AI · ML · DL Mystery 6 min

The Thousand Humans

01 · THE SETUP

In April 2024, Amazon pulled its most famous piece of in-store AI — the Just Walk Out checkout-free system — from its Fresh grocery stores. Grab your shopping, walk out, get billed automatically. Pure machine vision. Except reporting in The Information — as summarised by Axios — found the system had leaned on roughly 1,000 workers in India reviewing shopping sessions — and that in 2022, as many as 700 of every 1,000 sales reportedly needed a human check.

Amazon publicly disputed the framing — no one was “watching you shop live,” it said — but confirmed the humans were there, labelling footage and reviewing what the model wasn't sure about.

Pause. If “AI-powered” can legally describe a system with a thousand people inside it, what would you check before believing the label on anything?

02 · YOUR CALL ⏸ YOUR CALL — PICK ONE TO CONTINUE

What was the real tell that this system wasn't yet what the label promised?

If you pick A

Reasonable — hardware feels like the hard part. But cameras are exactly the right input for a vision system; the world's best models see through ordinary lenses. The hardware wasn't the tell. Something about who was doing the deciding was.

If you pick B — the mechanism

That's the one. Every serious AI system uses humans to label data and audit edge cases — that's normal, even healthy. The tell is the fraction. When most decisions still route through people years after launch, the machine hasn't finished learning the job.

If you pick C

A fair instinct — receipts sometimes took hours, and latency often betrays a human in the loop. But slowness is a symptom, not the disease. The question is what the delay was for: people were reviewing decisions the model couldn't make alone.

If you pick D

That answer makes sense — companies retire failures. But Amazon kept the technology for smaller stores and stadiums, where it works fine. The retreat tells you about economics at grocery scale, not about whether the thing was AI. Look at who was making the decisions.

Pick one — committing first is what makes the answer stick.

the lesson continues after you choose

03 · NOT SO FAST

The obvious reading is a scandal: fake AI, secretly people. It's satisfying, and it went viral for exactly that reason.

But it misses something more useful. Human review isn't the opposite of AI — it's how AI gets built. The interesting question is never “are there humans?” It's whether the humans are teaching the system or being the system.

04 · THE MECHANISM

So what separates “AI” from ordinary software? One thing: where the behaviour comes from. In classic software, a person writes the rules — if total > 500, flag it. In AI, the rules are learned: you show the system examples, and it works out its own internal rules for mapping input to output. Nobody wrote down what a shopper picking up yoghurt looks like. The model inferred it from millions of labelled clips.

The three terms people swap around are actually nested: deep learning is a kind of machine learning, which is a kind of AI.

Artificial intelligence is the umbrella: machines doing tasks that normally need human judgment. Machine learning is the dominant way of getting there: learn the rules from data instead of hand-coding them. Deep learning is machine learning with many-layered neural networks — the approach behind essentially everything you'd call AI in 2026, from Claude and GPT to the vision models in a checkout-free store.

Every real deployment sits on a spectrum: hand-written rules at one end, a fully learned model at the other, and human-in-the-loop review covering the gap between what the model can do and what the job requires. Just Walk Out wasn't fake — its gap simply stayed too wide, for too long, at grocery scale. Seven hundred human reviews per thousand sales is a classroom, not a checkout.

05 · BACK TO THE OPENING

So the opening wasn't really a story about Amazon faking AI. It was a rare public X-ray of the pipeline every AI system goes through: humans label → the model learns → humans verify what the model is unsure about. The system was genuinely learning. The mystery of the thousand humans is just what learning looks like before it's finished — and the economics ran out first.

06 · TAKE THIS WITH YOU

Your rule: when anything is sold to you as “AI-powered,” ask one question — “What did it learn from, and who checks it when it's unsure?” If nothing was learned, it's rules with a marketing budget. If everything is checked, it's people with a software costume. Real AI lives in between, and the human fraction tells you exactly how far along it is.

REFERENCES