When AI output looks too much like existing work

AI tools can repeat things they've seen. Nothing will warn you, so the check falls to you. Here's how to catch it and what to do next.

Why it happens

• Models can replicate what they were trained on. Wikipedia says that in some cases models can copy items from their training set into what they generate.¹

• Images can come out near-identical. Wikipedia says some image models can give results very close to pictures in their training set.¹

• Text and code can too. A class action against GitHub and others said a coding tool was "able to generate source code which matched its training data verbatim". That's an allegation, not a finding.¹

• Fixes have limits. As of August 2023, the page says, tools had tried to cut the problem. Still, "researchers have still been able to prompt leakage of copyrighted material".¹

• Others may get your output. OpenAI's terms say outputs aren't always unique. A near match may simply be a common result.²

Warning signs

• A sentence or paragraph that reads like a published source. Sudden polish or an odd, specific turn of phrase is a clue.

• A picture with a signature, a watermark or a logo. Don't crop it out or reuse it. Make a fresh image instead.

• A recognisable character, mascot or product. If you'd know it from a film, game or shop shelf, so will others.

• Code with comments, names or licence lines you didn't ask for. They may have come from a project with its own licence.

• A "style of" request that gave a very close result. An artist's overall style generally isn't protected by copyright, Wikipedia says. A near copy of one picture is still a different matter.¹

What to do

• Search for the source. Put a distinctive sentence in quotation marks in a search engine. For pictures, use a reverse image search.

• Check more than once. OpenAI's terms tell you to evaluate output before sharing it. Anthropic's say "You should not rely on any Outputs or Actions without independently confirming their accuracy." Apply that to originality too.²³

• If you find a match, don't publish it. Rewrite from your own notes, commission new work or choose a licensed image.

• If it's already published, take it down and swap it. Then keep the old version and your notes in a folder.

• If someone complains, don't argue by email. Answer politely, pause the use, and ask a lawyer. Your country's copyright office or the copyright council can explain the process.

Why it matters to a claim

Wikipedia says that under US law a plaintiff must also show "the AI had access to their work".¹ That's a US point. Rules differ by country, and the details aren't something to guess at.

Keep a record of your prompt, the tool and the date. Related reading: the ownership page and using AI-made images in business materials. This is general information, not legal advice.