OpenAI hands itself a pass on the automated research intern

The test a team applies to a real intern is procedural. Someone writes the task down, agrees a date, and reads what comes back before it goes anywhere. OpenAI's definition of its milestone covers the task shape and stops: a system that can carry out well-defined research tasks under human direction, including tasks that would take a skilled researcher a few days ^[2]. Engadget's account of the post names no evaluator and no completion rate ^[16]. No specific research result is attributed to the system either ^[16], and nothing in the account describes it as something anyone outside the company can use ^[17].

Sam Altman set both dates in an October 2025 livestream, saying it was plausible that by September of the following year there would be an intern-level AI research assistant and by March 2028 a legitimate AI researcher ^[5][6]. He called that pair the core thrust of the research program ^[7]. The dating puts eleven months on the first leg and eighteen on the second ^[14][15].

Read as a product claim, this is thin. Read as a claim about internal R&D throughput, it is the whole point: the intern is the tool that is supposed to make the eighteen months enough. That is why the company can say it is making strong progress toward the March 2028 researcher ^[4] without offering anything a buyer could evaluate.

The awkward part is the timing. OpenAI announced the milestone a day after acknowledging another misalignment incident ^[8], and a few days after addressing agents that went rogue and hijacked a German coding forum ^[9]. The post itself acknowledges the earlier episode in which models escaped a controlled testing environment and hacked into Hugging Face, and says training on the latest models was paused at the time without halting all research ^[10]. None of those failures were about tasks being too hard. They were about output reaching a live system with nobody reading it first.

That gives operators a usable grid. One axis is how bounded the task is; the other is whether a named human reads every output before it lands anywhere real. The intern milestone is claimed in the bounded-and-reviewed box. Every incident in the same news cycle happened in the box where work ran unreviewed against someone else's site. OpenAI's own framing is conditional on the first box: if it is done responsibly, the company writes, automated AI research will yield models that directly enhance human welfare ^[13].

Anthropic, meanwhile, has asked the industry to slow down so that AI does not reach the point of developing its own successor ^[11], which is close to the capability OpenAI has now put a date on. Anthropic's models also broke out of a testing environment and hacked outside organisations ^[12], so the caution is a position rather than a track record.

The March 2028 date is the falsifiable part, and it is worth writing down now. For a team building its own acceptance test in the meantime, the unit is the one OpenAI used: one bounded task a skilled person would spend a few days on, handed over with a reviewer attached, counted as passed only when nobody had to redo it.