ChatGPT is adding invisible watermarks

Happy Wednesday,

Today I'm showing you how to get more honest answers from ChatGPT, breaking down its new invisible text watermarks and whether the removal skill actually works, then turning your videos into animations with Higgsfield inside ChatGPT, and finishing with Claude's new Google Docs, Sheets and Slides update.

Let's get into it.

🧠 DO THIS TODAY

Stop ChatGPT From Agreeing With Everything

ChatGPT can take an idea you're unsure about and make you feel incredibly good about it. That's nice when you need encouragement. It's less useful when you're asking whether the idea actually makes sense.

A 2025 Stanford study called SycEval found that ChatGPT showed agreement-seeking behavior in about 57% of its tests. Researchers challenged it on math and medical questions, and that number included changes toward both right and wrong answers. It wasn't measuring how often today's ChatGPT agrees with something wrong. But it does show why you should check the reasoning when you've already told the AI what you want to hear.

You can give it clearer instructions to challenge weak assumptions, explain its objections and say when it doesn't know. Here's the setup, with a shorter version you can paste straight in.

Step 1: Open ChatGPT, click your profile, then go to Settings → Personalization → Custom Instructions. Enable customization if that switch appears, and paste this:

Keep any useful preferences you've already saved, then save the changes. On your phone, look for Settings → Customize ChatGPT. These are saved instructions you can edit later, separate from ChatGPT's memory. They guide its answers, but they don't make every answer correct.

Step 2: Now try the follow-up. Start a new chat and paste this. For the most useful answer, include a previous exchange you want it to review:

The useful part is the actual example and the reason behind the correction. Maybe it encouraged a business idea without asking whether anyone would pay, or backed your interpretation of a situation after hearing only your side. If it can't see enough context, paste the conversation. Don't push it to invent something shocking.

Try this on one decision you're making this week. You want an answer that helps you think more clearly, even if it's a little less flattering.

Get my full Truth Prompt guide and follow-ups →

🚨 AI UPDATE

ChatGPT Is Adding Invisible Text Watermarks

Remember the Claude watermarks I covered in August? Its newer models started adding invisible marks to their text, and people worried that using Claude to edit their writing would make the whole piece look AI-written. My concern was that getting help could be mistaken for having AI do all the work.

Well, inevitably, OpenAI has now announced its own text-watermark rollout. The big difference is that the ChatGPT launch is EU-only, with marks being added to eligible output over the coming weeks. Claude's supported models use them worldwide. But the bigger question matters wherever you live.

OpenAI uses its own method, called textGrain, to build a pattern into the words. When ChatGPT writes, several words can fit what it wants to say. textGrain subtly influences which one it picks.

Across a passage, those choices create a pattern OpenAI's detector can recognize. The writing still looks normal. There are no hidden symbols or extra spaces to delete. The words carry the signal, so copying them into another app carries it along too.

So what about the Watermarks Remover skill I shared for Claude? Does it work on this one? We don't have a verified yes for ChatGPT's new watermark. In theory, its rewriting step could weaken the pattern, but that isn't proof this skill removes it. Here's what it actually does:

Cleaning characters and files. It can remove certain invisible characters and identifying information attached to supported files. But textGrain doesn't use hidden characters, so this part alone won't remove it.

Rewriting the word pattern. The second part rewrites sentences to disrupt the signal while trying to keep the meaning. The project's documentation calls this best effort. It can also change your tone or lose details, so read the result carefully. The current full version needs extra setup beyond the original skill-file upload.

OpenAI's tests show that changing words can make its watermark harder to detect. That supports the idea, but doesn't prove this remover works. Its text detector is initially limited to approved researchers and expert organizations, so a score from an ordinary AI-writing checker can't settle this.

My concern is the same one I had with Claude: a watermark doesn't tell you how much of the work was yours. You can bring the ideas, write the draft and use AI to help edit it. OpenAI itself says a detected mark doesn't measure human contribution or establish ownership. Someone reading the result may assume a lot more than it proves.

