AI-generated content is text, images, audio or video produced by a generative model from a prompt, with human input that can range from none to a complete rewrite. That range, not the technology, decides which rules apply to whoever publishes it.
On 26 September 2026 we read the top ten results on Google UK for "ai generated content" (DataForSEO). They are definitions, detection guides and tool pages: IBM, Google Search Central, Conductor, Capitol Technology University, GIJN, Vtiger, QuillBot, Wikipedia and TechTarget. None of them brings together the three layers that actually affect a UK publisher, and none of them deals with what UK law says.
What counts as AI-generated content
The rules don't treat three very different situations in the same way.
- Writing assistance. The model corrects, rephrases, shortens, translates. The content remains what you had. The EU AI Act calls this an "assistive function for standard editing" and excludes it from the marking obligation.
- Substantial generation. The text, image or video comes from the model. This is where both platform policies and the EU transparency obligations apply.
- Alteration of real material. A person saying things they never said, a place that didn't look like that in the shot. This is deepfake territory, and in the UK it is the only category where some uses are already criminal offences.
The distinction matters for SEO too. An article written by a person and tidied up with a model is not the same as a thousand pages churned out automatically, and the data on how AI content is indexed and ranked shows that Google treats the two cases differently.
How much AI-generated content is really online
The most solid count available comes from Graphite, which in April 2026 analysed 55,400 URLs drawn at random from Common Crawl, published between January 2020 and March 2026, running them through three different detectors (Pangram, GPTZero and Copyleaks). The result: 49.9% of articles are mostly AI-generated and 50.1% are written by people. The share has held at around half for five quarters.
Two caveats that usually disappear when the figure is repeated. The sample is English-language articles, so it describes the English-speaking web as a whole, not the UK specifically. And the authors warn that they did not measure how much traffic those articles receive: in an earlier study of 43,000 URLs they found that this content largely does not appear in Google or ChatGPT.
The UK figure on the same trend exists and looks at it from the producers' side. According to the ONS article Artificial intelligence in UK businesses: 2023 to 2026 (July 2026), self-reported AI use among UK businesses with 10 or more employees rose from around 12% in late 2023 to around 35%, with larger firms more likely to have adopted it. The gap by sector is wide: 58% in information and communication, 13% in construction.
Put together, the two numbers say something counter-intuitive: in the UK, AI is still used by a minority of businesses, while the rules on declaring it are tightening, especially for anyone whose content reaches the EU.
Google's rules are about purpose, not the tool
Google Search's spam policies contain no violation called "AI content". What they do contain is scaled content abuse, defined as "when many pages are generated for the primary purpose of manipulating search rankings and not helping users". Among the examples, Google lists using generative AI tools "to generate many pages without adding value for users". The subject of the sentence is the purpose; the tool is a detail.
The part almost nobody quotes is in the documentation on creating helpful, reliable, people-first content, where Google asks three questions of anyone producing content with automation: "Is the use of automation, including AI-generation, self-evident to visitors through disclosures or in other ways?", "Are you providing background about how automation or AI-generation was used to create content?" and, on authorship, "Do pages carry a byline, where one might be expected?".
These are recommendations, not ranking requirements. But they are the same things the AI Act asks for, and the same things E-E-A-T rests on: who signs, with what experience, through what process. If your author structured data is already in order, you start ahead.
The rules that apply to UK publishers
Anyone publishing AI-generated content from a UK site or profile answers to several sets of rules at once, with different parties bound and different consequences.
