AI-generated content is becoming almost impossible to distinguish from human-written content.
Now, OpenAI is taking a step that could change how we identify it.
The company is introducing an invisible ChatGPT watermark for eligible users in the European Union. The watermark will be added to text generated by ChatGPT and Codex as OpenAI responds to transparency requirements under the EU AI Act.
You won’t see the watermark.
But a machine might.
And that raises a bigger question:
If AI can create content that looks human, how do we know where it came from?
What Is the ChatGPT Watermark?
OpenAI’s new AI watermark is called textGrain.
Unlike a traditional watermark, it doesn’t add a visible logo, symbol, hidden characters or unusual formatting to the text.
Instead, textGrain subtly influences the model’s choice of words and tokens, creating a statistical pattern within the generated text. A specialised detector can then look for that pattern.
So, when someone reads a ChatGPT response, it should look completely normal.
The watermark isn’t attached to the document. It is embedded statistically in the generated text itself, through patterns in the model’s word and token choices.
Think of it as a statistical fingerprint rather than a visible mark.
That means the watermark isn’t sitting at the top or bottom of a document, nor is it simply stored as metadata.
If you copy AI-generated text from ChatGPT and paste it into a Word document, for example, the watermark signal can potentially remain because it is associated with the text itself.
Invisible to people. Detectable by machines.
Why Is OpenAI Adding an AI Watermark?
The move is closely connected to the EU AI Act.
The regulation introduces transparency requirements for AI-generated content, including requirements for generated text to be identifiable in a machine-readable way.
OpenAI says its ChatGPT and Codex watermarking will roll out to eligible users in the EU over the coming weeks. Meanwhile, API customers around the world can choose to enable text watermarking for selected models, although it remains off by default.
This makes the ChatGPT watermark more than just a new technical feature.
It is part of a wider shift toward AI transparency and content provenance.
How Does AI Text Watermarking Work?
Think of textGrain as a statistical fingerprint hidden in the way an AI model generates language.
When generating a sentence, the model has multiple possible words or tokens it could choose next. The watermarking system subtly influences those choices according to a secret pattern.
Across enough text, these choices create a recognisable statistical signal.
A specialised detector can then analyse the writing and look for evidence of that signal.
Importantly, the watermark does not identify the person who used ChatGPT, establish ownership of the text, or show exactly how much of the content was generated by AI.
So an AI watermark can provide a clue about where text came from, but it isn’t proof of authorship.
Can the ChatGPT Watermark Be Removed?
This is where things get interesting.
The watermark isn’t foolproof.
Detection can become less reliable when text is substantially rewritten, paraphrased or translated. Shorter passages can also be harder to identify.
That means the technology shouldn’t be treated as a perfect AI content detector.
Instead, it is better viewed as another layer of AI content provenance.
The important distinction is that changing the format of a document doesn’t necessarily remove the signal. Significant changes to the actual wording, however, can weaken it.
What Does This Mean for AI-Generated Content?
For years, the biggest question around generative AI was:
What can AI create?
The question is now evolving into:
Can we trust, verify and understand what AI creates?
As businesses use AI for customer service, software development, research, decision-making and automation, the origin of AI-generated content becomes increasingly important.
This becomes even more relevant when AI systems interact with sensitive business data or become connected to real-world workflows.
AI adoption therefore brings more than productivity gains.
It also brings questions around:
- AI security
- AI governance
- Data privacy
- Regulatory compliance
- Human oversight
- AI transparency
- Accountability
The ChatGPT watermark doesn’t solve all of these challenges.
But it points toward a future where AI provenance becomes part of responsible AI adoption.
AI Transparency Is the Bigger Story
OpenAI’s move is not happening in isolation.
The company already uses provenance technologies for AI-generated images and audio and is now extending its approach to text. It also plans to make textGrain available as open-source technology.
The bigger shift is clear.
AI-generated content is becoming easier to create.
At the same time, the technology used to identify and trace that content is evolving.
For businesses, this means AI strategy can no longer stop at:
“How can we use AI?”
The more important questions are becoming:
Where is AI being used?
What data does it access?
Can its outputs be trusted?
How is its use governed?
And can the organisation demonstrate responsible AI adoption?
The Future of AI Isn’t Just About Intelligence
The ChatGPT watermark may be invisible.
Its impact may not be.
As AI becomes deeply embedded in business, trust, transparency, security and governance will become just as important as the intelligence of the technology itself.
The next phase of AI won’t simply be about creating smarter systems.
It will be about creating systems businesses can confidently trust.
For a deeper look at how AI systems are evolving beyond single-agent models, explore Multi Agent Orchestration: AI’s Microservices Moment and see how multiple specialised AI agents can work together as one intelligent system.
Building AI You Can Trust
As AI adoption grows, businesses need more than powerful models. They need the right data, security, governance and workflows to make AI work reliably in the real world.
Evvo Technology helps businesses turn AI opportunities into practical solutions through AI consulting, AI agents, GenAI, data analytics and cybersecurity.
Because adopting AI is easy. Building it securely, responsibly and for real business impact is what matters.
