Under the EU AI Act (2024), AI companies are obliged to mark text generated by their LLMs as of August 2nd 2026. This means that companies who signed the EU Code of Practice on Transparency are introducing ways to do so. Failure to comply with text-marking could result in fines of up to €15 million, or 3% of worldwide annual turnover. Much coverage has been given to Claude’s announced rollout of watermarking. In reality there’s actually very little to say.

Somewhere this quarter, in a talent review, a manager will describe an engineer as T-shaped. Heads will nod and a box will be ticked. The engineer, who is not in the room, will spend the next year receiving work shaped like the letter rather than shaped like their ability. Nobody will check whether the label was accurate, because it does not feel like a claim. It feels like an observation, but it’s neither.

Just about every senior leader has already decided to adopt AI. The all understand the risks of not doing, and a large proportion of them understand the risks of doing so. The question some aren't asking is what happens when one of the risks becomes real, and whether they can recover from it.

Developing concepts with Claude means understanding how Gen AI works and working with it's limitations. During an exploratory workflow today I ask it several questions as starting points to develop the nucleus of a concept. It came back with systems logic: structure, audience, constraints, examples. I told it the thinking would emerge from exploration instead. The AI pushed back. We went sideways anyway.

Evaluating Claude Sonnet 4.6 for 'reasoning redline', defensive hallucinations and structural bias.