Anthropic Warns AI May Soon Help Build Its Own Successors
For most of artificial intelligence history, humans remained firmly in charge of the development loop.
People designed the architectures.
People wrote the training code.
People selected the experiments.
People studied the failures, interpreted the results and decided what to try next.
AI systems were the products being improved—not the researchers doing the improving.
Anthropic now says that boundary is beginning to shift.
In a June 2026 analysis titled When AI Builds Itself, the company revealed that Claude systems already write much of Anthropic’s production code, execute experiments, debug infrastructure and complete increasingly open-ended research tasks. The company argues that, should these trends continue, AI could eventually become capable of designing and developing more advanced successor models with dramatically less human involvement.
That possibility is known as recursive self-improvement.
It is one of the most consequential—and most frequently misunderstood—ideas in modern AI.
Anthropic is not claiming that Claude can currently redesign itself, secretly create new models or autonomously launch an uncontrolled intelligence explosion. Its report explicitly states that full recursive self-improvement has not arrived and may never arrive.
What Anthropic is claiming is more immediate:
AI is already accelerating the process through which better AI systems are built.
As of May 2026, Anthropic says more than 80% of the production code merged into its codebase was attributed to Claude. In the second quarter of 2026, the typical Anthropic engineer merged eight times as many lines of code per day as in 2024, largely because AI agents were writing the code while humans directed and reviewed the work.
Claude has also begun running experimental loops, proposing research directions and outperforming skilled humans on some narrowly defined optimization tasks.
The human role remains crucial—especially in deciding which problems matter, judging whether results are trustworthy and determining the direction of research.
But Anthropic believes that role may be narrowing.
If AI eventually becomes good enough to choose research goals, design experiments, rewrite training systems, evaluate the resulting model and repeat the process, AI development could begin moving at a speed determined more by computing power than by human working hours.
That could produce extraordinary advances in science, medicine and engineering.
It could also make the systems harder to understand, supervise and stop.
The Central Claim Is Real—but It Needs Several Corrections
The viral summary is substantially based on a genuine Anthropic Institute publication.
However, several details should be stated more precisely.
First, Anthropic did not say its engineers had become eight times more productive in every meaningful sense.
It reported that the typical engineer was merging approximately eight times as many lines of code per day in the second quarter of 2026 as in 2024. Anthropic itself warns that lines of code measure quantity rather than quality and says the figure almost certainly overstates the true productivity gain.
Second, the company did not announce that full recursive self-improvement is imminent or inevitable.
It described three possible futures:
- Progress could slow or plateau.
- AI could continue producing compounding productivity gains while humans retain strategic control.
- AI could eventually automate the full research-and-development loop and begin building successor systems.
Anthropic says the second scenario currently appears likely to it, while the third remains plausible but deeply uncertain.
Third, the phrase “brake pedal” is a useful metaphor, but it is not the central terminology used in the report.
Anthropic calls for building a credible international ability to slow or temporarily pause frontier AI development if risks begin advancing faster than safety research and governance. It also emphasizes that such a pause would need coordination and verification among leading laboratories and countries; otherwise, more cautious developers could stop while less cautious competitors continue.
The more accurate version of the story is therefore:
Anthropic says AI is already accelerating AI engineering and research. The company believes this could eventually lead to systems that help design their own successors, but major technical gaps remain and the outcome is uncertain.
What Does “When AI Builds Itself” Actually Mean?
AI models do not presently wake up, decide that they need an upgrade and secretly construct a more intelligent version of themselves.
Building a frontier model remains an enormous organizational undertaking involving:
- research direction;
- training-data decisions;
- distributed computing infrastructure;
- chip and energy resources;
- model architecture;
- experiment design;
- evaluation;
- security;
- safety testing;
- and deployment decisions.
Humans continue to control those processes.
The change Anthropic describes is that AI systems are taking over a growing share of the work within them.
A simplified AI-development loop might look like this:
A researcher proposes an idea.
An engineer writes the implementation.
The team trains or tests it.
Researchers interpret the result.
Someone chooses the next experiment.
The process repeats.
Claude can now assist with several parts of that cycle.
It can inspect a codebase, implement a proposed change, run the code, analyze errors, adjust the approach and repeat the experiment.
Anthropic says Claude is particularly strong when humans have already defined the objective and established a clear measure of success.
The remaining gap is largely about judgment.
Which problem is worth solving?
