Resemble AI, a Mountain View company that sells software for detecting synthetic media, on October 22, 2026 published a study that scanned 37,768 image and video files from Facebook and Instagram ads and flagged 3,379 of them as AI-generated or manipulated. That is 8.9%, or roughly one file in 11. Of those 3,379 files, two mentioned AI anywhere in the ad copy.

The 21-page document, titled the 2026 Synthetic Advertising Report, draws on ads collected from Meta's public archive in September 2026. It covers 509 companies in 26 industries, from Temu and Wayfair to Lufthansa, Teladoc and Turkish Airlines. According to Resemble AI, 70% of the fakes it labelled were photoreal - built to pass as a real photograph or real video - and 46% showed a person. None of the 971 files checked for provenance metadata carried any.

The timing matters for advertisers. New York's synthetic performer disclosure law has applied since June 9, 2026, and Article 50 of the EU AI Act since August 2, 2026. Both put most of the duty on whoever makes or deploys the ad, not on the platform carrying it. Resemble describes its report as the first large-scale, independent audit of AI-generated media in the ads people see on Facebook and Instagram. It is also the work of a vendor with a direct commercial interest in the answer, and the document makes no secret of that: its final page advertises Resemble's detection models.

In Short

A company that makes AI-detection software checked tens of thousands of Facebook and Instagram ads and says about one in 11 images or videos contained AI-made or AI-altered material. Most of those were made to look like real photos or real footage, including selfie-style "customer" clips, and almost none said in the ad text that AI was involved. So when you scroll past a realistic face in one of these ads, the words around it will almost never tell you whether it was generated, even though New York and the EU now require some ads to say so.

How the scan was built

Resemble AI searched each company's name in Meta's Ad Library during September 2026 and ran every image and video it found through Resemble Detect, its own detection product. A file counted as AI when Detect returned a "Fake" verdict. The haul came to 15,433 ads published by 2,298 Facebook and Instagram pages.

A name search pulls in far more than a brand's own advertising. Resemble therefore sorted every file by who ran it. Searched brand means the company's own page. Other advertiser covers retailers, affiliates, creators and unrelated pages that matched the search. Bait page covers clickbait, search arbitrage and scam pages that use a brand's name as a lure. Before any rates were calculated, 718 files were moved from the brand column to "other advertiser" because the page belonged to a different company or to a standalone creator.

The headline brand figures rest on 28,463 files from 389 brands. The remaining 120 of the 509 companies had no files from their own pages in the sample, a gap the report does not explain (the 120 is a PPC Land calculation). Resemble removed 100 files from a Fanatics merchandise app, 34 duplicate media URLs, nine subsidiary pages and 139 failed scans from all rates.

Every fake in the main collection - 3,216 of them - was then labelled for what it showed and how it was styled. The four style groups are photoreal (built to pass as a real photo or video), footage imitating UGC, the user-generated content a customer might shoot on a phone (labelled UGC-style realism), graphic or UI (designed layouts, text promotions, app screens) and obvious AI or CGI (stylised output no viewer would take as real).

A second collection, also gathered in September 2026, covered nine prediction market, sportsbook and crypto brands. Those ads were scanned for audio as well as visuals. The main collection was not, and that difference runs through the whole report. According to Resemble AI, "Every rate is a floor," because audio alone caught 62 of the 142 fakes in the betting and crypto set.

One thing the report could not measure is where the ads ran. "Meta keeps stopped US commercial ads out of the public Ad Library, so placements for most ads in the study could not be recovered after the scan," according to the methodology page. The study also contains no impression, reach or spend data. A flagged file counts the same whether it reached a thousand people or ten million.

