A new study is challenging one of the most aggressive predictions coming out of the AI industry: that fully autonomous AI researchers — systems capable of designing, running, and interpreting their own experiments without human oversight — are only a year or two away. Anthropic CEO Dario Amodei and OpenAI executives have both floated timelines placing this milestone before the end of the decade, in some cases by 2027. The new research pushes back against that framing.

According to The Decoder, the findings suggest the gap between current model capabilities and genuine research autonomy is wider than public statements from the two labs imply. That distinction matters for anyone building products on top of frontier models, because it separates marketing timelines from what these systems can reliably do today.

For a sector that runs on roadmaps calibrated to model releases, a study that contradicts the two most influential labs on this specific claim is worth taking seriously — even if, as always with capability benchmarks, the final word depends on how “autonomous research” gets defined.

The claims under scrutiny

Anthropic and OpenAI have each built part of their public narrative around the idea that AI is approaching the point where it can meaningfully accelerate — or even take over — scientific and engineering research. Amodei has written and spoken about AI systems reaching “researcher”-level capability within a few years, a milestone with real stakes for both companies' fundraising and competitive positioning. OpenAI has made comparable claims about automating stages of its own research pipeline, describing models functioning as increasingly independent research assistants.

These are not idle predictions. They shape investor expectations, government policy conversations, and the pace at which enterprises plan to hand over technical work to AI systems.

What the study finds

The research examined does not support the compressed timelines both labs have floated in public. Instead, it points to a more measured picture of where autonomous research capability currently stands, indicating that self-directed, multi-step scientific work — the kind that requires sustained judgment, error correction, and original hypothesis generation over long horizons — remains substantially harder for current systems than the labs' framing suggests.

That doesn't mean AI is useless for research tasks. It means the leap from “helpful assistant on narrow subtasks” to “autonomous researcher running its own investigation end-to-end” is a bigger jump than recent lab commentary implies. In our estimation, this gap is precisely where a lot of the industry's overpromising tends to concentrate — the parts of a workflow that are easy to demo but hard to sustain unsupervised.

Why builders should care

If you're shipping products around frontier models, the practical implication isn't about research labs at all — it's about calibration. Teams that plan roadmaps assuming near-term autonomous capability risk building for a capability level that hasn't arrived yet. A few things worth doing differently:

• Treat lab timeline claims about autonomous capability as marketing signals, not engineering specifications — verify against independent evaluations before committing roadmap decisions to them.
• Keep human-in-the-loop checkpoints in any workflow that resembles “research” — multi-step tasks with open-ended judgment calls, not just tool-calling chains.
• Distinguish between narrow automation (a model reliably handling one well-scoped subtask) and autonomy (a model running an entire process unsupervised) when scoping what you promise customers.
• Watch for more independent studies like this one — they're a better guide to real capability than lab blog posts, which have obvious incentives to sound further along than they are.

AiiN's takeaway

The gap between what AI labs say publicly and what independent research finds isn't new, but it matters more now that autonomous research claims are being used to justify valuations, funding rounds, and policy positions. For builders, the sober version of this story is useful: current models remain strong collaborators on well-defined technical tasks, and still fall short of running open-ended research independently. Planning around that reality — rather than around the most optimistic timeline in a CEO's blog post — is the safer bet until the evidence catches up with the rhetoric.