Keep Every Published AI Image, Voice and Video Labelled, Cleared and Traceable

T09-L05 · Create media · Level 5 Operator · 18-minute read, plus 4 hours of doing

A journalist, a lecturer and a question nobody can answer

A year ago the institute's communications office started using AI for media. The seminar videos get captions from the pipeline in T09-L03. Press releases get AI illustrations. Since spring the public lecture series has German versions, dubbed with each lecturer's own cloned voice: an AI copy of the voice, made from recordings, that can speak text the person never said. Everyone likes the results. About 300 items are online.

On Monday a science journalist writes: the image in last week's press release, showing organoids under a microscope, is that a real microscopy image? It isn't. It was generated, and it carries no label. The head of communications asks you how many other images like it are online. You don't know. On Wednesday a lecturer who has moved to another university asks that her cloned voice no longer be used. Her voice is in 14 lectures. Which ones, and who can take them down?

In a company it's the software firm whose how-to videos are now dubbed into five languages with the voice of a support colleague. She leaves in December. A product teaser turns out to use a voice model whose licence forbids commercial use.

You are at Level 5 Operator. You run AI media for everyone who publishes through you, over months and years. When it breaks, the damage is public, and it has to work when you're on holiday. This book gives you a register, label rules, a monthly check, a takedown route and a deputy.

After this you can

• Keep one register of every AI-made or AI-changed asset you publish, with its tool, inputs, rights, consent, label and approver.
• Set label rules that meet the EU AI Act's disclosure duty, and check every month that labels and Content Credentials survive on your channels.
• Handle the three common incidents (an unlabelled asset, a withdrawn voice consent, a licence that doesn't allow your use) with a tested takedown route.
• Hand the operation to a deputy who can answer an auditor's question about any asset in five minutes, without you.

Before you start

• Books: T09-L04 for pinned model versions and licence checks. T12-L05 for the organisation-wide inventory this register feeds. The four-quadrant method is in T01-L01.
• A list of your channels: website, video page, social accounts, newsletters, slide templates. Ask whoever can publish to each.
• Your organisation's contacts for data protection, legal or contracts, and, in a company with a works council, whoever speaks to it.
• A deputy, named by your head of department, with two hours in their calendar.
• About 4½ hours: 18 minutes to read, about 4 hours of doing, over three weeks.

Last verified 23 September 2026. Since 2 August 2026 the EU AI Act requires you to disclose deep fakes you publish, including cloned voices. ElevenLabs clones only your own voice; its team plans start at $299 a month. XTTS-v2's licence forbids commercial use. Details for experts at the end of this book.

The idea in one page

One register, not a folder of files. Every AI-made or AI-changed asset you publish gets one row. The row is what you show when anyone asks, so it must answer six questions:

Label by kind, not case by case. The EU AI Act's transparency rules apply since 2 August 2026. Two matter here. Providers of generators must mark outputs in a machine-readable way. And whoever publishes a deep fake, meaning AI-made or AI-changed image, audio or video that resembles real people, objects, places or events and would falsely seem authentic, must disclose it. A dubbed lecture in which a real lecturer speaks German she never spoke is one. A generated "microscopy image" of organoids very likely is too. Write one label rule per kind of asset, stricter than the law where your audience needs it, so nobody has to decide at the upload button.

Content Credentials are the backup, the visible label is the rule. Content Credentials is the common name for C2PA, an open standard for signed information inside a file that records which tool made it and what was changed. Several generators, such as Adobe Firefly, add it automatically. But websites, social platforms and slide exports often strip it when they resize or convert a file. So the visible label or spoken line is what your audience relies on. The credentials, and the register row, are what you show an auditor.

Consent for a voice or face is a contract with an end. It says what the voice may be used for, in which languages, for how long, how the person can withdraw, and what you'll do then. At a company, consent from an employee is only valid if they could genuinely say no, which is why the works council belongs in the conversation.

Four jobs make it an operation.

• Watch: once a month, compare what's online with the register, check labels and credentials on each channel, and check consents, licences and vendor terms for changes.
• When it breaks: a written takedown route. Who removes an asset from each channel, within how many hours, and what replaces it.
• Who takes over: organisation-owned accounts, never personal ones, and a deputy who has run the monthly check and a takedown drill.
• What you show an auditor: for any asset, the register row with links to the consent, the licence, the approval and the credentials check.

