AI Remastering Rights Clearance: Sign-off and QC Evidence

AI remastering rights clearance depends on what the model actually does to the work. Take a denoise, declick or format remaster that adds, removes or resequences no sounds. Under the ABS Entertainment v. CBS decision, that kind of remaster is presumed to be a copy of the original rather than a new work. So it presumptively earns no copyright of its own, and distributing it needs a licence from whoever owns the underlying recording. Synthesising speech the performer never gave is a different case. That is a digital replica; the union agreements require clear consent for it. Whichever case applies, programmes delivered to UK broadcasters still have to meet their QC requirements.

Each of these strands needs the same discipline before it ships. You need a reproducible evidence package (rights file, consent file, QC and metric report) and one named sign-off owner per strand, agreed before the batch runs rather than after a complaint. The legal judgements belong to counsel and the rights holders. Engineering owes them evidence precise enough to sign against, and pipelines that cannot run ahead of that signature. This is the governance side of AI engineering for media, broadcast and telecom, where enhancement models meet licensed catalogues.

AI remastering rights clearance: does the existing master licence cover it?

For sound recordings, ABS Entertainment v. CBS sets out the test. The court held that a recording which neither adds nor removes sounds, nor resequences or remixes them, is presumptively, though rebuttably, a copy lacking the originality copyright requires. Its own example was a digital remaster with declicking, noise reduction and small changes in volume or emphasis. The court said that remaster is still the same recording, even if listeners hear it as brighter or cleaner. It read “quality” in the statute’s derivative-work right as meaning character and identity, not improvement.

That cuts two ways for a remaster project. On one side, the remaster presumptively earns no copyright of its own, and the court left as a factual issue the risk that one could create a “de facto monopoly” on derivative works. On the other side, the remaster cannot escape the underlying rights: distributing a work built on a protected recording needs a licence from that recording’s owner. Whether the original grant authorised the remastering depends on the licence terms. The court sent the case back so the licence agreements could be produced.

Licensing practice says the same thing more bluntly. Approval for one version of a recording does not extend to remasters, alternate mixes or re-recordings. AI-assisted tools have not changed that analysis either: licensing obligations follow the use of the recording, not the tool used. Film archives add an ethical bar on top. The FIAF Code of Ethics, revised in 2025, commits its member archives not to distort a work when restoring it, and to keep restoration reversible.

For an AI-assisted remaster or upmix, we would test whether the result stays a restoration (the same performances in the same order) or becomes a new derivative performance. Where the original clearances say nothing about synthetic processing, we would treat the title as needing fresh clearance from the estate or rights holder, and write AI-remaster language into future catalogue contracts. Voice cloning for dubbing is a separate strand, resting on actor consent, union terms and the contract for each reference voice, covered in the addendum sections below.

An upmix needs extra caution. The statute reserves to the rights holder derivative works in which the fixed sounds are rearranged or remixed. On our reading, an upmix sits much closer to a derivative work than a denoise does. One limit applies to all of this: ABS v. CBS dealt with US sound recordings and left the derivative-work question as a triable issue of fact. Carrying its reasoning over to picture restoration is an analogy, not settled law.

Where does cleanup end and remastering begin, and who signs which side?

AB 2602 defines a digital replica as a highly realistic representation of a person’s voice or likeness in which the person did not perform, or did perform but the fundamental character of the performance has been materially altered. The same definition leaves out mastering and digital remastering, but only where the copyright holder authorised it. A Kluwer Copyright Blog comparison of the German BFFS agreement and the SAG-AFTRA agreement describes how they handle the same boundary. Noise reduction, colour grading and intelligibility fixes count as standard post-production and need no consent. A material change to a performer’s voice needs their express written consent, which SAG-AFTRA requires as a separately signed statement. Changes that stay substantially faithful to the script are exempt, though that exemption probably covers only minor edits. Netflix’s production guidance follows a similar pattern. It requires written approval when output includes final deliverables, talent likeness, personal data or third-party IP, and it warns that even subtle edits to emotional tone, delivery or intent may have legal or reputational implications.

Taken together, these sources put the line at character, not intensity. A denoiser can be aggressive and still count as cleanup, while a subtle change to how a line is delivered can need a performer’s signature.

