Right now, nobody knows who used AI to write what.
A resume lands in an inbox. A cover letter reads a little too polished. An essay comes back with perfect structure. Somewhere behind it, AI may have done some or all of the work. Nobody can tell. Not the employer. Not the lecturer or teacher. Not the person reviewing the application.
This is not a small gap. It is the biggest blind spot in hiring and education today.
The problem isn’t AI. It’s the silence around it.
AI writing tools are not going anywhere. Used well, they help people communicate more clearly, prepare faster, and present with confidence. That is a good thing.
The problem is that almost none of this use is visible. There is no standard for disclosure. No shared signal for “this was AI-assisted” versus “this was AI-generated” versus “this is entirely human.” Everyone is guessing.
That silence has real costs.
Employers can’t tell if a cover letter reflects the candidate or the chatbot. Universities can’t tell if an essay reflects the student’s thinking or a prompt they wrote. And honest candidates, the ones who use AI responsibly and transparently, get lumped in with the ones who don’t disclose anything at all.
Unverified use looks exactly the same as responsible use. That’s the real failure.
A real step toward visibility
This week, something relevant happened at the infrastructure level.
Anthropic, the company behind Claude, announced it will begin applying a text watermark to outputs from Claude Opus 5, starting September 9, 2026. The watermark is an invisible statistical pattern in word choice. It doesn’t change what the AI writes, doesn’t add extra characters, and doesn’t slow anything down. It exists purely so that AI-generated text can be identified later if needed.
This isn’t Anthropic acting alone by choice. It’s a response to the EU AI Act’s transparency requirements, which apply to every major AI provider serving the EU market. New models launched after August 2026 carry the watermark from day one. Older models, including Opus 5, are being updated to carry it too, ahead of a December deadline. Because the mark sits at the model layer, it follows the model everywhere it’s served, including through major cloud platforms.
This is a genuinely useful development. For the first time, there’s a technical foundation that makes some AI-generated content traceable, not just guessable.
But a watermark on one model from one provider is not a system. It won’t tell an employer whether a candidate understands AI or just used it to produce a polished paragraph. It won’t tell a university whether a student can think critically about a tool they are oeprating with no safe guidelines. Detection is not the same as capability.
That’s the gap GAFI® (Gradstack AI Fluency Index) is built to close.
Where this needs to go
Watermarking is a good sign that the industry is moving toward transparency. It shows that even the biggest AI labs recognise that invisible AI use is a problem worth solving at the infrastructure level.
But hiring and education need more than detection. They need a standard that measures how someone uses AI, not just whether they did. Responsible use should be something people can prove and be recognised for, not something they have to hope goes unnoticed.
That is the shift Gradstack is building toward. Verified capability. Verified fluency. A clear, honest signal that replaces guesswork with proof, for every pathway into work, whether that’s university, TAFE, a bootcamp, or a career change.
The tools are catching up to the problem. Now the human verification standard needs to catch up too.