Does an AI label cost you clicks? The 31.5% study and the no-effect study tested different things

KEY TAKEAWAYS

1. The 31.5% figure comes from static display ads and is measured against the human-made ad. The no-effect result comes from a video survey with no click data.
2. Platforms, viewers and lawmakers draw the same line. An AI-made person is what gets a visible label.
3. A production saving is fixed. A click penalty scales with spend. The break-even is a spend level.
4. Use AI presenters to test concepts cheaply. Re-shoot winners with a person before scaling.

AI-made video is now cheap enough to use for every ad concept. The open question is the label. One study says disclosing AI cuts clicks by nearly a third. Another says labels do no harm. Both get quoted as settled. They measured different things, and neither tested the label your viewers actually see.

What changed in the last few months?

The rules converged on one thing: an AI-made person.

• 1 June 2026. Meta updated its ad labelling. AI info labels now sit in About this ad, inside the three-dot menu on every ad. Meta also said it would begin detecting ads made with third-party AI tools and label them there.
• 9 June 2026. New York's synthetic performer law took effect. Ads with an AI-generated human performer must carry a conspicuous disclosure. Penalties are $1,000 for a first violation and $5,000 after that.
• 2 August 2026. The transparency duties in Article 50 of the EU AI Act began to apply.
• 31 August 2026. Instagram renamed its AI creator label to AI-generated profile. Profiles that feature an AI-generated person and do not carry it may have their reach limited. Creators who only use AI tools in their process do not need it.

TikTok's ad policy already required disclosure for significantly altered content. Its help page for Symphony avatars, updated in August, says: "An AI-generated label is added to all exported videos."

Why do two studies give opposite answers?

Because almost nothing about them matches.

Start with the 31.5%. NYU Stern's summary reports two findings. Fully AI-generated ads raised click-through rate by up to 19% against ads made by human experts. Disclosing AI involvement lowered it by 31.5%. Summaries of the paper give the field figures as 105,999 impressions and 4,026 clicks, with click-through rates of 3.73% for the human ad, 4.44% for the AI ad and an estimated 2.56% with disclosure. We could not open the paper itself, so treat those as reported.

Do the division. From 3.73% to 2.56% is a 31% fall. The headline is measured against the human-made ad, not against the same AI ad without a label. Against the undisclosed AI ad, the same figures imply a fall of about 42%. That is our arithmetic on reported numbers, not a finding from the paper.

Either way the study covers banners, one product category and a disclosure whose exact placement the summaries do not describe. It is not a TikTok or Instagram video.

Now the no-effect result. MediaScience tested four label formats against an unlabelled control: text in the first three seconds, text in seconds four to six, text for the whole ad, and an icon for the whole ad. It reports no significant difference on brand recall, recognition, attitude or liking. It did not report clicks or purchases. Its partner in the study is an AI video platform the firm itself developed.

The detail that matters is how many people noticed. In the control group, with no label, 36% said the ad was AI-generated. With text on screen for the entire ad, 49% did. The strongest label moved awareness by 13 points. Half the people shown a continuous label still did not register it.

A null result under those conditions is partly a result about attention. The release does not give group sizes. If the 900 were split evenly, each group had 180 people, and a gap between two groups would need to exceed about 10 points to be distinguishable from chance. The largest attitude gap reported is 7 points.

So what does a label cost?

Somewhere between nothing measurable and about a third of clicks, and the product decides where you land.

A peer-reviewed study in Sage Open explains why. Across three experiments with 681 participants in China, a small text label on an AI-generated ad image did two things at once. It raised perceived novelty, which helped attitude and purchase intention. It lowered perceived authenticity, which hurt them. For a product framed as functional, the novelty path held and the authenticity penalty was not significant. For a product framed as pleasurable, the reverse.

That fits the two headline studies. Beauty banners are appearance-led, and they lost clicks. The limits are real: scenario experiments, one country, self-reported answers.

Surveys bracket the same range. The IAB found 73% of US Gen Z and Millennial consumers said knowing an ad was AI-made would raise their likelihood to buy or make no difference. YouGov found 45% of Australian adults would trust a brand less if it used mainly AI-generated advertising. These do not conflict. One counts neutral with positive among under-44s. The other asks all adults about trust. Both are stated attitudes. In the IAB data, 39% of Gen Z feel negative about AI ads, up from 21% in 2024.

No platform has published a labelled against unlabelled comparison for its own ads that we could find.

Which AI use produces a label viewers can see?

This is the part most guides skip. A label inside a menu and a label on screen are different exposures.

Two cautions on the Meta column. The June 2026 update describes About this ad and does not restate the next-to-Sponsored rule, so open your own live ads and look. And Meta's page says nothing about whether labels affect delivery.

Viewers draw the line in the same place. In the MediaScience survey, 60% wanted disclosure when AI simulates humans. For backgrounds it was 35%, and for colour and lighting 21%.

On organic Instagram, the profile label is the one case where the platform has stated a reach rule, and the rule penalises the missing label, not the label.

When is an AI presenter worth it?

