Anthropic recently disclosed that DeepSeek, Moonshot, and MiniMax used roughly 24,000 fake accounts to distill Claude's outputs into training data for competing models. Twenty-four thousand identities, none of them real, extracting value at industrial scale before anyone could tell the difference.
The postmortems focused on the input side: rate limits, account verification, abuse detection. Fair enough. But the uncomfortable question for anyone buying human data is what the episode says about the other side of the pipeline.
The human side has the same hole
Every RLHF pipeline depends on people: graders, demonstrators, red-teamers, native speakers. Ask a vendor how they know who those people are and you'll hear about resumes, profile checks, and qualification tests. Ask how they know the person who passed the test is the person doing the work — this week, at 2am, on the medical eval — and the answers get quiet.
- A qualification test verifies a session, not a person.
- A marketplace account can be shared, sold, or operated by someone else entirely.
- A self-reported title costs nothing to invent and nothing to keep.
Without ground-truth identity data, a vendor has no better way of confirming who's actually doing the work than anyone had of confirming who was behind those 24,000 accounts.
What ground truth looks like
The strongest identity signal that exists is employment: a contract that names the role, payroll that paid the person, tenure that accumulated over years, and a performance history attached to all of it. That's the data we hold, because we generated it — as the employer — for millions of professionals over a decade.
Label quality gets audited constantly. Label provenance almost never. That ordering is backwards.
Provenance-first data supply is what we built Rainbow to offer. The alternative is hoping your vendor's account list looks nothing like that other account list.
We were the employer