The Maze: Resemble AI, a company that sells software for detecting AI-made media, found written AI disclosures alongside just two of 3,379 ad files its software flagged on Facebook and Instagram. Its September sample raises a useful question for advertisers: what does a disclosure check actually see?
The Details:
Files are the unit. Resemble published its study on October 22. It searched 509 company names in Meta's public Ad Library, a searchable advertising archive. The collection contained 37,768 image and video files from 15,433 ads across 2,298 pages. Its own software classified 8.9% of those files as generated or manipulated with AI. A file is one piece of an ad, so that percentage does not mean 8.9% of ads shown to consumers contained AI. Brand-name searches also retrieved other advertisers' pages. The sample therefore extends beyond creative directly controlled by the named companies.
A caption search misses the picture. Resemble searched the words accompanying ads for mentions of AI generation. Only two flagged files had such a mention. That is a striking result for written ad copy. It is a narrower result than a count of everything viewers were told. In a separate check, seven of eight AI-made presenters in Kalshi ads had an on-screen AI notice. Those words appear inside the image or video, where the copy search cannot see them. The report did not systematically count these notices across the sample or establish which ads carried labels applied by Meta itself.
Detection is an estimate with limits. Every classification came from the vendor's own detector. The report supplies no published false-positive or false-negative rates for this audit, meaning readers cannot quantify how often it might flag real media or miss generated material. It also provides no impression or spending data. Each file receives equal weight regardless of how often people saw it. This is an observational snapshot, without an experiment testing how disclosure changes shopper behaviour. The study cannot tell an advertiser whether a label increased trust, reduced clicks or changed sales.
MarketMaze Take: The useful lesson is about how advertisers check their creative. A detector's verdict and a viewer's disclosure are separate pieces of evidence. Treating them as interchangeable can make a narrow audit sound conclusive while leaving the actual viewing experience unexamined.
Ask what the AI changed. The Interactive Advertising Bureau, a digital-advertising trade association, distinguishes potentially misleading synthetic people from routine production edits and generic synthetic voices. Its guidance is not law, but the distinction helps explain why an AI flag alone cannot establish a disclosure failure. Agencies need to understand what a supplied asset represents before deciding how to describe it to customers.
Check the delivered notice. Advertisers in any market can separate the question of how an asset was made from whether its final placement tells viewers. The Kalshi example shows why inspecting the video matters alongside reading the caption. This sample does not establish a worldwide disclosure rate or prove that particular advertisers broke a rule. It does show why counting caption mentions alone cannot settle either question.
Sources: Resemble AI, PPC Land, IAB
