Do AI UGC ads work? AI avatars, real creators and a third option
What an independent Ipsos test found when human and AI-made ads went head to head, the AI-label rules on TikTok, YouTube and Meta, and a third option.
By Abondy ResearchPublished 10 min read
In short: AI UGC ads, where an AI actor reads your script on camera, can work for plain product messages, but the strongest independent test we found has human-made ads ahead. In an Ipsos study with Syracuse University, published in May 2026, 10 human-made ads and 10 AI-made versions built from the same briefs were tested with 3,000 US consumers. On average, the human-made ads were 14% stronger on Ipsos's short-term measure and 17% stronger on its long-term one, and most viewers couldn't tell which ads were AI. A third option sits between the two: keep real footage of real people, and use AI only to make many versions of it.
We make Abondy, whose video model makes versions of your own footage rather than AI actors, so Abondy competes with the AI-avatar tools named here. Read our part with that in mind.
What counts as an AI UGC ad?
"UGC" means a video that looks like an ordinary user made it. There are three ways to get one:
- AI-avatar UGC. A tool generates a person who reads your script. Arcads' home page offers "1,000+ Captivating AI Actors" who can "hold your product, show your app, and wear your clothes", and its FAQ says they "are AI actors based on real actors, who gave consent and are paid for the use of their likeness." Arcads alternatives compares it with the other AI-actor tools.
- Creator UGC. A real person films the video for you. Billo's pricing page, read November 21, 2026, says its videos "start at $99, which covers a 15-second video from a regular creator with basic editing."
- Versions of real footage. AI makes many new cuts of a video you already have, such as a demo, a screen recording or a creator clip you have the rights to. The people in it stay real and do only what they did on camera.
"Do AI UGC ads work" is usually asking about the first kind against the second.
Do AI UGC ads work? What an independent test found
The most careful public test we found is Ipsos's AI Ads Are Good Enough — And That's the Problem, run with Syracuse University's S.I. Newhouse School and published May 18, 2026. We read it on November 21, 2026.
The study took 10 real 30-second ads from major brands. Google's Gemini reverse-engineered each one's creative brief and wrote a shot list, which went "directly into OpenAI's Sora 2 to generate the AI counterpart", with "no human creative intervention". All 20 ads were tested with 3,000 US consumers in Ipsos's Creative|Spark test and scored on measures Ipsos calls "sales validated". What it found:
| Finding | Ipsos's figure |
|---|---|
| Short-term effectiveness (Creative Effect Index) | Human-made ads 14% stronger on average |
| Long-term effectiveness (Equity Effect Index) | Human-made ads 17% stronger on average |
| Viewers of an AI ad who were at least somewhat confident it was AI | 25% |
| All viewers who were unsure either way | 40% |
| AI ads that used storytelling | 30%, against 49% of ads in another Ipsos report |
| People who say companies using AI should have to disclose it | 79% |
Some AI ads still worked: Ipsos says AI "can, and did, create some effective advertising". It says the strongest AI ads "tend to be product-driven, direct", such as the AI versions for Febreze and Herbal Essences, with "a clear problem, a clear solution, and a clear reason to believe". Where the brief asked for emotion or a creative leap, the AI versions "fell short". Ipsos concludes that AI has "an important supplemental role" but "is not a replacement for the human-led creativity" that sets ads apart.
What the study doesn't tell you
- Polished brand ads, not avatar UGC. The human ads were 30-second brand spots and the AI ads were made to match them, not an AI actor talking to a phone camera.
- Predicted, not measured. The scores come from a survey test, not sales or installs in a live ad account.
- No person in the loop. The AI ads came straight from Gemini and Sora 2, though Ipsos says "some ads required multiple iterations" before they were good enough to test. It doesn't show how AI ads do when a person writes the script and picks the takes.
- Ten pairs, US viewers only. Enough to show a pattern, not to predict how your product will do.
