OpenAI has announced that it will begin watermarking text generated by ChatGPT and Codex in the European Union. According to the company, the change is intended to support compliance with the EU AI Act. OpenAI also said that editing generated text can make these invisible marks harder to detect, which is an important limitation for businesses relying on AI output controls.

What happened

According to the reported announcement, OpenAI plans to apply invisible watermarking to text created by ChatGPT and Codex for EU users. The company has presented this as a compliance-related measure linked to the EU AI Act.

Based on the information available, this concerns text output rather than visible labels shown directly to end users. OpenAI also indicated that watermark detection may become less reliable after the text is edited, rewritten or otherwise changed.

At this stage, businesses should distinguish between an announced product or policy change and a general legal conclusion about all AI-generated text. The source reports OpenAI's intended implementation and rationale, but it does not establish that watermarking alone is sufficient for compliance in every business context.

Why it matters for European businesses

For European companies, the announcement matters because many teams already use generative AI for marketing copy, internal communications, software development, customer service drafts and knowledge work. If a major AI provider changes how outputs are tagged or traceable in the EU, that can affect governance, procurement and documentation practices.

Watermarking may help businesses in several ways:

• Internal traceability: teams may have an additional technical signal to help identify whether text originated from an AI system.
• Governance and risk management: organisations building AI use policies may be able to use watermarking as one supporting control when reviewing how AI is used in content and coding workflows.
• Compliance readiness: companies preparing for AI-related regulatory obligations may see watermarking as part of a broader evidence and accountability approach.

However, OpenAI's own caveat is commercially important: if normal editing can weaken detectability, companies should not treat watermarking as a guaranteed or standalone control. In practice, edited marketing content, support responses, translated text or code comments may become harder to classify as AI-generated later in the workflow.

This is especially relevant for businesses that need reliable records of how content was produced, whether for quality assurance, contractual requirements, regulated processes or internal review.

Who may be affected

• Marketing teams: businesses using ChatGPT for campaign copy, SEO drafts, email content or social media text may need to review approval and disclosure workflows.
• Software and IT teams: organisations using Codex or similar tools for development support may want to assess how AI-assisted output is tracked in coding processes and documentation.
• SMEs adopting AI: smaller businesses that rely on consumer-style AI tools without formal governance may need clearer rules on when AI-generated content can be published or reused.
• Regulated businesses: companies in sectors with stricter documentation, accountability or audit expectations may need stronger evidence than watermarking alone.
• Agencies and service providers: digital agencies and outsourced content teams serving EU clients may need to clarify whether deliverables include AI-generated material and how that is managed.

What companies should consider

• Review AI usage policies: define where generative AI can be used for content, code, research or support workflows, and who is responsible for human review.
• Do not rely only on watermarking: if AI-origin tracking matters to your business, combine technical controls with process controls such as logging, version history and staff declarations.
• Check vendor documentation: businesses using OpenAI tools directly or through third-party platforms should monitor product documentation and contract terms for changes affecting EU usage.
• Assess customer-facing disclosures: where relevant, review whether users, clients or partners should be informed when material is AI-assisted, especially in sensitive or regulated contexts.
• Prepare for broader AI governance: watermarking can support oversight, but it does not replace risk assessment, content review, data protection analysis or legal review where these are required.

For European businesses, the practical takeaway is that AI output traceability is becoming a more operational issue, not just a technical one. Companies using generative AI at scale should expect more provider-level compliance features, but they should also assume that internal governance will still be necessary when AI-generated text is edited, redistributed or embedded in business processes.