Who used it, and how much

On brands' own pages, 1,971 of 28,463 files were flagged, a rate of 6.9%. Use was spread thin across many advertisers and concentrated in a few. According to the report, 222 of the 389 brands (57%) ran at least one AI file from their own page, and 167 ran none. The distribution looks like this:

  • 167 brands: no AI files
  • 86 brands: under 5%
  • 50 brands: 5% to 10%
  • 44 brands: 10% to 20%
  • 29 brands: 20% to 40%
  • 13 brands: more than 40%

Four large advertisers sat above half. Temu used AI in 64 of 100 sampled files, Wayfair in 56 of 100, Motel 6 in 50 of 80 and Lufthansa in 50 of 94. The 185 brands with two or more AI files had a pooled rate of 12.5%. Resemble's announcement describes this as "one in eight of their ads," though the report's appendix calculates it across files rather than ads.

Rates were far higher on pages brands do not control. Bait pages ran 649 files, of which 463 (71.3%) were flagged. Other advertisers came in at 11.0%, or 924 of 8,428 files.

By industry

All 26 industries ran AI material. Consumer packaged goods led at 13.4%, followed by consumer electronics at 12.6%, fintech and crypto at 11.2%, and prediction markets and betting at 10.6%. Fitness and wellness reached 9.7%, with energy and utilities and automotive both at 9.3%.

At the bottom sat the regulated categories. Insurance came in at 5.0%, banking and financial services at 3.8%, nonprofits at 3.3% and pharmaceuticals and health at 2.8%. Quick service restaurants were lowest of all, at 2.4%. The report links the top of the table to "categories built on product and lifestyle imagery."

A caution applies to the two betting and crypto categories. Their bars in the industry chart include files flagged on audio alone, because only those brands received an audio scan. For prediction markets and betting, most of the 10.6% came from audio. The other 24 industries were measured on visuals only, so the comparison is not like for like.

Company size made little difference

Resemble matched 313 parent companies, covering 363 brands, to employee and marketing headcounts from Apollo's company data. The expectation, according to the report, was that smaller firms or leaner marketing teams would lean harder on AI. That did not show up. AI rates stayed between 6% and 9% at every size:

Employees Companies AI rate Ran any AI Marketing staff Companies AI rate Ran any AI
Under 1,000 37 7.5% 65% Under 100 99 7.5% 56%
1,000 to 10,000 86 8.9% 55% 100 to 299 84 7.3% 58%
10,000 to 100,000 141 6.1% 61% 300 to 999 73 7.1% 64%
100,000 or more 49 6.4% 65% 1,000 or more 57 5.9% 67%

"AI use isn't a resourcing story," the report states. It contrasts Temu, which it describes as a $54bn business with 232 marketers that used AI in 64 of 100 ads, with Amazon, which has nearly 23,000 marketers and used none. Larger marketing teams were somewhat more likely to try AI at least once (67% against 56%) but had a slightly lower overall rate (5.9% against 7.5%).

Photoreal, and often a person

What did the fakes actually show? Of the 3,216 labelled files, 2,259 (70%) were photoreal. Nearly half (46%) depicted a person. Only about one in nine was stylised enough that a viewer would read it as AI.

The ten most common subjects, as a share of all labelled fakes:

  • Person, realistic model: 26.1%
  • Product shot: 23.6%
  • Scene or place: 14.6%
  • Text or graphic promotion: 10.1%
  • Illustration, CGI or key art: 7.2%
  • Person, UGC or selfie style: 4.6%
  • App, UI or device: 4.3%
  • Food: 2.4%
  • Animal: 1.9%
  • Celebrity or public figure: 1.6%

Style varied by who ran the ad. On brand pages, 72% of the 1,829 fakes were photoreal, 2% UGC-style, 16% graphic and 9% obviously AI. Other advertisers' fakes were 66% photoreal and 7% UGC-style. Bait pages stood apart: 71% photoreal and 11% UGC-style, used, according to the report, "to make clickbait look like a real person's story." The report sums it up in five words: "Brands fake polish. Bait pages fake people."