Worked example: from 300 unknowns to a register, four quadrants

Both versions follow the four quadrants from T01-L01. The lab one is told in full. Names, numbers and dates are illustrations.

Lab: the institute's communications office

Structure it. Miriam, who runs digital media in the communications office, writes down what the head of communications needs. "For every AI-made asset we publish, we can say within five minutes what made it, what went in, whose voice or face is in it, how it's labelled and who approved it. When something has to come down, it comes down the same day, even when I'm away."

Understand it. She starts with what's online, not with what people remember. The website team exports a list of every image and video uploaded since last September; the video platform exports its catalogue. She asks the eight people who can publish one question each: which of these did you make with AI, and with which tool? The picture after a week:

• About 300 AI-assisted items: 180 captioned videos, 70 illustrations, 40 dubbed lectures, a few social clips.
• 60 of the illustrations have no record at all: no prompt, no tool, no approver.
• The cloned voices live in the ElevenLabs account of a former colleague who set up the dubbing. ElevenLabs is the voice service the office uses for dubbing. The account is paid with his personal credit card and reimbursed monthly.
• The website's image service strips Content Credentials from every picture it resizes. The Firefly illustrations had them when they were downloaded, and lost them on upload.
• The lecturer consent form for dubbing is one sentence: "I agree that my voice may be used for translated versions of my lecture." No end date, no way to withdraw, no list of languages.

The finding that changes her problem is the account. She had thought of this as a labelling job. It turns out the institute doesn't control the service that holds 40 lecturers' voices. If the former colleague closed his account, every voice would be gone, and nobody could delete them either. Back to Structure it: "First the institute takes control of the accounts and the consents. Then the register. Then the labels. The two open cases (the organoid image and the lecturer's withdrawal) are handled through the new route, as its first test."

Choose tools.

• Register: a list in the office's SharePoint site, one row per asset, because the institute backs it up and a new colleague gets access on day one. No new tool.
• Voice: ElevenLabs stays, because its dubbing keeps each speaker's voice across languages and the lecturers already know it. The account moves to a Scale workspace, ElevenLabs' plan for a shared organisation account, owned and paid by the institute, with Miriam and her deputy as administrators.
• Voice consent in the tool: ElevenLabs lets people make a Professional Voice Clone, its high-quality copy trained on long recordings, only of their own voice. So each lecturer makes the clone in their own ElevenLabs account, verifies it by reading a few lines aloud, and shares it with the institute's workspace through a private sharing link. The institute pays for their plan, the smallest one that includes a professional clone. The voice stays in the lecturer's hands, which suits a consent that can be withdrawn.
• The alternative is German captions from the T09-L03 pipeline instead of dubbing. That's cheaper, needs no voice consent, and it's what a lecturer gets who says no.
• Credentials check: c2patool, a free command-line program from the Content Authenticity Initiative that reads Content Credentials from a file, run on the website export once a month. For a single file, the Content Credentials Verify website, verify.contentauthenticity.org, does the same in a browser for a file you upload.
• Labels: text, written once per kind of asset.

Build it: take control of accounts and consents (week 1). IT and the head of communications approve the institute workspace. Each lecturer with a cloned voice gets an email: a new consent form, and the choice to make and share their own clone as described above, or to switch to captions. The new form, one page:

VOICE CONSENT · Public lecture series · v2
Speaker        [name]
Use            German (and, if ticked, French / Spanish) versions of my lectures
               in the institute's public lecture series. No other content.
Where          Institute website and video page. No advertising.
How long       Until I withdraw, at the latest 3 years from signing.
Withdrawal     Email to media@[institute]. Within 30 days, dubbed versions are
               replaced by versions with captions, and my voice is removed
               from the institute's voice account.
Label          Every dubbed version says, at the start and in the description:
               "German version with AI voice dubbing, in the speaker's own voice."
Signed         ______________  date ______

The old account's voices are deleted once each lecturer has either shared a new clone or chosen captions. The former colleague confirms the deletion in writing, and the confirmation goes into the register.

Build it: the register and the label rules (week 2). Miriam fills the register from the exports, oldest first. For the 60 illustrations without a record, the rule is simple: if nobody can say what made it and what went in, it comes down or is replaced. Twelve are still in use and get regenerated through Firefly with a recorded prompt; the rest are removed.

