Supervising AgentsNovember 8, 2026·4 min read

Altman, Hassabis, and Amodei All Said the Same Thing This Week. The Human Job Isn't Checking AI's Work — It's Knowing What to Ask For.

Altman, Hassabis, and Amodei all reframed the human role in AI work within days of each other, and Mollick brought research to back it up. The habit of checking AI's work isn't the skill that matters now — scoping and checkpoints are.

By Forge Team · Examples are illustrative composites

If your team's AI policy still reads "have someone check the output before it goes out," you're already behind. In the same week, the people running Anthropic, OpenAI, and Google DeepMind each said that mental model is wrong, and a researcher brought evidence that backs them up. The job was never going to be catching mistakes forever. It just became obvious sooner than most policies assumed.

Here's what happened. At ADIPEC on November 2, Sam Altman said there is "no room for debate" about keeping a human in control of AI, while in the same remarks calling AI "mostly smarter than the top scientists" he knows. Five days later, Demis Hassabis told X that "one person who really understands AI will outperform an entire startup team," predicting companies running at a tenth of today's headcount within five years. Dario Amodei's Anthropic S-1 filing, circulating that same week, described a future of "a country of geniuses in a data center." Then on November 7, Ethan Mollick went further than any of the three: he argued that treating human review as a check on AI's mistakes "is likely not true today," pointing to evidence that frontier models now catch factual errors better than human experts do. Days earlier, Intuit CEO Sasan Goodarzi told Stratechery that AI only works in professional services when it's paired with someone who has real domain expertise — not someone proofreading line by line, someone directing.

None of this means skip review. It means the review that matters isn't "does this sentence look right" — it's "did I give this the right scope, the right quality bar, and the right checkpoint before it ran." Reading AI output closely enough to catch a typo or a wrong number is a skill AI is closing in on fast. Deciding what AI should be allowed to do without you in the loop, and where a human still has to sign off before it ships, is not.

Take a composite case: a marketing manager at a 40-person SaaS company who used to read every AI-drafted customer email before it sent. That habit caught typos. It didn't catch the thing that actually went wrong last quarter — an AI support tool that answered a pricing question with a discount code that had expired, because nobody had written down which categories of question needed a human first. The fix wasn't more proofreading. It was deciding in advance, by topic, which answers the AI could send alone and which had to wait for a person.

A second composite case, pointed the other way: an operations lead at a 200-person logistics company who required a manager's sign-off on every AI-generated shipment note, regardless of the shipment's value or history. Nothing ever went wrong — but the backlog did. Three managers were spending an hour a day rereading routine paperwork that had never once needed a correction. The lesson isn't "trust AI more" or "trust it less." It's that supervision has to scale with the actual risk of the task, not get applied as one flat rule across everything AI touches.

The three people building the most capable AI systems in the world just told you, within the same week, what they think your job is turning into. The useful response isn't to argue with them. It's to write down, this month, which decisions still need you — and let the rest go.

Reading about it only gets you so far

Patin turns this into five-minute drills that score what you write and tell you why. It's in closed beta — join the waitlist and we'll email you when your cohort opens.

Just want the writing? .