Google DeepMind has built a systematic measurement tool to answer a question that increasingly divides the AI community: Is research automation actually accelerating the pace of progress? According to Haydn Belfield, a policy researcher and engineer at the company, the data so far says no.
Belfield presented findings from an internal tracking effort called the General Capability Index (GCI), which aggregates performance across 92 benchmarks and 123 models. The index shows Gemini models improving at a steady rate of roughly 0.1 GCI points per month, with no statistically significant change in that speed over time. The latest model, Gemini 4 Argon, sits slightly above the trend line, but the team attributes that bump to greater compute allocation and organizational investment, not to any self-accelerating loop driven by AI itself.
For now, however, the data tells a story of steady, hard-won progress — not explosion, not plateau, but a line that keeps climbing at the same pace it has for months.
The index was designed specifically to detect the kind of explosive growth that some researchers fear. If AI systems began automating the process of AI research itself, the thinking goes, the curve could bend sharply upward, compressing decades of progress into months. The GCI is meant to serve as an early-warning system for that inflection point. For now, the curve remains stubbornly linear.
What the General Capability Index actually measures — and what it misses
The GCI v1 covers a broad sweep of model capabilities across language understanding, reasoning, coding, mathematics, and multimodal tasks. By normalizing scores across benchmarks, it produces a single metric meant to track aggregate capability growth. The team behind it acknowledges limitations: no index can capture every dimension of intelligence, and benchmark saturation can mask stagnation. But the tool’s purpose is directional rather than absolute. If AI-driven research automation were real, the index should eventually show a detectable departure from the previous trend.
That departure has not arrived. The steady climb of 0.1 points per month has held across multiple model generations, with no acceleration visible in the data. Belfield’s presentation of the GCI is the first public accounting of an internal monitoring system that Google DeepMind has built to track this exact question, suggesting that the company takes the risk of recursive self-improvement seriously enough to invest in measurement infrastructure, even while concluding the risk remains low for now.
The team categorizes the current risk from AI-driven research automation as low, but has built the measurement framework to detect signs of acceleration going forward. In effect, Google DeepMind is preparing for a scenario it does not yet see unfolding.
The gap between automation hype and reality inside DeepMind
Belfield also detailed the extent to which AI already permeates the company’s own engineering workflows, and the data reveals a more nuanced picture than the headline numbers suggest. AI now writes more than 75 percent of production code at Google DeepMind. But that figure comes with a critical caveat: approximately 88 percent of code changes still require human oversight or review, and every change is reviewed by a human before submission.
The share of code changes fully delegated to AI has risen from 7.5 percent to 11.9 percent, a meaningful increase but far from the autonomous workflow that would signal a fundamentally different mode of operation. The pattern suggests that AI is most useful as a drafting tool for well-understood, bounded tasks, while humans remain indispensable for reviewing, validating, and deciding.
That dynamic becomes even clearer when task complexity increases. Across 1,389 internal runs of Argon, the company’s agent system, the success rate without any human input dropped sharply as task complexity grew. On DeepMind’s in-house RE-Bench, which comprises 74 realistic machine learning engineering tasks, Argon solves roughly half of them on its own. That is a significant jump from the 27 percent solved by Gemini 3.7 Flash, but it still leaves the system far from the near-perfect performance that full automation would require.
Where humans remain irreplaceable — for now
Belfield identified specific areas where human involvement remains essential: creative research, novel experiment design, diagnosing subtle training instabilities, and long-term planning. A survey of twelve senior DeepMind researchers confirmed the pattern, with automation running highest for well-defined, verifiable tasks and lowest for open-ended scientific ideation.
This finding aligns with broader research on AI agent reliability. Studies have shown that AI agents tend to overstate their results and struggle with tasks that require sustained reasoning, adaptation to novel circumstances, or judgment about what constitutes a meaningful result in the first place. The gap between what agents can do in controlled, benchmarked settings and what they can be trusted to do autonomously in real research environments remains wide.
Why the intelligence explosion hypothesis persists despite the data
Belfield’s findings place him and Google DeepMind squarely on one side of a growing debate. On the other side, more than 20 leading AI researchers including Geoffrey Hinton, Yoshua Bengio, and OpenAI’s Jakub Pachocki recently published a paper warning that AI systems could automate all AI research and development within a few years, potentially triggering an “intelligence explosion” that compresses decades of progress into months.