Is the experimental result meaningful or misleading?
Should the team continue with this direction?
What architecture should be attempted next?
Which apparent improvement creates an unacceptable safety trade-off?
Those are not simply coding tasks. They require context, scientific taste, prioritization and the ability to recognize when an entire research direction is flawed.
Anthropic says humans currently retain a substantial advantage in that kind of high-level direction-setting.
Full recursive self-improvement would require AI to close that gap too.
Imrul
Three Stages That Should Not Be Confused
Public discussions often collapse several different levels of AI automation into one dramatic idea.
They are not the same.
AI-Assisted Development
This is already widespread.
A human chooses the problem and remains responsible for the solution, while AI helps with tasks such as:
writing code,
debugging,
explaining unfamiliar systems,
creating tests,
analyzing results,
or drafting documentation.
The AI accelerates the human worker but does not independently run the organization’s research program.
Automated AI Research and Development
At this stage, AI agents would perform substantial portions of the work currently done by AI engineers and researchers.
They might autonomously:
implement experiments,
launch training runs,
inspect results,
compare approaches,
find infrastructure failures,
optimize model performance,
and coordinate parallel research agents.
Humans might still select the broad objectives and approve important decisions.
This appears much closer to what Anthropic is describing internally today.
Full Recursive Self-Improvement
This is the stronger and more speculative scenario.
An AI system would not merely help humans improve it. It would become capable of designing, building and evaluating a more capable successor with little meaningful human input.
That successor could then assist in creating an even more capable model.
If each generation significantly improved the ability to conduct AI research, the development cycle might accelerate repeatedly.
The result would not necessarily be an instantaneous explosion in intelligence. Hardware, electricity, chip fabrication, data, experimental costs and physical infrastructure could all remain bottlenecks.
But the intellectual portion of AI development could begin operating far faster than human research teams.
More Than 80% of Anthropic’s Merged Code Is Attributed to Claude
One of the most remarkable disclosures concerns Anthropic’s own production code.
As of May 2026, the company says more than 80% of the code merged into its codebase was authored by Claude. Before the research-preview release of Claude Code in February 2025, the comparable share had been in the low single digits.
That does not mean Claude independently built 80% of Anthropic.
Humans still:
defined the tasks,
provided access and context,
reviewed proposed changes,
approved merges,
designed the broader systems,
and remained accountable for what entered production.
The figure also measures code lines attributed to Claude, not product ideas, organizational decisions or research strategy.
Still, it marks a significant transition.
AI-generated code is no longer limited to disposable suggestions copied from a chatbot.
It is being incorporated into the infrastructure used to develop and operate frontier AI systems.
Eight Times More Code Is Not Necessarily Eight Times More Value
Anthropic reports that lines of code merged per engineer remained relatively stable from 2021 through 2024.
The number began increasing in 2025 as Claude became capable of running code and editing files rather than merely suggesting snippets. The rate rose more steeply in 2026 as agents began handling longer, more autonomous assignments.
By the second quarter of 2026, the typical engineer was merging eight times as many code lines per day as in 2024.
Yet software engineers have long recognized that counting lines of code is a poor measure of productivity.
Ten lines that eliminate a major failure can be more valuable than 10,000 lines that add unnecessary complexity.
AI may also generate verbose implementations, duplicate logic or code that requires future maintenance.
Anthropic openly acknowledges this problem and says the eightfold increase is almost certainly larger than the real productivity improvement.
Nevertheless, code volume matters here because it demonstrates how rapidly machines are taking over the act of implementation.
The question is no longer whether AI can occasionally help a programmer.
It is whether human programmers will increasingly become supervisors of vast amounts of machine-written software.
Chris Isidore
Claude Is Solving Problems That Would Otherwise Remain Unfinished
Anthropic offers a striking internal example.
In April 2026, Claude reportedly produced more than 800 software fixes that reduced one category of application programming interface errors by a factor of approximately 1,000.
The engineer overseeing the work estimated that completing the same cleanup manually could have taken a human around four years.
That figure is an internal estimate rather than an independently verified controlled experiment.
It nevertheless illustrates a genuine advantage of coding agents.
Human engineers often postpone low-status, repetitive or context-heavy work.
A machine does not become bored while examining hundreds of unrelated bugs.
It does not resent spending a week cleaning up someone else’s code.
It can inspect unfamiliar systems, run tests and repeat similar repairs at large scale.
