When a company uses artificial intelligence (AI) to produce a marketing message or client proposal, there’s now a way to tell if the text is machine-generated even in instances where there are no obvious signs.
Technology companies are embedding invisible watermarks on AI-generated content as regulators and businesses seek ways to establish where digital material came from.
OpenAI this week said it will begin adding an invisible watermark to text generated by its popular chatbot ChatGPT, extending a technology that has already been used to mark AI-generated images, audio and video.
The watermark will roll out over the coming weeks in the European Union (EU) to ChatGPT users and those using its coding agent Codex.
Developers using OpenAI’s application programming interface (API) anywhere in the world can also activate it for selected models.
Anthropic announced in August that it would watermark text generated by its Claude chatbot worldwide.
Google has also embedded its SynthID technology across content generated by Gemini, including text, images, video and synthetic audio. The tech giant has even integrated this verification into the Gemini app and Google Search, and users can check whether content was generated or altered using its AI systems.
Why watermark?
The push is partly driven by new EU rules requiring companies to make AI-generated and AI-manipulated content identifiable by other systems.
The EU AI Act recognises the risks of AI-generated material being used for impersonation, deception, fraud and misinformation, because AI can produce large amounts of synthetic content that is increasingly difficult for people to distinguish from human-created material.
For businesses, these watermarks could become another layer of accountability as they increasingly use AI in marketing, customer service, internal communications and content production.
Knowing whether a student, job applicant or contractor produced work themselves or relied heavily on an AI system can affect how that work is assessed, for instance.
‘These guardrails are important because innovation cannot outpace trust,’ Google’s director of public policy for Sub-Saharan Africa, Charles Murito, told the Business Daily.
‘AI will have its upsides and downsides. Misinformation and disinformation are huge concerns, which is where these watermarks come into play.’
These text watermarks do not disappear when text is copied into another document, unlike metadata tags. They are built into the way the AI chooses words.
How watermarking works
Large language models generate text one word at a time, choosing the next word from several possible options. A watermarking system uses a secret key to subtly steer those choices toward a particular statistical pattern.
The result reads normally to a person, but a detector with the corresponding key can identify the pattern across a sufficiently long passage.
However, OpenAI says its watermark can be weakened through editing. Replacing about 10 percent of the words with synonyms reduced detection in its tests from about 92 percent to 66 percent.
‘Detection rates were substantially lower for content such as mathematics, where there is less flexibility in word choice,’ the US company added.
Another limitation is that there is no universal watermark detector. Anthropic can detect the watermark it puts into Claude-generated text, while Google can verify its own SynthID marks.
Still, experts warn that the presence or absence of a watermark should not automatically be treated as proof that a human or AI created a piece of content.
A watermark can be weakened or removed through editing, while content generated by systems without watermarking will not carry the same signal in the first place.
Kenyan push
For Kenyan businesses, the changes come as companies increase their use of generative AI while the country develops its own rules around the technology.
Kenya’s Artificial Intelligence Bill, 2026, is not as explicit on text watermarking as the EU framework. It requires AI systems that generate or manipulate images, voice or likeness to label their output as AI-generated, especially where there is a risk of harm, misinformation or infringement of rights.
The proposed framework also requires providers and deployers to disclose to users and affected people the nature, purpose and limitations of AI systems, the degree of automation involved, any human intervention and measures taken to mitigate bias and ensure fairness.
As global technology companies adapt their products to meet requirements in major markets such as the EU, Kenyan users are likely to encounter some of these features even where local rules do not yet impose the same requirements.
Analysts say companies that rely on trust — from banks and insurers to retailers, media organisations and professional services firms — need to know where AI has entered their communication workflows.
While an AI-generated policy explanation, customer response or marketing message may be legitimate, a company needs human review to ensure that the material is accurate and does not give customers a misleading impression about how it was produced.