8 Lessons on AI UGC vs Real UGC for Meta Ads

Published 6 November 2026 · Last updated and fact-checked November 2026

Short answer: On the only public same-script test I found, the human ad beat the AI ad 6 add-to-carts to 1 (B&T, 2026), but 7 events cannot settle AI UGC vs real UGC. I found no larger published test as of November 2026. Use AI to screen hooks, then test on your own account.

Key takeaways

• The Thinkerbell test used one script and equal spend, and produced 7 add-to-carts in total, which by my calculation is too few to call (p about 0.13) (B&T, 2026).

• The most quoted $115k test compared 5 AI platforms and had no human arm (Segwise, 2026).

• Fully AI-made display ads beat expert ads on click-through by up to 19%, but disclosing the AI cut effectiveness by up to 31.5% (NYU Stern, 2025; Compass Lexecon, 2025).

• Meta puts its AI info label next to "Sponsored" when its own tools add a photorealistic human (Meta, 2026).

• The FTC has no blanket ban on AI avatars, but a fake testimonial is still prohibited (FTC).

• Billo videos start at $99, and at a $40 cost per purchase my test math needs $15,840 of media to detect a 30% gap (Billo pricing, read 6 November 2026).

What is the difference between AI UGC and real UGC?

AI UGC is a creator-style video in which a generated presenter delivers your script. Real UGC is the same format filmed by a person, usually a paid creator with the product in hand. The gap that matters is who can truthfully claim an experience, not only the $99 versus $11 price.

A real creator can say what happened when they used the product, if it did happen. An AI presenter can read product facts you can substantiate. The formats also differ in production time and in how Meta may label them, and the evidence on which one sells more is thin.

1. What does the only public same-script test show?

Thinkerbell ran adman Adam Ferrier against an AI copy of himself on Meta, with one script and equal spend. The human ad logged 6 add-to-carts and the AI ad logged 1. It is the only public test I found with this design, and at 7 events it points in a direction without proving one.

• Who ran it: Thinkerbell behavioural scientist Maddy Ross

• Presenter: Adam Ferrier, real and AI-generated

• Impressions: just under 2,000 for AI, just over 2,200 for human

• Conversion: 0.05% for AI, 0.26% for human

• Cost per result: $64.26 for AI, $10.86 for human, currency not stated

• Published: 27 February 2026

"Both ads used the same script, the same setting and the same call to action," B&T reported in its write-up of the Thinkerbell test, 2026.

A same-script design is what the rest of the market leaves out, so the Thinkerbell result deserves a fair reading. The human ad's cost per result was about 6 times lower. That looks decisive until you count the events. I ran a Fisher exact test on 1 of about 2,000 against 6 of about 2,200 and got a two-sided p value of about 0.13, so a gap this size could plausibly come from chance. That is my calculation from rounded impressions, not a figure in the article.

The test has two other limits for a DTC buyer. The product was The Bachelor's Handbag, fronted by a known advertising figure, and my read is that viewers who know him may react in a way strangers would not react to a stock creator. It also measured add-to-carts, not repeat purchase or refund rate. I read it as a reason to run your own test and as a warning that AI is not automatically equal.

2. What do the vendor numbers leave out?

Most ranking pages sell AI or UGC production, and their numbers rarely compare the same script. Before trusting one, check for a human arm, a shared script, disclosed spend and disclosed conversions. Of the 5 public claims I checked, 1 passes the first 2 checks, and none passes all 4.

• Misfit test: $115k across 5 AI platforms, no human arm

• APXlab: six months on unnamed supplement brands, script match not stated

• Skincare split: $15K, relayed secondhand with no AI versus human ROAS figure

• Hoox field data: AI reaches 85% to 110% of human click-through, as cited by Segwise

• Thinkerbell: shared script and human arm, with spend undisclosed

The Segwise write-up of the $115k test is useful if you are choosing between AI tools: its winner, Mirage Studio, reported 2.42x ROAS at a $70 CPA. It is often cited as AI against human, and it is not that, because all 5 arms were AI. A test with no human arm cannot tell you anything about AI versus real.

Segwise's main comparison, updated 2 November 2026, repeats the Hoox range and a $15K skincare split, then gives no ROAS or CPA figure for either side of that split. APXlab describes six months of side-by-side ads for unnamed supplement brands, but it does not say the scripts matched, and it sells this production. Zeely is the most candid of the group: it warns that AI UGC can lose badly and reports no direct test of its own. None of this makes these companies dishonest. It means their pages answer a tool buyer's question, not yours. Ask any vendor for the script, the spend and the purchase count before you repeat its number.

