This article was AI-generated as part of an experimental historical-content project. The date reflects the period being analyzed rather than the date the article was originally written.
Last Thursday OpenAI introduced GPT-5.2 as “the most capable model series yet for professional knowledge work.” The launch leans heavily on GDPval, OpenAI’s evaluation of “well-specified knowledge work tasks” across 44 occupations. By OpenAI’s account, GPT-5.2 Thinking beat or tied top industry professionals in 70.9% of comparisons judged by experts, on tasks that include “making presentations, spreadsheets, and other artifacts.” Instant, Thinking and Pro began rolling out to paid ChatGPT plans the same day, and all three are available in the API.
Set the benchmark aside. The more interesting statement is about where the output is meant to land: not in a chat window, but in files that get sent to other people. That changes the mechanics of how a wrong fact about a company travels.
Claims, responses and artifacts
OpenAI reports errors in two ways. The launch post says that on de-identified ChatGPT queries, GPT-5.2 Thinking’s responses with errors were 30% less common than GPT-5.1 Thinking’s. The system card update measures both the share of individual claims with a factual error and the share of responses with at least one major error, and says that with browsing enabled the model stayed under 1% across five domains, one of which is business and marketing research.
Those are real gains, and the distinction between claims and responses is the useful part. A chat answer contains a handful of claims. A market overview deck, a competitor table or a due diligence memo contains hundreds. Even at a low error rate per claim, a long document is more likely than a short answer to contain at least one wrong statement, and the reader has less reason to check any given cell.
What an artifact strips away
A chat answer that used search usually shows its sources beside the text. Once that content becomes row 14 of a spreadsheet or the third bullet on slide nine, the link is usually gone. The fact now carries the authority of the document it sits in, and of the colleague who sent it.
Consider which facts tend to fill these files: revenue ranges, headcount, the current chief executive, which subsidiary belongs to which parent, whether a lawsuit is pending or settled, who launched a product first. These are exactly the details where public sources disagree or go stale, and where the model has to choose between them. I wrote about misattribution two years ago as a legal theory. Inside a spreadsheet it is simply a cell.
Well-specified tasks, under-specified companies
OpenAI is careful to say GDPval measures well-specified tasks, and the qualifier matters. “Build a comparison of the top five vendors in this category” is well specified as a task. It says nothing about sources. The model decides which pages describe each vendor, and a company with a stale profile on a data aggregator, or an About page that contradicts its annual report, leaves that decision wide open.
OpenAI frames the speed and cost advantages as useful “when paired with human oversight.” That is the right caveat. In practice, the person reviewing a 40-row table checks whether it looks plausible, not every figure about every company in it.
What this means for reputation teams
The audience for your company’s facts now includes analysts, consultants, procurement teams and journalists who ask a model for a document rather than an answer. Most of the remedy is upstream: keep the basic facts consistent wherever they are published, put dates on them, and resolve contradictions between your own pages and the large business databases. It also helps to test the way these users work. Ask for a comparison table or a one-page briefing that includes your company and its competitors, then read every cell about you.
A wrong sentence in a chat window is a nuisance. The same sentence in a spreadsheet that gets forwarded to a committee becomes a fact.