A survey compiling interviews with 25 AI researchers by Severin Field and colleagues translates concerns about automated AI research into concrete, observable indicators. In a write-up published on July 8, 2026, and updated on July 28 of the same month, Field argues that some of the milestones cited by participants have already appeared. Indeed, on a narrow machine-learning task, under conditions where humans screened the candidates, one out of three AI-generated drafts reached workshop acceptance standard. Still, this falls short of recursive self-improvement, in which AI creates a strong successor that then goes on to improve itself further.
The Distribution of Views Captured by the 25-Person Survey
"AI Researchers' Views on Automating AI R&D and Intelligence Explosions," by Field, Raymond Douglas, and David Krueger, is a 26-page arXiv preprint revised on March 5, 2026. Its DOI is 10.48550/arXiv.2603.03338. It is not a peer-reviewed paper. Nor is the survey based on simulations or capability experiments—it consists of 40-to-60-minute semi-structured interviews conducted in August–September 2025.
The research team invited 182 people, of whom 25 participated. Recruitment channels included 7 authors of related literature, 8 people from NeurIPS- and ICLR-affiliated workshops, and 10 through researcher networks and snowball sampling. Participants included 7 current researchers at frontier AI companies and 4 former researchers. Nine were affiliated with universities, 3 with major tech companies or startups, and 2 with nonprofit organizations. Because this was not random sampling, and only 25 of the 182 invitees participated, the resulting proportions cannot be generalized to AI researchers as a whole.
Of the 25 participants, 20 named the automation of AI research and development by AI itself as one of the serious and urgent AI risks. Seventeen predicted that models with advanced coding and R&D capabilities would tend to be kept in-house within companies or governments. Among the 20 who gave a clear answer regarding disclosure policy, 10 predicted internal use only, 4 predicted public release, and 6 predicted a conditional or mixed approach.
What these figures capture is not causation but the perceptions of the 25 selected individuals. They do not measure how much research automation actually accelerates development speed or social risk. Because the first author single-handedly created the inductive coding scheme—using Claude to assist with the categorical classifications used in the figures—no inter-rater agreement among independent evaluators was calculated. The authors themselves do not claim the results represent a consensus among AI researchers, but rather a description of views held at a specific point in time.
Re-reading the "Crossed Milestones" in Chronological Order
Statements about milestones appeared in 22 of the interviews. References to coding capability appeared in 8, concrete proofs of existence in 5, and productivity growth as labor supply and researcher productivity gains each in 3. Examples included an agent that could complete tasks equivalent to 40 hours of human work with a 50% success rate, a model that generates 10,000 lines of correct code, top-tier performance at the International Mathematical and Informatics Olympiads, and a system capable of training a 7-billion-parameter model nearly autonomously.
These are not a pre-registered list of predictions with deadlines and judgment criteria. Rather, they are observational indicators that participants cited during the interviews as evidence that could substantially update their own views. Therefore, one cannot evaluate the predictive accuracy of the 25 participants simply by counting how many items were realized.
Chronology also matters here. Before the interviews even began, on July 21, 2025, Google DeepMind had announced that Gemini Deep Think scored 35 out of 42 points at the International Mathematical Olympiad, solving 5 of 6 problems with full marks. The solutions were produced in natural language, completed within the same 4.5-hour time limit as the competition, and certified by IMO graders as gold-medal level. The introduction to the original preprint itself notes that by August 2025, OpenAI's GPT-5 series models had also reached gold-medal level. While IMO-level mathematics constitutes proof of existence for a capability, it is difficult to call it a future prediction realized only after the interviews.
The ability to solve difficult math problems is not the same as the ability to select research questions, design experiments, and discard false hypotheses. In the interviews, too, 15 people distinguished between the ability to implement an idea and the ability to discern which ideas are promising. Sixteen interviews noted that constraints such as computing resources, data, and evaluation capability could hinder rapid acceleration. Crossing a milestone is material evidence that automation of the research pipeline is advancing—but it is not itself an observation of recursive acceleration.
One in Three Drafts and a PGR of 0.97: Automation Progress on Narrow Tasks
The most rigorously verified post-interview result is the peer-reviewed paper "Towards end-to-end automation of AI research," published in Nature by Chris Lu and colleagues on March 25, 2026. It appears in Nature, vol. 651, pp. 914–919, DOI 10.1038/s41586-026-10265-5. The AI Scientist generates machine-learning ideas, writes code, and runs computational experiments. It then analyzes the results, compiles them into a paper, and links the entire process through to automated peer review.