The warning reflects a theoretical possibility grounded in the logic of recursive self-improvement: if an AI system can design a better AI system, and that system can design an even better one, the feedback loop could produce rapid, discontinuous progress. The argument is mathematically plausible, but the empirical evidence for it happening in practice remains thin, at least according to the data Google DeepMind has collected internally.
That does not mean the scenario is impossible. Belfield’s own presentation acknowledged that the GCI is a first attempt, and that more sophisticated tracking tools will be needed as model capabilities grow. What the data does suggest is that the transition from AI as a productivity tool to AI as an autonomous research engine is not happening yet, and may require breakthroughs beyond current architectures and scaling approaches.
The structural barriers to self-improving AI
Several factors appear to be keeping the acceleration at bay. One is the fundamental nature of research: it involves formulating hypotheses, designing experiments to test them, interpreting results that often contradict expectations, and making judgment calls about which directions are worth pursuing. These are not well-structured tasks with clear success metrics. They require context, intuition, and the ability to recognize when a negative result is actually informative.
Another barrier is reliability. Even the most capable current systems, including Argon, fail a significant fraction of tasks when left to operate without human intervention. In research settings, where errors can cascade into wasted compute time, misinterpreted results, or lost insights, that failure rate is unacceptable. Until agents can operate with near-perfect reliability on open-ended tasks, humans will remain in the loop.
A third factor is the nature of the code being written. While AI can generate syntactically correct code for well-known patterns and libraries, research code often involves novel implementations, unusual combinations of techniques, or experiments at the boundaries of what existing frameworks support. In those cases, human review catches not just bugs but conceptual errors that an AI system trained on existing code would not recognize as mistakes.
What 75 percent AI-written code actually means for the research pipeline
The statistic that 75 percent of production code is now written by AI has been widely cited as evidence of automation’s reach, but it deserves careful unpacking. Production code at a company like Google DeepMind includes infrastructure, testing, data pipelines, and the scaffolding around experiments, much of which follows well-established patterns. AI excels at generating such code quickly, reducing the time researchers spend on boilerplate and routine implementation.
The fact that 88 percent of code changes still need human oversight, however, indicates that AI’s role is closer to that of a highly productive intern than an autonomous engineer. It drafts; humans refine and approve. The 11.9 percent of changes fully delegated to AI represent tasks where the system’s output is trusted to be correct without modification, likely in narrow, well-audited domains.
This pattern of selective delegation mirrors what other large tech companies have reported. AI raises the ceiling on what individual engineers can produce, but it does not yet eliminate the need for senior judgment. The bottleneck has shifted from writing code to validating code, a change that is real but not revolutionary.
Featured snippet: What is Google DeepMind’s General Capability Index?
The General Capability Index (GCI) is an internal Google DeepMind benchmark aggregator that combines performance data from 123 AI models across 92 separate benchmarks into a single metric. It is designed to track the growth of overall model capabilities over time and to detect any acceleration that might indicate AI-driven research automation. According to the index, Gemini model capabilities have been growing at a steady rate of about 0.1 GCI points per month with no statistically significant change in speed, suggesting that AI systems are not yet improving themselves in a self-reinforcing loop.
The RE-Bench results and what they reveal about agent limitations
RE-Bench, DeepMind’s in-house benchmark for realistic ML engineering tasks, provides a concrete illustration of where current agent capabilities stand. Of 74 tasks that range from data preprocessing and model training to debugging and hyperparameter tuning, Argon solves about half autonomously. That represents a near-doubling of performance from Gemini 3.7 Flash, which solved 27 percent, but it also means half the tasks still require human intervention.
The benchmark is designed to reflect actual research workflows, not toy problems. Tasks require multiple steps, tool use, and adaptation to unexpected outputs. The fact that the best-performing system still falls short on half the tasks underscores the gap between narrow AI capabilities and the kind of general research competence that would enable autonomous discovery.
Importantly, the improvement from Flash to Argon came primarily from increased compute and organizational investment, not from a fundamentally new architecture or training approach. That mirrors the GCI data: progress is real, but it follows a predictable scaling curve rather than a sudden leap.
