AI detectors, watermarks and what they prove

What it is good at

• Explains clearly why detector percentages are not reliable evidence, which prevents both wrongful accusations and false reassurance.
• Introduces content credentials (C2PA-style provenance metadata) as the more durable alternative to guessing.
• Gives practical steps for protecting your own authorship record, not just spotting others' AI use.
• Covers what happens to provenance metadata when a file is edited, compressed or re-uploaded, which most guides skip.

Where it falls short

• No detection method, including provenance metadata, is universally adopted yet, so absence of a watermark proves nothing either way.
• Metadata and content credentials can be stripped, deliberately or accidentally, by common editing and social platforms.
• Accusing someone of AI use based on a detector score alone can be unfair and, in some contexts, legally risky, treat scores as a prompt to ask, not a verdict.

How to use it for creative work

1. 01

   Stop treating detector percentages as proof

   If you use an AI-detection tool, treat its output as a rough signal at best. Human-written text is regularly flagged as AI-generated and vice versa, so a single score should never be the sole basis for a serious decision.

2. 02

   Understand what content credentials actually verify

   Standards such as C2PA attach a tamper-evident record to a file noting what tool created or edited it. This tells you provenance, when preserved, but does not itself judge whether the content is "good" or "true."

3. 03

   Check whether your tools preserve or strip this metadata

   Some generation tools embed content credentials by default; many editing tools and most social platforms strip metadata on upload or export. If provenance matters for your work, check each step of your pipeline, not just the source.

4. 04

   Build your own authorship record

   For work where your authorship might later be questioned, keep your own record: source files, layered project files, timestamps and screenshots of your process, independent of any platform's watermarking.

5. 05

   Ask, don't accuse, when a detector flags something

   If a detector flags a piece of work as likely AI-generated, treat it as a reason to have a direct conversation, not as settled proof, given how often these tools are wrong in both directions.

Ready-to-use prompts

Explain, in plain English for someone with no technical background, what content credentials or C2PA metadata actually record about an image or video, and what they do not prove.

I want to be able to prove I made [type of work] myself if it is ever questioned. List the specific files, timestamps and process records I should keep as I work, for [design software / writing tool] specifically.

An AI-detection tool has flagged this piece of writing as [percentage]% likely AI-generated. Explain what that percentage does and does not tell me, and what independent checks I could still do before drawing a conclusion.

I create images with [tool] and then edit them in [tool] before posting to [platform]. At which of these steps is embedded provenance metadata most likely to be stripped, based on how these kinds of tools generally handle metadata?

I need to raise, calmly and fairly, a concern that a piece of submitted work might be AI-generated without proper disclosure. Write a short, non-accusatory message that asks about it directly rather than stating it as fact.

Questions people ask

Can I trust an AI-detection tool's percentage score?

Not as a standalone verdict, these tools produce meaningful rates of both false positives and false negatives, so a score is a prompt for further checking, not proof of AI or human authorship.

What is C2PA and does every tool support it?

It is a technical standard for attaching tamper-evident provenance information to media files; adoption is growing but far from universal, and the metadata can be stripped by editing or by many upload platforms.

If I accuse someone of using undisclosed AI based on a detector, could that cause a problem?

Potentially, especially in an academic, employment or contractual setting, treat a detector result as a reason to ask questions and gather more evidence, and get institutional or legal guidance before making a formal accusation.