The generated customer

The section most likely to concern compliance teams covers testimonial-style creative. Resemble published four brand examples that Detect flagged with high confidence:

  • Teladoc: a telehealth patient in a waiting room, UGC style, flagged with 99.9% certainty
  • MoneyLion: a phone chat video in UGC style, 99.8%
  • Dave: a driver's selfie testimonial for a cash advance, 99%
  • Tarte Cosmetics: a selfie video for a tinted serum, 99.9%

These are detector scores, not findings of wrongdoing. The report includes no response from any of the four companies, and it does not say whether any of the clips carried an on-screen label.

Real footage, synthetic voices

In the main collection, brand images were flagged at more than twice the rate of brand videos: 8.3% (1,489 of 17,857 images) against 3.5% (340 of 9,706 videos). Resemble cautions that the video figure is the weaker of the two. "Without an audio scan, a real video with a generated voiceover reads as authentic," according to the report.

The betting and crypto collection shows how much that matters. Of its 142 fakes, 75 were caught on visuals only, five on both, and 62 on audio only. In crypto and brokerage, 37 of 88 fakes were audio-only. For sportsbooks it was 20 of 38, and for prediction markets five of 16.

Of the 67 files flagged on audio, 57 were synthetic voices and ten were music or sound effects. Fifty-four of the 57 voices played over video that Detect rated as real. The report traced 55 voices to ElevenLabs, the voice-generation company. Every one of Robinhood UK's 21 flagged ads used an ElevenLabs voice. Fanatics and eToro each had 11 synthetic voice files, Novig five, OKX four voices plus six music or effects files, FanDuel three voices and four music or effects files, and DraftKings two voices.

Resemble published four examples with their audio scores. A Robinhood UK crypto recurring-buy ad ("Get more crypto for your cash and safely set up recurring buys.") scored 99.9%. A Novig creator-style ad ("Drop $10 in Novig, you'll get an extra $25 in trade credits.") scored 86.9%. An eToro app launch video scored 99.2%, and a Fanatics sportsbook promotion 68.7%.

The report's own reading of this audio is more measured than the announcement that accompanied it. Resemble's release states that "the fakes you can't see are the ones you hear." The report itself says of the voiceovers: "By eye and ear, these were generally not deceptive or intended to be deceptive, and were used primarily as an editing technique." It adds that "a synthetic voice presented as a real person has a lot more potential to come across as deceptive or misleading than a voiceover describing a scene."

That distinction mirrors industry guidance. The IAB's AI Transparency and Disclosure Framework, version two, issued on August 18, 2026, requires disclosure for synthetic voices of living or deceased people and photorealistic AI influencers, but exempts generic synthetic voiceovers and background music. Under that framework, generic voiceovers of the kind the report describes as an editing technique would not require a label.

Disclosure: two files out of 3,379

Resemble searched the copy of every ad for any mention of AI generation. Two of the 3,379 flagged files had one, a rate of 0.06%. Only one brand disclosed on its own ad: Turkish Airlines, in an International Cat Day video for its Touristanbul programme, which Detect flagged with 99.9% confidence. The ad copy ends with a footnote saying the visuals "were created using artificial intelligence technology." The note sits below the caption, after the hashtags in most feeds, according to the report. Resemble does not identify the second disclosing file.

The method has a blind spot that the report acknowledges. A copy search cannot see a label burned into the video itself. Resemble checked on-screen labels for one advertiser only, Kalshi, and found that seven of eight AI-made presenters in Kalshi's ads showed an on-screen AI notice. The 0.06% figure is therefore a measure of written disclosure in ad copy. How many of the other flagged files carried a visual label is not known.

Provenance metadata told a starker story. None of the 971 files checked carried C2PA content credentials, the industry standard for recording how a file was made, including every confirmed fake. The report does not say whether Meta preserves such metadata when it processes uploaded creative, which would determine whether the absence reflects what advertisers uploaded or what the platform served. Nor does it say whether Meta attached any AI label of its own to the flagged ads.

Meta has had an AI labelling policy for content on its platforms since 2024. In April that year the company said it would apply a "Made with AI" label to AI-generated or significantly altered video, images and audio from May 2024, whether detected by Meta or disclosed by the creator. In July 2026, Meta signed the EU's Code of Practice on Transparency of AI-Generated Content, the voluntary route to compliance with Article 50.

