Is AI video good enough for ads yet?

"Good enough" means two different things, and mixing them up is how people waste money. An ad can be credible, meaning viewers do not notice anything off. Or it can be compelling, meaning viewers remember it and act. AI video is much closer to the first than the second. This page gives a verdict by ad type, the best available evidence and its limits, a 10 minute check for any clip, and a small test you can run with five people.

The short answer: yes for some ads, not yet for others

Think by job, not by tool. AI video works best for a short, direct product message in a familiar format. It works worst when the ad must make someone feel something, show a real person doing a real thing, or show exact packaging. The table below gives one line per ad type.

"Good enough" also depends on the comparison. Against no video at all, a clean AI clip is a clear step up. Against a 20 second phone video of the real owner, it may lose.

What the best evidence says (Ipsos, May 2026) and its limits

The most useful study was run by Ipsos together with the Newhouse School of Public Communications at Syracuse University, and published in May 2026. It tested 20 ads: 10 made by humans and 10 made fully by AI (the AI ads were built from briefs written with Gemini, Google's AI model, and the video was made with Sora 2, a video model from OpenAI). About 3,000 US consumers took part. Here is what Ipsos reported:

• Human-made ads scored 14% higher on short-term effectiveness and 17% higher on long-term effectiveness.
• Among people who saw an AI ad, 25% were at least somewhat confident it was made by AI. Across all viewers, 40% were unsure.
• The AI ads that did best were product-driven, direct and used familiar formats.
• AI fell short when the brief asked for a creative leap, emotion or a point of view. Ipsos also found the AI ads used storytelling only 30% of the time, against 49% of ads in earlier research by MISFITS Stories (a creative effectiveness study Ipsos cites).
• 79% of people said companies using AI in ads should say so.
• Ipsos recommends mixing human and AI work rather than choosing one.

Now the limits, because they matter. These scores are predicted effectiveness from a survey. Nobody bought anything. Twenty ads is a small set. All the AI ads came from one set of tools, and one of those tools, Sora, is now closed: OpenAI's help page says the Sora app ended on 26 April 2026 and the API on 24 September 2026. A different tool might score differently. So read the study as a careful signal about how people react today, not as a promise about your sales.

A trade report on a different study, by Taboola (an ad network that places ads on news and content sites), found roughly equal click rates for AI and human ads on its own platform. That study was run by a company that sells AI ad tools, so I treat it as one input and cover it in how to test whether AI ads perform as well as real ads. Studies disagree partly because they measure different things: Ipsos asked how effective an ad would be, Taboola counted clicks.

I would not rely on vendor blog posts that say AI ads match or beat real ones. They give no method, so there is nothing to check.

Where AI video works well today

• Product close-ups and b-roll. Short clips of an item on a surface, a pour, a turn, a reveal. Fewer moving parts means fewer faults.
• Social variants. Ten versions of one idea with different openings, to see which hooks hold attention. Volume is where AI is strongest. See making 20 ad variations.
• Animated explainers and simple scenes. Abstract or stylised visuals do not have to look like real life, so viewers judge them less harshly.
• Presenter clips with a consented avatar, when the script is short. Check lip sync in each language.

Where it still fails

• Text and labels. Garbled letters, a logo with the wrong shape, small print that changes between frames. Fast motion also makes on-screen words hard to read in the first second.
• Hands and faces. Extra fingers, faces that drift over the clip.
• The same character across shots. A person who looks different in each cut.
• Exact products. If the real item has a specific shape, colour or pack, the AI version may be close but wrong. A customer who receives a different item has a fair complaint.
• Stories and feelings. Ipsos names narrative as the weak point. A family moment or a customer story still lands better when real people tell it.
• Anything that claims a result. A testimonial, a before and after, a real person's endorsement. Do not make these with AI. They can mislead buyers, and rules on fake endorsements apply. See fake testimonials and AI reviews.

