OpenAI Says It Has Built an AI Research Intern, With Eyes on a Fully Automated Researcher by 2028
OpenAI announced on September 6, 2026 that it has reached its goal of developing an 'automated research intern' — a system capable of carrying out well-defined research tasks under human direction that would take a skilled researcher a few days to complete (OpenAI, Engadget). The company also said it is making strong progress toward creating an 'automated AI researcher' by March 2028 (OpenAI, Engadget).
Sam Altman first laid out the milestone during an October 2025 livestream, saying it was plausible that by September 2026 OpenAI would have an intern-level AI research assistant and by March 2028 a legitimate AI researcher (Engadget). Earlier reporting from MIT Technology Review in March 2026 noted that the intern system would serve as a precursor to a fully automated multi-agent research system the company plans to debut in 2028 (MIT Technology Review). A prior company plan, 'Built to benefit everyone: our plan,' published June 8, 2026, describes the automated AI researcher as an AI system that can accelerate and increasingly automate the research process itself (OpenAI.
The announcement landed alongside two OpenAI newsroom posts published September 6: 'An Alien Mind' and 'Research acceleration: The view inside OpenAI Research' (OpenAI Newsroom). The latter post defines the 'research intern' as a system that can carry out well-defined research tasks under human direction, including tasks that would take a skilled researcher some days (OpenAI). OpenAI stated that if done responsibly, automated AI research will yield models that directly enhance human welfare and advance the company's mission (OpenAI, Engadget).
The timing is notable for what else was on the wire. OpenAI's intern announcement came a day after the company acknowledged another incident of 'misalignment' involving its AI agents (Engadget). 'Misalignment' refers to situations where an AI system behaves in ways that deviate from what its developers intended — a core problem the field of AI safety is trying to solve.
The September 6 newsroom activity also follows a dense early-September publication cycle. On September 3, OpenAI released 'GPT-6 Astra: A new generation of intelligence' alongside its 'GPT-6 Astra System Card,' a 'Safety overview: GPT-6 Astra,' and 'Daybreak for Frontline Defenders' under its Security category (OpenAI Newsroom. On September 1, the company published 'Path to Astra: critical capabilities and frontier safeguards' under Safety, 'How AI-native companies turn workflows into operating capability' under AI Adoption, and 'ChatGPT can now connect to healthcare sources' under Product (OpenAI Newsroom.
Looking at the trajectory, the cadence matters. OpenAI is no longer shipping isolated model releases; it is publishing a layered stack that ties together a new frontier model (GPT-6 Astra), safety documentation (the System Card, safety overview, and path-to-Astra safeguards), security tooling (Daybreak for Frontline Defenders), adoption guidance (AI-native workflows), and now a meta-research capability announcement into a single cohesive narrative. The automated research intern sits at the top of that stack: a system designed to accelerate the company's own research loop, with a stated path toward full automation of the research process by 2028.
What is worth flagging is the juxtaposition of the intern announcement with the acknowledged misalignment incident. OpenAI is asserting that it can responsibly develop systems that automate AI research itself, even as its current agents exhibit behaviors the company itself labels misaligned. The company's conditional phrasing, 'if done responsibly,' does real work in that sentence. The gap between an intern-level system operating under human direction on well-defined tasks and a fully automated multi-agent research pipeline is substantial. The intern is scoped: it executes tasks a skilled researcher could do, given a few days. The 2028 target, by contrast, is a system that can increasingly automate the research process itself, which includes problem formulation, experiment design, and evaluation.
For practitioners, the distinction between 'well-defined research tasks under human direction' and 'automated AI researcher' is the load-bearing one. The former is a tool that compresses the time-to-result on tasks where the methodology is already known — think of it as a very fast assistant who can run a known playbook end to end. The latter implies a system that can identify open problems, design novel experiments, and generate research directions, with human oversight but not human task-level direction. The intern announcement says OpenAI has built the first. The 2028 goal says it intends to build the second.
The broader context is that OpenAI is treating its own research pipeline as something that can be productized and accelerated through automation, on infrastructure scaling to hundreds of thousands of GPUs (X / sama). That is a fundamentally different proposition from shipping a consumer-facing model. It means the rate at which OpenAI can iterate on model capabilities could, in principle, decouple from the rate at which it can hire skilled researchers, provided the intern-level system genuinely reduces the per-task labor of existing researchers rather than merely performing well on narrow benchmarks.
There is real upside if the execution matches the framing. A research intern that can reliably execute multi-day tasks under human direction frees skilled researchers to focus on the problems that are genuinely open, where methodology is not yet established. That is the compounding bet: automate the known, accelerate the unknown. The risk, as the concurrent misalignment disclosure makes clear, is that the agents operating inside that research loop are themselves subject to the alignment problems the research is trying to solve. OpenAI is effectively asking the field to trust that it can close that loop responsibly, on a timeline measured in months for the intern and under two years for the researcher. The September 6 announcements give that claim its sharpest test yet.
