More than 20 AI researchers warn that systems helping to develop successor AI could speed up progress and make human oversight harder. Their paper calls for greater visibility into how far companies have automated AI research and development, alongside independent oversight and plans for slowing or pausing dangerous acceleration.
Watch Desk analysis
What happened
The paper, published on 28 September by the University of Cambridge’s Programme on AI Science & Policy, argues that AI systems could increasingly help conduct the research that produces more capable systems. Its authors include Yoshua Bengio, Geoffrey Hinton, OpenAI chief scientist Jakub Pachocki and Anthropic co-founder Jack Clark. They wrote in their personal capacities, according to The Next Web’s account of the paper.
The authors say tentative extrapolations point to some months-long AI research projects being automated by mid-2028. They stress that this is uncertain: diminishing returns, limits on computing power and data, and tasks that remain hard to automate could slow progress.
Key findings
- A possible acceleration loop
AI could help develop more capable successors, potentially compressing years of progress into months, the authors warn. - A request for visibility
The authors propose standard reporting on how much companies automate their own AI research, with independent auditors among the measures they suggest. - Several brakes remain
The paper points to compute and data limits, diminishing returns, hard-to-automate work and long training runs as possible constraints. - Benefits and risks
Faster research could bring advances such as medical treatments sooner, the authors say, but they also warn of weaker human oversight and loss of control.
Why it matters
The question is no longer only what AI can do for users. It is also how much AI is becoming part of the research process that builds the next generation of systems, and whether people can still monitor that process effectively. That makes transparency about research automation a concrete policy question, not just a distant debate about superintelligence.
The paper describes a possible trajectory, not a prediction that an intelligence explosion will happen by a set date. Its authors themselves point to significant uncertainties and constraints. The Next Web reports the paper’s argument and proposals; the forecasts remain the authors’ analysis, not established outcomes.
Our read
This deserves attention because it puts a practical request beside a high-stakes warning: measure how much AI is doing inside AI research. That would not settle the safety debate, but it would give policymakers and the public a clearer starting point than trying to infer the pace from product launches. The important distinction is between tracking a plausible risk and treating its most dramatic outcome as inevitable.
What to watch
- Whether AI companies disclose how extensively they use AI in research and development.
- Whether independent audits or common reporting standards gain support.
- How the paper’s projections hold up against real-world evidence of automation.
- Whether proposed limits or pause mechanisms become specific policy proposals.
Discussion spark: Should companies developing advanced AI be required to disclose how much AI they use in their own research, or would that reveal too much about commercially sensitive work?
Sources and evidence
- AI researchers warn of risk of losing human control – Anadolu Agency – UA.NEWS (28 September 2026, 21:45 UTC)
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.