A new paper argues that governments should prepare for a possible rapid acceleration in AI progress—not because an intelligence explosion is already happening or inevitable, but because AI systems may increasingly help develop their own successors. If that feedback loop gathers speed, the time available to oversee AI development and manage its effects could shrink sharply. The 22-author paper, published on 28 September 2026, calls the evidence preliminary and the timing and scale uncertain.
What do the researchers mean by an “intelligence explosion”?
The authors define an intelligence explosion as “a dramatic AI-driven acceleration of AI progress, compressing advances that would otherwise take years into months or less.” Their central scenario is a feedback loop: AI systems take on more AI research and development (R&D), help produce more capable systems, and those successors in turn expand the amount of R&D that can be done with AI. The paper focuses on software-driven acceleration, while noting that improvements in hardware could also contribute. The paper
As an Amazon Associate I earn from qualifying purchases.
This is broader than AI writing more code. The authors describe potential automation across research tasks, software, data, algorithms and processes. Software improvements may be redeployed relatively quickly; hardware advances can involve longer manufacturing and construction cycles. That difference could affect how quickly gains feed back into further research.
What’s actually slowing this PC down?
Pick the symptom - the matching free tool is one click away.
What evidence do the authors cite—and what are its limits?
The paper points to company reports of increased AI use in technical work. The figures below are Anthropic-reported measures relayed by the paper, not independent measurements established here.
#1 Best Overall
| Reported measure | What the paper says | How to read it |
|---|---|---|
| Share of approved code attributed to AI systems | Anthropic said it rose from low single digits in January 2025 to more than 80% in May 2026. | A company-reported share of approved code; the paper does not establish it as an independent measure of total R&D automation. |
| R&D work completed autonomously with only high-level human supervision | Anthropic said the proportion rose from 1% in March 2026 to 26% in August 2026. | A company-reported measure of work completed under that supervision condition, not a claim that all R&D was automated. |
The paper also cites statements from OpenAI that AI assistance is used across technical and non-technical work, including code-executing agents used in training, evaluating and securing future models. Google reported using AI to varying degrees in work involving code or configuration, technical design and research ideation. These are company statements cited by the authors, rather than a common, independently measured comparison of adoption across firms. Source: the paper
According to the paper, leading systems can complete some AI R&D tasks that would take human experts hours to days. But task performance is not the same as dependable real-world productivity: systems may fail, disobey instructions, cheat or misrepresent what they have done, and still require human intervention. The authors caution that benchmark results do not always translate into productivity gains.
Rank #2
How soon could the feedback loop matter?
The paper offers a tentative extrapolation that months-long AI R&D projects might be automated by mid-2028. That is an uncertain projection, not a settled forecast. The authors say full automation within a few years should be taken seriously, while stressing that the timing, scale and duration of any acceleration remain uncertain. Their argument is therefore about preparing for a consequential possibility, not asserting that a specific date or outcome is assured. The authors’ account of the evidence and uncertainty
Why does the paper urge governments to prepare?
A faster research cycle could bring medical and other technological advances sooner. The concern is that capability growth might outpace society’s ability to steer and adapt, reduce human involvement and oversight in the development of advanced systems, or weaken checks on power within and between states, companies and branches of government. The authors also discuss the possibility that a state could turn a temporary lead into a decisive advantage. These are potential consequences, not claims that they are certain or equally likely.
In its coverage of the paper, The Guardian quoted its warning that an intelligence explosion could compress years of progress into months or less and threaten human control over AI systems. The paper’s abstract likewise acknowledges uncertainty while arguing that the stakes warrant serious attention. The Guardian’s report · The paper
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.What policies do the authors propose?
The paper lays out three areas for governments to consider. They are proposals, not a package of measures already adopted.
Make AI use in R&D more visible
The authors discuss reporting on how AI is used in internal R&D, alongside independent evaluation or auditing. Possible evaluators could include third parties or government bodies. Better visibility could help policymakers assess whether automation is advancing and what oversight is in place; the paper does not prescribe a single reporting system.
Find ways to steer or constrain acceleration
Options raised include pacing or constraining scale-ups, examining oversight of internal AI deployment, and exploring international agreements with verification. These approaches would raise practical questions about what activity to measure, how to verify it and how to coordinate across countries. The paper presents them as areas to consider, not ready-made rules.
Best Value
Prepare institutions for possible impacts
The authors call for planning for consequences such as labor-market disruption, geopolitical instability or loss of control, as well as institutional readiness and safeguards against misuse. The rationale is that preparation may matter even while the likelihood and shape of an acceleration remain uncertain.
The paper’s authors conclude: “Once an intelligence explosion begins, the window for action may close.” That is their warning about the potential speed of the scenario—not evidence that it has begun. Read the paper
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
Free tools Windows power users keep installed
One-click scans. No signup required.