3. Do large ad studies favor AI creative?

One working paper found AI-made display images up to 19% ahead of expert ads on click-through. Another, covering 16 billion impressions, found they do best when they do not look like AI. Both measure images in display ads, so neither tells you how an AI person holding your product performs in a Reel.

• Ghose, Lee, Todri and Adamopoulos: lab experiment plus field study, working paper

• Field result: up to 19% higher click-through than human expert ads

• AI-edited human ads: no significant improvement

• TUM, Columbia and Harvard paper: over 16 billion impressions, 7,000 advertisers

• Seen as potentially human-made: about 59% of AI ads, per the TUM summary

The NYU Stern summary of the Ghose paper holds the finding most relevant to a media buyer: fully AI-created ads won, while AI edits to human ads showed no significant gain. If you plan to patch a creator video with AI fixes, this gives you no support, though the study did not test video.

"Disclosing AI involvement in ad generation significantly reduces advertising effectiveness by up to 31.5%," wrote Ghose, Lee, Todri and Adamopoulos in their working paper abstract, 2025.

The Technical University of Munich summary of the second paper adds the condition that matters.

"perform best when they don't look AI-generated," the Technical University of Munich wrote of AI ads in its summary of the working paper, 2025.

UGC asks for more than a good image. It asks a viewer to believe a person tried a product, and a display banner never makes that claim. So I would use these papers to justify AI for product shots and static hooks, and I would not stretch them to cover testimonial video. Neither was described as journal-published on the pages I read, so treat them as strong working evidence that may still change in review. A click is also not a purchase, and UGC buyers pay for purchases.

4. What happens when viewers know an ad is AI?

Peer-reviewed experiments mostly find that knowing lowers trust, though not in one simple line. A University of Amsterdam study of 304 people found both a negative route and a positive route. These studies measure attitudes and intent, not CPA, so read them as risk signals.

• Koning and Voorveld, 2025: N = 304, Instagram ad for a fictional boxing studio

• Yıldırım and Dursun, 2026: 762 Gen Z women in Türkiye, human vs disclosed AI vs undisclosed AI

• Sands and colleagues, 2022: no difference in intention to follow AI or human influencers

• Wang, Xu and Lin, 2026: 50 studies, 17,642 participants

• Measured: trust, perceived manipulation and purchase intention, not sales

"AI disclosures increased conceptual AI knowledge and attitudinal persuasion knowledge, resulting in a decrease in trust towards the advertisement and organization," wrote Koning and Voorveld of the University of Amsterdam in the Journal of Interactive Advertising, 2025.

The same abstract reports that disclosure also raised trust through a second route, so the label is not a pure penalty. Hiding it does not obviously help either, as a larger experiment with 762 participants found, though that sample was Gen Z women in one country.

"both disclosed and undisclosed AI influencer conditions significantly increased perceived manipulation," wrote Yıldırım and Dursun in the Journal of Theoretical and Applied Electronic Commerce Research, 2026.

The picture is not one-sided. Sands and colleagues found people equally open to following AI or human influencers. Wang, Xu and Lin's meta-analysis found a stronger association between credibility and purchase intention for virtual influencers (r = 0.52) than human ones (r = 0.32). It does not show that virtual influencers sell more, because a correlation across 50 studies cannot compare two ads for one product, and none of these studies used your audience or your offer. The stronger caution comes from Sands and colleagues in 2024, who found that a brand's intention to use AI-generated rather than human models lowered brand attitude across four studies of underrepresented consumers.

5. Does Meta label AI UGC ads?

Yes, in some cases, and placement depends on how the ad was made. Meta's own tools can put AI info next to "Sponsored" when they add a photorealistic human. Ads made with outside tools get a label inside About this ad when Meta detects them, under a policy updated 1 June 2026.

• Meta's tools, major edits: AI info in the three-dot menu

• Meta's tools, photorealistic human: AI info next to "Sponsored"

• Meta's tools, minor edits: no label if no photorealistic human appears

• Outside tools: detected through industry signals, label in About this ad

• Rollout: Meta says not every eligible ad carries the label yet

"next to the Sponsored label (not behind the three-dot menu)," Meta stated of photorealistic humans in its generative AI transparency update, updated 1 June 2026.

Segwise describes this as a disclosure rule for ads with photorealistic AI humans. That is broader than what I can confirm from Meta's own pages. The Meta help page describes labels Meta applies, and the visible position it names is tied to AI-generated photorealistic humans.

The Stern figure gives this a practical edge. If a visible disclosure can cut click-through by up to 31.5% on display ads, an AI label near the top of your video could change a test result, so it belongs in your test notes as a variable. Watch which version actually ships. An AI video made outside Meta and then edited with Meta's tools may not land in the same place as a video made wholly inside Ads Manager. Check the live ad preview and About this ad before you read the numbers, since a label on one arm and none on the other is a difference between your two ads.

