Safety & Design
AI watermarking
A hidden or visible mark added to AI-generated text, images, or audio so a tool can later tell that the content was made with AI.
- Businesses
- Policymakers and staff
What parents should know
An AI watermark is a mark that says a machine helped make a file. Some marks are visible. Some are hidden patterns a detector looks for. Marks get cropped, edited, and skipped by tools that never added one. A missing watermark does not prove a student cheated, and a present one does not end the conversation.
On this page
What is AI watermarking?
Visible watermarks are labels you can see, like a corner badge. Hidden watermarks are patterns in the pixels, the audio, or the choice of words. The goal is the same as content provenance: a later check. The methods fail in different ways. A badge can be cropped. A hidden pattern can be too faint after a screenshot. A text watermark can vanish when a student edits the paragraph.
The EU AI Act includes marking duties for some generated content, with dates that have already shifted once. As of the Commission's timeline after the 2026 omnibus, do not quote an old August 2026 date for every watermark rule without checking. A school policy should not wait on Brussels to ban detector-only discipline.
Why AI watermarking matters
Vendors sell watermarks and detectors as certainty. Families experience them as accusations. The honest use is a clue beside other clues: the draft, the prompt transcript, the student's explanation. UNESCO's student guidance is about competence and judgment, not about a mark.
Businesses that generate ads or support docs may want a mark so they can disclose AI use. That is a labeling choice. It is not, by itself, a child-safety program.
How it shows up in practice
- A publisher badges generated images, and a screenshot on social media loses the badge.
- A school refuses to fail a student based only on a watermark detector.
- A class labels generated art in the caption, which survives better than a hidden mark.
- A policy team rechecks the EU date before it writes a compliance slide.
How HeyOtto helps
HeyOtto does not claim a hidden watermark on Otto's replies or images, and it does not offer a detector that reads other people's watermarks. Honesty in schoolwork is handled in the open: homework help shows the steps, and a teacher or parent can read the session. If your school needs a visible label, put it in the assignment rules. Do not treat the absence of a HeyOtto watermark as evidence that a child did not use AI, or that they did.
- The transcript is the record HeyOtto actually keeps for families and schools.
- AI detection scores are not a HeyOtto feature.
- A business that needs a different labeling setup should say so in a partnership conversation, not assume a watermark exists.
For families
Try freeFAQs
If I cannot find a watermark, did a human write it?
No. Many AI tools add no mark, and many marks do not survive a copy. Human writing also has no AI watermark. Use drafts and a conversation. See AI detection tools for why a score is not proof.
Are watermarks required yet?
Some EU transparency and marking duties are phased, and the 2026 omnibus moved at least one watermark date for systems already on the market. Check the European Commission AI Act page before you cite a deadline. US schools are not waiting on that date to set cheating rules. This is general information, not legal advice.
Is a visible label better?
For schoolwork, yes. A sentence the reader can see beats a hidden pattern a vendor can miss. HeyOtto's session view is in that spirit. The help is visible to the adult who is supposed to judge the work.
Can I rely on a watermark to spot a deepfake?
Not alone. Assume a harmful clip might have no mark. Call the person, avoid forwarding intimate fakes, and use NCMEC when a minor is pictured. A watermark is a bonus clue, not the plan.
Sources
Last reviewed September 26, 2026. This entry is reviewed twice a year.