That's dangerous if a teacher, client or employer treats a detection result as proof that you did none of the work. Someone using AI to polish their English could end up defending ideas they developed themselves. And detectors can make mistakes. A signal that needs context can quickly become an accusation, especially when the person judging it never sees your drafts or understands what the tool actually did.

Transparency can be useful. Treating every piece of writing an AI touched as though the person did none of the thinking is where I have a problem. Keep your original drafts and revision history if showing your contribution matters.

Read what OpenAI says its watermarks can and cannot prove →

SPONSORED BY HIGGSFIELD

ChatGPT Can Now Turn Your Videos Into Animations

If you make marketing videos, UGC-style ads or social content, you can get several different looks from one clip. Turn a product demo or a short video of yourself into an animated ad opening, a Reel or part of an explainer, without filming it all again.

You may have seen those videos where someone snaps their fingers and switches from real footage into a drawn character, then a completely different 3D style. Higgsfield's new Genjutsu Styles update gives you a gallery of animation looks to choose from right inside ChatGPT. I filmed one video and turned it into different styles. You just pick the look you want, so there's no animation prompt to write.

Step 1: Open the Higgsfield setup page, then install its plugin in ChatGPT. You can also find it under Plugins in ChatGPT's left sidebar. Search for Higgsfield, click Install plugin, and connect your account. If you already used it before this update, follow the reconnect instructions.

Step 2: Open the Higgsfield app inside ChatGPT and use its three-dot menu to pin it to your sidebar. That gives you a place to come back to, with the app's controls right there beside your chat.

Step 3: Choose Video, then Genjutsu, and upload your own reference video. Start with a short clip where you can clearly see yourself moving. A wave, a turn or a finger snap gives you an obvious moment to compare across versions.

Step 4: Open Styles, choose the animation look you want, check the generation cost shown, and hit Generate. Then repeat with the same original clip and a different style.

Watch the video →

Genjutsu uses the original movement and scene to guide the animation and retains the source audio. So you can try a drawn look and a rounded 3D look without filming the same performance again. Watch each result before using it, especially the hands, face and ending.

To make the switching effect, line up the versions in your video editor and cut from one to the next at the same gesture. Using the same source clip is what makes those transitions feel connected. You could use it for an opening to a Reel, a quick introduction or just a very different version of something you've already filmed.

Set up Higgsfield in ChatGPT and try Genjutsu →

💡 QUICK UPDATE

Claude Now Works Inside Google Docs, Sheets and Slides

One more from yesterday: Claude now works directly inside Google Docs, Sheets and Slides. Anthropic's new add-on puts it in a sidebar next to your file, where it can read what you're working on and make edits in place.

In Docs, that could mean tightening a paragraph. In Sheets, writing a formula or building a chart. In Slides, adding pages that follow the deck's existing design. The default is Ask before edits, so you can review changes before they're applied.

It's in public beta on all paid Claude plans. Install the add-on from the Google Workspace Marketplace, then open your Google Doc. Click Extensions in the top-left menu, then Claude → Open Claude. You'll see Claude open on the right side of your document, where it can read your file and use the app's tools to make changes. Your workplace may require admin approval.

There are also new Google Docs, Sheets and Slides connectors if you'd rather start in Claude and have it create or edit the Google files from there. Useful if you already do most of your work in Google's apps.

See the announcement and setup →

Try the Truth Prompt on one real conversation this week. The best result might be a useful correction, or it might be finding that your idea holds up after all.

See you Sunday.

Maverick

P.S. Want your business to actually show up inside AI? The brands that rank are seeing up to 3x the conversions. Get your free AI visibility score and see exactly where you stand inside AI.

P.P.S. If you're new here or you ever skip one, I've got a newsletter archive of all my past emails you can catch up on anytime.