AI has a data problem. It’s bad, but everyone is too scared to fix it

In 1972, a Stanford psychologist sat children down in front of a marshmallow and made them an offer. Wait fifteen minutes without eating it, and you’ll get a second one. The children who waited, the study concluded, were more disciplined. More successful. The marshmallow test became one of the most cited experiments in psychology – proof that willpower predicts outcomes.

It was misread for forty years.

In 2013, Celeste Kidd and her colleagues at Rochester ran the same experiment with one change. Before the marshmallow appeared, some children were given a small gift by a researcher who then came back with a bigger one as promised. Others were given the same promise and then let down. One broken promise was the only variable.

In the reliable condition, 64% of children waited the full fifteen minutes. In the unreliable condition, just 7% did. The difference was a single broken promise. The children who didn’t wait had learned, correctly, that the adults in this system didn’t keep their promises. Stopping was the rational move.

I spent years working inside the AI training data industry before founding Poindexter Labs. The industry has made exactly the same misread as those early researchers. It has looked at bad data and blamed the people producing it. It has never seriously asked whether the system itself is the problem.

It is.

When the cost of approving something is higher than the cost of rejecting it, rejection becomes the default. Good work gets discarded. Bad work becomes the norm – because when rejection feels inevitable, why would anyone try?

This is a logical consequence of how these platforms are structured. On the major data annotation platforms, I’ve seen rejection rates routinely run between 30% and 80%. The incentive structure makes rejection the safer choice. Approve something and your capability is on the line if it fails further up the chain. Reject it and nothing happens. There is no recourse for an unfair rejection. There is no reward for an accurate approval.

A quality control system this is not.

I know this as a participant. Plenty of times I would wake up two or three times a night to check my phone – to see if I was still on a project. I would receive tasks I thought were perfect and skip them. The risk of sending them forward was too high. Someone above me might find a reason to reject it – or invent one. The reasoning was rarely clear. Work that was objectively flawless became a victim of a system where rejection was always the safer call.

This behaviour is rational. It is happening at scale, across every major data annotation platform, every day.

When a platform has a rejection rate above 50% it becomes a lottery. Submit perfect work and it comes back. Submit work that barely passes and it comes back. The bar isn’t high – it’s arbitrary. And when the bar is arbitrary, the only rational response is to stop trying to clear it.

The consequences compound. Earlier this month, Meta drafted 6,500 of its own engineers into generating training data – puzzles and coding problems to train its AI models. Engineers with advanced degrees, describing the work as soul-crushing. As a gulag. This is worth pausing on. These are people who, in all likelihood, enjoy puzzles and coding problems. That is probably why Meta hired them. The factory workflow is what made it soul-destroying – the monitoring, the metrics, the incentive structure applied to work that was never designed to be measured that way. Meta got exactly the result that structure produces.

This keeps happening because the industry has never been forced to confront what its workflows actually produce. When we went through one customer’s reject pile – tasks their annotation platform had discarded – we were able to save 70% of them with minor or no edits. Seventy percent. That work was going to be thrown away. The people who produced it were going to be penalised for producing it. The customer was going to pay to have it done again.

See Also

Community: the most undervalued currency in purpose-led startups

The obvious challenge to this is: if everyone in the system wants the task to go through, what stops bad work getting waved forward? It is the right question and it deserves a direct answer.

At Poindexter Labs, we have removed the disincentive to reject work. A reviewer on our platform doesn’t carry personal risk for sending something forward. If a task is ambiguous, it triggers a conversation. The expertise in the room is used to resolve ambiguity rather than provide cover for discarding it. That is a structural difference. You cannot fix this with better management or stronger values or more rigorous hiring. The workflow itself has to change.

This problem will not go away and is set to get worse by some order of magnitude as models get more advanced and labs demand more. The tasks are getting harder. Every additional layer of complexity gives reviewers more reasons to reject. This isn’t a problem that stabilises – it compounds. The industry is running an exponential rejection machine on problems that require exponential expertise, and wondering why the data keeps getting worse.

AI is the most heavily capitalised industry in history. Billions flow in every quarter. And a significant portion of it is being spent paying people to throw away good work, redo it badly, and throw it away again. The inefficiency isn’t a side effect of the workflow. It is the workflow. The waste is being funded at scale because the people at the top of the chain have built systems where they believe they don’t need to tell the difference.

The contributors producing AI training data have learned exactly what the children in Kidd’s study learned. When the system doesn’t keep its promises, the rational move is to stop waiting for the second marshmallow. The industry built that system. Until it changes it, it will keep getting the data it deserves.