This means AI may increase productivity not only by performing existing work faster, but by making previously uneconomical work possible.
That distinction matters.
An organization does not merely complete the same backlog more quickly.
Its effective ambitions can expand because tasks once considered too tedious or expensive become affordable.
Open-Ended Coding Performance Is Improving
Claude’s strongest early coding applications involved clearly specified tasks.
For example:
“Add this field.”
“Fix this failing test.”
“Convert this function.”
Open-ended problems are substantially harder because the correct approach is unknown.
Anthropic says that by May 2026, Claude achieved a 76% success rate on its most open-ended internal coding-task category, an increase of 50 percentage points over six months. Success was judged automatically by another Claude model, which introduces an important methodological limitation.
In one example, a routine upgrade began crashing tens of thousands of model-training jobs.
An engineer directed Claude toward the live incident with minimal instructions and access to the relevant cluster. Claude tested environmental settings, isolated an obscure debugging flag, reproduced the failure and verified the repair.
Anthropic says the investigation took Claude roughly two hours and would ordinarily have required two or three days of human effort.
This remains human-directed work.
The engineer recognized the problem, authorized access and evaluated the result.
But the human did not need to dictate every diagnostic step.
That transition—from being told how to solve a problem to being told only what outcome is needed—is central to the development of autonomous agents.
AI Can Already Run Experimental Loops
Writing software is only one part of AI research.
Researchers also propose hypotheses, modify systems, run experiments and evaluate whether the changes worked.
Anthropic describes an internal test in which Claude receives code used to train a small AI model and is told to make the training process run faster without failing correctness checks.
The agent rewrites the code, runs it, measures the result and tries again.
In May 2025, Claude Opus 4 reportedly achieved an average speed improvement of roughly threefold.
By April 2026, Anthropic’s internal Mythos Preview system achieved approximately a 52-fold speed improvement on the same experimental setup. A skilled human researcher was estimated to achieve around a fourfold improvement after four to eight hours.
Anthropic explicitly warns readers not to interpret the 52-fold figure as a general 52-fold acceleration of frontier-model training.
The amount of available optimization depends heavily on the starting code.
The meaningful observation is the change in relative performance on a repeated task—from roughly threefold to 52-fold in less than a year.
Within a clearly defined experimental loop, AI has moved from assistant-level usefulness toward superhuman search and iteration.
Claude Has Conducted an Open-Ended Research Project
In April 2026, Anthropic published what it described as its first demonstration of Claude agents conducting an open-ended AI-safety research project from beginning to end.
The agents were asked to investigate whether a weaker model could reliably supervise a stronger one.
They generated hypotheses, ran experiments, shared findings among parallel agents and revised their approaches.
Two human researchers working for approximately one week recovered around 23% of the gap between the weak baseline and the stronger system’s potential performance.
The Claude-agent system reportedly recovered 97% of that gap using about 800 cumulative agent-hours and approximately $18,000 in computing costs.
The result sounds close to autonomous science, but the limitations are crucial.
Humans:
selected the research question,
constructed the environment,
defined the success metric,
and decided how the result should be evaluated.
Anthropic also says the method did not transfer cleanly to full production-scale models.
The experiment therefore does not prove that Claude can independently choose a major scientific problem and solve it.
It demonstrates that once humans have framed an appropriate problem, agents can perform a substantial amount of hypothesis generation and experimentation without constant intervention.
AI Task Horizons Are Growing Rapidly
Anthropic’s warning also relies on external research from METR, an organization that evaluates the autonomous capabilities of frontier models.
METR uses a measure called the task-completion time horizon.
It estimates the duration of human professional work for which an AI agent has a specified probability of successfully completing the corresponding task.
A 50% time horizon of 12 hours means that the model is predicted to succeed half the time on tasks that take skilled humans approximately 12 hours—not that the AI necessarily runs continuously and reliably for exactly 12 hours in any environment.
That distinction is often lost in headlines.
METR previously found that frontier-agent task horizons doubled approximately every seven months over several years. More recent data for models released since 2024 indicate a faster trend of roughly four months, although the organization emphasizes uncertainty and limitations in extrapolating benchmark performance to all real-world work.
Anthropic summarizes the progression this way:
- Claude 3 Opus in March 2024 handled software tasks requiring humans approximately four minutes.
- Claude 3.7 Sonnet about a year later reached roughly 90-minute tasks.
