OpenAI Just Warned That Its Own Agent Makes 'Finished Mistakes.' Here's How to Catch Them.

OpenAI published a safety warning for its own new agent product last week. The warning, buried in the ChatGPT Work release materials: the system can produce "finished mistakes at a scale that is harder to catch" (OpenAI, July 9). That phrase — from the team that built the tool — is worth sitting with before you hand it your next project.

What happened this week

Three findings landed in the same week, and they point at the same gap.

OpenAI launched ChatGPT Work on July 9 — an autonomous agent that connects to Slack, Google Drive, and SharePoint, runs multi-step tasks unattended for hours, and delivers finished deliverables. In the same release, their own safety guidance flagged the "finished mistakes" problem (OpenAI, July 9). Ethan Mollick published findings from Wharton's GAIL research group showing that every prompting tactic people have been relying on — chain-of-thought instructions, tipping, threatening, expert persona framing — returns zero measurable gain in output quality (Mollick, July 7). And Cursor's Developer Habits Report revealed a 46x productivity gap between P99 AI power users and the median user, with a Gini coefficient of 0.77 (Cursor, July 8).

Three findings, one conclusion: the gap between skilled and unskilled AI users is not about which tools they have access to. It is about whether they can tell good AI output from bad.

What to do differently Monday morning

The "finished mistake" problem is not new — AI has always produced confident wrong answers. What has changed is the volume and the packaging. An agent running unattended for two hours delivers output that looks complete. The formatting is professional. The logic holds together internally. The specific error is buried somewhere it takes attention to find.

Three things are worth treating differently:

Verification is the skill, not the afterthought. If you cannot explain why an AI output is right — not just that it looks right — you are accumulating risk that compounds when agents run for hours without human review. "Looks right" is not a check; it is an assumption.

Specification beats prompting. Mollick's research did not just kill prompt tricks — it identified what works instead: clear goals, defined constraints, explicit output shape, and named acceptance criteria. These are writing skills, not technical ones.

The 46x gap is a verification gap. Cursor's report tracks users of the same tool — the 46x difference between P99 and median users is not explained by better access. The gap is about the ability to evaluate, critique, and iterate on output — to distinguish "looks polished" from "is actually right" (Cursor, July 8).

Cassie: the content director who almost shipped a ghost pricing tier

Cassie runs content at a 40-person SaaS. Part of her job is a monthly competitor summary — three named competitors, five dimensions each — that goes to the product team. She started using Claude to draft the summaries from a web search before she reviewed them.

The first three months came back clean. In month four, one competitor's pricing section included a tier that had been removed from their site six months earlier. The presentation was indistinguishable from the accurate sections — same tone, same formatting, the same confident specificity.

Cassie caught it because when she set up this workflow, she wrote one line into her personal checklist: "Verify current pricing against each company's pricing page before sending." She had named that specific check when she started because pricing is the one dimension she knew could go stale silently. Without that named criterion, she would have been reviewing by feel — which means she would have caught some errors and missed others at random.

Daniel: the legal ops manager who runs every comparison twice

Daniel manages contracts at a 200-person professional services firm. He uses AI to produce first-draft clause comparisons — given two contracts, flag differences in indemnification, limitation of liability, and IP ownership. It saves him two hours per document set.

Six weeks ago he ran the same clause comparison twice on the same two contracts, in two separate sessions, with no context shared between them. The outputs were similar but not identical. In one, a limitation of liability clause was flagged as one-sided in the vendor's favour. In the other, the same clause was described as "standard." One reading was right. The other was a reasonable-sounding interpretation that would have gone unquestioned in isolation.

He now runs every high-stakes comparison twice before it goes to a partner. Not because he distrusts the tool — because two passes on this specific type of work takes fifteen minutes and has a specific failure mode it catches. That is a verification protocol, not a workaround.

The one-sentence version

OpenAI's warning is not a bug disclosure — it is a usage instruction: agents produce confident, polished, finished-looking output, and the question is not whether you can spot mistakes in general, but whether you decided in advance which specific things you are going to check.

<BlogPracticeSection />