A synthetic presenter saves production money once per ad. A click penalty, if there is one, costs media money in proportion to spend. So the saving wins at low spend and loses at high spend.

Assume the penalty acts through click-through rate, with cost per thousand impressions and conversion rate unchanged. If clicks fall by a share p, cost per result rises by p / (1 - p). Then:

• Break-even spend = saving x (1 - p) / p
• Largest penalty the saving can absorb = saving / (saving + spend)

A hypothetical example. A creator-filmed video costs $1,500. The avatar version costs $100. The saving is $1,400.

Read it the other way round. With $2,000 behind an ad, the saving survives any penalty up to 41%. With $20,000 behind it, a penalty above 6.5% wipes the saving out.

One caution. The 19% uplift for undisclosed AI ads in the NYU and Emory study is not available on a platform that labels the ad. Do not net it against the penalty.

What to do, by situation

A seven-step method

1. Sort every live and planned ad by AI use: edit only, background, full visual without a person, voice, or person.
2. Find out what viewers see. On Meta, open the three-dot menu on each live ad and read About this ad. On TikTok, confirm the AI disclaimer is on, and check again after duplicating a campaign.
3. Label any synthetic person on purpose, everywhere. For reach into New York or the EU, treat an on-screen disclosure as the safe default and take legal advice. The New York statute does not define conspicuous.
4. Run the break-even sum before producing. As a rule of thumb of ours, if the absorbable penalty is under 10%, use a person. If it is above 30%, the AI version is defensible on cost.
5. Test person against avatar with the same script, hook, offer and landing page. Compare cost per result, not clicks alone. For how many versions your budget can read, see how many ad concepts to test.
6. Keep results in two buckets, appearance-led and functional. Do not apply one bucket's result to the other.
7. Re-shoot scaled winners with a person. Whose handle that video runs from is a separate decision, covered in creator handle or brand handle. Re-check label placement each quarter, because it is still changing.

The line that matters

AI in the workflow is not what gets disclosed. An AI-made person is. Platforms, viewers and lawmakers have each arrived at that line separately. Most uses of AI in creative carry no visible label and no evidence of a cost. One use carries both, and the cost grows with spend.

Trust in what a brand shows is part of the same picture we track for search and AI answers. If you want to see where yours stands, the waitlist is open.

Questions

Does labelling an ad as AI-generated reduce performance?

The evidence is split because the tests differ. A New York University and Emory study of display ads reports 31.5% fewer clicks with disclosure. A MediaScience survey of 900 people found no significant change in recall or attitude for labelled video ads. Neither tested a Meta or TikTok label in a live feed.

Where does the AI label appear on Meta ads?

According to Meta's June 2026 update, AI info labels appear in About this ad, inside the three-dot menu. Meta's earlier rule placed the label next to Sponsored when its own tools generated a photorealistic human. The update does not restate that, so check your live ads.

Do I have to disclose AI in TikTok ads?

Yes, when image, video or audio is fully AI-generated or significantly altered. TikTok's ad policy exempts minor edits such as lighting, colour, denoising and background changes. Undisclosed AI content can be rejected or restricted. The Ads Manager disclaimer shows at the bottom of the video for the whole ad.

Does the Instagram AI-generated profile label reduce reach?

Instagram says no. Since 31 August 2026, profiles that feature an AI-generated person and carry the label see no change to reach. Profiles that should carry it and do not may have reach limited. The stated penalty is for not labelling.

Sources

1. NYU Stern, Research Highlights: The AI Advertising Paradox stern.nyu.edu
2. Dataïads, Impact of visual generative AI on advertising effectiveness dataiads.io
3. The Decoder, Telling consumers an ad is AI-generated cuts clicks by 31 percent, study finds the-decoder.com
4. Marketing Dive, AI disclosure labels don't hurt ad performance: Here's what the numbers say marketingdive.com
5. MarTech Cube, Study: AI Labeling Does Not Hurt Video Ad Performance martechcube.com
6. Sage Open, Consumer Responses to AI Disclosure Labels: The Role of Novelty and Authenticity journals.sagepub.com
7. IAB, The AI Ad Gap Widens iab.com
8. YouGov, 45% of Australians say AI-generated ads would make them trust a brand less yougov.com
9. Instagram Creators, Bringing More Transparency to AI-Generated Profiles on Instagram creators.instagram.com
10. Meta Newsroom, Expanding GenAI Transparency for Meta's Ads Products about.fb.com
11. TikTok Business Help Center, Misleading and false content ads.tiktok.com
12. TikTok Business Help Center, About ad disclaimers in TikTok Ads Manager ads.tiktok.com
13. TikTok Business Help Center, How to create avatar videos with Symphony Creative Studio ads.tiktok.com
14. Reed Smith, Fake performer, real penalty: what advertisers need to know before June 9 reedsmith.com
15. Stephenson Harwood, EU AI Act Update: European Commission Adopts Guidelines on Article 50 Transparency Obligations stephensonharwood.com

NEXT STEP

See where your brand stands.

Your Digital Authority Score, who AI recommends for your target questions, and the placements that close each gap. Your first report is free.

Get your free report →