The debate over whether automation is accelerating or plateauing
Belfield’s findings align with a growing body of research suggesting that AI agents remain far from autonomous research. Recent studies have shown that agents overstate their own results, perform poorly on out-of-distribution tasks, and require significant human oversight even in controlled environments. The pattern is consistent enough that some researchers have begun to question whether the current scaling paradigm will ever produce the kind of recursive self-improvement that the intelligence explosion hypothesis assumes.
Others argue that the data from Google DeepMind is too narrow to settle the question. The GCI covers only Gemini models, and only within a specific time window. If a future model architecture or training method produces a genuine breakthrough, the curve could still bend upward. The infrastructure that Belfield’s team has built is designed exactly to detect that moment when it arrives, but its absence so far does not prove it will never come.
What it does prove is that, as of late 2025, the much-feared or much-hoped-for self-improving AI has not materialized within one of the world’s leading AI research labs. That is a significant data point in a debate that has been dominated by theoretical arguments and extrapolations from narrow benchmarks.
Why Google DeepMind built the tracking infrastructure in the first place
The decision to create the GCI and to run the Argon deployment studies reflects a deliberate effort to replace speculation with measurement. Belfield’s presentation of the data is notable not just for its conclusions but for its existence: a major AI lab is publicly sharing its internal monitoring frameworks, along with the sobering finding that the feared acceleration is not happening yet.
This transparency serves multiple purposes. It provides a baseline against which future progress can be measured. It informs policy discussions about AI risk by grounding them in empirical data rather than hypothetical scenarios. And it sets a precedent for other labs to share similar data, which would allow the broader field to track whether acceleration is occurring anywhere, even if it has not yet appeared at DeepMind.
The risk of AI-driven research automation is currently rated low by the team, but they have built the tools to track it going forward. That is a prudent approach to a problem that is too important to be left to intuition or extrapolation.
What the steady GCI curve means for AI governance and safety research
For policymakers and safety researchers, the GCI results offer both reassurance and a caution. The reassurance is that the most alarming near-term scenarios — sudden, runaway improvement leading to superhuman capabilities within months — are not supported by the available data. The caution is that steady growth can still produce transformative capabilities over years, and that a regime change in the trend could happen with little warning.
Safety research that assumes a gradual, predictable scaling path may be adequate for current conditions, but it must remain flexible enough to adapt if the signal changes. The GCI provides a way to detect that change in near real-time, but only if the monitoring infrastructure is maintained and expanded as models evolve. Belfield’s work suggests that Google DeepMind is treating this as an ongoing commitment rather than a one-time study.
The steady 0.1-point-per-month climb also implies that the timeline for advanced AI capabilities is longer than some of the more dramatic forecasts suggest. That does not mean there is no urgency, but it does mean that the urgency should be based on realistic projections rather than worst-case theoretical possibilities. The difference matters for regulation, for investment, and for public understanding.
Balancing the two narratives: automation is real, but not recursive
The evidence from Google DeepMind can be summarized in a single, nuanced statement: AI is already automating significant portions of the research and development workflow, but it is not accelerating the rate of AI progress itself. The distinction is crucial.
Seventy-five percent of production code is written by AI, and agents can solve half of realistic ML engineering tasks without help. Those are real advances that increase productivity and compress timelines for routine work. But none of this has produced a detectable increase in the rate at which models improve. The gains from automation are being absorbed into the existing scaling curve, not amplifying it.
That could change. A breakthrough in agent reliability, a new training method that is itself discovered by an AI system, or a hardware advance that removes a bottleneck could all trigger a departure from the current trend. The GCI will capture that departure when it happens. For now, however, the data tells a story of steady, hard-won progress — not explosion, not plateau, but a line that keeps climbing at the same pace it has for months.
Belfield’s findings do not settle the philosophical debate over whether self-improving AI is possible, but they do provide the most detailed empirical look yet at how far the current technology is from achieving it. The infrastructure to detect acceleration is in place. The acceleration itself has not arrived. That is a fact worth sitting with, even as the field continues to push the boundaries of what AI can do.
- Is research automation accelerating the pace of AI progress?No, according to Google DeepMind's General Capability Index, AI progress remains linear at 0.1 GCI points per month with no acceleration.
- What does the General Capability Index v1 measure?It measures model capabilities across 92 benchmarks and 123 models, including language understanding, reasoning, coding, mathematics, and multimodal tasks.
- What percentage of production code at Google DeepMind is written by AI?AI now writes more than 75 percent of production code at Google DeepMind, but this has not accelerated the rate of model improvement.