Ad type table: AI, real footage or a mix

Ad type	AI today	What to film for real
Product b-roll for a social post	Good	Only the final product shot with the real pack
Hook variants of one offer	Good	One real clip as the control
Spokesperson or avatar ad	Fair, needs consent and a label check	The owner on a phone often beats an avatar for trust
Local service ad (plumber, dentist)	Fair for b-roll only	The real team, the real van, the real shop
Testimonial or review	No	A real customer, with permission
Before and after result	No	Real photos or video of the real job
Emotional brand story	Not yet	Real people and a real story
Packaging, labels, fine print	Risky	The real pack, filmed or photographed

"Fair" means it can work if you check carefully. "Risky" means I would film the real thing unless you can prove the AI version matches it. For a longer discussion of the never-fake cases, see when not to use AI video, and for cost and speed see AI video vs stock footage vs filming yourself.

A 10 minute pre-flight check for every clip

Run this on every clip before it goes into an ad account. It is a short version of the 12-point list in the AI video quality check, so use that page when you need the full pass and fail rules.

1. Pause on the product. Pause three times at random. Does the label, logo and shape match the real item? Put the real item next to the screen.
2. Count fingers. Pause any frame with a hand. Count them.
3. Read all text at full screen. Every word, including small print and signs in the background.
4. Watch faces on mute for five seconds. Does the face stay the same person? Do the eyes and teeth stay steady?
5. Play it with sound. Does the mouth match the words? Does the voice sound human, with breath and pauses?
6. Watch it on a phone. Most people will.
7. Check the claim. Does the clip say or show anything you cannot prove? Remove it.
8. Check who appears. Does any face or voice belong to a real person? See the consent checklist.

Any fail on item 1, 3 or 7 is a stop. The others are fix-or-regenerate decisions. If a clip needs a fourth round of fixes, film it.

A small viewer test before you spend on media

You do not need a lab. Here is a test that costs an afternoon and no ad budget. It is a plan I suggest, not a method from a study, so adjust the numbers to your business.

1. Make three versions of the same offer: one AI, one filmed for real (a phone is fine), and one mix of both.
2. Find five people who do not work with you and who fit your customer. Show each person the three clips on mute, one at a time, in a different order for each person.
3. After each clip, ask two things: "What is this ad offering?" and "Would you trust this business with your money? Why or why not?"
4. Do not say AI is involved. Note whether anyone brings it up on their own.

A reasonable pass for any version: at least four of five describe the offer correctly, and nobody mentions "AI" or "fake" unprompted. If the AI version fails and the real one passes, you have learned something for free. Five people cannot tell you how a campaign will perform, so this test is a filter for bad clips, not a forecast. When a version passes, the next step is a paid head-to-head, which I cover in the test page linked above.

Labels and disclosure

Expect that a label can appear whether or not you add one. A trade report from 7 July 2026 says Meta, the company behind Facebook and Instagram, shows an AI label in the "About this ad" panel for ads made or heavily edited with its own AI tools, and for outside AI content it detects through C2PA metadata (a standard way of marking where a file came from). The report describes Meta applying the label; Meta's own help page is not confirmed here, so check Ads Manager and Meta's current rules before you rely on any of this. More on that mark in why you should not strip provenance marks.

Rules differ by platform, country and ad type. Do you have to label AI content is a starting point. This is information, not legal advice: check local law and each platform's current rules. Do not promise a client that "no one will know". Ipsos found 25% of people who saw an AI ad were at least somewhat confident, and 79% want disclosure.

When to film it for real instead

• A real person, a real customer or a real result is the reason the ad works.
• The product has detail that must be exact, such as a label, a size mark or a safety warning.
• The clip keeps failing the pre-flight after a few tries.

A mix is often the right answer. Film the people and the proof. Use AI for b-roll, variants and the pieces that do not need to be exact. Tools such as Higgsfield (a web platform that gathers several AI image and video models in one place) and many others can make the AI side. The tool matters less than the review you do after it. For a first project with a small budget, see your first AI video ad.

How I'd approach this

I would pick one ad type from the "good" rows of the table, make the three versions, and run the viewer test before paying for any media. If the AI version passes, I would run the pre-flight on it, write down what was checked, and only then put a small budget behind it. If it fails, I would film the real version and use AI only for b-roll. I would never use AI for a testimonial or a result.

Blur Studio builds the repeatable side of this for small businesses and agencies: a review checklist, an approval step and a shared library that records which clip came from which tool and who signed it off. See the business page if that would help.