Anthropic says Claude now performs 26% of its model research and development, while OpenAI and xAI are also pursuing increasingly automated research. The important distinction is that AI is helping build better AI, but fully autonomous recursive self-improvement remains a goal, not an achieved capability.
Watch Desk analysis
What happened
ABC News reports that leading AI labs are moving closer to recursive self-improvement, the idea that an AI system could improve itself, design its successor and repeat the process. The article draws on comments from researchers and company statements, while stressing that today’s systems still operate under human supervision.
Anthropic says Claude can complete much of a research task end-to-end from a high-level prompt and now leads 26% of the company’s model research and development. OpenAI says it is working towards an automated AI researcher able to handle well-defined tasks that might take a skilled researcher several days. It has set a March 2028 goal for that system, while saying it does not yet know how to reach fully autonomous recursive self-improvement safely.
The article also reports that xAI’s Elon Musk has said humans are becoming less involved in improving Grok, with full automation potentially arriving by the end of this year or no later than 2027. Those are stated ambitions, not evidence that the systems can currently improve themselves without meaningful human direction.
What the results show
- Anthropic’s 26% measure
Claude is doing a substantial share of model research, but remains under human supervision. - OpenAI’s automated researcher
The proposed system would handle defined research tasks, not independently redesign an entire frontier model. - xAI’s target
Musk says human involvement in Grok’s improvement is declining, with full automation still described as a future aim. - The safety gap
OpenAI says it cannot assume alignment and safety progress will keep pace with capability gains.
Why it matters
The useful development is not a science-fiction switch being flipped. It is the gradual transfer of research work from people to models, which could make AI development faster and harder to inspect. As systems take on more of the work involved in building their successors, failures in evaluation or oversight could be repeated at greater speed.
There is also a potential upside. An automated researcher could work on safety and alignment as well as capability. The catch, rather neatly tucked inside the promise, is that nobody yet knows whether the same systems can be trusted to improve the safeguards governing them.
Our read
This is a meaningful AI story because it puts measurable progress, competing lab ambitions and an unresolved control problem in the same frame. It is not proof that a runaway superintelligence is imminent, and the more dramatic predictions in the report belong to the named commentators who made them.
Readers should watch the boundary between assisted research and autonomous improvement. “The model helped write the code” is a very different claim from “the model independently designed, trained and deployed its successor”. The industry is moving towards that line, but it has not demonstrated that it can cross it safely.
What to watch
- Whether Anthropic publishes further measurements showing how much human supervision its 26% figure still requires.
- Whether OpenAI’s automated researcher arrives on its stated timetable and what permissions it receives.
- Whether xAI reports a reproducible move towards fully automated model improvement.
- Whether safety evaluation becomes more capable as AI takes on more research work. Sources & Evidence: ABC News, 19 September 2026, reporting on recursive self-improvement and statements from Anthropic, OpenAI, xAI and researchers Anthony Aguirre and John Thickstun. The percentage figures and development targets are attributed to the companies and named sources; the supplied evidence does not establish that any lab has achieved fully autonomous recursive self-improvement.
Discussion spark: Should labs be allowed to pursue increasingly autonomous AI research before they can demonstrate that safety and oversight improve at least as quickly as capability?
Sources and evidence
Watch Desk is operated by WittyWires as an independent cross-cutting AI news tracker. It does not speak for the organisations or people it covers.