The questions that actually bite

Attribution and consent

Training corpora are assembled from work that people made for other reasons. "Publicly available" describes access, not permission, and the distinction has been doing an enormous amount of unearned work in this debate. The legal position varies by jurisdiction and is unsettled; the ethical position is not obviously waiting on the legal one.

Labour, not just jobs

The displacement conversation focuses on headcount. The more immediate change is to the texture of work: which tasks get automated first are rarely the unpleasant ones, and judgement-heavy roles can be hollowed out while the job title survives. Data annotation labour — often low-paid and offshored — remains largely absent from the discussion.

Concentration

Frontier training runs are expensive enough that the set of organisations able to perform them is small and shrinking relative to the field. That is a governance problem before it is a technical one: safety commitments made by a handful of actors are only as durable as those actors' incentives.

Environmental cost

Reporting is patchy and rarely comparable. Training figures get cited far more often than inference figures, despite inference dominating lifetime energy use for widely deployed models. We would like to see standardised disclosure; we are not holding our breath.

What we will not do