AI generated video statistics at a glance
| Category | Statistic | Source |
|---|---|---|
| Adoption | 63% of video marketers have used AI tools to create or edit marketing videos, up from 51% a year earlier | Wyzowl, 2026 |
| Adoption | More than a third of video teams already use AI in their workflow; nearly a quarter more plan to start (900+ professionals) | Wistia, April 2026 |
| Adoption | 93% of creators say creative AI helps them produce content faster | Adobe, June 2026 |
| Quality | 57% of creators say AI outputs need moderate or extensive editing before publishing | Adobe, June 2026 |
| Output | Google has watermarked more than 100 billion images and videos with SynthID, up from 10 billion a year earlier | Google, May 2026 |
| Cost | Synthesia's Creator plan lists at $59 a month billed yearly for 360 minutes a year, which is $1.97 a finished minute | Synthesia, September 2026 |
| Cost | US film and video editors earned a median $75,420 in May 2025; BLS puts the occupation group's median pay at $36.10 an hour | US Bureau of Labor Statistics, May 2025 |
| Detection | 0.1% of 2,000 UK and US consumers correctly identified every real and fake item shown to them | iProov, February 2025 |
| Disclosure | YouTube began applying AI labels automatically in May 2026, after two years of self-disclosure | YouTube, May 2026 |
| Trust | 62% of people across 48 markets worry about fake news online, up 4 points year on year | Reuters Institute, June 2026 |
How many marketers use AI to make video?
63% of video marketers have used an AI tool to create or edit a marketing video, according to Wyzowl's 2026 report, but only just over a third of video teams have AI inside their workflow, according to Wistia.
| Statistic | Source |
|---|---|
| 63% of video marketers say they have used AI video tools to help create or edit marketing videos. The prior year's figure was 51%. Survey of 266 respondents, marketers and consumers, fielded in late 2025. | Wyzowl, 2026 |
| More than a third of video teams already use AI in their workflow and nearly a quarter more plan to start. Based on a survey of 900+ professionals plus 13 million videos and 79 million hours of viewing data. | Wistia, April 2026 |
| Teams that publish video daily are 58% more likely to use AI, and AI users are 57% more likely to produce 50 to 100 videos a year, twice as likely to produce 100 to 250. | Wistia, April 2026 |
| AI users are 82% more likely to publish multilingual subtitles on their videos. | Wistia, April 2026 |
| 93% of creators say creative AI helps them produce content faster, and 75% call it integrated or essential to how they work. Survey of more than 16,000 creators run with The Harris Poll. | Adobe, June 2026 |
Wyzowl asks video marketers whether they have ever used an AI video tool. Wistia asks video teams whether AI sits inside their workflow. Different questions, different samples, different units. The honest reading is two facts side by side, not a subtraction: trial is now the majority case, workflow adoption is not.
The adoption number everyone quotes is a trial number, and trial is cheap when every tool has a free tier. What Wistia measured is harder: teams that put AI in the pipeline publish far more video than teams that did not. That is a throughput effect, not a quality effect. If your bottleneck is volume, this already pays for itself; if your bottleneck is that nobody watches your videos, it will make the problem arrive faster.
How much published video is actually AI-generated?
Nobody publishes a credible measurement of the share of online video that is AI-generated, and any page giving you one has invented it. What exists is volume from the companies running the models: a floor, not a share.
| Statistic | Source |
|---|---|
| Google has watermarked more than 100 billion images and videos with SynthID, plus 60,000 years of audio, as of May 2026. | Google, May 2026 |
| Twelve months earlier Google reported over 10 billion pieces of content watermarked with SynthID, covering text, audio, image and video from Gemini, Imagen, Lyria and Veo. | Google, May 2025 |
| SynthID verification inside the Gemini app has been used 50 million times globally as of May 2026, and Google is extending it to Search and Chrome. | Google, May 2026 |
| HeyGen states that 118 million videos have been created on its platform. This is vendor data, from a company post rather than a survey. | HeyGen, June 2026 |
| HeyGen reports $200 million in annual recurring revenue, doubled in eight months, across more than 30 million users in 196 countries and 85% of the Fortune 100. Vendor data. | HeyGen, June 2026 |
HeyGen's 118 million is the best platform level evidence of AI video shipping inside businesses, and it is the company's own count, so treat it as a floor. The pattern matches short-form vs long-form content statistics: cheaper production pushes the median piece shorter and the count higher.
Google's 100 billion is the number I would put on a slide, with an asterisk. It counts images and video from Google's generative models, and now from partners like NVIDIA and OpenAI, so most of it is a private experiment nobody watched rather than media competing for attention. Going from 10 billion to 100 billion in a year is roughly 247 million watermarked items a day. When someone quotes a precise percentage of online video that is synthetic, ask what the denominator was.