6. Can an AI actor give a testimonial?

An AI actor can present a product, but it cannot truthfully report an experience it never had. The FTC's rule has been in effect since 21 October 2024. It does not ban avatars, and it does prohibit a testimonial that is fake or false. Ask counsel about your script, since this is not legal advice.

• Rule: 16 CFR Part 465, in effect 21 October 2024

• Avatars: no blanket ban

• Prohibited: testimonials that misrepresent the speaker's existence or experience

• Penalties: courts may impose civil penalties for knowing violations

• Also relevant: the FTC Act and the Endorsement Guides

"The rule has no blanket prohibition on the use of AI-generated avatars in marketing," the FTC stated in its Consumer Reviews and Testimonials Rule Q&A.

The FTC's announcement names testimonials that misrepresent that they are by someone who does not exist, and section 465.2 covers misrepresenting whether the speaker used the product and what the experience was. The Q&A says a stock avatar's use could be a testimonial, prohibited only if the underlying testimonial is fake or false, and it adds that avatars may still be deceptive under the FTC Act.

My working line for scripts follows that logic. "This serum has 12% vitamin C" is a claim you must substantiate, delivered by a presenter. "I used this for 3 weeks and my skin cleared up" from an avatar describes an experience no one had. A real customer's words read by an avatar is the gray case: the experience exists, but viewers may think the avatar is that customer, so get counsel's view before it runs. Those example lines are mine, not the FTC's.

The script is also where the test can quietly break. AI UGC scripts often lean on first-person claims because that is what makes UGC work. If your compliant AI script has to drop them and your creator script keeps them, you are testing two different ads.

7. How much cheaper is AI UGC, really?

Production is about 9 times cheaper on list prices: Billo publishes a $1,000 pack for up to 10 videos, and Segwise reports Arcads at $110 a month for 10. But production is the small part of a fair test. At a $40 cost per purchase, a test sized for a 30% gap needs $15,840 of media.

• Billo: videos start at $99, $1,000 pack covers up to 10 (Billo pricing, read 6 November 2026)

• Arcads: $110 a month for 10 videos as reported by Segwise; I could not read it on Arcads' site

• HeyGen: Creator plan $29 a month, $24 billed annually (HeyGen pricing, read 6 November 2026)

• Human add-ons: usage rights, whitelisting and rush fees add 25% to 150%, per a report cited by Segwise

• Vendor range cited by Segwise: $5 to $30 AI, $150 to $500 human

Those are list prices. Neither figure includes your time on scripts, briefs and reviews, and an AI video that needs five regenerations costs more of that time than the plan price suggests. Billo's $99 buys a 15-second video from a regular creator; longer cuts and Premium creators cost more.

Where the saving earns its keep is hook volume. If you can screen 30 opening lines for about what 3 human videos cost (30 x $11 against 3 x $99), you learn which message to pay a creator for. That is the case for a hybrid, and it is what Zeely and Segwise recommend. I agree with it as a working plan, while noting that I found no published proof that a hook which wins in AI also wins when a creator says it. A saving of $890 on 10 videos is real, but it is 6% of the $15,840 test, so it should not decide the question by itself.

8. How do you run a fair AI vs real test?

Change one thing: the presenter. Keep the script, setting, product, offer, landing page and conversion event the same, then buy enough conversions to read the result. At 1% of landing-page visitors converting, detecting a 30% gap needs 198 conversions in the weaker arm, and detecting a doubling needs 23.

• Variable: presenter only, with the same script read word for word

• Event: purchase, not add-to-cart or click

• Budget: weaker-arm conversions x target CPA x 2 arms

• Stop rule (mine): end an arm early only at 3 x target CPA with zero purchases, a budget guard rather than a statistical test

• Record: AI label position, spend, impressions, purchases, refunds

The sample sizes use the standard two-proportion formula at 5% significance and 80% power, with a 1% visitor conversion rate as a placeholder. They are my calculation and they are rough. They do show why the 7-event test could not settle anything.

Two pieces of research should keep you humble about the result. Lewis and Rao analyzed 25 field experiments representing $2.8 million in digital spend and found a median confidence interval on return on investment over 100 percentage points wide. And Braun and Schwartz showed that platform A/B tools can deliver each ad to a different mix of users.

"The estimated comparisons confound the effect of the ad content with the effect of algorithmic targeting," wrote Braun and Schwartz in the Journal of Marketing, 2025.

So a win tells you which ad worked better as Meta delivered it, which is still useful for buying decisions, but it is not proof that viewers prefer the presenter. Run the winner again against a fresh version before you shift the budget. If you sell to a small audience and cannot afford the 30% version, test for a doubling at $1,840 and accept that smaller gaps stay invisible.