We looked for an independent, published test of AI-avatar UGC against creator UGC, with its method shown, and didn't find one. So: AI ads can work, they did best on simple product messages in the one careful test we found, and on average they trailed human-made work.
When do AI avatars make sense, and when don't they?
| AI avatar | Real creator | Versions of real footage | |
|---|---|---|---|
| Who's on screen | An AI actor | A real person you hired | Whoever is in your footage |
| What you supply | A script, product images or an app link | A brief, often a product shipped to them | A video you already have |
| Cost per video, list prices read Nov 21, 2026 | Under $1 to about $12 on the plans we checked (Arcads, Creatify, MakeUGC, HeyGen) | Billo from $99; Insense UGC from $100 plus plan and fee | Cents on Abondy (below) |
| AI label needed | Usually, if it looks realistic (rules below) | No | Depends on what the edit changes |
| Fits best | Plain product messages, many languages, quick script tests | Testimonials, product in hand, a face people trust | Scaling a video that already works |
The prices come from each company's own pricing page, and the units differ (a 15-second clip, a minute, a whole video), so treat them as a range. An AI actor is the wrong tool for a testimonial: one saying "I've used this for a month" describes an experience nobody had (see the FTC rule below). It's also a poor fit where trust in real people is the point. Ipsos found human-made ads more "entertaining", "unique" and "talkable".
Is AI UGC legal? What the platforms and the FTC say
This is a summary of what each organization's own page says, read November 21, 2026, not legal advice. For a script that makes claims, ask a lawyer.
- TikTok, for posts. Its Community Guidelines (the Integrity and Authenticity section, released August 25, 2026) say "we require creators to label AI-generated or significantly edited content that shows realistic-looking scenes or people", and "Unlabeled content may be removed, restricted, or labeled by our team".
- TikTok, for ads. Its ad policy on misleading and false content (last updated November 2026) allows AI-generated content if you "Apply the AIGC label, or by adding a clear disclaimer, caption, watermark, or sticker of your own". It counts "completely AI-generated" content and making someone "say something they didn't actually say, using AI voice-cloning" as needing it, and lists lighting changes and background edits as minor. It warns: "If we identify AI-generated content that has not been disclosed, your ad will be rejected or restricted."
- YouTube. Its help page says "We require creators to disclose AI-generated or altered content that seems realistic." It lists "Cloning one's own voice to create voice overs or dubs" as something you don't need to disclose, and says "Disclosing AI content alone won't limit a video's audience or impact its eligibility to earn money."
- Meta. Its February 6, 2024 announcement says it will "require people to use this disclosure and label tool when they post organic content with a photorealistic video or realistic-sounding audio that was digitally created or altered, and we may apply penalties if they fail to do so." Check Meta's help center for the current wording.
- The FTC. Its rule on fake reviews and testimonials, announced August 14, 2024, covers testimonials that "misrepresent that they are by someone who does not exist, such as AI-generated fake reviews", or by someone "who did not have actual experience with the business or its products or services".
In practice: label realistic AI people, and don't script an AI actor as a customer describing results.
The third option: real footage, many versions
Ipsos's numbers point to a middle route. Keep the part people respond to, a real person and a real product, and use AI for the part it's good at: volume. A tool that makes versions reframes, recolours, re-captions and re-paces one real video into many. Nobody is made to say anything new.
Volume matters because short-video views are very uneven. Here is what versions of one upload got in our own posts, made by our video model, Flagship, and posted by real people on their own accounts:
| Flagship uploads with 20 or more read posts, Oct 19 to Nov 12, 2026 | Result |
|---|---|
| Uploads | 136 (10,674 read posts) |
| Versions posted per upload, median | 23.5 |
| The typical post's views (median of each upload's median) | 42.75 |
| Each upload's best post, median | 1,533 views |
| Uploads with at least one post past 1,000 views | 83.1% |
Our reader never got a view count back for 13.9% of the posts on these uploads (1,728 of 12,438), so the table uses only the posts it read.