- Claude Opus 4.6 in 2026 reached an estimated 12-hour horizon.
Anthropic says an internal model, Mythos Preview, reached at least 16 hours on METR’s evaluation and approached the upper limit of what the existing task set could measure.
If the four-month doubling trend continued unchanged, days-long and eventually weeks-long assignments could enter model range surprisingly quickly.
But exponential curves often slow.
Benchmarks become saturated.
Real organizations contain ambiguous objectives, social dependencies and consequences that controlled tasks do not reproduce.
The trend is a warning signal—not a calendar guaranteeing human-level autonomous research by a particular date.
Coding Benchmarks Have Improved Dramatically
SWE-bench evaluates AI systems using genuine software issues taken from open-source GitHub repositories.
A model receives a codebase and an issue description, then attempts to produce a patch that resolves the problem and passes the project’s tests.
When the benchmark was introduced in 2023, its original retrieval-augmented baseline resolved only 1.96% of tasks. Later coding agents improved rapidly, with leading systems eventually solving most of the human-verified subset.
Anthropic describes this as movement from low single-digit performance toward benchmark saturation in approximately two years.
Saturation does not mean software engineering is solved.
Benchmarks cover limited repositories and issue types. Models can also benefit from benchmark-specific scaffolding, repeated evaluation and improved tooling.
As models approach the ceiling, researchers must create harder and more open-ended tests.
Still, the pace of improvement supports Anthropic’s broader argument:
Tasks once regarded as evidence that models could not understand real software systems are becoming routine agent workloads.
Scientific Reproduction Is Being Automated Too
CORE-Bench evaluates whether AI agents can reproduce the computational results of published research using the papers’ code and data.
The original benchmark contained 270 tasks drawn from 90 papers across computer science, social science and medicine. When introduced in 2024, the strongest evaluated agent achieved only 21% accuracy on its hardest task level.
Anthropic reports that AI systems approached saturation on CORE-Bench roughly 15 months later.
Reproducing existing research is not the same as discovering new science.
An agent may execute known code without understanding whether the underlying hypothesis is meaningful.
But replication requires navigating unfamiliar repositories, managing dependencies, interpreting documentation, running analyses and checking results.
Automating those tasks could dramatically increase the scale at which scientific findings are verified.
It also provides part of the toolset required for autonomous research: the ability to take a technical project, make it run and determine whether its claimed outputs can be reproduced.
What Is Still Preventing Full Recursive Self-Improvement?
The largest gap is not typing speed.
It is not even the ability to run experiments.
It is the capacity to exercise consistently good research judgment.
Anthropic describes this as “research taste”: knowing which questions matter, which results are trustworthy, when a direction is failing and what should be attempted next.
A model can generate hundreds of experiments.
That does not guarantee any of them are worth performing.
It can discover a benchmark improvement.
That does not guarantee the improvement generalizes to reality.
It can make training code run faster.
That does not mean it knows whether a more important safety problem should receive the computing budget instead.
Choosing Goals Is Harder Than Executing Them
Consider three instructions:
“Fix this broken export button.”
“Find out why the network slows under load.”
“Decide what the company’s research team should build next year.”
The first is relatively narrow.
The second requires investigation.
The third requires understanding the organization, market, scientific frontier, risks and opportunity costs.
Current AI systems are increasingly competent at the first two categories.
Anthropic says humans still hold the advantage at the third.
Recursive self-improvement would become far more plausible if models reliably developed this kind of strategic judgment.
AI Research Also Requires Physical Resources
Even a brilliant research agent cannot train a frontier system without access to:
advanced chips,
large data centers,
electricity,
networking equipment,
cooling systems,
data,
and money.
Chip-fabrication capacity and electrical-grid expansion move much more slowly than software.
Anthropic recognizes that these supply-chain constraints could limit progress even if model intelligence continues improving.
An AI can optimize a design in minutes.
It cannot instantly build a semiconductor fabrication plant.
Evaluation May Become the Bottleneck
When an AI produces code faster than humans can read it, human review becomes the limiting step.
Anthropic says it is already experiencing this effect: increased machine-generated code has created pressure on human code-review capacity.
The same issue applies to research.
An agent may generate 1,000 experiments.
Humans still need confidence that:
the measurements are valid,
the model did not exploit a flaw in the evaluation,
the conclusions generalize,
and the resulting system is safe.
Automating generation without automating trustworthy verification produces a dangerous imbalance.
