Hi - I answer from the OpenSmartRoute documentation: routing, the API, plans and quotas, self-hosting. Ask away, or open a support ticket if you need a person.
Grounded in the docs - follow a source before acting on it.
OpenAI Adds Invisible Watermarks to ChatGPT Text in the EU - OpenSmartRoute
OpenAI Adds Invisible Watermarks to ChatGPT Text in the EU
OpenAI will add invisible watermarks to ChatGPT and Codex text in the European Union starting this month. The company also released a technical report on its method called textGrain.
Key points
OpenAI rolls out invisible watermarks for ChatGPT and Codex in the EU only.
The watermark works by subtly shaping word choices rather than adding symbols.
Replacing 10% of words drops detection accuracy from 92% to 66%.
Short passages, math answers, and translated text are harder to detect.
Why it matters: Users can verify if their text was generated by OpenAI systems without seeing a symbol.
By OpenSmartRoute editorial · written through the router by writer-small
From TechCrunch AI - “OpenAI will start watermarking ChatGPT’s text in the EU”
OpenAI will start adding an invisible watermark to text generated by ChatGPT and Codex in the European Union. The company said this happens to comply with the EU AI Act. This law requires AI companies to mark AI-generated content. Other systems must be able to identify that mark. OpenAI rolled out the change starting this month.
The rule took effect on August 2 last year. It applies to all AI content in Europe now. Users see no visible symbols or icons on their screens. The mark lives inside the words themselves instead. A detector can pick up the pattern later. Readers cannot see the difference when they read normally.
Developers using OpenAI's API anywhere can turn it on today. They must select specific models to apply the watermark. It is off by default for everyone else right now. OpenAI will not make this a global default at launch. This means only EU users get the feature initially.
The company plans to roll out the feature over coming weeks. It covers eligible ChatGPT and Codex users on all plans. The rollout happens only in the European Union region. Users outside Europe do not see this change yet. OpenAI keeps the decision for global rollout separate from this launch.
How the Invisible textGrain Watermark Works
OpenAI calls its method textGrain. It is a technical report published alongside the announcement. Researchers from the University of Pennsylvania and Yale co-wrote the paper. They explain how the watermark shapes word choices subtly. The model predicts the next word in a sentence differently. This creates a hidden pattern in the output text.
The pattern travels when users copy and paste text. It lives inside the words themselves, not as an extra tag. A detector needs the secret key to find the pattern. Without the key, the text looks normal to human eyes. The system does not add visible icons or symbols to the page.
OpenAI says the watermark does not identify the user directly. It marks that a model generated part of the passage. The company saw no meaningful change in performance with it on. Users do not notice slower speeds or lower quality output. Engineers can run tests to check for these subtle shifts.
The secret key helps sort next-word predictions during generation. Adding hundreds of nudges makes the pattern detectable. A detector uses only the text and the key to spot AI content. It does not need access to the original model files. This keeps the method practical for external researchers to use.
OpenAI uses a secret key to configure the watermarking process. The key tells the model which words to nudge slightly. It changes the probability of certain words appearing next. These small shifts accumulate across many sentences in a paragraph.
The sorting happens before the final text is sent to users. The model still generates fluent and natural-sounding English. Humans cannot tell the difference without running a detector. OpenAI keeps the exact key details confidential for now. They want to protect the integrity of the detection system.
Researchers can access detectors if they are approved by OpenAI. The company limits initial access to expert organizations only. This helps them evaluate reliability and responsible uses of the tool. OpenAI wants to prevent bad actors from exploiting the system.
The method relies on statistical patterns rather than unique identifiers. Each model run creates a slightly different text grain. No two generations are exactly identical in their hidden structure. This makes it hard to reverse engineer the exact key easily.
Limitations of Detection and Editability
OpenAI tested removing the watermark by editing the text. Replacing 10% of words with synonyms dropped detection from about 92% to 66%. The company admits that some edits can hide the mark effectively. This suggests users might try to clean up AI content before sharing it.
Short passages are harder for detectors to find reliably. Math answers and translated text also resist detection well. OpenAI noted these limitations in their technical report clearly. They do not claim perfect detection is possible yet.
The company decided to limit detector access due to these risks. They want approved researchers to help evaluate the system properly. Unapproved users might misuse the detectors for other purposes. OpenAI prioritizes safety alongside transparency in this rollout.
A missing watermark does not prove human authorship automatically. The text could be too short or heavily edited by humans. It could also come from another company's AI model entirely. OpenAI warns against assuming all unmarked text is human-written.
Comparison with Anthropic's Global Approach
Anthropic announced it would watermark text generated by Claude worldwide. This decision came two months after OpenAI's EU announcement. Some Claude users argued they supplied the instructions and context themselves. They felt Claude was just a tool in their workflow.
