Your DAM Metadata Is Now Public: Captions, AI Search, and the 2027 Deadline

Metadata used to be an internal problem. It is not anymore. Caption files, transcripts, alt text, and asset descriptions now travel out of the DAM onto public web pages, where AI answer engines read them and repeat them to people who never visit your site. For healthcare organizations, that shift arrives with a compliance date attached: recipients of HHS federal financial assistance with fifteen or more employees must meet WCAG 2.1 Level AA by May 11, 2027.

A cardiologist records a patient education video and says "watch for signs of hypotension." The automatic caption writes "hypertension." Two letters. The opposite condition, and the opposite thing a patient should do about it.

Nobody catches it, because captions are not really anybody's job. The video goes live. Six months later an AI answer engine reads that caption file and repeats it to somebody searching for information about their own care.

That is a metadata problem. It just stopped being an internal one.

This is a post about what actually leaves your DAM, who is reading it now that it does, and what to do about it before an answer engine decides what your organization meant to say. Healthcare is where this got real first, but it affects all of us.

For as long as we have been doing this, metadata has been an inside-the-building argument. Someone says search is broken. We look under the hood and find search working exactly as designed. The tagging is what's broken. We fix the metadata, search starts behaving, everybody goes back to work.

That argument is over. Not because we won it, but because the stakes moved outside. What lives in your DAM does not stay in your DAM, and the things reading it are no longer just your own employees.

Start with what actually leaves the building

Ask a simple question about your own library: what lives in here that ends up somewhere public?

The list is longer than most teams expect. Transcript and caption files ride along with video. Alt text follows images onto your website. Social copy usually sits right beside the asset it describes, so whoever grabs the post grabs the words too. Product descriptions, campaign names, usage terms, all of it travels.

Every one of those fields is a place where your organization describes itself in its own words, and those words are increasingly read by machines that then describe you to other people. When AI trains on something wrong, you have not made one mistake. You have started a game of telephone you no longer control.

The part most teams get wrong about AI search

Here is the uncomfortable version, and it surprises people who have already paid for AI.

"One thing people are always surprised at is, if they invested in AI search, how come when they use the search bar it's still not accurate?" said Andrea Barrera, Head of the Management Team at Stacks. "It's because AI, as of August 2026, is going based off of visuals, or for audio, what it hears. It is not yet paying attention to your metadata and learning your terminology."

Yet. That last word is doing a lot of work.

That is the gap. AI search today is largely reading pixels and audio. It is not reading your controlled vocabulary, your product nicknames, or the fifteen years of institutional shorthand your team uses to find things.

That is changing. Platforms are adding business-term features so the system can build a working vocabulary, and the MCP connectors rolling out across DAM platforms are what will finally link the AI agents organizations already built to the assets in the DAM. But we are at the beginning of it, and Andrea's read on what that means is the whole point of this post: "It's important to start actually paying attention to your metadata now, so that when we get to that point, and I'm sure it's not going to be very long from now, your data is ready and you're not scrambling."

If you want the longer view on where platform AI is heading and what it can and cannot do today, we covered it in our guide to AI features in DAM platforms.

Workflow and governance are the actual work

Before anyone buys another AI feature, two unglamorous things need to exist.

The first is workflow. Even with no AI in the picture, someone has to be accountable for the alt text getting written, the copy getting attached, the transcript getting tied to the right video. AI does not remove that need. It relocates it. "I don't think anyone, at least in the near future, is planning on removing humans from that workflow," Andrea said. "They're always going to want a human to review and validate that whatever the alt text or the transcription the AI produces is accurate."

The second is governance, and it matters more the more sensitive your content is. "There's opportunity for so many issues when you're working with patients and HIPAA," said Becca Browning, DAM Librarian at Stacks. "That's probably one of the most important things in making sure your metadata is ready." If AI is going to read your library and speak on your behalf, deciding who can publish and what gets reviewed is not a nice-to-have.

Governance is the layer most programs skip and then rebuild under pressure. It is also the layer that makes everything above it hold, which is why we treat it as structural rather than optional in our guide to building a DAM strategy that outlasts the platform.

Where this got very real: healthcare video

The clearest example we are seeing right now sits in healthcare, and it has a deadline attached.

In May 2024, HHS finalized a rule requiring recipients of its federal financial assistance to make web content and mobile apps accessible under WCAG 2.1 Level AA. It reaches further than hospitals: doctors, dentists, clinics, emergency rooms and other providers, child welfare agencies, medical and nursing schools, and childcare and social service providers.

Compliance was originally set at May 11, 2026 for recipients with fifteen or more employees. HHS extended that by a year in an interim final rule effective May 7, 2026, so the dates are now May 11, 2027 for recipients with fifteen or more employees and May 10, 2028 for those with fewer than fifteen.

WCAG 2.1 Level AA requires captions for prerecorded video, and not just any captions. The standard expects them to carry the dialogue, identify who is speaking, and convey meaningful non-speech audio. Captions that actually match what happened.

The clients we work with did not panic. "They were actually pretty all right with the change and ready for it," Becca said. "Any apprehension was just trying to figure out the platform, and getting frustrated with the limitations of the platform." That frustration is worth naming, because as she put it, "web accessibility is becoming very, very important, so it's something that our DAMs should be really developing."

Why just turning on auto-captions is not the answer

Automatic transcription is a useful starting point, not a finish line, and medical vocabulary is where that shows fastest.

A systematic review published in the Journal of the American Medical Informatics Association, "Speech recognition for clinical documentation from 1990 to 2018," looked at nearly three decades of speech recognition in clinical settings. Across the studies it evaluated, reported word error rates ranged from 7.4 percent to 38.7 percent, and the share of documents containing at least one error ranged from 4.8 percent to 71 percent. That research covers clinician dictation rather than captioning, but the problem underneath is the same.

The hypotension example at the top of this post is one version of it. Here are two more. All of these are illustrative, not real client examples.

A pharmacist says "take 0.5 milligrams." The caption drops the decimal and publishes "take 5 milligrams." Ten times the dose, live on your website, with your logo on it.

And the one that would be funny if it were not going public: a clinician mentions a patient with a history of ileus, a bowel obstruction. The caption confidently introduces a man named Elias.

None of these are exotic. They are one letter, one decimal point, and one homophone. Then an answer engine reads the file and passes it along, and the game of telephone starts with a machine that nobody checked.

What to do about it

None of this requires a heroic project. It requires deciding that metadata is now an external-facing asset and treating it that way.

Inventory what actually leaves your DAM. Decide where transcripts and caption files live and how they stay tied to the right video, language, and version. Build the review step into the workflow instead of hoping someone catches it, and write down who owns it. Then get your vocabulary in order, because when platform AI starts learning your terminology, it will learn from whatever you have.

If your metadata model was built for internal search and has not been revisited since, that is the place to start, and it is most of what we do when we help teams build their program.

Andrea put the timing plainly: "AI is going to be able to tag your company language, and we're not there yet. But once we do get there, it needs to learn somehow. So you've got to have your metadata there."