Compare the evidence by what it can prove

Public claims differ on the four things that make a comparison readable: a human arm, a shared script, disclosed spend and disclosed conversions. This matrix records what each source stated when I read it on 6 November 2026, so a "not stated" means the page did not say, not that the fact is false.

Work out your own test budget

Multiply the conversions you need in the weaker arm by your target cost per purchase, then by 2 arms. The example uses a $40 target as a placeholder, so swap in your own number. The smaller the gap you want to detect, the faster the media bill climbs.

Conversions come from the two-proportion formula at a 1% conversion rate, 5% significance and 80% power. These are my calculations and they ignore delivery effects, so treat them as a floor.

Frequently asked questions

Does AI UGC work for Facebook ads?

It can, but the public evidence is thin. Hoox field data cited by Segwise puts AI at 85% to 110% of human click-through, a vendor figure. The only public same-script test favored the human ad, on just 7 add-to-carts. Treat AI as something to test on your account.

Do I have to disclose AI UGC on Meta?

For ordinary commercial ads, Meta's pages describe labels that Meta applies, not a blanket declaration by the advertiser. Check your own ad preview to see which label appears. Your legal duty on testimonials is separate, and it comes from the FTC.

Is AI UGC cheaper than real UGC?

Yes, on production. Billo lists $1,000 for up to 10 videos, while Segwise reports Arcads at $110 a month for 10. That gap is about 9 times. Media decides the bigger number, though, because a test sized for a 30% gap needs $15,840 at a $40 target.

Can I use an AI avatar for a testimonial?

Not for an experience nobody had. The FTC says its rule has no blanket ban on AI avatars, but a testimonial that is fake or false is prohibited. An avatar presenting product facts you can substantiate is a different case. Ask counsel to review any first-person script before it runs.

Should I replace my UGC creators with AI?

Not on current evidence. Keep creators for the ads you scale, and use AI to screen hooks and angles cheaply. I found no published proof that a hook winning in AI also wins with a creator, so test the transfer on your own account before cutting a creator budget.

How I would set this up for a DTC brand

I would start with your target CPA and work back to a budget, then pick 2 hooks from an AI screen and film both with one creator and one AI presenter. This article reports no test of mine, and my own proof comes from Meta creative and funnel work, such as Escape Velocity.

On Copy Chief's Escape Velocity campaign, Meta Ads and a mobile-first funnel reached a 55.03x average ROAS as CPA moved from $42 to $22. That was not an AI test, and I name it to show the kind of account this reasoning comes from. If you want the weekly decisions handled, my Conversion Creatives plans run $999 or $2,449 a month.

Before you spend on the test, run your current ads through my free Kill Sheet. It gives you five signals and three verdicts to use each Monday, so the budget you free up can pay for a test big enough to read.

Sources

Peer-reviewed research

• Koning and Voorveld (2025). Disclaimer! This Content Is AI-Generated. Journal of Interactive Advertising, 25(3), 240-253.

• Yıldırım and Dursun (2026). Disclosure Matters. Journal of Theoretical and Applied Electronic Commerce Research, 21(6), 194.

• Wang, Xu and Lin (2026). A Meta-Analysis of Influencer Credibility and Purchase Intention. Social Sciences, 15(11), 762.

• Sands, Campbell, Plangger and Ferraro (2022). Unreal influence. European Journal of Marketing, 56(6), 1721-1747.

• Sands, Demsar, Ferraro, Campbell and Cohen (2024). Inauthentic inclusion. Psychology & Marketing, 41(6), 1396-1413.

• Lewis and Rao (2015). The Unfavorable Economics of Measuring the Returns to Advertising. The Quarterly Journal of Economics, 130(4), 1941-1973.

• Braun and Schwartz (2025). Where A/B Testing Goes Wrong. Journal of Marketing, 89, 71-95.

Working papers (not peer reviewed on the pages I read)

• Ghose, Lee, Todri and Adamopoulos (2025). The Impact of Visual Generative AI on Advertising Effectiveness, via NYU Stern and Compass Lexecon.

• AI in disguise (2025), Technical University of Munich, Columbia Business School and Harvard Business School, via TUM.

Government and platform sources

• US Federal Trade Commission: Rule Q&A, final rule announcement, 14 August 2024 and 16 CFR 465.2.

• Meta: Help Center on AI labels and generative AI transparency update, updated 1 June 2026.

Industry and vendor sources (not peer reviewed, prices read 6 November 2026)

• B&T on the Thinkerbell test, 27 February 2026.

• Segwise comparison, updated 2 November 2026, and Segwise on the $115k test.

• APXlab and Zeely.

• Billo pricing and HeyGen pricing.