| Source | Who is bound | What it requires | Since | What happens if you don't comply |
|---|---|---|---|---|
| Google Search spam policies | Anyone publishing on the web | Don't generate pages at scale to manipulate rankings; make the use of automation evident | In force | Manual action or algorithmic demotion of the pages |
| EU AI Act, Art. 50(2) (Regulation (EU) 2024/1689) | Providers of generative AI systems, including UK providers whose systems are used in the EU | Mark outputs in a machine-readable format so they are detectable as artificially generated | 2 August 2026 | Administrative fines under the regulation |
| EU AI Act, Art. 50(4) | Deployers who publish, including UK businesses whose output is used in the EU | Disclose deepfakes and AI-generated text published to inform the public on matters of public interest | 2 August 2026 | Administrative fines; exemption where there is human review and editorial responsibility |
| Copyright, Designs and Patents Act 1988, s. 9(3) | Authors and rights holders | For computer-generated works with no human author, the author is the person who made the arrangements necessary for their creation; protection lasts 50 years | In force (repeal proposed by the government in March 2026) | Uncertain protection for output with no human creative input |
| Sexual Offences Act 2003, as amended by the Online Safety Act 2023 and the Data (Use and Access) Act 2025 | Anyone who shares or creates the material | Prohibits sharing, and now creating, sexually explicit deepfakes of a person without consent | In force | Criminal offence; up to two years in prison for the most serious forms of sharing |
| CAP Code (ASA) | Advertisers | No blanket duty to disclose AI, but ads must not mislead; disclose AI use where the audience would otherwise be misled | In force | ASA ruling and withdrawal of the ad |
| Platform policies (example: YouTube) | Creators and advertisers | Disclose realistic content that is altered or generated with AI; filters, colour correction and AI-written scripts are excluded | In force | Label applied by the platform, removal or suspension from monetisation |
Three points in the table deserve an extra line.
The first is the division of responsibilities in Article 50 of the EU regulation. Technical marking falls on the model provider, and it must be "effective, interoperable, robust and reliable as far as this is technically feasible". Disclosure to the public falls on whoever publishes. They are not the same obligation and they don't fall on the same party. The UK is outside the EU, but the regulation also covers providers and deployers in third countries when the output of the AI system is used in the Union, so a UK business writing for EU customers cannot ignore it.
The second is the exception that covers most company blogs. For text published "with the purpose of informing the public on matters of public interest", the disclosure obligation falls away when the content has gone through human review or editorial control and someone holds editorial responsibility for it. In plain terms: a real byline is worth more than a label.
The third is UK law itself. There is no general obligation in the UK to label AI content on the web, and anyone who says otherwise is oversimplifying. The government's Report on Copyright and Artificial Intelligence (March 2026) proposes removing the protection for computer-generated works in section 9(3) of the 1988 Act, while keeping copyright for works created with AI assistance, and leaves labelling to industry best practice for now. On deepfakes, UK criminal law targets sexually explicit material; there is no general offence for AI-falsified images or voices. Italy, by comparison, introduced one in 2025, punishable by one to five years in prison.
How to tell whether content is AI-generated
There are three technical approaches, and they work very differently.
A watermark inside the content. Google DeepMind's SynthID embeds an imperceptible signal in the model's output. It survives cropping and recompression better than a visible watermark, but only tools that know about it can read it.
Provenance metadata. The C2PA standard records inside the file the chain of who created and edited it. Google joined the coalition's steering committee in 2024 and announced that it would bring C2PA metadata into "About this image" in Google Images, Lens and Circle to Search, as well as into its ad systems. This is the approach Article 50 of the AI Act makes mandatory for providers. The limitation is well known: a screenshot is enough to strip the metadata.
Statistical detectors. These are what everyone tries first, and they are the least reliable of the three. Graphite's 2025 study reported a false positive rate of 4.2% for its detector on articles published before ChatGPT: out of a hundred texts written by people, four were flagged as artificial. On a single article that margin isn't a statistical error, it is a wrong accusation. We have compared AI detectors and the picture hasn't changed: they are useful for assessing an archive, not for judging one piece.
On YouTube the question is already operational. The platform's rules ask creators to disclose realistic content that makes a real person appear to say or do something they didn't do, alters footage of real events or places, or shows realistic scenes that never happened. Minor changes are explicitly excluded: beauty filters, colour correction, upscaling, cloning your own voice and scripts written with AI. The label appears in the player for photorealistic content and in the expanded description for the rest, and YouTube can apply it automatically when it finds C2PA metadata in the file.
