Anthropic says its Claude models “lead” 26% of the company’s AI research and development as of August 2026, up from under 1% in February 2026. Leading means the AI can complete most of a task end to end while a person supervises. In Anthropic’s example, Claude finds and fixes a failing data pipeline, and an engineer reads the write-up and decides if it ships. The 26% is Anthropic measuring itself. Our article this week argued that once AI leads a task, the person’s job becomes the check.
Why It Matters
In Anthropic’s example, the one checking is the engineer who no longer has to “stay actively tuned in” while Claude works. In 1983, Lisanne Bainbridge, a psychologist at University College London, published “Ironies of Automation,” about operators who watch over automated plants. The knowledge they need for the unusual case, she wrote, “develops only through use and feedback,” and recalling it “depends on frequency of use.” She also named the exception: operators who get frequent alarms keep that knowledge as part of their normal work. She wasn’t writing about AI at a desk. Our reading is that a check stays fed while it keeps catching things, and thins when the AI rarely needs correcting.
The Decision
The question is whether a check’s owner keeps doing some of that work start to finish. Doing it may keep the check fed but costs part of the time the AI saved. Skipping it keeps that time, and where a miss is cheap to catch later, a thinner check may be enough.
What To Do This Week
- Pick one task AI now leads on your team and ask whoever checks it what they last sent back and why.
- Ask when they last did that task start to finish themselves.
- Read both answers together. A check that keeps sending work back with reasons is likely still getting practice. One that sends nothing back, whose owner hasn’t done the task since AI took it over, is where this decision applies.
What Not To Do
Don’t read a check that never sends anything back as proof the work is fine. Don’t pass a busy owner’s check to someone who never did the work. And don’t cut training on a task because the AI leads it. Bainbridge argued the most successful automation may need the most training.
Signal Boost
Ironies of Automation, Lisanne Bainbridge, Automatica, 1983, on the publisher’s page, which may charge for access. A free copy sits on Gwern.net, an independent third-party site. Bainbridge’s five pages explain why automating most of a job can make the human part harder.