Ghost AI: what we know about Builder.ai and Amazon

OPINIÃO

The "ghost AI" story: what the press reported about Builder.ai and Amazon’s checkout, what was disputed, and what it teaches about transparency in AI.

In 2024 and 2025, two stories became headlines as examples of "ghost AI": Amazon’s cashierless checkout and the British startup Builder.ai. They share one thing, the distance between what is advertised as AI and the human work behind it. But the versions that circulated were, in part, disputed. This post separates what is documented from what isn’t.

What was reported and what was disputed

Amazon Just Walk Out. According to Axios, the claim that Amazon’s cashierless checkout relied on more than a thousand people in India watching and labeling videos came from a Gizmodo report. Business Standard, on April 4, 2024, reported "as many as 1,000" workers. Amazon disputed the characterization. On April 17, 2024, Jon Jenkins, the vice president overseeing Just Walk Out, told Axios that the idea of live human review is "completely not true": employees in India watch some videos after the fact to help label actions and products, and check automatically generated receipts "in a small percentage of cases," with a team of "way less than 1,000."

Builder.ai. The version that spread the most, that the company faked AI with 700 engineers in India, doesn’t hold up as fact. Gergely Orosz, in The Pragmatic Engineer (June 12, 2025), traces the number to an X post by an author with no reporting history, which the Financial Times later found to be false, and he writes that he and other outlets repeated the version at first. According to the piece, the company had an outsourced developer network sold as a service, and what brought it down were accounting-fraud allegations, not the discovery of engineers hidden behind an AI.

What the two cases show

• Human review in AI systems is common. Labeling, auditing and checking uncertain cases are part of many legitimate pipelines. What’s debated in the Amazon case is the proportion and how it was communicated.
• An “AI” label can hide something else. In the Builder.ai case, the problem pointed to by the cited source was accounting, not hidden engineering. The sentence that went viral was simpler and easier to repeat than the real story.
• The viral version isn’t always the verified one. Even serious outlets repeated the 700-engineer number. Reading the primary source, or at least whoever disputed it, changed the picture.

The limit of this reading

The Amazon accounts come from press and from the company’s own statements, not from an independent audit. The Builder.ai information comes from an analysis piece that relies on Financial Times reporting (which we did not consult directly), and the author himself marks part of what he writes as inference. This post does not claim that any AI product is manual work in disguise, nor that human review is a mistake.

What changes with a controlled scene

This section is our reading, not the sources’. In a synthetic data pipeline, humans also come in: they define what to generate, review samples and validate on real data. The difference is that these decisions can be recorded. The scene parameters (light, material, camera, class proportions) are declared, each sample can be reproduced from them, and the label comes from the render itself, with no manual step estimating what the image contains. That makes the pipeline easier to audit. It doesn’t make it automatic or remove validation on real data.

What to take to your project

• Ask where the human is. Who labels, who reviews, what is truly automatic and in what proportion. A vague answer is a warning sign.
• Read the source, not the headline. Before repeating a “fake AI” story, or an “AI that works” story, check who originated it and who disputed it.
• Ask for the data’s origin and measure on your case. How the dataset was collected or generated, who checked the annotations, and how the result behaves on your real data.

Does your product’s data pipeline have manual parts that aren’t documented? Tell us about it in a diagnostic call. If synthetic data isn’t the answer, we’ll say so.

Sources

• Orosz, G. Builder.ai did not “fake AI with 700 engineers”. The Pragmatic Engineer, June 12, 2025 (analysis citing the Financial Times).
• King, H. Amazon pushes back on perception of Just Walk Out, announces expansion in third-party stores. Axios, April 17, 2024.
• Amazon’s ‘just walk out’ checkout tech was powered by 1,000 Indian workers. Business Standard, April 4, 2024.