OpenAI had built a text watermark before but held off on releasing it. The Wall Street Journal reported concerns that users would switch to rivals. Those rivals did not have watermarks at the time. OpenAI feared losing customers if they revealed the feature too early.
Anthropic applied its watermarking globally instead of regionally like OpenAI. This drew different reactions from their user base and partners. Both companies are among those committed to following the EU's code of practice on AI-generated content. The EU Act sets a baseline for transparency across member states.
Other tech giants have also committed to similar codes of practice. Google, Meta, Microsoft, and OpenAI all signed up recently. They aim to standardize how they mark their AI outputs. This creates a shared language for detecting synthetic content globally.
Why it Matters for Model Runners and Agents
Model runners need to know if text is AI-generated for compliance reasons. The EU AI Act mandates marking AI content in many sectors. Companies face fines if they fail to comply with these new rules. OpenAI's watermark helps them meet this legal requirement easily.
Agents building on OpenAI models might inherit the hidden patterns automatically. Developers do not need to manually add tags to their applications. The watermark travels with the text when it is copied and pasted. This ensures downstream systems can still detect the origin of the content.
Engineers should test their own pipelines for detection capabilities soon. They need to know if their tools can read the secret key format. Some third-party detectors might not support the new textGrain method yet. OpenAI will provide access only to approved researchers initially.
Safety teams must update their verification protocols to handle this change. False negatives become more likely when content is edited or short. Teams need to balance trust with skepticism when evaluating AI outputs. They cannot rely solely on the presence of a watermark for final decisions.
OpenAI Announces EU Watermarking for ChatGPT and Codex
The company will add an invisible watermark to text from ChatGPT and Codex in the European Union. This follows the EU AI Act transparency rules that took effect on August 2. The rules require AI companies to mark AI-generated content in a way other systems can identify.
OpenAI said the watermark will roll out over the coming weeks. It will reach eligible ChatGPT and Codex users on all plans, but only in the EU. Developers using the OpenAI API anywhere in the world can turn it on for select models starting today. It is off by default, and OpenAI is not making text watermarking a global default at launch.
The watermark is not an actual symbol. It works by subtly shaping the model’s word choices. This leaves a pattern readers cannot see, but a detector can pick up. Because it lives in the words themselves, it travels with the text when it is copied and pasted. OpenAI said the watermark does not identify the user. The company saw no meaningful change in its models’ performance with it switched on.
OpenAI also published a technical report for its method, called textGrain. Researchers from the University of Pennsylvania and Yale co-wrote it. The report walks through an example of using a secret key to sort next-word predictions. The model nudges the next word to fit a pattern. Add hundreds of these nudges together, and the detector can spot AI-generated content using only the text and the key.
Can the watermark be removed by editing? OpenAI’s tests suggest yes. In one test, replacing 10% of words with synonyms dropped detection from about 92% to 66%. The company also said short passages, math answers, and translated text are harder to detect.
“These limitations contribute to our decision to provide initial detector access only to approved researchers and expert organizations,” said the company.
OpenAI also cautioned that a missing watermark “does not prove human authorship.” The text could be too short or too heavily edited, or it could come from another company’s AI.
“[Watermarks] can indicate that an OpenAI system generated or processed part of a passage, but not how much human judgment, editing, or creativity went into it,” the company said.
The announcement comes two months after Anthropic said it would watermark text generated by Claude, a move it’s applying worldwide. That decision drew backlash from some Claude users. They argued they had supplied “the instructions, context, decisions” while Claude was just “the tool.”
OpenAI had built a text watermark before but held off on releasing it. The Wall Street Journal reported in 2024 that they did so partly over concerns that users would switch to rivals that didn’t watermark.
Anthropic, Google, Meta, Microsoft, and OpenAI are among the companies that have committed to following the EU’s code of practice on AI-generated content.
What to Do When Evaluating AI Content
Readers should check if their detectors support textGrain before relying on them. OpenAI has not released public documentation for the secret key format yet. Engineers might need to wait for official API updates or SDKs. Third-party tools may struggle to parse the new pattern correctly.
Evaluate short passages and math answers with extra caution during testing. These content types are harder to detect according to OpenAI's own tests. Do not assume a missing watermark proves human authorship automatically. Always cross-reference multiple signals when making final judgments on content origin.
Contact approved researchers if you need help evaluating the new watermark system. OpenAI will provide access only to expert organizations initially. This ensures the tool remains safe and reliable for everyone using it. The company wants to avoid misuse of their detection technology.
Try running your own models with similar constraints to see how they behave. Compare performance metrics between watermarked and non-watermarked generations. Look for any subtle shifts in word choice or sentence structure that might indicate AI generation. This hands-on testing helps build internal expertise on the new standard.
OpenAI is rolling out textGrain watermarking for users in the European Union. API customers worldwide can opt into watermarked text outputs starting today.
