New · See every ThinkChat feature in action. Take a look →
ThinkChat
Start free trial

Knowledge AI

Can anyone tell your text came from ChatGPT? What OpenAI’s new watermark actually does

AI 3 min read

On 5 October 2026, OpenAI said it will add an invisible watermark to text written by ChatGPT and Codex, starting with users in the European Union. It’s OpenAI’s response to the EU AI Act, which requires AI providers to make AI-generated text identifiable by machines.

Many small businesses now use AI to write product descriptions, social posts and customer replies. So a fair question is: can someone now tell that your text came from ChatGPT? The short answer is “sometimes, and not reliably.” Here’s what OpenAI actually announced.

What OpenAI announced

  • ChatGPT and Codex in the EU only. Over the coming weeks, an invisible watermark is being added to eligible ChatGPT and Codex text for users in the EU, on all plans. OpenAI says it is not making this a global default at launch.
  • Optional for developers everywhere. From 5 October, businesses using OpenAI’s API anywhere in the world can choose to turn watermarking on for some models. It stays off by default.
  • The checker is not public. OpenAI’s watermark detector is, for now, only open to approved researchers and expert organisations. You can’t paste text into a public website to check it.

How the watermark works

OpenAI calls the technology textGrain. It doesn’t add a visible label or hidden characters. Instead, it slightly shapes which words the AI picks, leaving a statistical pattern a person can’t see but a detector can look for. OpenAI says the watermark made no meaningful difference to the quality of its model’s answers in its tests.

Why it’s easy to weaken

OpenAI is open about the limits. In its own tests:

  • Short text is harder to detect. At a 1% false-positive target, for content such as psychology, the detector found the watermark in about 80% of 200-token passages (about 150 words by OpenAI’s rough rule of thumb for English), compared with about 95% of 400-token passages. Detection was substantially lower for content such as mathematics, where there is less flexibility in word choice.
  • Editing weakens it. In 400-token passages, swapping 10% of the words for synonyms cut detection from about 92% to 66%. Swapping 25% cut it to 17%.

That’s why OpenAI is keeping the detector limited for now: it can miss watermarks, and it can also wrongly flag text that has none.

What a watermark can’t tell you

OpenAI lists several things a detection result does not prove:

  • It doesn’t show how much a human edited or added.
  • It doesn’t say who owns the text or who is responsible for it.
  • It doesn’t identify the user, their account or their prompts.
  • It doesn’t say whether the text is true.
  • No watermark doesn’t mean a human wrote it. The text may be short, edited, translated, or made by another company’s AI.

What this means for a small business in Malaysia

1. Nothing changes overnight for ChatGPT users here. The ChatGPT watermark is for EU users. If you sell to customers in Europe or work with EU partners, it’s worth knowing about.

2. “Undetectable” isn’t the goal. Accuracy is. A watermark only says whether AI was involved. It doesn’t check facts. A wrong price or a made-up promise hurts you whether a human or an AI wrote it.

3. Be honest about AI with customers. Watermarks are one sign of a wider push for AI transparency. A simple habit is to let customers know when they’re chatting with an AI assistant, and give them an easy way to reach a person.

4. Check before anything goes out. Read AI-written posts and replies before you publish them. For customer chats, the AI should answer only from information you approved, and say “I don’t know” when it doesn’t know.

Where ThinkChat fits

ThinkChat is an AI Chat Agent and multi-channel AI customer service platform. WhatsApp is the main channel in Malaysia, and ThinkChat also works on Instagram, Facebook Messenger, Telegram, website chat and more. It is built to answer customers from information you approved, and hand over to a person when a customer needs one.

Soft next step

Want an AI Chat Agent to answer your customers 24/7? Talk to Lyia on WhatsApp: +60 11-5515 9317.

Sources

Checked 27 October 2026.

← All articles