Captions vs the Article 50 label
Captions are speech on screen. Article 50 is a first-exposure AI mark. One does not satisfy the other.
Captions are not the AI label. Article 50 of the EU AI Act wants a visible or audible mark at first exposure on certain synthetic media. Burned-in speech text is a transcript of what was said. Putting words on screen does not disclose that a face was generated, and an EU icon in the corner does not caption the take.
The two objects share a frame and none of a legal test. Teams collapse them because both are "text we added." That collapse is how a photoreal talking-head ships with perfect karaoke and no disclosure, or how a real founder video gets an "AI-generated" stamp because someone ran auto captions.
This page is the split. It is not the copy-paste label pack — that lives in the Article 50 disclosure template. It is not TikTok's AIGC toggle, which fires on realism of people and scenes rather than on whether you captioned. Platform toggles and Article 50 stack. Captions sit in a third bucket.
What Article 50 actually marks
Article 50 has applied since 2 August 2026. The Commission's FAQ on transparency obligations splits the work into four areas: people must know they are talking to an AI; providers of generative systems must machine-mark synthetic audio, image, video and text; emotion-recognition and biometric-categorisation uses must be disclosed; deployers who publish deepfakes — content resembling real people, places or events that would appear authentic — must put a visible or audible label at first exposure. AI-generated text on public-interest matters is labelled unless it had substantive human review, not a spell-check.
Guidelines were adopted on 20 July 2026. Provider and deployer are different jobs. Providers owe machine-readable marking, with a grace period to 2 December 2026 for systems already on the market before 2 August, and watermark-detection interoperability to 2 February 2027. Deployers who publish the deepfake owe the label the viewer sees. Artistic, creative, satirical or fictional works may label in a way that does not spoil the work. Standard assistive editing is exempt. Content generated and published before 2 August 2026 is not retroactive; generated before and published after is covered.
Three official EU icons exist for the visible layer: fully AI-generated, AI-modified, and a basic icon with an interactive second layer (EU icons for labelling AI-generated content). Fines can reach €15 million or 3% of worldwide turnover.
The test for a creator who publishes a clip is usually the deployer test: would a reasonable viewer take this person, place, or event as authentic? If yes, label at first exposure — first frame or first second of audio, not the description field, not a caption cue at 0:08. If no — obvious animation, obvious fiction — the deepfake duty may not fire. Other duties might.
None of that paragraph is about subtitles.
What captions actually are
Burned-in captions are speech (and, if you authored SDH, non-speech) rendered into the pixels so a mute or deaf viewer can follow the soundtrack. The Versely path is transcription plus a look plus a burn: add captions to a video, 165 language codes, 21 basic presets at 1× credits, 9 dynamic presets at 2×. Caption Studio's 45 CSS presets are the drawing.
Captions answer "what was said." They do not answer "was this face generated." They do not answer "did a model write this script." A perfect transcript of a cloned voice is still a transcript of a cloned voice. The disclosure, if the clip needs one, is a different layer.
Assistive editing is the exemption that actually touches this page. A caption pass on real footage — transcribe the founder, burn Classic, ship — is the ordinary accessibility job. Treating that pass as "we used AI, so Article 50 applies" is how a real video gets a synthetic-media stamp it did not earn. The Commission's assistive-editing exemption exists so that spell-check, colour-grade, and captions do not become deepfake labels.
The inverse error is worse for viewers. A generated spokesperson with no label and with karaoke captions reads as a real person who happens to have good subtitles. The captions make the clip more legible as reality, not less. That is the opposite of first-exposure transparency.
Assistive editing vs synthetic media
Draw the line on the picture, not on the tool list.
Assistive, usually not an Article 50 deepfake label: captions and subtitles on footage of a real person saying things they said; a crop; a colour grade; a hook overlay you wrote; noise reduction; a transcript. The tool can be an AI caption generator. The output is still words that were spoken.
Synthetic media the deployer may have to label: a generated or face-swapped human who would pass as real; a cloned voice impersonating a real person; a photoreal scene presented as a camera original; real footage substantially altered so a person appears to say or do something they did not.
A caption on the second pile does not complete the duty. You still label at first exposure. Then you caption the speech so the mute viewer can read it. Two layers. Two jobs.
Public-interest AI text is a separate Article 50 branch. A voiceover script drafted by a model and read on camera can be "AI-generated text on a public-interest matter" even when the picture is a real reporter. Captions of that read are still captions. The label, if the text was not under substantive human review, is still a label. Human review here means editorial control of the claims, not a pass through a spell-checker, and not "we burned subtitles."
TikTok's AIGC label, YouTube's realistic-synthetic disclosure, and Meta's C2PA/self-disclosure rules are platform policy. They are not Article 50. They may help a viewer. They do not replace the legal test, which follows the audience and the content. Do not treat "we switched the TikTok toggle on because we captioned with AI" as compliance. Do not treat "we did not toggle because the captions are just text" as a reason to skip a deepfake label on a generated face.
Two objects on one frame
Place them so they do not erase each other.
The Article 50 mark has to be perceivable at first exposure. That usually means a corner badge, an opening slate, or a spoken line in the first second. It does not mean a caption cue that appears when the guest starts talking. A disclosure that lands after the hook is not first exposure.
Speech captions live in the mid-lower band, off the platform chrome and off the face. If you put the EU icon in that same band, the first caption line covers the legal mark, or the legal mark covers the first claim. Put the icon top-start or as a pre-roll slate. Put the transcript where captions go.
Do not write the disclosure as if it were a subtitle ("this video is AI-generated" timed to the waveform). Timed text that follows speech will vanish when the next sentence arrives. A legal mark that can be swiped past in a karaoke pop is not a mark.
Do not use Signature or Headline as a substitute for the icon. A pretty "made with AI" in Pacifico is not one of the three official EU icons, and it is easy to read as decoration. If you need the official marks, use them. If you need a spoken disclosure, put it in the first second of audio and caption that line like any other speech.
Fictional or satirical work may label in a way that does not spoil the piece. That is still a label, still at first exposure, still not a caption style. A Reveal-family typewriter of the plot is not a disclosure.
Ship both when both apply. Ship only captions when the clip is real speech on real picture. Ship only the label when there is no speech to transcribe — a silent generated plate still needs the mark, and auto captions will find nothing to burn.
This is not legal advice. It is a production split: the transcript is the transcript; the Article 50 mark is the mark. The AI caption generator does the first. The disclosure template does the second. Do not ask either object to do the other.
Label and caption in Versely.
FAQ
If I caption a generated UGC ad, have I labelled it?
No. Captions make the speech readable. Article 50, when it applies, wants a first-exposure mark that the clip is synthetic. Do both.
Does using an AI captioner trigger Article 50 on real footage?
Standard assistive editing is exempt. Transcribing a real person and burning subtitles is that job. A generated face is a different job.
Can the EU icon live in the caption track?
Do not time it as a cue. First exposure is the first frame or the first second of audio, not line three of the transcript. Keep the icon off the speech band.