The label rules, one line per kind of asset:

Then the two open cases go through the route. The organoid image comes down the same afternoon and is replaced with a real image from the group, which the group leader supplies; the press release gets a correction note. The lecturer's withdrawal: her 14 lectures get German captions from the T09-L03 pipeline, the dubbed versions come down on day 12, and her voice is removed from the workspace's voice list. Both cases get a line in an incident log with dates.

Build it: the runbook. Before anything can be pulled back without her, Miriam writes one page and puts it next to the register:

RUNBOOK · AI media · v1
Register     SharePoint site "Communications", list "AI media register"
Accounts     ElevenLabs workspace (admins: Miriam, Jonas); Firefly on the institute's Adobe licence
Takedown     website: web team, within 4 hours · video page: Miriam or Jonas, same day
             social: whoever posted it, same day
Replacement  captioned version for a dub; real image or no image for an illustration
Steps        1 remove  2 put the replacement up  3 update the register row
             4 line in the incident log  5 answer whoever asked
Monthly      first working day: exports, credentials check, four checks, check log

Build it: the monthly check. Once, she installs c2patool. On a Mac that's brew install c2patool with Homebrew, the usual installer for command-line programs on a Mac. On Windows or Linux she downloads the archive for her system from the project's releases page on GitHub, unpacks it and puts the program in a folder on her PATH, the list of folders the terminal searches for programs.

On the first working day of each month, Miriam downloads the website's image export and runs one command over it in a bash terminal (on Windows, Git Bash). It asks c2patool for a short report on each file and writes all reports into one dated file:

for f in export/*.jpg; do echo "== $f"; c2patool "$f" --info; done > c2pa-check-2026-11.txt

In the file she looks at one thing per image: whether the report found a manifest, the signed record inside the file, and whether it validated.

Then four more checks, 30 minutes in all: every new asset on a channel has a register row; three random assets still show their label as published; no consent has been withdrawn or has reached its end date; and the terms pages of ElevenLabs and Firefly haven't changed since last month (she keeps the date she last read them in the register). The result is one line in the check log: date, what she found, what she did.

The first check shows what she expected: images on the website have no credentials, because the image service strips them. She asks the web team whether the service can keep them. Until it does, the register row, with the original file's credentials report, is what she shows when asked. The visible label is on the page either way.

Understand it: the takedown drill. A route nobody has tried is a guess. Miriam's deputy, Jonas, takes one real dubbed lecture and, using only the one-page runbook, takes it down from the video page, puts the captioned version up, updates the register row and logs it. It takes him 50 minutes. He stops once: the runbook says "remove from the video page", but not that the lecture is also embedded in two news posts on the website. That line goes in. Then he runs the monthly check alone.

Structure it: what an auditor sees. The head of communications picks an asset at random, a dubbed lecture from June, and asks Jonas: what made it, whose voice, did she agree, where's the label, who approved it? He answers from the register row and its three links in four minutes.

What it taught her: the labels were an afternoon. Taking the accounts and the consents back into the institute took two weeks, and without that no label would have helped.

Company: five languages and a colleague who's leaving

Structure it. Jana's voice dubs 150 how-to videos in five languages. She leaves in December. "Every video keeps a voice the company may use, the licences fit a commercial product, and marketing and support use one register."

Understand it. Two findings. Jana's consent was an email saying "sure, go ahead". The data protection officer points out that consent from an employee to their employer is only valid if it's freely given, and a voice used in 150 company videos is hard to refuse; the works council wants an agreement on employee voices before anyone else is asked. And the register check finds a product teaser with an AI voice made by a marketer on her laptop with XTTS-v2, an open voice model whose licence allows only non-commercial use. A product teaser is commercial.

Choose tools. A professional voice actor, hired under a contract that licenses her voice for AI dubbing of the how-to videos for three years, with ElevenLabs as the service. Employees' voices are used only under the new works agreement, the written agreement between management and works council, which says it's voluntary, can be withdrawn, and says what happens to the videos then. The teaser is regenerated with the contracted voice.

Build it. The register lives in the marketing team's SharePoint site, shared with support. The runbook adds a column for the works agreement and a rule: no employee voice without a signed consent under it. Jana's videos are re-dubbed over six weeks, published before her last day, and her voice is deleted from the workspace on her last day, with her written confirmation. The deputy runs the takedown drill on the teaser.