What to sort out before publishing
Four checks, in the order it makes sense to do them.
- Decide who signs before you generate. The byline is what triggers the Article 50(4) exemption and it is the first question in Google's documentation on authorship. A real person, with a verifiable role and an author page, not an anonymous editorial team.
- Define where review happens. Who reads, what they check, where the record is kept. Without this step there is no exemption, because what the regulation asks for is editorial responsibility, not good intentions.
- Treat formats separately. A text reviewed by a person and an image generated from scratch follow opposite rules: the first may not need a disclosure, the second will be marked at source and should be disclosed if it depicts real people or events.
- Check the numbers one by one. This is where models get things wrong most often, and no label can save it. It matters most if you are aiming for citations in answer engines, where verifiable data is the selection criterion.
On the balance between machine and person, Conductor, a US SEO and AEO platform, proposes a clean split: around 70% of the heavy lifting to the model (initial research, first draft, workflow support) and the 30% that matters to the person, meaning strategy, subject expertise and quality control. It is a vendor's commercial recommendation, not research data, though "What is the 30% rule for AI?" does appear in Google UK's "People also ask" box for this search. The proportion describes well what stays human: choosing the angle and the evidence. The same two things that make content built for SEO different from a summary of what is already on page one.
If your goal is to appear inside generated answers and not just in the ten blue links, the discussion moves to a different level: AI Overviews, citations in ChatGPT and generative engine optimisation reward sources that bring original material. In our AI SEO projects the starting point is almost always the same: establish what the site knows that nobody else does, and put it in a readable form.
Frequently asked questions
No. Google Search spam policies contain no violation called AI content: they contain scaled content abuse, meaning many pages generated for the primary purpose of manipulating rankings. The test is the purpose, not the tool. A text generated with a model, checked and signed by a person, breaks no rule.
In the UK there is no general legal requirement. The ASA says there is no blanket requirement to disclose AI in ads, but disclosure is needed where the audience would otherwise be misled. If your content is used in the EU, Article 50 of the AI Act, applicable from 2 August 2026, requires disclosure for deepfakes and for text published to inform the public on matters of public interest, unless the text has gone through human review with editorial responsibility. Google recommends making the use of automation evident and adding a byline.
For now, partly. Section 9(3) of the Copyright, Designs and Patents Act 1988 protects computer-generated works with no human author for 50 years and treats as author the person who made the arrangements necessary for their creation. In its March 2026 report on copyright and AI, the government proposed removing this provision while keeping protection for works created with AI assistance, so output with no human creative input may lose protection.
Not reliable enough to judge a single text. Graphite reported in 2025 a false positive rate of 4.2% for its detector on articles written before ChatGPT. Across an archive of thousands of URLs that margin is acceptable; on a single article it means wrongly accusing one writer in twenty-five.
According to the Graphite study of April 2026, based on 55,400 English-language URLs drawn from Common Crawl and analysed with three detectors, 49.9% of articles are mostly AI-generated and 50.1% are written by people. The figure covers the English-speaking web and does not measure how much traffic those articles receive.
Some are. The Sexual Offences Act 2003, as amended by the Online Safety Act 2023 and the Data (Use and Access) Act 2025, makes it an offence to share or create sexually explicit deepfakes of a person without consent. There is no general offence covering other AI-falsified images, video or voices, which fall under the same laws as any other content, such as defamation, fraud or harassment.
One thing the industry rarely says: the marking required by Article 50 will move disclosure from a line of text at the bottom of the page to structured data inside the file. When that happens, the question a client, a journalist or a search engine asks you will no longer be "did AI write this?" but "who is responsible for what it says?". If you already keep a record of how your content is made, you will have the answer ready. To see how we work on this, the projects we run all start from the same inventory.
