70 years"
- Note what created each character, even the synthetic ones. It's the only evidence that exists
Under the NO FAKES Act's definition, a newly created, computer-generated, highly realistic electronic representation that is readily identifiable as the voice or visual likeness of an individual. It also covers real recordings of someone where the fundamental character of the performance has been materially altered.
Yes. Voice sits alongside visual likeness in the definition rather than beneath it, so a cloned voice is a replica of the person it belongs to whether or not any face appears. That applies to narration and dubbing as much as to on-screen performance.
Not reliably. The test is whether the output is readily identifiable as an individual. Composites don't blend evenly, and if one source dominates the output is identifiable as that person regardless of how many references you used.
Five: news, public affairs and sports where the replica is materially relevant; documentary, historical and biographical work representing the person as themselves; commentary, criticism, scholarship, satire and parody in the public interest; incidental or negligible use; and advertising that promotes any of those. Sexually explicit depictions override all five.
Under the bill, yes — but the structure matters. Ten years after death by default, renewable in five-year periods only where the right holder shows active authorised commercial use in the preceding two years, with a hard ceiling of 70 years. So most rights would lapse at ten unless actively used. Today it's a state question and it varies.
Only once it stops being both highly realistic and readily identifiable — both are required by the definition. A recognisable caricature is still recognisable, which is the point of a caricature.
Three tiers, one clear line
AI actors generated from a description have no real-world counterpart to identify. Human creators arrive with consent verified and payment on every render. Anything you upload carries whoever is in the photograph. The library is organised around the distinction this post describes.
See the actor libraryMore on characters, consent and reusable elements in Elements.