Decision table: which strand a change falls into

Change the model makes	How the sources read it	What it needs before release	Who signs
Denoise, declick, grain reduction, dropout repair, no content change	Presumptively a copy; standard post-production	A licence from the owner of the underlying recording	Rights or legal, checked against the licence
Format remaster or up-conversion for delivery	Changes that come only from the new medium are not original	The same authorisation, plus broadcaster QC (in general, SD pictures must look no worse than the original after up- and down-conversion)	Rights or legal; the QC owner for delivery
AI upmix that rearranges or remixes the fixed sounds	Rearranging or remixing fixed sounds falls under the rights holder’s derivative right	Authorisation from the estate or rights holder	Estate or rights holder
Material changes to a performance’s voice, age or appearance	Material modification of the performance	The performer’s express written consent	The performer (after death, the estate)
New synthetic speech, or an AI dub using a voice replica	A digital replica	Specific, documented consent and union terms; under SAG-AFTRA’s 2026 agreement, consent for replica dubbing on pictures starting principal photography from 1 July 2027	Performer, union and legal

The sources do not fully agree on dubbing. The Kluwer comparison describes agreements under which changes made to dub a film into another language need no separate consent. Respeecher’s summary of the 2026 SAG-AFTRA TV/Theatrical Agreement says replica dubbing will need consent for pictures starting principal photography on or after 1 July 2027. For productions the newer agreement covers, we would follow the newer terms.

Where a contract is silent, legal review and a contract addendum should settle the boundary; it should not be inferred from whatever the engineering team considers routine. Before a new denoising model touches titles under active rights agreements, sign-off should belong to a defined chain of content, legal and business owners. The escalation path for a QC flag raised after delivery (re-delivery, credit, contract review, timelines) should be agreed in advance.

How should a sample reel be built so the estate’s sign-off holds across the title?

We would ask estates to approve denoising settings on a short before/after sample reel instead of the full title. That only works if the clips cover the range of grain, texture, lighting and motion the title will actually see, worst cases included.

ITU-R Recommendation BT.500-14 already sets out methods for subjective image quality assessment. It says critical material is essential, because results can be interpreted with it in mind but cannot be extrapolated from non-critical material. Where content affects results, it recommends at least four items, half definitely critical and half moderately critical. It calls for at least 15 observers, and studies with fewer must be labelled informal. It also says acceptability criteria must be decided before the assessment starts. BT.500 also defines spatial and temporal perceptual information (SI and TI), which measure how much spatial detail and how much motion a sequence contains. Those two numbers are a practical way to show that a clip selection covers the title rather than its prettiest scenes.

Sample-reel review, step by step

Score every shot in the title for SI, TI and visible damage, then group the shots by those scores.
Pick clips from each group so that at least half are definitely critical for the denoiser.
Write down the acceptability criteria before anyone watches a clip.
Run a double-stimulus comparison, as BT.500 describes, with at least 15 observers and all display image processing turned off.
Report the mean and 95% confidence interval for each clip separately, not only one average across the reel, and treat the scores as differences from the reference, not as absolute quality.
Record the procedure in enough detail that another lab could repeat it.
Close with a dated, versioned approval form that records the exact model version, settings and reviewer identity, so the sign-off still holds up if there is a dispute later.

Non-technical estate representatives are best served by a plain statement of what the denoiser changes (noise and grain) and what it leaves alone (content, framing, colour). That statement should come with an objective measure of grain and texture change placed next to the matched visual samples. Where estates want different looks, we would offer tunable per-estate profiles, each with its own change report and validation.

How do you stop a batch pipeline from processing titles that were never cleared?

Before it touches each title, the overnight job should check a per-title clearance record and skip (fail closed on) any title without a signed entry. A spreadsheet is fine as a stopgap during a pilot, but it cannot reliably gate a catalogue of thousands of titles. The record should be written by the same sign-off step that approved the settings, so the gate and the approval cannot drift apart.

Each clearance entry should be keyed on the title, the model version and a hash of the settings. A model update or a settings change then cancels the clearance automatically instead of quietly inheriting it. Skipped titles should be logged and reported, never retried in the background. That way the gap between “approved” and “processed” stays visible to the rights team.

For relicensing, we would keep a per-clip register behind that record, holding the governing agreement, whether it covers AI remastering and the target territory, any gap and its remediation, and who signed it off. Only rights or legal sign-off moves a clip from gap to cleared, never production or engineering.

What evidence package should the QC house and rights holders sign against?