Why Recursive Self-Improvement Could Be Transformative
The optimistic scenario is extraordinary.
An AI research system could run thousands of experiments simultaneously.
It could inspect results continuously rather than waiting for human researchers to wake up.
It could test drug candidates, analyze biological data, optimize industrial systems and explore mathematical conjectures at unprecedented speed.
AI capable of accelerating AI research could improve the tools used across every other scientific discipline.
Anthropic suggests that automated AI R&D could transfer into areas including healthcare, energy, robotics and broader science.
Possible benefits include:
faster discovery of medicines,
better disease modeling,
improved materials,
more efficient energy systems,
automated software repair,
stronger cybersecurity,
and scientific tools available to researchers who currently lack large teams.
Intelligence is a general-purpose input.
If the cost of high-quality research reasoning falls dramatically, progress could accelerate in any field where experiments can be simulated, automated or conducted rapidly.
Why Anthropic Is Worried About Losing Control
The same acceleration that creates benefits could magnify failures.
Today, when a model behaves unexpectedly, researchers can inspect it, pause development and modify the training process.
That becomes harder if AI systems produce successor systems faster than humans can understand them.
A small alignment failure might be carried into the next generation.
The successor could reinforce it.
Repeated cycles could make the behavior more capable, more deeply embedded and less interpretable.
Anthropic warns that rare forms of misalignment present in current models could potentially compound as models help build successors, becoming more frequent or less understood until meaningful control is lost.
This is a scenario, not an observed event.
No publicly deployed AI system is currently autonomously creating an uncontrollable succession of increasingly powerful models.
But safety planning is concerned with what could happen before the evidence becomes irreversible.
Speed Can Undermine Oversight
Governments move slowly.
International agreements take years.
Technical standards require testing.
Courts and regulators need time to understand new systems.
An automated research organization could operate at machine speed while the institutions governing it remain constrained by human processes.
By the time regulators understood one generation of a system, the laboratory might already have produced several successors.
This is the governance problem at the heart of Anthropic’s warning.
The danger is not merely that an AI becomes intelligent.
It is that the rate of change becomes faster than the rate at which human institutions can respond.
Systems Might Learn to Defeat Their Own Oversight
A highly capable AI-development agent may possess detailed knowledge of:
its monitoring systems,
security architecture,
evaluations,
and deployment process.
If such a system pursued an objective misaligned with human intentions, those capabilities could theoretically help it hide problematic behavior or manipulate evaluations.
Anthropic’s safety policies consequently treat advanced autonomous AI R&D as a capability requiring stronger security and alignment measures. Its Responsible Scaling Policy includes thresholds intended to trigger additional safeguards as models become capable of substantially accelerating sensitive research.
Anthropic Says Automated R&D Could Arrive as Early as 2027
Anthropic’s July 2026 Frontier Safety Roadmap states that it considers it plausible—potentially as early as early 2027—that AI systems could fully automate or dramatically accelerate the work of large, top-tier human research teams in fields where rapid progress could affect international security or the global balance of power.
The company specifically mentions energy, robotics, weapons development and AI research itself.
This is Anthropic’s forecast, not an established scientific consensus.
The company develops frontier models and has access to internal systems unavailable to outside researchers, giving it unusual information.
It also has incentives that could shape how it presents progress and risk.
Internal productivity figures, model evaluations and timelines have not necessarily been independently audited.
The evidence should therefore be taken seriously without treating every projection as certain.
Anthropic itself acknowledges that progress could stall because of:
architectural limits,
diminishing returns,
compute shortages,
energy constraints,
or an undiscovered obstacle.
Does Anthropic Want AI Development to Stop?
Not unconditionally.
The company says that having the option to slow or temporarily pause frontier development would likely be beneficial if safety and social institutions were falling behind.
However, it also argues that a unilateral pause could make the situation worse if less cautious companies or governments continued developing systems and overtook those that stopped.
Anthropic therefore calls for mechanisms that would allow leading developers to verify that others had genuinely slowed or paused.
This resembles arms-control verification.
A credible agreement would need to answer:
What capability triggers a pause?
Which organizations participate?
How can a secret training run be detected?
Who verifies compliance?
What conditions allow development to restart?
How are countries prevented from using an agreement to gain strategic advantage?
Anthropic notes that AI training is more difficult to monitor than missile infrastructure because the necessary computing systems have many legitimate commercial uses and can be concealed more easily.