What does AI video cost compared with filming it?
A finished minute of AI avatar video costs $1.97 at Synthesia's published Creator price for 2026, against a US Bureau of Labor Statistics median of $36.10 an hour for film and video editors and camera operators, before any camera, studio or talent cost.
| Statistic | Source |
|---|---|
| Synthesia's Creator plan lists at $59 a month billed yearly, which is $708 a year, for up to 360 minutes of video and 3,600 credits a month. That works out at $1.97 a finished minute. | Synthesia, September 2026 |
| Synthesia's Starter plan lists at $14 a month billed yearly, $168 a year, for up to 120 minutes of video, which is $1.40 a finished minute. The free Basic plan includes 10 minutes a month. | Synthesia, September 2026 |
| US film and video editors earned a median $75,420 in May 2025 and camera operators $74,990. BLS lists median pay for the occupation group at $75,100 a year, or $36.10 an hour. | US Bureau of Labor Statistics, May 2025 |
| About 40% of companies spend under $5,000 a year on video production, and 46% of video teams are keeping production budgets flat for 2026. | Wistia, April 2026 |
| 24% of marketers who do not use video say the reason is cost, joint top barrier with "not needed". | Wyzowl, 2026 |
| 57% of creators say AI outputs require moderate or extensive editing before they publish them. | Adobe, June 2026 |
Two warnings. A software list price and a wage are not the same kind of number, and Adobe's 57% says the AI minute arrives needing human hours on top. In practice $100 buys about 51 minutes of finished AI avatar video or about 2 hours 45 minutes of one editor's time, and you still need the editor.
The cost argument is real and smaller than the pitch decks claim. Strip out the fantasy that a generated minute ships unedited and the saving lands on scripts you would never have filmed at all: internal training, product walkthroughs, the fourteenth language of an explainer. That is why Wistia found 46% of teams holding video budgets flat while output keeps climbing. Teams are buying more minutes with the same money, the pattern running through marketing productivity statistics too.
Can viewers tell when a video was made by AI?
Almost nobody can tell reliably, and video is the harder format to judge. In iProov's February 2025 study of 2,000 UK and US consumers, 0.1% correctly identified every real and fake item, and participants were 36% less likely to spot a synthetic video than a synthetic image.
| Statistic | Source |
|---|---|
| 0.1% of participants correctly distinguished real from fake across all stimuli, images and video, in a test where they had been primed to look for deepfakes. Sample: 2,000 UK and US consumers. | iProov, February 2025 |
| Participants were 36% less likely to correctly identify a synthetic video than a synthetic image. | iProov, February 2025 |
| Self-reported confidence stayed above 60% whether or not the answer was right, and was highest among 18 to 34 year olds. | iProov, February 2025 |
| 75% of creators believe their audiences can already tell when creative AI was meaningfully involved in their work. | Adobe, June 2026 |
| 89% of consumers say video quality affects their trust in a brand, down from 91% the previous year. | Wyzowl, 2026 |
| 62% of people across 48 markets worry about what is real and fake online, up 4 points year on year, in a survey of about 2,000 respondents per market. | Reuters Institute, June 2026 |
Put Adobe's 75% next to iProov's 0.1% and you have the most useful contradiction on this page. Creators are confident their audience notices; the audience, tested, does not. The iProov study is nineteen months old and still the largest published controlled test of its kind.
I no longer think "will they notice" is the right question. The evidence says they will not, not reliably. The better question is what happens when they find out later, a brand problem rather than a perception problem. Wyzowl's 89% is the closest proxy I have: people cannot detect synthesis, but they can detect cheapness, and AI makes cheapness very easy to mass produce.
How are platforms labelling AI-generated video?
YouTube moved from creator self-disclosure to automatic detection in May 2026, and the C2PA Content Credentials standard now has more than 6,000 organisations behind it, so provenance metadata rather than viewer judgement is becoming the disclosure mechanism.