Google has taken a more visible route for ads. On July 9, 2026, it added a "How this ad was made" panel to My Ad Center across Search, YouTube and Discover ads. Ads made with Google's own generative tools are disclosed automatically; ads made with outside tools rely on a manual advertiser control. YouTube in May 2026 moved its AI labels to more prominent positions and began applying them automatically to content carrying C2PA metadata.

Why might advertisers stay quiet? The IAB framework itself cites NYU Stern research finding that adding an AI disclosure reduced click-through rate by 31.5%. The Resemble report does not address that cost.

Bait pages

Searching a brand's name also surfaced pages the brand does not control. These pages ran 649 files, of which 71.3% were flagged - ten times the rate on brand pages. They made up 2% of all files but 14% of all fakes.

According to the report, the bait-page fakes broke down as follows:

  • Clickbait advertorial: 349 files (75.4%)
  • Affiliate or search arbitrage: 64 files (13.8%)
  • Scam or deceptive: 40 files (8.6%)
  • Standard ad: 10 files (2.2%)

Their subjects differed from brand creative. On brand pages the leading subjects were realistic models (27.4%), product shots (25.3%) and scenes (13.0%). On bait pages, scenes led at 30.5%, followed by realistic models at 22.0% and UGC or selfie-style people at 11.2%.

The pattern fits a long record of brand impersonation on Meta's platforms. Meta said in December 2025 that it had removed more than 134 million scam ads during 2025. In February 2026 it sued scam advertisers in Brazil, China and Vietnam over celeb-bait and cloaking. And on September 16, 2026, a Frankfurt court ordered Meta to stop distributing fake Finanzfluss ads, one of which used an AI-generated deepfake video of co-founder Thomas Kehl, with fines of up to 250,000 euros per future violation. That ruling is not final.

Betting and crypto

The second collection covered nine brands across three segments. Of 900 brand-page files, 142 (15.8%) were flagged, more than double the 6.9% brand-page rate for the whole study.

Segment Flagged files Rate
Crypto and brokerage 88 of 300 29.3%
Prediction markets 16 of 173 9.2%
Sportsbooks 38 of 427 8.9%

By brand, OKX had the highest rate at 48.5% (48 of 99 files), followed by ProphetX at 23.5% (4 of 17), Robinhood UK at 20.4% (21 of 103), eToro at 19.4% (19 of 98), FanDuel at 16.7% (17 of 102), Novig at 13.6% (8 of 59), Fanatics at 8.4% (19 of 227), Polymarket at 4.1% (4 of 97) and DraftKings at 2.0% (2 of 98).

Resemble's release calls crypto and brokerage "the most synthetic in the entire study at 29.3%." That rate, however, includes an audio scan the other industries did not receive. Counting only the 51 crypto and brokerage files flagged on visuals gives 17.0%, a PPC Land calculation. That would still exceed the 13.4% recorded for consumer packaged goods.

Several flagged ads featured real people. According to the report, FanDuel ads with Rob Gronkowski and the Bussin' With The Boys podcast were flagged, as were Kalshi images of Marshawn Lynch. The report does not specify which elements of those files Detect attributed to generation.

Kalshi and Elliot Choy

Kalshi, which was not among the nine brands in the betting collection, received its own scan. On September 24, 2026, Resemble checked every active US Kalshi ad. It flagged 52 of 511 files (10%), appearing in 50 of 397 ads (13%). Images were flagged at 5% and videos at 14%. Forty videos carried a synthetic voice, 23 traced to ElevenLabs and 16 to Cartesia. None of the 511 files carried C2PA credentials.

The scan came a day after a public complaint. On September 23, YouTuber Elliot Choy saw a Kalshi ad built from his 2025 apartment-move video. The script, the room and the framing matched his. The man on screen did not: AI had replaced Choy, who is Asian American, with a white man who does not exist, according to the report. "i guess kalshi saw this and thought they should steal my video, turn me into a white dude, and run it as an ad," Choy wrote in a post shown in the report. Detect returned an AI-generated verdict on the ad, with a score of 0.9011.

