OpenAI has announced that it will add an invisible watermark to text generated by ChatGPT and Codex in the EU, according to reporting by Search Engine Journal. The company also plans to offer an opt-in approach for its API. Reported testing indicates that editing the output can weaken the watermark, which is an important limitation for businesses considering how such markers may be used in practice.

What happened

According to the report, OpenAI plans to introduce invisible watermarking for text produced by ChatGPT and Codex for EU users. The watermark is intended to act as a hidden marker within generated text rather than a visible label shown to readers.

The report also says OpenAI will make watermarking available as an opt-in for API users. That distinction matters because many businesses use OpenAI models through APIs inside websites, customer service tools, internal assistants, content systems or software products rather than only through the standard ChatGPT interface.

A key point in the report is that OpenAI's own tests indicate the watermark can be weakened if the text is edited. That means the feature may help with provenance and internal governance in some cases, but it is not a guarantee that AI-generated text can always be identified later.

Why it matters for European businesses

For European companies, the development is relevant less as a marketing feature and more as an operational and governance issue. Many businesses are already using generative AI for website copy, product descriptions, support drafts, code generation, internal documentation and campaign production. If a model provider adds watermarking in one region, businesses may need to understand where it applies, how durable it is and whether it changes existing review or disclosure processes.

The announcement also highlights a wider issue for AI adoption in Europe: traceability is not the same as compliance. A watermark may support internal controls or help teams distinguish AI-assisted output in some workflows, but it does not automatically satisfy legal, regulatory or contractual obligations. Businesses should avoid assuming that watermarking alone is enough for transparency, authorship checks, copyright review, accuracy control or sector-specific compliance.

There are also practical implications for organisations operating across regions. If watermarking is applied in the EU but not everywhere else, companies with multilingual or cross-border content operations may end up with inconsistent governance rules depending on where content is generated and which tools are used.

For marketing and publishing teams, the weakening of the marker after editing is especially important. Most business content is edited before publication, translated, rewritten for tone, adapted for SEO or combined with human-written material. In these real-world workflows, an invisible watermark may be less reliable as evidence that text was generated by AI.

Who may be affected

• Marketing teams using ChatGPT or API-based tools for blogs, landing pages, ad copy, email campaigns or product content.
• Software and IT teams integrating OpenAI models into internal tools, support systems, search features or coding workflows using Codex or related services.
• E-commerce businesses generating large volumes of catalogue text, translations or customer support content.
• Founders and SME leaders setting AI usage policies and deciding how to document AI-assisted work.
• Regulated businesses that need stronger controls around provenance, record-keeping, human review or content approvals.

What companies should consider

• Review where OpenAI tools are used across your organisation, including direct ChatGPT use and API-based integrations embedded in third-party software.
• Do not rely on watermarking as a sole detection method. If OpenAI's reported testing shows edits can weaken the marker, businesses should treat it as a limited signal rather than a definitive proof mechanism.
• Update AI governance policies to clarify when teams may publish AI-generated or AI-assisted content, what human review is required and how outputs should be documented internally.
• Check vendor and workflow differences. If some tools apply watermarking and others do not, content provenance controls may become inconsistent across departments and markets.
• Assess API settings carefully if your business builds on OpenAI's API. An opt-in model means technical and compliance teams may need to decide deliberately whether watermarking should be enabled in each use case.
• Separate transparency from authenticity. Watermarking can be part of a broader governance approach, but it should sit alongside editorial review, logging, access controls and quality assurance.

At this stage, the reported change is best understood as an infrastructure and governance update rather than a complete solution to AI disclosure or detection. For European businesses adopting generative AI at scale, the practical question is not only whether text is watermarked, but whether internal policies, publishing workflows and risk controls are ready for mixed human and machine-generated content.