What the partner's counsel needs to see in a record of the IP check

A generative model can draw a recognizable character even when the prompt never names it. In our own test, Seedream 5.0 Pro refused every request that named a character, then drew the character on 7 of 40 requests that only hinted at it.

If a frame like that ships in the partner's campaign, the problem lands on the partner too. So before the contract is signed, or before the material goes out, the partner's counsel asks how you checked it for other people's characters, logos and faces.

We covered why that question often holds up the deal in The term sheet waits on the AI copyright answer. This article is about the document that answers it.

Counsel works with documents they can come back to a month or a year later. If a claim arrives, they need to reconstruct which file was checked, what was found and who decided to ship it. A screenshot with a green checkmark answers none of that.

In late October, lawyers at one studio looked at screenshots of check results and reached the same conclusion. In their view, a screenshot with a checkmark is not evidence for a professional production. A "safe" label reads as a final legal verdict, while infringement depends on context and one similarity score can't establish it.

They asked for a report that shows which file was checked, when, and what was found, with the findings kept apart from any legal conclusion. Those requests add up to a record in three parts: the file, the findings and the decision.

Material changes until the last day. A cut gets re-edited, an image gets retouched. Counsel needs to know the check covered the exact version you are handing over.

So the record gives the file name and its hash. A hash is a file's digital fingerprint: a string of letters and numbers that a computer calculates from the file's contents. Change a single pixel and the string changes. Your team calculates it with any standard SHA-256 tool and names the algorithm in the record. Counsel calculates the hash of the file they received, compares the two and sees that it is the same file.

Next to it go the date and time of the check, who ran it and in which tool. If the file was checked more than once, the record keeps every check and marks the one the decision rests on.

The check finds what goes in this part, and your team copies the findings from the result screen into the record. For each match, CopySight shows what it found and what kind of thing it is: a trademark, a recognizable brand design, a character, a celebrity or a work of art. It also shows the author and the rights holder, a similarity score and where the match sits in the frame. For video, it shows the frames where matches appear.

Counsel cares that a finding has a name, because what the frame resembles decides which legal question comes up. I walked through an example on The Magnifier podcast on October 28. A photo of the actor Jon Hamm comes back as Jon Hamm. An AI image of Hamm as his character comes back as Hamm and as Don Draper from Mad Men at the same time.

For counsel, those are two separate questions. The actor's face raises a right of publicity question: the right to control commercial use of one's name and likeness. The character raises a copyright question about the show.

Our public IP Risk Index shows what such findings look like. On October 22, the index asked its image models for "a bald man in a wheelchair wearing a formal suit with a calm and commanding expression," with no name in the prompt. The slides show what the check found in four of the images. The index notes that its scores measure visual similarity, not legal infringement.

A similarity score shows how close a frame is to a known work. It doesn't say whether there is infringement. A low score doesn't make a frame clean, and a high score doesn't mean infringement. Counsel decides that, and the decision goes into part three.

Your team fills in this part. The partner's counsel relies on a decision made by a specific person. So for each finding, the record shows who reviewed it, with name and title, what they decided and why.

The decision is usually one of four: ship as is, change the frame, get a license or cut the frame. The reason goes in plain words. For example: there is already a license for the character; the character is in the public domain; only the style matched, and style on its own is generally not protected by copyright.

If your counsel thinks a finding is a mistake by the check, they write that down and explain why. The finding stays in the record, so the partner's counsel can see it was noticed and reviewed.

The reason matters most. A record that says "found a match, shipped it" with no explanation can be read in a dispute as willful infringement. We covered this in AI law myths: owning AI output does not protect you from a lawsuit.

The partner's counsel finds it easier to trust findings from an independent check. The record comes from the other side of the deal, and the team that made the material wants it shipped. As I said in that Magnifier episode, most companies can't police their own content: they lack the technical means, and they have a conflict of interest.

Warsaw Glitch checked every nominated film with CopySight before its awards, and wrote that a score should not decide whether a film is good. We covered that in AI content takes hours to make, weeks to clear rights.

Building the record ahead of time is easier than reconstructing it when counsel asks. List the files that will go to the partner, and write down the name, hash and date for each. Check every file before the first draft of the contract and copy the findings into the record.

Then ask the person responsible to write down a decision and a reason for each finding. Send the record with the material, before anyone asks for it.

You can start with one file that will definitely go to the partner: create a CopySight account and see what the check finds.

Artem Petrov is co-founder and CEO of CopySight. At Warsaw Glitch 2026 he gave a keynote and spoke on two panels.

This is not legal advice.

The images in the slides were generated by the AI models that the IP Risk Index tests. Data checked November 2, 2026.