How can you tell if an image is AI-generated? Do you carefully watch the subject’s movements, count how many times they blink, or zoom straight in on their fingers, hoping to find that tell-tale 6th digit? And what about text: ‘delve into’, ‘in today’s fast-paced world’ and of course the dreaded em dash are all regarded as universal signifiers of AI, but there’s no real way to be sure… until now.
Determining whether content was made by a human or an AI model used to be an individual responsibility, but all that is about to change with the introduction of Article 50 of the EU AI Act, which requires AI platforms to label all artificially generated or manipulated content.
Here’s what we know, what we don’t (yet), and what’s to come as this new legislation is rolled out across Europe.
What Is Article 50?
Article 50 of the EU AI Act came into effect on August 2nd, and aims to make AI-generated content easier to both detect and to distinguish from human creations.
It sets out new transparency obligations for the providers of generative AI models, who must now:
- Make it clear to users that they are interacting with an AI model;
- Mark all AI-generated or manipulated content as such;
- Disclose all AI images, audios, or videos that constitute a deep-fake.
Some efforts have been made before to label content as AI-generated, but we’re not just talking about easily-cropped logos anymore. The Article directly stipulates that these disclosures must be made in a ‘machine-readable format’: one which is woven into the technical data of the output and remains detectable even if the original content is pasted elsewhere.
Already known as invisible watermarks, these digital disclosures might not be visible to you, but they are most certainly there – and, more importantly, search engines can see them even if you can’t.
What AI Content Is Covered By Article 50?
Both written and visual output is included in the scope of the EU AI Act.
These transparency rules cover all ‘synthetic audio, image, video or text content’, although exceptions are made for:
- Systems that are involved in the detection, prevention, investigation or prosecution of criminal offences, as authorised by law.
- Systems that ‘do not substantially alter the input data’ e.g. that are used for editing original, human-produced content.
- Artistic, fictional, or satirical content, which must still be disclosed as AI, but can be done so in a way that doesn’t detract from the work itself.
Is Human-Reviewed Content Permitted?
The EU AI Act doesn’t constitute a ban on AI-assisted content, provided that the final editing and verification process is done by humans.
It is false or misleading content that hasn’t been reviewed by a human that Article 50 is cracking down on – not all AI-assisted output. This is made clear in Paragraph 4 of the Article, which makes a direct exception for content that ‘has undergone a process of human review or editorial control’.
In cases such as these, whereby a human holds the final editorial responsibility for the content before it is published, the role of AI in its creation does not need to be disclosed.
Does AI Watermarking Have Any Limitations?
An AI watermark shows that AI was used to make content, but not how far that involvement went.
The presence of a watermark on a few words does not prove that AI wrote the finished article, or even the entire sentence. It does mean that AI had a hand in it – perhaps to research topics, reword ideas, and structure thinking – but the final piece may have been entirely written by humans, and therefore fall outside of the scope of Article 50.
Currently, more clarity is needed regarding exactly how AI watermarks will be used to categorise content that was created with the assistance of AI. This is especially true when the final output is edited and reviewed by a human, since (as we already touched upon) this type of content does not need to be disclosed.
The Age of AI Watermarking
Here’s how the big AI platforms such as Anthropic, OpenAI, and Google are complying with the transparency obligations set out in the EU AI Act’s Article 50.
Anthropic
Anthropic, the owner of Claude, has announced that all new Claude models launched in the EU will embed provenance meta data (AI watermarks) within the content they generate.
The organisation also stated its intention to update the models launched before August 2 so that they generate the watermarks as well, and to support users to detect them, since they are currently only visible to digital systems.
OpenAI
OpenAI will utilise a system called SynthID to automatically embed AI watermarks, and (just like Anthropic) also intends to release a public verification tool to make OpenAI-generated content easier to detect.
These new initiatives build upon OpenAI’s existing provenance efforts, such as the introduction of image Content Credentials in 2024, and the meta data watermarks that were trialled within both Sora and Voice Engine.
AI behemoths may only just be introducing provenance systems now, but Google has been quietly advancing the detection of AI content for years.
In fact, it is Google who introduced digital watermarking technology to the world with the release of SynthID three years ago. Today, SynthID has been successfully integrated into Google’s own generative AI models, and the search engine will continue to provide more ways to verify AI content in the coming months – including implementing C2PA Content Credentials across its generative tools.
Final Thoughts From Our Team
In an age of deepfakes and digital deception, it is in the public interest to watermark AI-generated materials that are created to deceive or mislead.
However, at SEO It Right, we believe that provenance markers should serve as evidence of AI input – not as a judgement on the quality or authenticity of the work.
Want to discuss this topic in more detail? Feel free to contact our team today, or check out our other publications on our website.