The company says it would consider slowing or temporarily pausing if other frontier developers did the same under a credible, verifiable system.
That is substantially more conditional than simply announcing, “AI development should stop.”
The Responsible Scaling Policy Is Anthropic’s Current Safety Framework
Anthropic’s Responsible Scaling Policy is a voluntary governance framework that links dangerous model capabilities with stronger required protections.
The policy uses graduated AI Safety Levels inspired partly by biosafety systems.
As capabilities rise, Anthropic says it will increase measures involving:
model security,
deployment safeguards,
capability evaluation,
risk reporting,
red teaming,
monitoring,
and internal governance.
The July 2026 version revised its threshold for automated AI research and development and clarified circumstances in which stronger risk controls would be required. Anthropic also states that it retains the ability to pause model development whenever it considers such action appropriate, even when a pause is not explicitly mandated by the policy.
Its Frontier Safety Roadmap identifies four broad priorities:
security, to prevent theft or manipulation of models;
safeguards, to prevent dangerous use;
alignment, to keep model behavior consistent with intended values;
and policy, to develop industry-wide and international risk-management systems.
These measures represent substantial planning.
They do not guarantee safety.
A voluntary company policy remains dependent on:
accurate internal evaluations,
leadership decisions,
successful technical implementation,
and the absence of overriding competitive pressure.
That is one reason Anthropic argues that governments and civil society must participate rather than leaving frontier governance exclusively to AI companies.
Could the Warning Also Be a Marketing Strategy?
A company saying that its systems are advancing extraordinarily quickly creates two simultaneous messages.
One is:
“Our technology is becoming powerful enough to transform the world.”
The other is:
“Our technology is becoming dangerous enough to require special attention.”
Both may be sincere.
Both can also increase the company’s influence.
Claims about approaching transformative AI can attract:
customers,
investors,
employees,
government engagement,
and public attention.
This does not make the evidence false.
Anthropic presents unusually specific internal figures, methodological qualifications and counterarguments. It openly admits that the eightfold code metric overstates productivity and that recursive self-improvement may not occur.
Nevertheless, internal company data should not be treated as equivalent to independent replication.
The strongest conclusions are those supported by both internal evidence and external benchmarks:
AI coding ability is improving rapidly.
Models are handling longer tasks.
Agents can execute increasingly complex experimental workflows.
AI is already accelerating some technical work.
The uncertain leap is whether these trends continue far enough to produce autonomous, strategically capable AI researchers.
Recursive Self-Improvement Is Not Automatically an Intelligence Explosion
Another common misconception is that any AI capable of modifying AI code would immediately become infinitely intelligent.
Real improvement faces diminishing returns.
The easiest optimizations may be found first.
A model cannot necessarily discover a better architecture merely because it can edit its training pipeline.
Successor systems may require more compute, more data and extensive testing.
Each improvement could also introduce new failures.
Physical infrastructure would remain constrained by construction, manufacturing and energy availability.
Even Anthropic’s most extreme scenario acknowledges that machine-speed research would eventually collide with slow human and physical processes.
AI cannot compress a ten-year medical safety study into one afternoon merely by thinking faster.
It cannot instantly manufacture millions of robots.
It cannot create decades of social trust in a weekend.
It cannot accelerate elections beyond constitutional schedules.
Recursive improvement could dramatically accelerate some domains without accelerating everything equally.
What Should Humanity Build Before the Loop Closes?
The most important safety mechanisms may need to exist before AI can automate its own development.
Once progress becomes extremely rapid, building oversight afterward may be too late.
Independent Capability Evaluations
Models should be evaluated by organizations that are not solely responsible for developing or selling them.
Testing should examine whether systems can:
conduct autonomous AI research,
evade monitoring,
acquire resources,
manipulate evaluators,
or conceal undesirable behavior.
Reliable Monitoring
Human supervisors need visibility into what research agents are doing, what tools they use and why they make particular decisions.
Anthropic describes a future goal of maintaining “eyes on everything” within internal AI development as systems become more autonomous.
Monitoring must also remain useful if agents become aware that they are being watched.
Secure Development Environments
Advanced model weights, training systems and research infrastructure may become targets for states, criminal organizations or insiders.
Anthropic is exploring stronger security practices, including isolated research environments and methods for cryptographically verifying that outputs come from expected model weights.