In both places the register was the easy part. What made it work was owning the accounts, having consents with an end, and a deputy who had pulled an asset back once.

What goes wrong

Voices and assets in a personal account. Symptom: the service that holds your cloned voices is paid by one person's credit card. Fix: an organisation workspace, paid by the organisation, with two administrators. Move everything, then delete the old copies with written confirmation.

Consent without an end. Symptom: someone withdraws, and nobody knows what was promised or how fast. Fix: a consent form with scope, duration, withdrawal route and what replaces the asset, and a check each month for consents that end.

Credentials assumed to survive. Symptom: you tell an auditor every image carries Content Credentials, and the website shows none. Fix: check what's actually online each month. Keep the visible label as the rule, and the register row with the original file's report as the record.

A licence that doesn't fit the use. Symptom: an open model's output turns up in a product teaser or a paid campaign. Fix: the register records the tool's licence and the intended use; the approver checks both. Regenerate with a tool whose terms allow it.

Labels decided at upload. Symptom: three colleagues label the same kind of image three ways, or not at all. Fix: one label rule per kind of asset, written in advance, and images that look like real results go to the group leader.

The operation lives in your head. Symptom: when you're on holiday, a takedown waits a week. Fix: a one-page runbook, a deputy who has done the drill, and the register where the organisation keeps records.

A register nobody updates. Symptom: the register lists 300 assets, and the website has 340. Fix: the monthly comparison with the channel exports, and a rule that nothing gets published without a row.

Do it yourself (4 hours, over three weeks)

1. Structure it (15 min). Write what your organisation must be able to say about any AI asset, and how fast a takedown must happen.
2. Understand it (60 min). Get exports from every channel. Ask everyone who publishes which items used AI, and with which tool. List every account that holds voices, faces or generated assets, and who pays for it. Read your current consent forms. Rewrite your problem sentence.
3. Choose tools (20 min). Decide where the register lives, which workspace becomes the organisation's, and how you'll check credentials. Name the alternative for people who say no to a voice clone.
4. Build it: accounts and consents (45 min, plus waiting). Move voices and assets into the organisation's workspace. Send the new consent form. Agree employee voices with the works council where you have one.
5. Build it: register, labels and runbook (60 min). Fill the register from the exports, remove or replace what nobody can account for, write the label rule per kind of asset, and write the one-page runbook: who takes what down from which channel, how fast, and what replaces it. Install c2patool.
6. Understand it: check and drill (40 min). Run the first monthly check and log it. Have your deputy do a takedown drill from the runbook and run the check alone. Fix the runbook wherever they stopped.

Exit check

Delivered artifact: a media register and a one-page runbook, in a place your organisation owns, with:

• one row for every AI asset published in the last quarter, with tool and version, inputs and rights, consent where a person is in it, label, approver and date;
• the label rules, one per kind of asset;
• the log of one monthly check, including what the credentials check found on each channel;
• the incident and takedown log with one drill, and the runbook changes it caused.

It passes when the deputy answers an auditor's question about a randomly chosen asset (what made it, what went in, whose consent, which label, who approved it) from the register alone within five minutes.

Rule to remember

If you can't say what made it, who agreed to it and who can take it down, it isn't ready to publish.

Details for experts

Checked on 23 September 2026. You don't need this to follow the book; it's for setting things up or answering a colleague. It isn't legal advice.

Further reading and tools

Next books

• T09-L04 · Transcribe Recordings That Must Stay in the Building: pinned model versions, licence checks and signed data routes.
• T09-L03 · Build a Caption Pipeline Your Team Can Rerun Every Week: the captions that replace a withdrawn dub.
• T12-L05 · Keep One List of Every AI System Your Team Runs, and Make It Survive You: the organisation-wide inventory of AI systems and its handover.
• T13-L05 · Keep Everyone's AI Skills Current, and Show It When Asked: training records, including for the people who publish.
• T15-L05 · Keep Your Group's AI Use Compliant and Running After You Leave: the same register-and-successor pattern for research.

Tools

• ElevenLabs: Taught. Dubbing with each speaker's own voice, in an organisation workspace.
• C2PA: Taught. Content Credentials, read with c2patool in the monthly check.
• Adobe Firefly: Catalogued. An image generator that attaches Content Credentials.
• XTTS-v2: Catalogued. An open voice model whose licence rules out commercial use.