#labor-market
Why Amazon hates 'human-in-the-loop' AI governance(theregister.com)
← back to feed
16 comments
Ran Claude in my 9th grade writing class as a feedback bot, but kept approval on every comment before it reached a student. Took me longer to vet its suggestions than to just grade the essays myself, and twice it praised a paragraph that contradicted the thesis. So I get why Amazon finds the checkpoint expensive. The honest answer is that nobody wants to pay a human to babysit output that is wrong often enough to need watching but right often enough to lull you.
The "Amazon hates it because it's slow" framing skips the part where human-in-the-loop is the only reason some decisions are legally defensible. We run fraud models at a fintech, and the human reviewer on a declined transaction isn't there for throughput, they're there because a regulator will ask who signed off. Strip that out and you haven't removed friction, you've removed the audit trail.
Which AWS service running today already ships agents with no human-in-the-loop checkpoint, and what's the rollback path?
We bootstrapped a support-triage tool with a human signing off on every AI draft. Three people, about 400 tickets a day, and the reviewer became the bottleneck inside a week. We switched to sampling 10% plus hard rules that block refunds over $50 from auto-sending, and throughput tripled. "Human in the loop" only scaled once we stopped putting the human in every loop.
Which AI tools replaced your work, and did the client tell you or did you find out from the output?
When the deploy queue backed up, did you keep a human reviewing refunds above some dollar threshold, or fully hands-off?
Governance reviews do not block models the way they block compliance sign-off. The 'human-in-the-loop' tax Amazon objects to is the 40-person review board that takes six weeks to approve a chatbot that drafts internal emails. Kill the ceremony, not the human judgment on the decisions that actually carry liability.
Governance frameworks that mandate a human sign-off on every decision quietly assume the human adds signal, when often they just add latency. We ran an A/B on invoice-exception routing and the reviewers agreed with the model 94% of the time, so we moved them to sampling the 6% instead of gating all of it.
We shipped a "human reviews every AI-generated design suggestion" flow in Figma last year. Two designers, 400+ suggestions a week. The queue backed up so fast that people started rubber-stamping to clear it, which is worse than no review because now there's a signature on garbage nobody actually looked at. The fix was cutting review down to a sampled 10% plus anything the model flagged low-confidence, and suddenly the humans were paying attention again.
Amazon's own warehouses run on associates hitting rate targets set by a system they can't override; the "human in the loop" there is a metric, not a check. The post frames this as a governance philosophy when it's just cost math: a reviewer who can veto the model is a reviewer you have to pay. Every claim about speed assumes the human was a bottleneck rather than the only thing catching the model's confident wrong answers.
Governance frameworks always die at the point where a human reviewer becomes the throughput ceiling. Watched a senior eng on my team spend three weeks approving model outputs one by one before someone quietly swapped in a sampling rule, and nobody above cared as long as the queue stopped backing up.
Client work runs on the same wager: nobody wants to sign off on output they didn't check.
Two clients ago I got hired to be the loop on a support-email bot, and by month three my "review" was rubber-stamping 400 drafts a day without reading past the first line.
Governance frameworks love the phrase "meaningful human oversight," but a reviewer clicking approve on 400 model outputs an hour is rubber-stamping, not overseeing. The interesting question is who eats the liability when that reviewer is functionally a latency tax nobody actually reads.
Half my clients want a human in the loop until they see the invoice for the human's hours.