Broadcast caption compliance in 2026 begins with deadlines that have already passed. The European Accessibility Act has applied since June 28, 2025. The EU AI Act's Article 50 transparency obligations started to apply on August 2, 2026. The FCC's closed-caption display-settings compliance date arrived on August 17, 2026.
Three dates suggest one compliance event. They are three different engineering problems.
Article 50 governs transparency around specified AI systems and AI-generated or manipulated content. The FCC's new rule governs how viewers find and use caption display settings on covered devices and multichannel video programming distributor experiences. The European Accessibility Act governs accessible access to audiovisual services, including the transmission, synchronization and control of accessibility components.
The practical result is easy to state. Caption compliance has become an end-to-end systems problem. An accurate transcript can still produce a failed viewer experience when the caption track disappears from one rendition, drifts after ad insertion, cannot be customized in the player or lacks an applicable AI disclosure.
How do the three rules differ?
The three rules regulate different owners and different stages of the media chain.
Rule | Live date | Directly regulates | What it means for a media chain |
|---|---|---|---|
European Accessibility Act, Directive (EU) 2019/882 | June 28, 2025 | Covered products and consumer services that provide access to audiovisual media services in the EU | Accessibility components must travel intact, render accurately, stay synchronized and remain under user control; electronic programme guides and access interfaces must be accessible |
EU AI Act, Article 50 | August 2, 2026 | Providers and deployers of specified AI systems | Providers may owe machine-readable marking; deployers may owe clear disclosure for deepfakes and certain public-interest text; exceptions and the provider/deployer distinction matter |
FCC 47 CFR § 79.103(e) | August 17, 2026 | Manufacturers of covered apparatus and multichannel video programming distributors | Caption display settings must be readily accessible under four factors: proximity, discoverability, previewability, and consistency and persistence |
Earlier caption rules remain in force. In the United States, 47 CFR § 79.1 continues to apply caption quality standards built around accuracy, synchronicity, completeness and placement. In Europe, caption availability also sits within the Audiovisual Media Services Directive and Member State law.
A facility can therefore pass one test and fail another. The ASR output may be accurate while device settings remain buried. The source caption track may be complete while one regional package drops it. Editorial review may be sound while a generated public-interest summary lacks a required disclosure.
What does EU AI Act Article 50 mean for broadcast captions?
Article 50 creates separate duties for AI-system providers and professional deployers. The correct analysis depends on the role, the transformation performed, the content published and the review completed before publication.
Role and duty | Trigger | Broadcast example |
|---|---|---|
Provider: machine-readable marking under Article 50(2) | An AI system generates or manipulates synthetic audio, image, video or text, subject to stated exceptions | A vendor supplies a system that generates synthetic voice, video or materially rewritten text |
Deployer: clear disclosure under Article 50(4) | A professional user publishes a deepfake or AI-generated or manipulated text intended to inform the public on a matter of public interest | A newsroom publishes an AI-generated story summary or alert without qualifying human review and editorial responsibility |
The European Commission published its final Article 50 guidelines on July 20, 2026. The guidelines list transcriptions of conversations among examples of standard editing or minor alteration that do not semantically change the source. A faithful transcription may therefore fall outside the provider's machine-readable marking duty in Article 50(2).
That provider-side example does not resolve every caption workflow.
The deployer's test in Article 50(4) is separate. A professional deployer publishing AI-generated or manipulated text to inform the public on a matter of public interest must disclose its artificial origin unless the text received substantive human review or editorial control and a natural or legal person holds editorial responsibility. The Commission's Article 50 Q&A says a grammar check, a procedural glance or review by another AI system is insufficient.
The final guidance contains no broadcast-caption example that settles every live-news case. Broadcasters using unreviewed live ASR captions should have counsel analyze the deployer-side disclosure question in each priority EU market. The provider-side transcription example alone cannot carry that conclusion.
What counts as substantive human review?
Substantive human review requires a knowledgeable person to examine the meaning of the text and possess authority to approve, alter or reject it. The Commission also requires a person or organization to hold editorial responsibility for publication.
Three operating cases deserve separate policies:
- Prerecorded captions reviewed before publication. A qualified editor reviews names, numbers, meaning and context, then approves the final version. This workflow may qualify for the human-review and editorial-responsibility exception when the other conditions are met.
- Live AI captions supervised during transmission. A supervisor may correct errors or switch to a backup, yet may not review every cue before viewers see it. A post-air log creates useful evidence. It does not become pre-publication review of text already displayed.
- AI summaries, headlines, alerts and rewrites. These transformations can create public-interest text that goes beyond faithful transcription. They need their own disclosure decision and approval path.