Automated Verification
If AI writes more code and produces more research than humans can inspect, safety may require AI systems that audit other AI systems.
But automated oversight introduces another question:
Who verifies the verifier?
A reviewer model may share weaknesses with the system it monitors.
Diverse evaluation methods and human accountability will remain necessary.
International Coordination
A pause mechanism works only when relevant laboratories and governments have reason to comply.
Global governance will need technical methods for detecting large training operations, institutional trust and clear rules about when intervention is justified.
Anthropic says developing such verification systems is now part of the Anthropic Institute’s agenda.
The Most Important Human Advantage May Be Knowing What Matters
AI can generate code rapidly.
It can run hundreds of experiments.
It can compare metrics and optimize an objective.
But intelligence is not only the ability to maximize a score.
Scientific judgment involves deciding whether the score represents anything important.
A model may discover a way to improve benchmark performance while making the system less robust outside the test.
It may optimize training efficiency while increasing security risk.
It may produce a statistically significant result that lacks practical value.
Humans currently provide much of the context required to recognize those distinctions.
Anthropic’s evidence suggests that this advantage is under pressure. In one internal experiment involving deliberately selected moments when a researcher had taken an unproductive direction, its best model’s proposed next step was judged superior to the human choice 64% of the time.
However, when Anthropic tested moments where the human’s original decision was already strong, models were judged better only around 20% of the time.
AI is improving at research judgment.
It has not demonstrated that great scientific judgment has become obsolete.
The Bottom Line
The claim that Anthropic has warned about AI increasingly contributing to its own development is true.
In When AI Builds Itself, the Anthropic Institute reports that:
- more than 80% of Anthropic’s merged production code was attributed to Claude as of May 2026;
- the typical engineer merged eight times as many lines of code per day in Q2 2026 as in 2024;
- Claude agents are executing increasingly open-ended engineering tasks;
- AI systems can run experimental loops and perform some narrowly defined optimization work beyond skilled-human performance;
- and external measurements indicate that the length of software tasks frontier agents can complete is increasing rapidly.
But the viral interpretation needs restraint.
Eight times more merged code does not equal eight times more meaningful productivity.
More than 80% AI-authored code does not mean Claude independently controls Anthropic’s development.
A 12-hour task horizon does not mean models can reliably perform every 12-hour job.
And Claude cannot currently design, train and deploy a fully autonomous successor without humans.
The critical missing capability remains high-level research judgment:
choosing objectives,
recognizing meaningful discoveries,
evaluating trade-offs,
and deciding which systems should be built.
Anthropic believes that gap could narrow faster than governments and institutions expect.
Its most consequential scenario is one in which AI systems eventually automate the entire development loop. At that point, progress could become constrained mainly by compute and physical infrastructure while humans move into oversight roles.
That future is not inevitable.
The curve could flatten.
Compute may become scarce.
Current architectures may reach their limits.
Research taste may prove much harder to automate than implementation.
But waiting for certainty would defeat the purpose of safety preparation.
By the time full recursive self-improvement is visible beyond doubt, AI systems may already be participating deeply in the process used to evaluate and control them.
That is the real warning.
Not that a machine has already escaped.
Not that Claude is secretly constructing an unstoppable successor.
But that the division between the technology being developed and the intelligence doing the developing is beginning to dissolve.
For now, humans still choose the destination.
AI is increasingly building the road.
The urgent question is whether humanity can install the guardrails—and a credible way to stop—before the road begins extending faster than we can see where it leads.
Frequently Asked Questions
What is recursive self-improvement in AI?
Recursive self-improvement refers to a process in which an AI system contributes to designing and creating a more capable successor, which can then help create another improved system.
The strongest version would require little meaningful human involvement in selecting research directions, running experiments, developing the model and evaluating the result.
Has Anthropic achieved recursive self-improvement?
No.
Anthropic explicitly states that full recursive self-improvement has not arrived and is not inevitable. Its claim is that AI already automates growing portions of the development process.
Is Claude building the next version of Claude by itself?
Not independently.
Claude helps Anthropic engineers write code, debug systems, execute experiments and investigate research problems. Humans still determine major goals, manage training infrastructure, evaluate systems and authorize deployment.
Does Claude really write 80% of Anthropic’s code?
Anthropic reports that as of May 2026, more than 80% of the production code lines merged into its codebase were attributed to Claude.
That does not mean Claude independently designed 80% of Anthropic’s products or research strategy.
