OP-ED: Enter, the age of provenance — or the week AI started tagging itself

AI models and platforms now tag and detect AI-generated content, forcing us to define authorship, disclosure and where human judgement must prevail.

Something’s changed in artificial intelligence (AI). You will love it or loathe it. The world appears somewhat divided on this one.

Anthropic ⁠ says new Claude models launched in the European Union (EU) on or after 2 August 2026 will support machine-readable marking from launch. That will fan out globally. Anthropic says greater transparency via provenance data, revealing where content comes from, can give people useful context, while also helping it meet its commitments under the EU AI Act .

Substack ⁠ has introduced “Scan for AI text”, powered by Pangram, which estimates how much eligible writing is human-written or AI-assisted. Creators can also add a “How I make this” statement explaining their use of AI.

Substack put its position rather neatly: “The problem isn’t AI, the problem is when you pretend something written by AI was written by a person.”

LinkedIn ⁠, meanwhile, is reducing the wider distribution of low-effort AI-generated material that lacks genuine perspective or substance. Importantly, LinkedIn is not rejecting AI-assisted writing. It says AI can help refine language, but that posts still need to represent the author’s own voice and perspective. Its conclusion is simple: “The ultimate value comes from the human behind the tool.”

These are three different interventions. Anthropic is working on provenance. Substack is introducing detection and disclosure. LinkedIn is trying to suppress generic synthetic content.

After several years dominated by generation and adoption, we appear to be entering the age of provenance. And that creates a much harder question. What exactly counts as human work?

If the original thought is mine, but AI improves the structure, who created the work?

If I develop the argument, build the framework and use AI to improve the language, where exactly does authorship sit?

I hate making PowerPoint presentations. If AI turns my thinking into a better-looking presentation because I would rather not spend six hours aligning boxes, has the machine created my work, or packaged it?

The client may not care. This week I presented my idea within a Claude deck to one of the largest brands on the planet. I told them it was a Claude deck, personally the idea was so robust it didn’t matter what wrapping it came in. They wanted a great idea, so I simply chose to box it up for presentation with speed. They loved it – they wanted the strongest thinking presented clearly, rather than an invoice reflecting my endurance and struggle with PowerPoint.

That exposes something AI is forcing knowledge workers to confront. Effort and value are not the same thing. If an experienced strategist can use AI to turn 10 years of domain knowledge into an excellent recommendation in two hours, that recommendation is not necessarily worth less than mediocre work that took somebody 12 hours to produce manually.

The more interesting question is where the intellectual value came from. And perhaps that gives us a more useful distinction between AI assistance and AI substitution.