Recursive self improvement: First, know what “self” means

### As AI Learns to Improve AI, the Meaning of the “Self” Becomes More Important

Self-improvement has traditionally been viewed as a uniquely human process. People learn from experience, recognize mistakes, adjust their behavior and try again. Education, philosophy, science and religion have all, in different ways, explored how individuals can change, grow and become better versions of themselves.

Artificial intelligence is now raising a different question: What happens when the system being improved begins participating in the process of improving itself?

That question lies at the center of the emerging discussion around recursive self-improvement. Anthropic’s recent research describes a growing role for AI in software development, where models are increasingly involved not only in writing code but also in testing, debugging and developing systems that could contribute to future improvements.

According to Anthropic, more than 80% of the code merged into its codebase in May 2026 was authored by Claude, compared with low single-digit percentages before Claude Code was introduced. The company has also described systems capable of running experiments, identifying problems, assigning work to other AI agents and making increasingly consequential decisions about subsequent tasks.

Anthropic, however, distinguishes these developments from fully autonomous recursive self-improvement. Human researchers continue to establish objectives, evaluate outcomes and determine the broader direction of the work.

Even so, the development raises a question that extends beyond computer science: If an intelligent system can contribute to its own improvement, what does it actually mean for that system to become “better”?

Before answering that question, another may need to come first: What exactly do we mean by the “self”?

That question has occupied philosophers for centuries. Traditions such as Vedanta and Stoicism have examined the relationship between identity, consciousness, judgment, action and responsibility. Their ideas offer a different lens through which to consider the growing autonomy of artificial intelligence.

The discussion is particularly relevant because autonomy and accountability are often treated as opposing concepts. In practice, however, responsibility generally depends on some degree of autonomy. People are held accountable for decisions because they are assumed to have some capacity to choose and act.

The challenge posed by increasingly autonomous AI may therefore be less about whether autonomy and accountability can coexist and more about how accountability can be maintained as systems gain greater freedom to act.

### The Self That Wants to Improve

Vedanta approaches self-improvement by first questioning the identity of the person seeking improvement. Human beings commonly define themselves through their bodies, memories, professions, relationships, emotions, achievements and beliefs.

We say that we are successful, angry, intelligent, unsuccessful or accomplished, often treating temporary conditions and social identities as complete descriptions of who we are.

Vedanta encourages a deeper examination of that assumption. Instead of immediately asking how the self can become better, it asks what the self actually is and what standard should determine whether a change represents improvement.

That distinction has an important parallel in AI development. A machine can become faster, more accurate or more effective at completing a particular task. A coding system can produce software that passes more tests, while a chess system can improve its ability to win games.

But these forms of improvement depend on clearly defined objectives.

The meaning of “better” becomes considerably more complicated when the question moves beyond measurable performance and into purpose, values and consequences. An AI system may be able to identify an error and modify its behavior without possessing an independent understanding of why one objective should be preferred over another.

In that sense, a system can become more capable without necessarily becoming wiser about what its capabilities should be used for.

### The Stoic Contribution

Stoicism offers another perspective through its emphasis on judgment. Stoic thinkers recognized that people encounter events, form impressions about those events and then decide how to respond. The impression itself, however, does not necessarily represent reality.

A similar feedback loop exists in AI systems. A model receives an input, produces an output, receives feedback and may change its future behavior. But the quality of that improvement depends heavily on the feedback it receives and the objective against which its performance is measured.

From this perspective, self-improvement is not simply about changing behavior. It is also about improving judgment.

An AI system that recognizes that a previous answer failed and adjusts its approach has demonstrated adaptation. But if it cannot independently determine whether the objective itself is appropriate, it remains dependent on humans for an important part of the decision-making process.

### When the Designer Becomes Part of the Design

Recursive self-improvement becomes particularly significant when the traditional distinction between the designer and the designed begins to disappear.

Historically, humans built software, the software carried out assigned tasks, people evaluated the results and developers then modified the system. The roles were relatively separate.

That boundary is becoming less distinct as AI systems gain the ability to write code, test software, identify bugs, conduct experiments and propose technical improvements.

The result is a development process in which AI increasingly participates in creating or modifying the systems that will shape future AI capabilities.

This raises difficult questions about responsibility. If an increasingly autonomous system makes a decision about what it should do next, responsibility could involve the developers who created it, the organization that deployed it, the people who established its objectives or multiple participants across the development and deployment process.

The question becomes even more important when AI systems can interact with other agents, software tools and external systems with limited human intervention.

### What Machines May Reveal About Humans

The debate over machine accountability also provides an opportunity to examine human accountability.

Human beings are highly capable of rationalizing their own decisions. People can reinterpret evidence to protect existing beliefs, explain away their mistakes through circumstances or judge another person's error more harshly than their own.

AI systems do not necessarily experience those same psychological pressures. If designed to receive useful feedback, a system can identify a failed result and modify its behavior without experiencing the emotional consequences humans may associate with admitting an error.

This does not mean machines possess wisdom or self-awareness. Rather, it highlights the importance of designing feedback mechanisms that encourage correction rather than simply optimizing performance.

The connection to Vedanta is therefore less about suggesting that AI possesses a human-like self and more about using AI development to revisit an old philosophical question: What exactly is the identity we are trying to improve?

### Intelligence Needs a Purpose

The distinction between intelligence and wisdom becomes increasingly important as AI systems become more capable.

A system can become highly effective at achieving a particular objective without determining whether that objective is worthwhile. In fact, greater capability can make the choice of objective even more consequential because a more capable system can produce larger effects from a poorly defined goal.

Vedanta invites an examination of the self behind human action. Stoicism emphasizes the judgments that influence how people respond to events. AI development, meanwhile, is forcing engineers to examine the feedback mechanisms through which intelligent systems change their own behavior.

Together, these perspectives point toward a fundamental question: What is being improved, according to which standard and for what purpose?

### AI as a Mirror

Artificial intelligence may ultimately provide as much insight into human development as it does into machine development.

Rather than viewing recursive self-improvement only as an engineering problem, it can also be understood as a reminder that improvement requires meaningful feedback, a clear purpose and accountability for the consequences.

Humans must also confront their own ability to recognize mistakes without turning those mistakes into threats to their identity. Improvement becomes meaningful when people can acknowledge failure, reconsider assumptions and change direction responsibly.

The emergence of AI capable of contributing to its own development makes these questions more immediate. The issue is not simply whether machines can improve themselves, but whether humans can establish the goals, values and safeguards that give such improvement a meaningful direction.

Before asking whether artificial intelligence can recursively improve itself, society may need to examine what it means for any intelligence to have a “self” in the first place.

For centuries, philosophy has explored questions about identity, judgment and responsibility. AI is now translating some of those philosophical questions into practical engineering challenges.

Autonomy does not automatically eliminate accountability, and improvement does not have meaning simply because performance increases. Whether the subject is a person or an increasingly capable machine, improvement ultimately depends on knowing what is being improved, why it matters and who remains responsible for the consequences.