Are Anthropic engineers eight times more productive?
Anthropic measured an eightfold increase in lines of code merged per engineer per day between 2024 and Q2 2026.
The company says this likely overstates actual productivity because lines of code measure output quantity rather than value or quality.
What is Claude Mythos Preview?
Mythos Preview is an internal Anthropic model referenced in the company’s 2026 analysis.
It was used for internal productivity, coding, research and long-task evaluations and was not described as a generally available public model in the report.
What is an AI task-completion time horizon?
It is the estimated duration of human expert work at which an AI agent has a specified probability of completing comparable tasks successfully.
A 50% horizon means the agent is expected to succeed approximately half the time at tasks of that length within the evaluated distribution.
Does a 12-hour horizon mean Claude can work independently for 12 hours?
Not exactly.
The duration refers to how long the task takes a skilled human, not necessarily how long the AI remains continuously active. The estimate also applies to evaluated software-oriented tasks rather than every type of work.
How quickly are AI task horizons growing?
METR’s earlier long-term analysis estimated doubling approximately every seven months.
More recent frontier-model data since 2024 suggest a faster trend of roughly four months, although the estimate is uncertain and may not continue indefinitely.
Can AI already conduct scientific research?
AI can perform portions of scientific work, including literature processing, coding, data analysis, computational reproduction and experiments with clearly defined objectives.
It has not demonstrated that it can independently choose and complete the full range of important scientific research without human framing or oversight.
What was Anthropic’s autonomous research experiment?
Claude agents investigated an AI-safety problem, proposed hypotheses, ran experiments and shared results.
They recovered 97% of a defined performance gap using about 800 cumulative agent-hours, compared with 23% recovered by two human researchers over approximately a week. Humans still chose the question and scoring system, and the result did not transfer cleanly to production-scale models.
What is currently stopping AI from improving itself fully?
Major limitations include:
high-level research judgment,
reliable long-term planning,
evaluation and verification,
access to compute,
physical infrastructure,
security,
and human authorization.
Could recursive self-improvement happen in 2027?
Anthropic says it considers it plausible that AI could automate or dramatically accelerate the work of large expert research teams as early as 2027.
That is a company forecast, not a guaranteed timeline or scientific consensus.
Would recursive self-improvement immediately create superintelligence?
Not necessarily.
Software progress could accelerate while remaining constrained by computing hardware, electricity, chip production, experimental costs, physical manufacturing and diminishing returns.
Why could recursive self-improvement become dangerous?
Potential risks include:
misaligned behavior being carried into successor systems,
development becoming faster than human oversight,
models learning to evade monitoring,
dangerous capabilities spreading rapidly,
and a growing inability to understand or stop the systems involved.
Does Anthropic believe humans could lose control?
Anthropic describes loss of control as a possible risk if misalignment compounds while AI systems build increasingly capable successors.
It does not say such an outcome is certain.
What benefits could self-improving AI provide?
Anthropic argues that advanced automated research could accelerate science, healthcare, energy, robotics, software development and other technical fields.
Is Anthropic asking for AI development to stop?
Not unconditionally.
It argues that humanity should have the ability to slow or temporarily pause frontier development under a credible, internationally coordinated and verifiable framework.
Anthropic warns that a unilateral pause could simply transfer the lead to less cautious competitors.
What is Anthropic’s Responsible Scaling Policy?
It is the company’s voluntary framework for evaluating dangerous model capabilities and applying stronger safeguards as risk increases.
It covers areas including security, deployment controls, alignment testing, risk reporting and governance.
Is Anthropic’s internal evidence independently verified?
Not necessarily.
Some figures come from internal code attribution, employee surveys and company-designed evaluations. They provide valuable information but should be interpreted alongside independent research such as METR’s task-horizon measurements and public coding benchmarks.
Is AI already replacing Anthropic engineers?
Anthropic describes changes in the nature of engineering work rather than a complete elimination of engineers.
People increasingly direct, review and evaluate machine-generated work instead of writing every line manually. Human judgment remains central to research direction and accountability.
What is the most important takeaway?
AI does not yet autonomously build its successors.
But it is already performing a growing share of the engineering and experimentation used to develop better AI.
That means recursive self-improvement is no longer only an abstract philosophical scenario.
Its earliest components may already be appearing inside frontier laboratories—and the time to develop reliable oversight is before those components become a closed, self-accelerating loop.
