47 figures on AI-generated images and deepfakes, from how many are made to whether people, detectors and platforms can spot them, and which labelling laws apply. Each figure was checked against the linked source on 22 October 2026. Estimates, company figures and secondary sources are marked.
Pixel 10 Camera attaches Content Credentials to every JPEG and reached the highest level the C2PA Conformance Program currently defines. Google Photos displays the credentials
Cloudflare Images added a one-click "Preserve Content Credentials" setting. Cloudflare says credentials had frequently been stripped when CDNs transformed images
Platform labelling and whether metadata survives upload
Figure
What it measures
Source
2024-02-06
Meta said it would detect C2PA and IPTC "AI generated" signals from Google, OpenAI, Microsoft, Adobe, Midjourney and Shutterstock. It acknowledged that people "can strip out invisible markers"
YouTube began requiring creators to disclose realistic altered or synthetic content. A more prominent label is used for news, elections, health and finance
LinkedIn shows the C2PA "CR" icon on credentialed images and video, and says it is "not yet possible to identify and label all AI-generated and modified content"
The Washington Post uploaded an AI video to 8 platforms. Only YouTube flagged it, in the description. All 8 stripped the Content Credentials (Secondary source.)
5 detectors (Hive, AI or Not, ZeroGPT, Sightengine, ScamAI) called authentic war photos AI-generated 13.33% of the time combined. The worst tool did so 40% of the time
Drop in open-source detector performance on deepfakes circulating in 2024 compared with academic benchmarks. Commercial tools did better but still trailed forensic analysts
Zero-shot accuracy of 23 open detector variants across 12 datasets (~2.6M images). Flux Dev, Firefly v4 and Midjourney v7 images held most detectors to 18–30%
According to a Hunton Andrews Kurth summary of the Digital Omnibus on AI (Reg. (EU) 2026/1744, in force 2026-07-27), marking obligations apply from this date to systems already on the market before 2026-08-02. Check the Official Journal text for compliance decisions
Organisations that had signed the EU Code of Practice on marking and labelling AI-generated content by end of July 2026. The code was finalised for the 2026-06-10 plenary
China's Measures for Labeling AI-Generated Synthetic Content (国信办通字〔2025〕2号) took effect. They require explicit labels plus implicit metadata (provider name, content ID), from 4 agencies. Dated 2025-03-07, published 2025-03-14
California AB 853 (signed 2025-10-13) delayed the AI Transparency Act to 2026-08-02. It added platform provenance-detection duties (2027) and capture-device latent disclosures (2028)
Under the US TAKE IT DOWN Act (signed May 2025), platforms must remove reported intimate imagery, including AI deepfakes, within 48 hours. Platform duties became enforceable 2026-05-19
India's IT Rules amendment (notified 2026-02-10) requires prominent labels plus permanent provenance metadata on synthetic content. The takedown window fell from 36 to 3 hours
South Korea's AI Basic Act took effect, requiring generative-AI output labels and notice for realistic AI media. Fines go up to KRW 30M (~US$21k), after a 1-year grace period
People and detectors both struggle with appearance alone, and platforms routinely strip the records that say how an image was made. That is why isGenAI reads recorded evidence, such as Content Credentials, AI labels and saved prompts, and says plainly when none survives. Our own platform metadata survival study measures what X removes, and the deepfake check covers source tracing when records are gone.
Cite this page
Please cite the original source for any figure you quote. To credit this compilation, use one of these:
APA citation
isGenAI. (2026, October 22). AI image and deepfake statistics (2026). https://isgenai.com/ai-image-statistics
HTML link
<a href="https://isgenai.com/ai-image-statistics">AI image and deepfake statistics</a> (isGenAI, sources linked per figure)
How this page was compiled
We included a figure only after opening the source page and finding it there. Two rows rely on secondary reports because the primary page could not be opened; they are marked. Regulation rows mostly cite law-firm summaries rather than statute text and are not legal advice; see the Article 50 guide for the EU rules.
Questions about AI image statistics
How many AI images are there?
Everypixel Journal estimated in August 2023 that more than 15 billion images had been created with text-to-image tools, about 34 million a day. That is an extrapolation, not a count, and no reliable 2026 total exists. Google says more than 10 billion pieces of content carry its SynthID watermark, and TikTok says it has labelled more than 1.3 billion videos as AI-generated.
Can people tell AI images from real ones?
Not reliably. In Microsoft’s “Real or Not” experiment, people were right 62% of the time across about 287,000 judgments, and a PNAS study found 48.2% accuracy on StyleGAN2 faces, about chance. Ofcom found only 9% of UK adults were confident they could identify a deepfake.
How accurate are AI image detectors?
Accuracy drops on real-world content. NewsGuard found five detectors called authentic war photos AI-generated 13.33% of the time combined, and the Deepfake-Eval-2024 benchmark measured AUC drops of about 45–50% for open-source detectors on deepfakes that circulated in 2024.
Do social platforms keep Content Credentials?
Usually not. In a Washington Post test, all eight platforms stripped Content Credentials from an uploaded AI video, and isGenAI’s own study found X removed every provenance record we uploaded.
Can I reuse these statistics?
Yes. Cite the original source linked in each row. The compiled table is available as a CSV download.