Kalshi told NPR the ad was old, that it no longer runs AI-generated ads, and that it was reviewing its relationship with the agency that made it, according to the report. The report does not date that statement. Its own scan of Kalshi's active US ads on September 24 found eight AI-made presenters, seven of them with on-screen AI notices.

Choy was not the first. Creator Pushpek Sidhu said a betting ad in June copied one of his videos almost line for line and also swapped him for a white man; the report does not name the advertiser. "In both cases the person whose work was used found out by recognizing it," the report states.

Creator footage has surfaced in court this month too, though not through AI. On October 9, 2026, content creator Amber Jordan Holston sued True Nutra over 12 Meta ads that she says reused her real videos without permission. In Denmark, a bill tabled on October 8 would require publishers to prove consent before releasing realistic digital imitations of a person from January 1, 2027, if passed.

The betting focus also tracks a fast-growing ad category. Google began accepting US prediction market ads from CFTC-regulated operators on January 21, 2026.

The rules the report maps

Resemble's report includes a table of jurisdictions that require AI disclosure in ads. "Most rules cover generated people, not AI music, voices or scenery, and most put the duty on the advertiser rather than the platform," according to the report.

Region Rule Requirement, per the report Status, per the report
South Korea KFTC endorsement guidelines Ads using an AI-generated virtual person must label it as virtual on screen In force, June 1, 2026
New York General Business Law section 396-b Conspicuous disclosure of a synthetic performer the advertiser knows about; audio-only ads and platforms exempt In force, June 9, 2026
European Union AI Act, Article 50 Machine-readable marking of AI output, visible labels on deepfakes August 2, 2026 (marked "verify")
China AI content labelling measures Visible and machine-readable labels on AI text, images, audio and video In force, September 1, 2025
India IT Rules amendment Labelling and takedown of synthetically generated content In force, February 20, 2026
California SB 942, AI Transparency Act Generative AI providers must offer detection tools and embed provenance data Marked "verify date"
Tennessee ELVIS Act Consent required to use a person's voice or likeness in an AI replica In force
US federal FTC Act, NO FAKES Act No AI ad disclosure rule; FTC deception rules apply; NO FAKES pending Partial
United Kingdom ASA advertising code No AI-specific disclosure law; misleading ads rules apply Marked "verify"

Three entries in the published table still carry "verify" placeholders. For the EU, PPC Land's earlier reporting confirms the date: Article 50 has applied since August 2, 2026, with generative systems already on the market given until December 2, 2026 to meet the machine-readable marking rules. Fines reach 15 million euros or 3% of worldwide annual turnover, whichever is higher. The same coverage details New York's rule, Chapter 617 of the Laws of 2025: fines of $1,000 for a first violation and $5,000 for subsequent ones, with media carriers shielded. In Europe, an IAB Austria guide published in September concluded that the agency using the AI tool is normally the responsible deployer, not the client.

Zohaib Ahmed, chief executive of Resemble AI, framed the findings around shared responsibility. "Generative AI is a powerful creative tool for advertisers, and more than half of the brands we scanned ran at least one AI-generated ad, which is to be expected. However, most of those ads were built to look like real photos and video, and almost none in that style disclosed that AI was used," he said in the announcement. "Disclosure should be a shared responsibility between the advertiser and the platform hosting the content. Until there is a clear federal rule for AI in advertising, brands that want to keep their customers' trust should watermark and clearly label AI-generated content."

The report proposes a three-layer structure. Model providers would embed imperceptible watermarks and provenance data that survive editing and compression. Platforms would check uploads for those marks and for synthetic media, and add an AI label at serving time. Advertisers would select an AI disclosure at upload, "the way political advertisers already declare digitally altered media." The last of these resembles the manual control Google introduced in July. The report argues that "waiting for advertisers to volunteer has not worked."

