01What Are ChatGPT Watermarks and How Do They Work?

ChatGPT watermarks are invisible statistical signals added to AI-generated text so that detectors can identify whether the text was likely produced by an OpenAI model. OpenAI is rolling out this feature to comply with new EU regulations, specifically the EU AI Act.

In this episode, Julian Goldie explains that the watermark does not add visible marks or special characters. Instead, it embeds a pattern in the word choices of the model. A detector can then look for that pattern to determine if the content came from ChatGPT or Codex.

The technology is called text grain. It works by coupling the generation of tokens from a secret key and the preceding text. Crucially, it does not identify who the user is, what account or conversation generated the text, or which company owns it.

The watermark also does not slow down the models. OpenAI says that it is an ongoing area of research, and only approved researchers will have access to the detector for now.

02Key Features of OpenAI's Watermark System

OpenAI's watermark system comes with several important features that differentiate it from other AI providers. Here are the main points from the episode:

  • Option to choose – API customers can turn text watermarking on or off for select models right now.
  • No visible marks – The watermark is invisible and does not alter the text visually.
  • No impact on speed – The models do not slow down when watermarking is enabled.
  • No ownership identification – It does not reveal who created the content, the company, the account, or the prompt.
  • Limited detection for short text – Passages under 200 tokens are often undetectable.
  • Not open to the public – Only approved researchers can access the detector for evaluation purposes.

In contrast, Anthropic's Claude watermarks, which were released earlier, are automatically applied with no choice for users. OpenAI gives users more control, especially for those using the API.

Additionally, OpenAI already has verification tools for audio and images. The new text watermarking is part of a broader content provenance initiative.

03How to Set Up or Avoid ChatGPT Watermarks

If you want to manage ChatGPT watermarks, here are the steps you can take based on the episode:

  1. Use the API – If you are an API customer, you can enable or disable text watermarking for select models at any time.
  2. Keep content short – For shorter passages, like posts on X, detection is unlikely because there is not enough text to form a clear pattern.
  3. Edit your content – Replacing some words with synonyms weakens the watermark signal significantly.
  4. Use another AI for editing – Running content through a different AI model, like Claude, can further reduce detectability.
  5. Check if you are in the EU – The watermark is currently only applied to text output in the EU. Users outside the EU may not see it.

Julian emphasizes that these steps are for those who want to avoid detection. However, he also notes that watermarks are not the main concern right now, because detection is not highly accurate.

04How to Use the Watermark Detector and Verify Tool

OpenAI has made a watermark detector available, but only for approved researchers. If you are a researcher, you can apply for access to help evaluate and improve the technology.

For regular users, OpenAI already offers a content provenance API and a verification tool at OpenAI.com/verify. You can use this to check if text, audio, or images were AI-generated.

Julian demonstrates the summarize feature in Google to get a quick summary of the technical report. He suggests running the report through another AI, like Claude, to better understand the complex mathematics behind text grain.

The detector works by using the generated text, a secret key, and the context window. It does not require access to the model itself, which makes it efficient. However, it cannot measure human contribution, establish ownership, or verify accuracy.

05Test Results: Detection Rates and Limitations

OpenAI's own tests reveal significant limitations in watermark detection. Here are the key numbers from the episode:

  • Short vs. long passages – At a false positive rate of 1%, the detector identified watermarks in about 80% of 200-token passages, but in about 95% of 400-token passages.
  • Editing impact – Replacing 10% of words with synonyms reduced detection from 92% to 66%.
  • Heavy editing – Replacing 25% of words dropped detection to just 17%.
  • Content type matters – More logical or factual content, like math, is harder to detect than creative content like psychology.

Julian points out that these numbers show the watermark is not highly accurate right now. He expects it to improve over time, but for now, editing content can make a big difference.

06Tips, Limits, and Mistakes to Avoid

Based on the episode, here are practical tips and warnings for dealing with ChatGPT watermarks:

  • Don't panic – The watermark technology has many limits and is not highly reliable at the moment.
  • Edit your content – Replacing synonyms, especially 25% or more, significantly reduces detectability.
  • Keep it short – For posts on X or brief comments, detection is unlikely due to lack of text.
  • Use the API if you want control – API users can turn watermarking on or off.
  • Don't worry about Google – Google does not have access to OpenAI's or Anthropic's watermark patterns, and Google's guidelines focus on quality, not AI detection.
  • Remember the absence of a watermark is not proof of human writing – The detector cannot confirm human authorship.

Julian warns that trying to avoid watermarks entirely may not be necessary, as the technology is still being refined and may not affect SEO rankings.

07Who Should Care About ChatGPT Watermarks?

ChatGPT watermarks are relevant for anyone who creates content with AI, especially for SEO or publishing. If you use ChatGPT to write articles, social media posts, or other text, you might wonder if your content will be flagged.

Julian explains that for SEO purposes, you don't need to worry much. Google does not care about AI-generated content as long as it is high-quality and fact-checked. Google updated its guidelines to emphasize quality control, but it doesn't penalize AI content automatically.

For users in the EU, watermarks will be added to ChatGPT and Codex text output in the coming weeks. If you are outside the EU, you may not see them yet. API users have the flexibility to opt out.

In short, if you are a content creator, marketer, or SEO professional, understanding watermarks helps you make informed decisions. Julian advises staying ahead by learning how the system works, but not overreacting because current detection accuracy is limited.

08Additional Tools: OpenAI's Content Provenance API

OpenAI is not just adding watermarks; it is also providing a content provenance API. This API helps detect AI-generated content across different types, including text, audio, and images.

In the episode, Julian highlights that you can use the OpenAI.com/verify tool to check if content was AI-generated. This is useful for researchers, publishers, and platforms that want to know the origin of content.

The content provenance API is part of OpenAI's broader effort to comply with EU rules. It allows third parties to integrate detection into their own systems, which could be valuable for content moderation or authenticity verification.

While the public doesn't have access to the text watermark detector yet, the verification tool offers a starting point for those curious about AI provenance.