A mixed week of reminders that the tools keep improving while the discipline around them lags behind. I have been thinking about who owns the output when AI is involved, how the economics of the work are shifting, and what young people are asking us to do about all of it.
Consultancies still are not fact-checking AI
It is astonishing that we still get frequent examples of consultancies not fact-checking their AI. The latest is a KPMG report that contained AI hallucinations on the benefits of AI, of all subjects. It would be funny if it were not so avoidable.
The lesson here is not that AI is useless for this kind of work. Used well, it can draft, summarise and synthesise at a pace no team could match on its own. The lesson is that a human has to own the output. When a report goes out under a firm's name, the accountability cannot be delegated to a model. Someone has to read every claim, check every citation, and be willing to put their own name behind it. That is not a limitation of the technology; it is just what professional work has always required.
An image model that treats pictures like code
Reve 2.0 is out, and it took second place on the Text-to-Image Arena, ahead of Nano Banana 2 and behind only GPT Image 2. What interests me is not the ranking but the approach.
Most image models expand your prompt into prose and then paint the whole scene in a single pass. If you want to change one element, you re-roll and hope the rest survives. Reve treats an image as a structured layout instead: every element gets a position, a size and a description, so the picture becomes addressable and editable in much the way code is. You can change one region and leave the rest of the scene intact.
That is a meaningful shift. It moves image generation from something you gamble on to something you can iterate on deliberately, and it hints at a more general pattern where creative outputs become structured and revisable rather than one-shot.
A cheaper default for coding work
I switched my default coding model to Composer 2.5. It is fast and cheap, and for the bulk of agentic work, multi-file edits, refactors, running tests, fixing CI, it gets through the task with a little babysitting. I move up to a larger model only for intense UI work where the extra capability earns its cost.
The signal here is not any single model. It is that the price-performance frontier for everyday engineering work keeps dropping. Tasks that a year ago justified the most capable and expensive model can now be handled well enough by something far cheaper. That changes how you budget, how you delegate, and how much you can reasonably attempt in a day. It is worth watching where that frontier settles, because it reshapes the economics of the work far more than any headline benchmark.
What teenagers actually do with AI
Oxford University Press released research with nearly 4,000 UK teenagers on how they actually use AI for schoolwork, and it is more measured than the usual panic. The full report is worth reading.
Young people are not reaching for AI as a default for homework. Instead they describe uncertainty about where the line sits, and they are asking their schools to draw it. That is a request for guidance, not permission to cut corners. Notably, 77% want their teachers using AI in lessons to support the class, while being clear that a teacher's value is not something AI replaces. They are also more excited than worried about what AI means for their education.
I find that reassuring and instructive in equal measure. The people closest to this are asking the adults to set the boundaries. The worst thing we could do is leave that gap unfilled.
Apple makes a case for caution
Apple spent its WWDC keynote making a case for caution, and it was a more interesting message than the product news around it. Siri is being rebuilt on Google's Gemini through a billion-dollar partnership and rechristened "Siri AI". But the framing mattered more: Craig Federighi drew a line, suggesting some companies appear to be pursuing AI for its own sake without much regard for the people it is meant to serve.
Then Apple shipped some of the most concrete child-safety controls I have seen from a platform vendor:
- parental permission before a child opens a new website
- proactive blurring of gore and violence in messages
- simplified screen time with built-in time limits aligned to American Academy of Pediatrics guidance
- an age-graded setup that expands what a child can access as they get older
Whatever one makes of the Siri rebuild, this is what taking the human seriously looks like in practice, not in a slogan.
The AI boom in charts
The Guardian walked through the AI boom in six charts, and the numbers are hard to hold in your head. Spending on AI infrastructure is heading from roughly $765B this year to $1.6T by 2031 on Goldman Sachs estimates. Forty-one AI-related stocks now make up close to half the entire S&P 500 by market value, and one Harvard economist calculates that datacentre investment accounted for 92% of US GDP growth in the first half of 2025.
In software engineering the returns already look worthwhile: where there is enough demand inside an organisation, the tools mean more work gets completed at lower cost. That is the one place the story clearly holds so far. The open question is whether the same pattern shows up in other kinds of work. 2026 is the year we find out whether the spending is buying real, broad productivity or a narrower gain than the market has priced in.