It also separates disclosure from consent. "A label would have told viewers the Kalshi ad was generated. It would not have asked Choy before his video was used," according to the report.

Resemble's release closes with recommendations for brands and agencies: audit creative including work produced by agencies and partners, treat a synthetic person differently from a synthetic backdrop, disclose by default, and watermark brand content while monitoring for lookalike ads. Several of those steps involve the kind of detection and watermarking products Resemble sells.

Meta's position

The report landed a day after a New Mexico judge ordered Meta to pay $25bn on October 21, 2026, following a jury verdict that found the company's public statements misleading. That case concerned statements about privacy and content moderation on Facebook, including the Cambridge Analytica episode. It did not concern ad creative or AI-generated media, and the Resemble report does not mention it.

Meta has pushed generative tools into its own ad products for more than a year. At Cannes Lions in June 2025 the company presented image-to-video generation, virtual try-on and brand-aware text and image tools for advertisers. The Resemble report does not separate creative made with Meta's own tools from creative made elsewhere. It contains no comment from Meta.

What the numbers do not establish

The study is large by the standards of published ad research, and its appendix shows how each headline figure was calculated. Several limits still shape how far the numbers can be read.

The detector is the vendor's own product. Every verdict comes from Resemble Detect. The report publishes no false-positive or false-negative rate, no external validation, and no score threshold for a "Fake" verdict. Its examples range from 99.9% down to a 68.7% audio score, all counted as fakes.

"Manipulated" is not defined further. A file counts as AI when Detect says so. The report does not separate fully generated images from photographs edited with AI tools. The IAB framework, for one, treats them differently: it exempts routine post-production edits.

The headline brand rate mixes two methods. The 28,463 brand-page files include the 900 betting and crypto files that received an audio scan, and the 1,971 brand fakes include those 142. Excluding them, the visual-only brand rate is 1,829 of 27,563 files, or 6.6%, a PPC Land calculation.

Some totals do not reconcile. The executive summary heading refers to "37,786 files," while every other page uses 37,768. The chart of rates by advertiser type sums to 37,540 files and 3,358 fakes, against 3,379 fakes overall. The report does not explain the gaps, which are small relative to the totals.

The sample is a snapshot of active ads. Many per-brand samples cited in the report cluster around 100 files, and DraftKings' figure is described as its "98 most recent ads." The report does not state a sampling rule, and with no impression data, a heavily served file and a barely served one carry the same weight.

No company responses. Apart from Kalshi's statement to NPR, the report contains no comment from the brands it names or from Meta.

Resemble says a follow-up focused on UGC and influencer content is next. The report gives no date for it.

Timeline

Summary

Who: Resemble AI, a Mountain View synthetic media security company led by chief executive Zohaib Ahmed, which scanned ads from 509 companies including Temu, Wayfair, Motel 6, Lufthansa, Teladoc, MoneyLion, Dave, Tarte Cosmetics, Turkish Airlines, OKX, Robinhood UK, FanDuel and Kalshi.

What: The 2026 Synthetic Advertising Report flagged 3,379 of 37,768 Facebook and Instagram ad files (8.9%) as AI-generated or manipulated, found 70% of labelled fakes were photoreal, and counted two files out of 3,379 that disclosed AI in ad copy. No file of 971 checked carried C2PA credentials.

When: Ads were collected in September 2026, with a Kalshi scan on September 24, 2026. Resemble AI published the report on October 22, 2026.

Where: Meta's Facebook and Instagram, via Meta's public Ad Library, with a regulatory map covering New York, the EU, South Korea, China, India, California, Tennessee, US federal law and the United Kingdom.

Why: Disclosure rules for synthetic people in ads now apply in New York and the EU and mostly place the duty on advertisers, yet little public data had measured how much synthetic media runs in live campaigns. The report's figures come from the vendor's own detector, with no published error rates, and measure written disclosure in ad copy rather than on-screen labels.