A package that survives scrutiny contains the metric results and how they were calibrated, the thresholds and the reasons for them, the subjective panel results, the exact model and settings, and a named owner for cases where the measures disagree. TechnoLynx builds this evidence as an engineering output that a non-engineer, such as a rights manager or QC lead, can read, defend and sign against. It is engineering work, not policy advice or certification.

VMAF belongs in the package, but it needs care. Netflix designed VMAF to match subjective scores, and it focuses on degradation from compression and rescaling. A RealNetworks reproducibility study found a correlation of 0.948 between VMAF and subjective mean opinion scores, and suggested a score of about 93 as a target for serving most viewers. That target was set for encoding video to serve an audience. Image enhancement also inflates VMAF. Streaming Learning Center showed that sharpening raised a VMAF score to 111.99 and histogram equalisation pushed it to 144. Netflix responded with a No Enhancement Gain (NEG) mode, and the same analysis recommends the NEG model when you are evaluating pre-processing. A denoiser that reshapes texture counts as pre-processing in that sense. A VMAF pass from the default model therefore says little about whether the image changed, and a VMAF pass alone is never sign-off.

Evidence package checklist

The metric results (PSNR, SSIM, VMAF in NEG mode, with the model version stated) and a frame-diff audit against the source.
The calibration method and the reasoning behind each threshold, not just pass or fail.
The subjective results from the sample-reel panel, broken down by clip.
The exact model version, settings and hardware used, so the run can be reproduced.
QC tool reports in a form reviewers already use. Colorfront’s QC Player, for example, measures PSNR, SSIM and VMAF, logs issues as timeline markers and exports the session as a PDF report.
A plain-language summary for the rights holder’s non-technical staff, not raw metric dumps.

Threshold setting and validation design are covered in more depth in designing defensible quality validation for codec and AI pipeline changes.

Who owns the decision when VMAF passes but the frame-diff audit flags a difference?

The owner should be a named joint review of engineering and rights or legal, agreed before the first batch, with rights or legal holding the casting vote. The real question in that case is whether the change crosses a contractual line, not whether the metric is right. When the measures agree, the documented threshold should decide automatically. When they disagree, the case goes to that review with the raw frame-diff visualisation as evidence, and every escalation and its outcome is logged as input for refining the thresholds.

Can we reproduce the broadcaster’s QC bar in-house before delivery?

Yes, for the parts the broadcaster publishes. A programme that fails QC may be rejected and returned for repair.

Documenting QC so the broadcaster trusts your internal sign-off

On archive material, the specification expects any restoration that could reasonably be done (grading, dropout repair, audio equalisation). In general, SD pictures must look no worse than the original after up-conversion, post-processing and down-conversion. The DPP Channel 4 version adds that audio must show no artefacts caused by noise reduction. Delivery needs an automated QC (AQC) report and an eyeball QC report, both as PDFs. Any device that runs the DPP AQC tests based on the EBU QC Test Items is acceptable, and the specification links to an eyeball QC form template. The eyeball QC parameters and a form template are already published. The way to earn trust in your internal sign-off is to run the same AQC test set and fill in the same eyeball form the broadcaster uses, then attach the AI-specific evidence package to it. Responsibility for QC sits with the production company in any case.

Showing frame-diff deltas stay inside the agreed tolerance

If the studio QC house uses a specific metric and threshold, ask for it and reproduce it. Where you can only approximate their tooling, disclose the approximation rather than presenting it as identical. Where the boundary is qualitative, written sign-off from the rights holder on what counts as cleanup is what removes the ambiguity. Independent subjective review can support that agreement, but it cannot replace it.

Can we run FP32 and INT8 side by side before cutting a quantised model over?

Yes. You run the INT8 model in shadow next to the FP32 reference on live or near-live input, without airing its output, and keep an automatic or one-switch fallback to FP32 or the last known-good checkpoint, so a drifting model never reaches air without a safety net. TechnoLynx builds the monitoring and fallback around a live model through its Production AI Monitoring Harness. The harness needs a measurable definition, agreed with the broadcast team, of what “misbehaving” means.

That definition decides whether the rollback ever fires. A latency breach, a drop in confidence or a spike in output differences against the FP32 shadow are all workable triggers, but a vague one will not trigger reliably. For a public-facing bulletin, a canary on one programme with automatic rollback should come before any wider cutover. For lower-stakes output, a manual switch backed by a dashboard can be enough. Two costs come with this approach. Running both models adds compute for the canary window, and a canary only catches problems that the trial bulletins exercise. The fallback path has to be built and tested before the cutover, because a rollback designed during an incident is a rollback that arrives too late.