Version boundaries matter. If an AI system changes text after editorial approval, the Commission says the output may lose the exception. The approval record should identify the exact final text or a verifiable hash, the reviewer, the approval time and any later processing step.
The Article 50 text requires applicable disclosures to be clear and distinguishable by first exposure and accessible to the audience. The Commission's AI-content icons are optional. An icon by itself does not establish compliance.
Does Article 50 still have a grace period?
One limited transition remains. Qualifying AI systems placed on the market before August 2, 2026 have until December 2, 2026 to meet the Article 50(2) marking and detection duty. The transition does not postpone deployer disclosure duties generally.
The enforcement ceiling is material. The Commission says Article 50 fines can reach €15 million or 3% of the preceding financial year's worldwide turnover. Proportionality can be considered for smaller companies. National market-surveillance authorities will handle most enforcement.
Teams should document why a system qualifies for the limited transition rather than treating December 2 as the effective date for Article 50 as a whole.
What changed at the FCC on August 17?
The FCC's 2026 deadline concerns ready access to caption display settings. It did not create a new speech-to-text accuracy threshold.
Under 47 CFR § 79.103(e), the FCC assesses access through four factors:
- Proximity. Caption display settings must be gathered in one area reached through a mechanism reasonably comparable to a button, key or icon.
- Discoverability. Covered entities must conduct usability testing with consumers and disability groups, make good-faith efforts to correct identified problems and train customer-facing staff.
- Previewability. Viewers must be able to preview captions while changing display settings.
- Consistency and persistence. Covered experiences must use device-level settings and application programming interfaces as the rule specifies, allowing preferences to carry into applicable applications.
Scope details affect the implementation plan. The rule addresses covered apparatus manufactured in or imported into the United States. Third-party pre-installed applications are excluded from this requirement. Screens under 13 inches receive an achievability qualification. The codified compliance language covers devices using next-generation operating systems deployed after August 17, 2026. The FCC's January 2025 public notice fixed the date.
This rule belongs on the player, device, application and customer-experience roadmaps. Caption generation remains important, but it is not the layer changed by this deadline.
Does the FCC require 99% caption accuracy?
The FCC does not mandate a general 99% caption-accuracy threshold. The familiar number came from a caption-vendor best practice, not the regulatory standard.
In its 2014 caption-quality order, the FCC reproduced a vendor example showing how 99% word accuracy might be calculated. Footnote 228 says the Commission was not adopting accuracy metrics as part of its captioning standards. The current rule remains qualitative and context-sensitive. Captions must satisfy accuracy, synchronicity, completeness and placement standards, with live and near-live programming evaluated case by case.
Internal measurement still matters. A 98%, 99% or 99.5% target can serve as an engineering or contractual control. It is not a federal safe harbor.
A useful quality-control program should measure:
- meaning errors involving names, numbers, negation and speaker identity;
- delay between speech or sound and the displayed cue;
- cue completeness through program and commercial transitions;
- placement relative to names, scores, warnings and other essential graphics; and
- track presence, language identity and decoder behavior in every distributed version.
Word error rate alone misses several of these failure modes. Procurement language should keep the distinction clear. A claim of 99% FCC-compliant captions combines a model benchmark with a legal conclusion the FCC did not make.
What does the European Accessibility Act require of a media service?
The European Accessibility Act regulates the access, delivery and interface layer for audiovisual-access services. It does not publish one EU-wide caption word-error or latency threshold.
Directive (EU) 2019/882 applies to covered services provided to consumers after June 28, 2025, including services that provide access to audiovisual media services. The definition can include websites, online applications, set-top-box applications, mobile apps, media players and connected-TV services.
For those services, the Directive requires:
- electronic programme guides that are perceivable, operable, understandable and compatible with assistive technology, including information about available accessibility features;
- accessibility components such as subtitles for deaf and hard-of-hearing viewers, audio description, spoken subtitles and sign-language interpretation to be fully transmitted;
- adequate quality for accurate display;
- synchronization with sound and video; and
- user control over display and use.
The wording is outcome-based. Caption availability also interacts with the Audiovisual Media Services Directive and national law. A service distributed across several EU countries still needs a jurisdiction-by-jurisdiction matrix.
Some transitional provisions run longer than the headline date. Article 32 allows service providers to continue using certain products already lawfully used for similar services until June 28, 2030. Some service contracts concluded before June 28, 2025 can continue unchanged until expiry, capped at five years. These are narrow provisions. A launch before June 2025 does not create a general exemption.
Where does each rule land in the caption pipeline?
