ChatGPT to follow Claude’s watermark pledge – but Grok to swerve it
Elon Musk’s Grok is poised to swerve rivals’ plans to ‘watermark’ all artificially-generated content, City AM understands, after refusing to sign up to a European push forcing firms to make AI content traceable.
xAI was the only major large language model maker to shun a European Union-led move to tighten AI transparency rules that means the likes of Anthropic, Google and OpenAI’s responses will contain an invisible watermark.
On Monday, Claude-maker Anthropic became the first major AI firm to announce plans to place machine-readable signals indicating a text had been generated using its model. The company said its machines would “mark AI-generated content from day one” as part of its commitment to the EU’s AI Act.
But other AI firms – including Microsoft, Meta and Mistral – also signed up to the charter, paving the way for others to follow in Anthropic’s footsteps.
City AM understands that OpenAI plans to watermark text outputs as part of their EU AI Act commitment, and that the company is currently working through the details.
Sam Altman’s AI firm has said its goal is to expand its existing content provenance measures to text, indicating that ChatGPT is set to follow Claude in making its written output easier to identify.
Google is also expected to introduce measures covering AI-generated content under the code, although exactly how the requirements will be applied to text produced by Gemini has yet to be set out.
“Consistent with our commitments under the European Commission’s Code of Practice on Transparency of AI-generated content, our goal is to expand provenance signals to all modalities including text,” OpenAI said in a post on its own support website this month.
The ChatGPT maker added that the measures were intended to give customers and developers “clear ways to meet their own transparency obligations as standards and tooling continue to mature”.
OpenAI has not said exactly how it will mark text or confirmed that it will use the same invisible watermarking method announced by Anthropic this week.
But xAI, Elon Musk’s company behind Grok, did not sign the code, meaning it has made no equivalent commitment through the agreement for Grok.
However, Musk’s AI firm will still have to comply with applicable requirements under the EU AI Act, and so could eventually be made to introduce its own system for marking AI-generated material despite staying outside the code.
City AM approached OpenAI, Google and xAI for comment.
Claude announces watermarking plans
As part of its update on Monday, Anthropic said new Claude models will embed an invisible, machine-readable watermark directly into generated text. Anthropic said the mark should remain when text is copied and pasted and may survive some editing.
Supported images and other files will carry signed provenance information recording that they have been processed by Claude.
Despite the change being driven by European rules, Anthropic said it will apply the marking worldwide wherever supported Claude models are available.
Even human-written text could carry a Claude watermark if someone subsequently asks the chatbot to proofread, translate or edit it. But AI-generated text that is heavily edited or paraphrased may lose a detectable mark.
That is partly because the EU rules distinguish between content generated by AI and cases where the technology is merely assisting with standard editing. Article 50 provides an exception where AI performs an assistive editing function without substantially changing the original material or its meaning.
Grok scrutiny continues
Grok faced regulatory scrutiny earlier this year over its ability to create sexualised images of real people on its platform.
The European Commission opened an investigation into X in January over concerns that Grok had been used to create manipulated sexually explicit images and whether the company had properly assessed and addressed the risks involved.
In the UK, Ofcom separately launched an investigation after reports that Grok was being used to digitally remove clothing from images of real people, including concerns around sexualised images involving children.
X subsequently said it had stopped Grok from altering images to remove people’s clothing in jurisdictions where such content is illegal, saying users prompting Grok to produce illegal material would face the same consequences as those uploading it themselves.
The investigations are separate from the AI Act’s transparency requirements, but put Grok’s handling of AI-generated material under scrutiny months before the new rules took effect.
Regulators in Australia, France and Germany also examined the chatbot, while Grok was temporarily restricted in Indonesia and Malaysia.
The European Commission’s investigation sits under the Digital Services Act and could result in penalties if it ultimately finds X breached the rules, yet no such finding has yet been made.