A free spot-check on AI-generated supplier photos — and the three things it misses

There's now a free Google site where you can upload an image, video, or audio file and check whether it carries Google's AI watermark — and the useful part isn't the tool, it's the rule you build around it.

What the tool actually does

Google has launched a SynthID verification site that lets you upload media and check whether it carries SynthID, and it covers images, video, and audio, according to TechCrunch's October 2026 write-up. As of October 2026, that piece reports it as free.

SynthID is Google's own watermarking technology, embedded into output from Google's generative AI tools; for images it's embedded directly into the pixels rather than sitting in the file's metadata, per Google DeepMind.

We haven't independently tested the site. Everything in this section is what's reported, not what we've verified in our own hands — so confirm the current scope before you build a process on top of it. Follow the link from the TechCrunch piece rather than guessing at a URL; our view, not a sourced claim, is that lookalike "AI detector" domains are a crowded and unreliable genre.

Mechanically it's an upload-and-read-the-result tool. Which means the only reason it won't happen in your business is that nobody has decided when it happens.

The three things it won't catch

This is the part to internalize before you lean on it, because the failure mode here is expensive confidence.

1. It only knows Google's watermark. SynthID is Google's mark, embedded by Google's tools (Google DeepMind). Our reading of that, stated plainly as inference: a product shot generated in a non-Google tool has no SynthID in it to find. "No watermark detected" is not "made by a human" — it's "not made by the one vendor this checker knows about."

2. A clean result is weak evidence either way. Google says SynthID is designed to stay detectable through common modifications like cropping, compression, and added filters (Google DeepMind) — but that's Google's own claim about Google's own system, and we have no independent test of where it breaks. Files in the real world get screenshotted, re-rendered, pasted into decks, and run through third-party resizers. Our position: write your process so a negative result changes nothing. It is absence of evidence, not evidence of absence.

3. A hit isn't automatically fraud. Plenty of legitimate suppliers use AI for background cleanup, lifestyle mockups, or catalog filler while the actual product is real and fine. The watermark tells you how the image was made, not whether the goods exist. The right response to a positive result is a question, not a cancellation.

So the honest summary — our judgment, not something either source claims: this is a cheap positive detector and a near-worthless negative one. Use it to catch obvious cases, and use contract language and sampling for everything else.

Where to put the check so it actually gets done

Three moments, each one a decision point where money is about to move:

• New supplier, before the first PO. Pull three to five photos from their catalog — the hero shot, a detail shot, and anything that looks suspiciously well-lit — and run them. Our estimate: a few minutes.
• Creative deliverable, before you release payment. Photography, video, voiceover. Run a sample of the files on handoff, not three months later when it's in a brochure.
• Anything going into a regulated or claim-bearing context. Product imagery showing a feature a customer will expect, before-and-after photos, testimonial-style video or audio.

That's it. Don't check everything — you'll stop within a week. Check at the points where being wrong costs real money.

A worked example — hypothetical, every number below is invented for illustration

Say you run a small outdoor gear shop. A new supplier shows up on a sourcing marketplace with a clean catalog and a good price. The opening order is 400 units at $18 = $7,200, plus roughly $600 of inbound freight. Again: these are made-up figures to show the shape of the decision, not data from a real client.

If the photos are generative mockups and the real product has different hardware, you're looking at a return fight with an overseas vendor, dead inventory you've already paid freight on, and weeks of a shelf slot earning nothing. Call it the $7,800 you can see, plus the margin on whatever should have been in that slot.

The check is five photos and a few minutes. If one comes back carrying a SynthID watermark, you haven't proven anything damning — you've earned the right to send one email: "Can you send three unretouched phone photos of the actual unit on a bench, with a tape measure in frame?" A real supplier does it in a day. The response to that email is the real test; the watermark check is just what prompted you to send it.

Same logic on the other side of the ledger: if you've paid a freelancer for lifestyle photos under a contract that says no generative imagery, a positive result on a few files is leverage for a redo before you release the final invoice — not after the catalog print run.

What we'd actually build

For most small businesses, nothing. Honestly — a sticky note on the purchasing screen that says "check the photos" will outperform any software you buy for this, because the bottleneck is attention, not capability.

It's worth automating when volume justifies it: you're onboarding new suppliers most weeks, or you're running a marketplace, agency, or print operation where creative lands in bulk. Then what we'd build is small and boring:

• A shared intake folder or form where new supplier and freelance media gets dropped — nothing enters the approved asset library without passing through it.
• A checklist item in the vendor onboarding record: verification run, date, result, who looked at it. The audit trail is worth more than the result, because it's what you point at when a customer complains.
• An escalation path with exactly one branch: positive result → the "send me unretouched photos" email template → sign-off or reject.

Our own estimate of the effort is a couple of hours of setup in tools you already pay for — an estimate, not a quote. Anyone selling you a platform for this is selling you a form.

The clause that does more work than the detector

The durable fix is contractual. Add a line to supplier terms and freelance SOWs along these lines: product photography must depict the actual goods supplied; any use of generative AI in deliverables must be disclosed in writing at handoff. We're not lawyers — run the actual wording past yours.

A clause like that works against every AI tool on the market, not just the one with a watermark, and it doesn't depend on where detection technology lands over the next two years. The verification site is a spot check. The clause is the policy. Run both, and a positive detection stops being a crisis and becomes a routine disclosure conversation you already have terms for.

One caveat on everything above: the product facts here reflect what was reported as of October 2026. Verification tooling moves fast — re-check the scope before you build a process that assumes more coverage than it has.

Not sure where AI is already in your supply chain

If you're finding out about AI-generated media the same week you're wondering which of your own workflows should be automated, start with the map rather than the tools. Our AI readiness audit maps where AI is worth using in your business — and where it's a liability you should be writing rules about instead.