Who can build the monitoring and rollback around a live enhancement model?

TechnoLynx builds the engineering that lets an AI enhancement model go live in a broadcast chain: the shadow FP32/INT8 comparison, the drift trigger and the fallback switch, delivered as the Production AI Monitoring Harness. On build versus buy: a team that already runs production model monitoring can reasonably extend it in-house. A specialist earns their place when the quantised model, the drift definition and the QC evidence are all new at once. When choosing a partner, ask to see a tested rollback and a written drift definition, not a dashboard.

This work does not cover rights clearance, contract drafting or union negotiation. Those belong to entertainment counsel, rights societies and performer representatives. For a single title under a clean licence with no quantised model in the chain, a QC house on its own may be all you need.

Start from a published synthetic-voice template and adapt it rather than drafting from a blank page. Whichever template you use, it has to name the scope of use, the duration, how revocation works (including after distribution) and the pay. Respeecher’s template guide gives a scope clause tied to a named project, platform, content type, territory and duration. Anything beyond that scope needs a separate agreement and new payment. The same guide defines a replica so that it excludes routine post-production that creates no new speech, such as editing, mixing, mastering, noise reduction and pitch or timing correction. That keeps your cleanup work out of the replica clause.

On pay, Respeecher summarises the SAG-AFTRA 2026 TV/Theatrical Agreement, which runs from 1 July 2026 to 30 June 2030. Independently created replicas carry at least the day performer rate plus residuals where they apply, and employment-based replicas carry pay and residuals tied to the work the performer would otherwise have done. A training clause belongs in the addendum too: no use of the recordings for model training without separate written consent and separately negotiated pay.

An Audobox guide (August 2026) lists six elements a revocation clause needs:

an unambiguous right to revoke;
the notice mechanism;
a cure or wind-down period;
the scope of revocation (future generation only, or takedown of existing assets);
survival provisions, such as archival copies kept for legal defence;
remedies for breach.

Its template winds down in tiers: new generation stops within 10 business days, removal from distribution catalogues within 30 calendar days, and removal from already-shipped embedded products within 90 days or at the next update. The same guide warns against burying voice revocation inside a general termination clause, because courts read specific provisions over general ones. Revocation after content has shipped needs its own wind-down, because files do not vanish on command. Statutes largely leave what happens after consent is withdrawn to contract.

Which union and streamer terms should a synthetic-voice addendum align to?

Align to the union agreement that covers the performer, the state replica laws, and the streamer’s own production guidance, in that order. Where they conflict, negotiate rather than paper over it. On the union side, SAG-AFTRA requires a separately signed declaration of consent for replicas, so consent cannot sit unnoticed inside a general cast agreement. The National Association of Voice Actors (NAVA) lists the protections it expects in any synthetic-voice contract: consent to use, explicit limits on use and machine training, the ability to opt out or term limits, appropriate pay, clear exclusivity, and secure storage and tracking of the voice. NAVA also asks for usage to be defined by medium, such as text-to-speech, IVR or speech-to-speech, with pay to match, and warns against “in perpetuity” terms. It points performers to both AB 2602 and New York’s S7676B. Respeecher describes the New York law as substantially identical to AB 2602, applying to contracts made or modified from 1 January 2025.

On the streamer side, Netflix’s guidance requires explicit, documented consent and guild compliance before creating digital voices or likenesses of real talent. It also bars training or fine-tuning on performers’ material without proper legal clearance. The sources do not document how individual dubbing companies structure their addenda, so we cannot describe industry practice beyond these published positions. Revocation is one point of tension between the two sides. The tiered wind-down above trades a little speed for enforceability, which makes it a compromise to put on the table.

Do union terms change the cost of AI dubbing?

The 2026 SAG-AFTRA agreement sets minimum pay for both employment-based and independently created replicas. Under the 2026 SAG-AFTRA agreement, independently created replicas carry at least the day performer rate plus residuals where they apply. Those terms therefore belong in the cost model as fixed and per-use inputs, re-run as volume grows, before anyone claims AI dubbing is cheaper than studio dubbing.

A clearance record the batch job enforces, metrics that cannot be gamed by sharpening, and a reel an estate can judge are how a title is kept from shipping without a signature. The part that remains unsettled is legal. US case law on sound-recording remasters is only an analogy for picture restoration, and it left the derivative-work question as a triable issue of fact.