AI is helping build AI. It has not yet closed the loop.

If an AI laboratory says its system is “building the next AI,” the phrase could describe anything from autocomplete with excellent publicity to a machine that independently creates a more capable successor. Those are not remotely the same achievement.

The distinction matters beyond the laboratory. Claims of self-improving AI influence how people understand the speed, usefulness and manageability of the technology. Fortunately, they can be translated into observable stages—and checked without accepting either dazzling promises or frightening forecasts as evidence.

Stage one: writing the code

AI systems are already doing substantial implementation work inside at least one frontier laboratory. Anthropic reports that Claude authored more than 80% of the code merged into its codebase in May 2026. It also says the typical engineer merged eight times as much code per day in the second quarter of 2026 as in 2024.

That is evidence of a large change in how the company produces software. It is not, by itself, evidence of an eightfold productivity gain. Anthropic explicitly warns that lines of code measure quantity rather than quality. Extra code can represent valuable software, unnecessary complexity or work that requires later correction; the count alone cannot tell us which.

The stronger observation is that AI has moved from suggesting snippets to editing files, running tools and carrying out longer sequences of engineering work. In Anthropic’s account, people increasingly provide a goal while Claude supplies much of the method. The company’s examples and measurements are informative, but they remain mostly company-reported evidence rather than independent replication. Anthropic describes its engineering measurements and caveats here.

Stage two: executing a human-chosen experiment

Research involves more than code. Someone must decide what question to ask, design a useful test, interpret the result and decide what deserves another attempt.

AI can now perform meaningful portions of that loop when the destination is supplied. Anthropic describes a recurring optimization test in which Claude receives code for training a small model and a fixed goal: make it run faster without failing predetermined correctness checks. The system rewrites the code, runs it, measures the result and tries again. Anthropic reports that its result improved from roughly a threefold speedup in May 2025 to roughly 52-fold in April 2026.

This is a real experimental capability within a bounded task. It still leaves people responsible for choosing the problem, defining success and deciding whether the result matters outside the test.

That boundary is easy to miss. A machine can be exceptionally good at searching a landscape after humans have drawn the map. Recursive self-improvement would require it to help redraw the map—and eventually decide which landscape deserves exploration.

Stage three: proposing what to try next

Anthropic also reports early movement toward research judgment. In one analysis, researchers selected 129 moments when a human investigation had taken an unhelpful detour. Models were shown the work preceding the detour and asked what should happen next. A separate Claude judge, which could see the eventual outcome, preferred the April 2026 model’s proposed next step over the human’s choice in 64% of those deliberately selected cases.

That result is narrower than “AI has better scientific judgment than people.” The sample was constructed from moments when the human decision was already known to have room for improvement, and another model supplied the judgment. It is better read as evidence that the system sometimes identifies promising next steps in realistic research situations.

Anthropic’s broader automation index makes the current boundary unusually plain. For August 2026, the company classified Claude as “leading” 26% of its weighted AI research and development work and at least “collaborating” on more than 90%. It classified no measured subset as fully autonomous. The index was also built and rated largely with Claude; exact agreement between model and human ratings was 59%, although ratings were within one level 97% of the time. Anthropic says shared methods and independent checking are needed before laboratories can be compared confidently. The methodology and its limitations are published with the index.

Stage four: improving how models learn

A system that merely writes more training code has not necessarily improved AI. A more revealing test asks whether it can change the learning algorithm itself: the rules that turn computation and data into a trained model.

A recent preprint introduces AI4AI-Bench, a collection of ten frozen research repositories designed around that distinction. Agents receive four hours on one accelerator to rewrite a training algorithm. Their submissions are then rerun from scratch and evaluated using a fixed test hidden from the agent.

Across 29 configurations of six systems, the mean normalized score was 0.166 and the best reached 0.250. On the authors’ scale, 0.1 represents the original repository algorithm and 1.0 the task optimum. Most submissions did not change how the model learned at all. The minority that did make such changes scored better on average than those that did not.

The experiment suggests that current agents can occasionally improve a bounded learning procedure. It also exposes the remaining distance: the strongest tested system closed less than one-fifth of the gap between the original algorithm and the stated optimum. Because this is a preprint and a particular benchmark, it should be treated as one repeatable measurement, not a universal verdict. The paper’s abstract reports the task design, submissions and scores.

Stage five: closing the loop

Genuine recursive self-improvement needs more than one clever optimization. The improvement must alter the process that creates the next AI, survive a fresh training run, produce a more capable successor and enable that successor to repeat the process.

A persuasive demonstration would therefore need receipts for the entire loop:

1. The system chooses or justifies a consequential research goal, rather than receiving one fully specified.
2. It changes the training method, not merely presentation, parameters or supporting code.
3. Independent evaluation shows that the resulting model is more capable under fixed tests.
4. The improvement persists when training is repeated from scratch and is not a scoring trick.
5. The successor uses the inherited improvement to make another verified improvement, with humans no longer supplying the crucial research decisions.

None of the cited evidence completes this chain. Anthropic itself says recursive self-improvement has not arrived and may not be inevitable.

A small test for the next large claim

When another laboratory says its AI is “building AI,” replace the slogan with five questions:

• Who chose the goal?
• Who defined success?
• Did the system change how the model learns?
• Was the claimed improvement reproduced from scratch?
• Did a more capable successor inherit it and improve the process again?

Also ask how reliability was measured. METR’s “time horizon,” for example, is not simply the number of hours an agent stays switched on. It estimates the duration of a task, measured by expert human completion time, at which an agent reaches a given probability of success. The current suite mainly covers well-specified software engineering, machine-learning and cybersecurity work; METR cautions that ordinary jobs are messier and that measurements beyond 16 hours are unreliable with its present tasks. METR publishes its definition, data and methodological limits.

The evidence today is neither “nothing has changed” nor “AI now invents its successors.” A more accurate—and more interesting—description sits between them: AI is doing much more of the implementation and experimental labor involved in building AI, while people still provide much of the judgment that determines what gets built and why.

That may change. The useful habit is to keep watching the joins in the loop, where a human still supplies the goal, the standard or the final interpretation. Those joins tell us more than the loudest verb in a laboratory’s headline.