In human evaluation, three AI-generated papers were submitted to ICLR 2025's "I Can't Believe It's Not Better" workshop. One paper received review scores of 6, 7, and 6—an average of 6.33—meeting the acceptance threshold, while the other two fell short. The AI-generated submissions were included among the 43 papers reviewed at the workshop, but the authors note that this venue has a lower acceptance bar than the main conference. Because the research team, following their pre-registered experimental protocol, withdrew all three papers, what was actually achieved was passing peer review at an acceptance-level standard—nothing more.
Human involvement also remains embedded in the experimental conditions. The team selected promising candidates from the generated output at each stage, checking whether they fit the workshop's theme, whether the code ran as intended, and whether there were formatting issues. The scope was limited to machine-learning research that could be completed entirely within a computer. The authors themselves cite naive ideas, implementation errors, methodological weaknesses, and inaccurate citations as failure modes, and they do not regard the one-in-three result as evidence of stable, top-conference-level research capability.
The same paper also reports a correlation in which newer generations of foundation models tend to receive higher automated review scores. Each non-template setting was repeated 6 times and each template-based setting 3 times, with the correlation's p-value below 0.00001. However, this is not a causal experiment with random assignment of conditions other than model generation. It is not grounds for concluding definitively that newer models alone caused the improvement in paper quality.
Anthropic's Jiaxin Wen and colleagues went a step further, publicly releasing an autonomous research agent called AAR on the company's official research blog. While not a peer-reviewed paper, the code and experimental conditions have been made public. Running nine Claude Opus 4.6 agents for five days, totaling 800 hours, on a task where a weaker model supervises a stronger one, they raised the performance gap recovery rate (PGR) to 0.97. The best result achieved by two humans over seven days, tuning four existing methods, was 0.23. The cost, combining compute and API usage, was approximately $18,000—about $22 per agent-hour.
However, what AAR explored was an environment using a weak supervisor, Qwen1.5-0.5B-Chat, and a strong student, Qwen3-4B-Base, where answers could be automatically graded. Humans designed the problem, the evaluation API, and the nine initial research directions. When the top-performing approach was transferred to a production-scale training environment, the improvement was limited to 0.5 points on the evaluation metric—within the range of noise. Within the confines of a limited experiment, it surpassed the human baseline. Whether this holds up at different scales or with different tasks remains unconfirmed.
The Extension of the Time Horizon and Four Gaps Separating It From Recursion
The "50% time horizon" metric from METR, repeatedly referenced by interview participants, is not a figure indicating how many hours an AI has continuously operated. Rather, it is the value X such that an agent is estimated to succeed with 50% probability on a task that would take a human expert X hours to complete. The current Time Horizon 1.1 fits a logistic curve to performance on 228 tasks centered on software engineering, machine learning, and cybersecurity.
According to METR's long-term estimate, the 50% time horizon doubled roughly every 196.5 days—about 6.5 months—between 2019 and 2025. Using only data from 2024 onward yields a doubling time of 88.6 days, but METR cautions that estimates using this current task set are sensitive to analytical method due to a shortage of long-duration tasks, and that measurements above 16 hours are not reliable. A doubling of the time horizon does not mean automation rates or research productivity have doubled—because unless one also measures the human verification and repair required upon failure, the figure cannot be translated into practical speed gains.
Corporate observations carry the same boundaries. In May 2026, Anthropic reported that internally, more than 80% of code lines merged into production could be attributed to Claude. This figure was reportedly in the low single digits before the research preview launch of Claude Code in February 2025. The typical engineer's daily merged-line count, indexed to 2024 as 1, rose to 8 in the second quarter of 2026. However, absolute line counts, the number of people involved, and how the surveyed employees were selected have not been disclosed. The company itself explicitly states that lines of code do not represent quality and that the figure almost certainly overestimates true productivity gains. Even with before-and-after observations lined up side by side, this is not a controlled experiment that can isolate Claude as the sole cause of the increase.
What has been confirmed at present is a rapid improvement in the ability to write code and iterate on experiments against evaluable goals. Four gaps still remain before recursive self-improvement can be verified. Can AI itself choose research questions and evaluation methods? Can small-scale successes be transferred to production scale? Can an improved successor actually be measured to accelerate the pace of its own subsequent improvements? And can that outcome be reproduced across multiple generations even as human intervention decreases? The next decisive milestone will not be a count of papers or lines of code—it will be an experiment in which this entire chain can be replicated under identical conditions.