The three regimes become manageable when the media chain is divided into auditable stages.
Stage | Control to implement | Rules it supports |
|---|---|---|
Source | Feed the captioner clean program audio; supply rundowns, names, places and scripts when available | FCC quality and downstream accuracy |
Generate | Record model or vendor, version, language, timestamps, confidence and transformations performed | Article 50 scoping and incident reconstruction |
Review | Define which outputs require substantive human approval before publication and who holds editorial authority | Article 50 deployer exception where applicable |
Package | Validate service descriptors, language tags, cue continuity, positioning data and every output format | FCC quality and EAA full transmission |
Distribute | Check every rendition after encoding, ad insertion, manifest generation, CDN handoff and regional packaging | EAA transmission and synchronization; FCC completeness |
Render | Test caption controls, preview, persistence, operating-system preference handoff and assistive-technology access | FCC § 79.103(e) and EAA user control |
Prove | Preserve time-aligned samples, alarms, operator actions, approvals, disclosure state, corrections and final hashes | Regulatory response, complaint handling and internal audit |
The viewer consumes the rendered output. A passing mezzanine file or green status on an ASR console proves little about the final device experience.
This distinction also affects architecture and cost. Teams comparing deployment models can use the same stage-by-stage approach described in our analysis of cloud ASR versus on-prem GPU captioning. Location changes control over media, models and logs. It does not change the legal test.
What evidence should a broadcaster keep?
A useful evidence trail must reconstruct the text viewers received and the system decisions that produced it.
For each service and affected program-time window, retain enough information to answer:
- Which source produced the text?
- Which model, vendor and version generated or changed it?
- Was the transformation transcription, translation, summary, rewrite or synthetic media?
- Which person reviewed it, when did that happen and what authority did the person hold?
- Did an AI step alter the text after approval?
- Which disclosure decision applied, and did the disclosure survive clipping, syndication and download?
- Did the caption track survive packaging, ad insertion and every regional rendition?
- What did the viewer see, including timing, position and display controls?
- Which alarm fired, who acknowledged it and what corrective action followed?
Deployment location affects how easily a team can preserve this record. On-prem systems may give operators more control over retention and versioning. Cloud services may reduce infrastructure work and provide managed audit features. Neither location establishes compliance on its own.
Where does this analysis remain unsettled?
The largest open question is how Article 50(4) will be applied to unreviewed live AI captions for public-interest programming across EU Member States. The Commission's final guidance gives useful rules for providers, deployers, transcription and editorial review, but no example resolves the full broadcast workflow.
The EAA also depends on national implementation and the service's distribution footprint. The FCC display-settings rule has detailed boundaries around covered apparatus, MVPD experiences, operating systems and third-party applications.
Teams should resist a single global compliance flag. Maintain a matrix by service, territory, device class, legal role, transformation type and editorial workflow. Have counsel validate the matrix for priority markets. That honest boundary is more defensible than turning a broad regulatory summary into a product-level compliance claim.
What should broadcast teams do this week?
Start with an exposure register and a delivered-output test. Both can begin without buying software.
- Map legal roles and markets. Identify the provider, deployer, video programmer, MVPD, device manufacturer, application owner and editorially responsible entity for every service.
- Classify every AI transformation. Separate faithful transcription from translation, summary, rewrite, synthetic voice, manipulated imagery and generated video. Give each class its own review and disclosure decision.
- Test what viewers receive. Verify caption presence, language, continuity, timing, placement and display controls after the final CDN on representative devices. Include ad breaks and regional variants.
- Define review and evidence. Name the people with substantive approval authority, prevent unreviewed AI changes after sign-off and preserve time-aligned samples with alarms and operator actions.
- Ask every vendor the same questions. Which stages can the system observe? Can it distinguish transcription from transformation? How are model versions and approvals recorded? Can it prove what reached the viewer? What happens when a track disappears after the vendor's monitoring point?
The last question often reveals the real boundary. Compliance evidence has to extend to the point where the audience receives the service.
Sources
- European Commission: Guidelines on Article 50 transparency obligations
- European Commission: Article 50 questions and answers
- AI Act Service Desk: Article 50 text and explanation
- European Commission: EU icons for labelling AI-generated content
- FCC: January 2025 notice confirming the August 17, 2026 compliance date
- FCC: Closed Captioning Display Settings Small Entity Compliance Guide
- 47 CFR § 79.103: caption display and settings requirements
- 47 CFR § 79.1: television closed-caption quality requirements
- FCC 14-12: 2014 closed-caption quality order and best practices
- EUR-Lex: Directive (EU) 2019/882, the European Accessibility Act