| Statistic | Source |
|---|---|
| YouTube requires creators to disclose AI that meaningfully alters or generates photorealistic content, in three categories: a real person appearing to say or do something they did not, altered footage of a real event or place, and a realistic scene that did not occur. | YouTube Help, 2026 |
| From May 2026, YouTube applies the label automatically when a creator does not declare AI use but its systems detect significant photorealistic AI. The label moved directly below the player on long-form video and became an overlay on Shorts. | YouTube, 27 May 2026 |
| YouTube has labelled disclosed AI content since 2024, a dated start date that sets the timeline, and disclosure does not affect recommendations or monetisation. | YouTube, 27 May 2026 |
| Content made with YouTube's own AI tools, and content carrying C2PA metadata that marks it fully generative, gets a disclosure the creator cannot remove. | YouTube, 27 May 2026 |
| The Content Authenticity Initiative passed 6,000 members in January 2026, with the CAWG specification reaching version 1.2 in 2025 and Sony shipping the PXW-Z300, a video camera that writes Content Credentials at capture. | Content Authenticity Initiative, 18 January 2026 |
| The C2PA steering committee has 11 organisations: Adobe, Amazon, BBC, Google, Meta, Microsoft, OpenAI, Publicis Groupe, Sony, TikTok and Truepic. The specification is at version 2.3. | C2PA, 2026 |
| 49% of creators always or often disclose their AI use and 18% rarely or never do; 85% say audience expectations around disclosure are rising or steady. | Adobe, June 2026 |
| Deloitte predicts that generative video will provoke a US regulatory response in 2026, potentially including a labelling requirement for AI content on social platforms. | Deloitte, TMT Predictions 2026 |
Self-disclosure was never going to hold at 49% compliance, so the machinery moved underneath it. A video carrying C2PA metadata or a SynthID watermark gets labelled whether or not the uploader ticks the box, which turns the planning question from "should we disclose" into "what will the file say about us".
If you publish AI video for a brand, assume the label is automatic and write the plan from there. The worst outcome is not a label, it is a label you did not expect on a piece you presented as filmed. Adobe's 49% against 18% says the norm is still forming, and norms formed during an enforcement wave settle conservatively. Disclose on purpose and in your own words.
Which AI video models shipped, and what can they do?
OpenAI released Sora 2 on 30 September 2025 and withdrew the Sora product on 26 April 2026, while Google DeepMind's Veo generates video with natively generated audio and marks its output with SynthID.
| Statistic | Source |
|---|---|
| OpenAI released Sora 2, its flagship video and audio model with synchronised dialogue and sound effects, on 30 September 2025. The same launch post now states that as of 26 April 2026 the Sora product is no longer available. | OpenAI, 2025 and 2026 |
| Veo generates dialogue, sound effects and ambient noise natively alongside the picture, with extended videos and expanded creative controls. Veo output carries a SynthID watermark, as does output from Gemini, Imagen and Lyria. | Google DeepMind, 2026 |
Read launch posts as evidence of what shipped and when, never of adoption or quality. The Sora line is the sharpest reminder that a category leader can be a withdrawn product nineteen months later.
Model output is a shot, and a shot is not a video, so every published AI piece is still an edit: shots stitched, audio fixed, captions added, brand applied. That is the same fact as Adobe's 57%, and it says the job that shrinks first is shooting, not editing.
How we calculated the original numbers
Four figures on this page appear nowhere else. Here is the arithmetic, so you can check it.
- 23.5%: the one-year rise in AI video tool use. Wyzowl's 2026 report puts video marketers who have used AI video tools at 63% and states the prior year's figure was 51%. The gap is 12 points, and 12 / 51 = 0.235. The 2026 wave sampled 266 respondents, so the precision is low and the direction solid.
- $1.97 a finished minute, and the bigger plan costs more per minute. Synthesia's Creator plan lists at $59 a month billed yearly, which is $708 a year for up to 360 minutes of video: 708 / 360 = $1.97. The Starter plan lists at $14 a month billed yearly, $168 a year for up to 120 minutes: 168 / 120 = $1.40. Tripling the minute allowance raises the unit price by 41%, because the extra money buys features and credits rather than cheaper minutes. List prices checked in September 2026.
- $100 buys 51 AI minutes or 2 hours 45 minutes of an editor. At $1.97 a finished minute, $100 buys 50.8 minutes of Synthesia avatar video. The US Bureau of Labor Statistics lists median pay for film and video editors and camera operators at $36.10 an hour (May 2025), so $100 buys 2.77 hours of one editor's time. This compares a published software list price with a published wage, not two production budgets.
- About 247 million watermarked items a day. Google reported over 10 billion pieces of content watermarked with SynthID on 20 May 2025 and more than 100 billion images and videos on 19 May 2026. The difference is 90 billion over 364 days, which is 247 million a day, or roughly 2,860 a second. The counts are not strictly like for like, because the 2025 figure includes text and audio, so read it as an order of magnitude.
What the 2026 numbers actually say
Adoption is real and shallower than the headline. 63% of video marketers have tried it, more than a third of video teams run on it, and those that do publish far more video. If you are choosing where to put AI first, the evidence points at volume work: subtitles, dubbing, variants of what exists, which is where Wistia found AI users ahead by 82%.