So for the same video, cut different ways and posted on different accounts, the typical post got about 43 views while the best post typically got about 1,500. That says why many tries help. It doesn't say versions beat AI avatars, and we didn't test that. The data is in more detail in AI video variations.
We did run a small AI-avatar test of our own. In October we ran AI-avatar reaction videos, with the avatar clips made in outside tools, in a campaign for Abondy's own app, posted by real people on their own accounts: 1,110 posts from October 11 to 15, 2026. Our reader never got a count back for 202 of them (18.2%); the 908 it read had a median of 20 views and a mean of 186. The other videos in that campaign that week came from sets that also had AI-made reaction clips, so there was no real-footage group to compare them with, and we draw no conclusion from it.
How to read our numbers: the sample is mostly gaming clips, largely Roblox, on small accounts. A few campaigns dominate it: in the Flagship set above, one brand's campaign accounts for about 45% of the read posts, and Abondy's own campaigns for about a fifth. View counts are the count when we last read them, and a few were last read before the post was 7 days old. Posts our reader never got a count back for are left out of the figures, rather than counted as 0. Flagged, rejected and deleted posts are left out. Instagram counts are a floor, since our reader can under-count them.
How to test AI UGC against real UGC yourself
- Use one brief for both: same hook, offer and length, or you're testing the script, not the actor.
- Make several of each. One AI ad against one creator ad tells you about two videos. Meta and TikTok's advice on volume is in how many creatives per ad set.
- Run them side by side, in the same ad set or posting week, at the same budget.
- Judge on your goal, such as cost per install or purchase, not views alone.
- Label the AI versions, so a rejected ad doesn't spoil the test.
Organic posts can do a first round without ad spend: see ad creative testing without an ad budget.
Where Abondy fits
Flagship, the model brands use on Abondy, doesn't make AI actors. It makes new versions of a video you upload, each with its own look and caption, in phone, wide and square shapes. Real people post them on their own TikTok, YouTube Shorts, Instagram Reels and Facebook accounts. A post costs 12 credits, about 16¢ to 20¢ a post on a plan, and posts under 100 views 24 hours after the link is handed in aren't charged. Plans start at $20 a month and include 100 to 2,500 free Flagship videos a month (pricing). How the posting side works is in UGC distribution.
Questions people ask
Does AI UGC convert?
No public test we found measures it directly. Ipsos's May 2026 test predicted sales effects rather than measuring conversions: on its scores, human-made ads were 14% stronger on short-term and 17% stronger on long-term effectiveness on average, while AI versions of simple product ads did well. None of it was avatar UGC, so test it on your own product.
Can people tell an ad was made with AI?
Mostly not. Only 25% of people who saw an AI-made ad in the Ipsos test were at least somewhat confident it was AI, and 40% of all viewers were unsure either way.
Do I have to label AI UGC on TikTok?
For realistic AI-made people, yes. TikTok's guidelines require a label on AI-generated content "that shows realistic-looking scenes or people", and its ad policy says undisclosed AI content in ads will be "rejected or restricted".
Is AI UGC cheaper than hiring creators?
Per video, by a wide margin at list prices: under $1 to about $12 a clip on the AI tools we checked, against $99 and up per video from Billo or Insense. If you already have footage that works, versions of it cost less again.
Read next
- Creative testing · 20 Nov 2026 · 6 min read
AI video variations: turning one ad into hundreds of versions
What makes a video variation read as a new video, which edits don't count, and what 31,973 posts of versioned videos show about why volume matters.
- Creative testing · 20 Nov 2026 · 8 min read
Ad creative testing without an ad budget: organic posts as a test bed
Can organic posts test ad creatives before you spend on ads? What 31,973 posts of versioned videos say about noise, sample size and calling a winner.
- Playbooks · 17 Nov 2026 · 8 min read
How many creatives per ad set? What Meta and TikTok recommend
What Meta's and TikTok's own pages say about creatives per ad set or ad group, the hard limits, automated creative tools and how often to refresh.
