Human in the Loop, Done Properly

The demo was going well until the salesperson said the quiet part out loud.

"And here's the human-in-the-loop step," he said, and clicked a button marked Approve All. 312 listings went from Pending to Filed in one animation. He looked pleased. "So you stay in control."

Nadia, who runs brand for a mid-market homeware company and has been doing this for 11 years, asked how long the review had taken.

"That's the beauty of it," he said. "Seconds."

She didn't buy the product. She did keep the screenshot, because it's the cleanest example she's ever seen of a control that exists in the interface and nowhere else.

The phrase has been hollowed out

Human in the loop is now a compliance sticker. It gets applied to any workflow with a button in it, whether or not the person pressing the button could possibly have known what they were approving.

A loop that a person cannot meaningfully affect isn't oversight. It's a signature collection device, and in brand protection it has a specific cost, because the thing being signed is a sworn statement. A DMCA notice carries a penalty of perjury under 17 U.S.C. 512(c)(3), and 512(f) makes knowing misrepresentation actionable in damages. Approve All on 312 listings is 312 sworn statements that nobody read.

So it's worth being precise about what the loop is for, and the answer isn't "to make the customer feel safe". It's to put a human judgement in front of the one decision no model can make.

The two decisions software genuinely can't make

Everything in this workflow is automatable except two things, and they're the two that have consequences.

Is this listing a copy? A model can see that the photograph is 94% similar to yours, that the title repeats your product name, and that the price sits 68% below yours. It cannot know whether the seller is the distributor you authorised in March, or your own second shop under a trading name, or a legitimate reseller of genuine stock. Those facts live in your head and your contracts, not on the page. A high similarity score is evidence of similarity. It is not evidence of infringement.

Should a notice be filed? Even on a confirmed copy, filing is a judgement about whether this specific seller, in this jurisdiction, with this history, is worth the notice and the fortnight that follows a counter-notice. Sometimes the answer is to wait and build a repeat infringer case instead. A model has no view on that worth having.

Those two go to the human. The rest is machinery.

What the machine should do, and do hard

Drawing the line properly means loading the automated side up, not thinning it out. A loop with a human in it only works if the human arrives at a decision that takes 15 seconds rather than 15 minutes.

So the software should: search on a schedule, score each candidate on photo similarity, title overlap and price distance, dismiss the obvious non-matches, drop anything from the allowlist, rank what's left, and write a one-line reason for each. Then group by seller, so 20 listings from one shop arrive as one shop with 20 items rather than 20 separate decisions. Then capture the evidence at the moment of finding, with seller name, seller ID, listing detail, a screenshot and a timestamp. Then draft the notice, in the right form for that platform, with the evidence attached.

By the time a person looks, the question should be a single comparison on a single screen: here's your product, here's the listing, here's why we think it matters, copy or not a copy or one of yours.

That's a reviewable decision. Approve All is not.

Four properties of a loop that isn't theatre

The irreversible act is blocked in the code, not in the policy. There's a meaningful difference between a product that chooses not to file automatically and a product that cannot. The first is a setting someone can change under commercial pressure. The second is an architectural commitment. Ask any vendor which one they've built, and ask them to show you.

No bulk action on the irreversible step. Bulk-confirming every listing from one seller is reasonable, because the reviewer has made one judgement about one shop. Bulk-confirming 312 listings across 90 sellers is not, because no judgement occurred. The distinction is whether the batch shares the fact being judged.

Dismissals are visible and reversible. The riskiest automation in this workflow isn't the false accusation, it's the quiet dismissal. A model that drops a real copy into a hidden bucket has cost you a sale and you'll never know. Dismissed items belong in their own tab, reviewable, with a one-click path back to confirmed.

The overturn rate gets measured and published. How often does a human disagree with the model's dismissal? That single number tells you whether the triage is trustworthy, and it's the number almost no vendor will give you. If they publish it beside the base rate it should be compared against, take them more seriously.

On that last point, I'll hold our own product to it. Brand Protection Tools, which Verity AI runs, says on its own how-it-works page that it will publish the dismissal overturn rate once there are enough scans for the figure to mean something, and gives no number today. That's the honest position and it's also an uncomfortable one, because "no number yet" reads weaker than a made-up one. A number invented to fill the gap would be worth less than the gap.

The rest of the design matches what's above. Confirming and filing are blocked for the AI in the code. Matches arrive grouped by seller with a reason attached. Dismissed listings sit in a reviewable tab. And it's software, not legal advice, which is the sentence to look for on anything in this category.

Where this connects to actual AI governance

Two bits of law are worth knowing, and neither of them says what people assume.

The EU AI Act sets out human oversight obligations in Article 14, and they apply to high-risk systems as classified by the Act. A tool that finds marketplace listings and drafts a notice isn't one of them. So the obligation doesn't bind here, and any vendor telling you their brand protection tool is AI Act compliant in the high-risk sense is selling you a misunderstanding.

The design principle is still the right one to borrow. Article 14's underlying idea is that a person must be able to understand the system's output, interpret it correctly, and decide not to use it. Read that as a product spec rather than a legal test and it describes exactly the loop that works: a reason the reviewer can read, a comparison they can judge, and a real option to say no.

The second bit has sharper teeth. If the seller on the other side is a natural person, which on a handmade-goods marketplace they very often are, then a fully automated decision producing legal effects about them starts to look like the territory GDPR Article 22 governs. I'm not going to pretend that's a settled question for takedown workflows, because it isn't. I'd note that every reason to keep the human in the loop points the same way, and that a vendor who has never thought about it is telling you something.

The test to run in the demo

Nadia's version is 3 questions, asked in this order, and she asks them before pricing.

Show me what happens if I click file on a listing the system has not confirmed. If it files, the loop is decoration.

Show me the dismissed tab. If there isn't one, the model is making decisions you can't audit.

Tell me your dismissal overturn rate, or tell me you don't have one yet. Either answer is acceptable. A confident number with no methodology behind it is the one to walk away from.

She bought a different product in the end, after asking the same 3 questions of 4 vendors. 2 of them couldn't answer the second one.

For the governance framing of all of this, including the question about who signs the notice, read counterfeits as a governance issue. For the version that fits a brand with no budget line, the small-brand playbook covers the same control set by hand. And once the copies are down, the other half of the job is making sure the machine knows which store is actually yours: entity hygiene.

The wider argument is in When the AI shopping assistant recommends the fake.