OpenAI says Americans send 3 million daily ChatGPT messages asking about pay

What the report actually counts

OpenAI reports that people in the US send nearly 3 million messages a day, on average, asking ChatGPT about wages, compensation, or earnings. The company broke labeled wage-benchmarking messages into categories: pay calculation (26%), a specific role (19%), entrepreneurship (18%), a specific role at a company (11%), and occupation or career questions (11%).

The classification came from what OpenAI describes as a privacy-preserving analysis using automated classifiers, with no human viewing individual messages. That matters for reading the results: these are machine-labeled aggregate patterns, not a survey of what workers say they use the tool for.

The distribution points somewhere specific. OpenAI found wage search over-indexes in higher-skill and less transparent occupations — creative fields, management, healthcare, computer and mathematical roles — and rises where pay is more dispersed and higher. People reach for the tool most where a posted benchmark is hardest to find.

The interface, not the model, is the product being described

Strip away the economics and this is a piece about frontend behavior. The pitch is explicitly interactional: instead of searching across multiple websites, interpreting scattered salary pages, or asking a socially awkward question, a worker types one prompt and gets a benchmark back in seconds.

Rather than requiring a worker to search across multiple websites, interpret scattered salary pages, or ask a socially risky question, a model can synthesize wage information and return a benchmark in seconds.Montana Labs

The 'socially risky question' framing is the tell. Part of what the chat surface changes is not information availability but the social cost of asking — a text box has no colleague on the other end. That's an interface property, and it's why entrepreneurship questions concentrate in areas where 'there often is no posted wage benchmark' at all.

It also means the design does something salary websites don't: it collapses search, interpretation, and calculation into one turn. Pay calculation being the single largest category (26%) suggests people aren't just looking up a figure, they're asking the interface to do arithmetic and translation on it.

WorkerBench puts a number on the number

When a conversational surface returns a single figure, the figure carries the whole weight of trust — there's no results page to eyeball, no range to compare. OpenAI seems to recognize this, which is why the report introduces WorkerBench, evaluating GPT-5.4 against 2024 OEWS median wages at the national occupation and metro levels.

OpenAI reports the model is 'highly accurate' in the observed sample: high coverage, small bias, and nearly all numeric estimates falling very close to the benchmark. The honest caveat is baked into the scope — this first benchmark tests against published national and metro medians, exactly the case where the answer is already knowable. The company concedes the real questions are elsewhere: 'the geography, firm, level, and compensation questions workers actually ask every day.'

So the accuracy claim and the demand pattern point in opposite directions. Workers over-index toward the least transparent occupations, while the benchmark validates the most transparent ones. The gap between where the tool is trusted and where it's proven is the part still to be closed.

Why a synthesized answer changes what frontends owe the user

The specific implication here is about presentation, not retrieval. A search engine hands back sources and lets the user judge; a chat answer hands back a conclusion. When 3 million daily messages ask a text box what a role pays — and disproportionately in negotiable, high-stakes fields — the interface is being treated as authoritative on questions where it is least tested.

For anyone building on top of these models, the takeaway is concrete: a single confident number in a chat bubble needs to carry the uncertainty OpenAI itself flags. WorkerBench is a start at measuring that, but until the benchmark covers firm, level, and geography, the reliability of the answer lags the confidence of the format. The frontend advantage — one clean synthesized figure — is also the frontend liability.