Nobody knows what share of published video is AI-generated, and that is the biggest hole in this dataset. Google's 100 billion watermarked images and videos is a supply number and HeyGen's 118 million is a vendor number. Neither tells you what fraction of what a person watched this week was synthetic, and no independent measurement exists.
Detection is not coming from viewers, so provenance is the mechanism left. 0.1% of tested consumers got every item right, YouTube now labels automatically, C2PA membership passed 6,000, and Deloitte expects US regulation to follow. Plan as if the file will declare itself, because increasingly it does. That shift towards machine-readable provenance also sits underneath generative engine optimization statistics, where being identifiable as a source now beats being visible.
FAQ
What percentage of marketers use AI to create video?
63% of video marketers have used an AI tool to create or edit a marketing video, up from 51% a year earlier (Wyzowl, 2026, 266 respondents). Wistia's narrower April 2026 measure, based on 900+ professionals, puts video teams using AI inside their workflow at more than a third. The two surveys asked different questions of different samples, so read them as two facts, not one gap.
How much does AI-generated video cost per minute?
$1.97 a finished minute at Synthesia's published Creator price for 2026, which is $59 a month billed yearly for 360 minutes a year. The smaller Starter plan works out cheaper per minute, $1.40. Those are list prices for avatar video, not production costs: 57% of creators say AI output needs moderate or extensive editing first (Adobe, June 2026).
Can people tell if a video is AI-generated?
Only 0.1% of 2,000 UK and US consumers correctly identified every real and fake item in iProov's February 2025 study, and participants were 36% less likely to spot a synthetic video than a synthetic image. Confidence stayed above 60% whether or not the answer was right, which is the part that should worry you.
Does YouTube require you to label AI-generated video?
Yes. YouTube requires creators to disclose AI that meaningfully alters or generates photorealistic content, and since May 2026 it applies the label automatically when its systems detect photorealistic AI the creator did not declare (YouTube, 27 May 2026). Content carrying C2PA metadata marked fully generative gets a disclosure the creator cannot remove.
How much AI video is being produced?
Google has watermarked more than 100 billion images and videos with SynthID as of May 2026, up from 10 billion a year earlier, which works out at about 247 million watermarked items a day. HeyGen reports 118 million videos created on its platform (vendor data, June 2026). No published source measures what share of online video is AI-generated.
What is C2PA and who supports it?
C2PA is the open standard behind Content Credentials, the provenance metadata that travels with a file to record how it was made and edited. Its steering committee has 11 organisations including Adobe, Google, Meta, Microsoft, OpenAI, Sony and TikTok, and the spec is at version 2.3. The related Content Authenticity Initiative passed 6,000 members in January 2026.
Sources
- Wyzowl, Video Marketing Statistics 2026 (266 respondents, fielded late 2025)
- Wistia, State of Video Report: Video Marketing Statistics for 2026 (April 2026)
- Adobe, 2026 Creators' Toolkit Report (June 2026, 16,000+ creators, with The Harris Poll)
- HeyGen, HeyGen Surpasses $200M ARR (June 2026, vendor data)
- Synthesia, Pricing (list prices checked September 2026)
- US Bureau of Labor Statistics, Film and Video Editors and Camera Operators (May 2025 wage data)
- iProov, Study Reveals Deepfake Blindspot, press release via Silicon UK (February 2025, 2,000 UK and US consumers)
- Google, Making it easier to understand how content was created and edited (19 May 2026)
- Google, SynthID Detector: Identify content made with Google's AI tools (20 May 2025)
- Google DeepMind, Veo (2026)
- YouTube, Improving AI labels for viewers and creators (27 May 2026)
- YouTube Help, Disclosing altered or synthetic content (2026)
- OpenAI, Sora 2 is here (30 September 2025, updated April 2026)
- C2PA, Coalition for Content Provenance and Authenticity (specification 2.3, 2026)
- Content Authenticity Initiative, The State of Content Authenticity in 2026 (18 January 2026)
- Reuters Institute, Digital News Report 2026 (June 2026, 48 markets)
- Deloitte, TMT Predictions 2026 (2026)
Methodology. Numbers were collected in September 2026 and checked against the original publisher's page, report or press release, never a statistics roundup. Vendor numbers appear only where the publisher states what is being counted, and they are labelled vendor data in the sentence. The only figure older than 24 months is YouTube's 2024 labelling start date, marked DATED in the source data. Where a publisher blocks automated fetching (Wyzowl, bls.gov, openai.com, iProov), the text was verified through a text-rendering proxy or a verbatim wire syndication and marked verified in the data file. Full stat list: `research/ai-generated-video-statistics-stats.tsv`. See the full statistics hub for related pages. Last updated 21 September 2026.