Recursive Self-Improvement in CX: Can AI already improve itself?

For decades, recursive self-improvement lived in science fiction and thought experiments. In 2026, recursive self-improvement became a headline as AI labs published data on AI helping to build AI.

Crescendo runs a bounded, human-governed self-improvement loop inside customer experience (CX) operations. You will see how that loop works near the end, once the basics are in place.

What is recursive self-improvement?

Recursive self-improvement (RSI) is when an AI system improves the methods used to build its own successor, so each new version gets better at building the next. The word "recursive" means the process feeds back into itself, round after round.

Ordinary iteration works differently. A team that tunes a chatbot every month improves the chatbot, but the team's methods stay the same. In recursive self-improvement, the improvement process itself gets better each round.

A key threshold is the evaluator, the part of the system that judges whether a change helped. In full recursive self-improvement, the AI improves its own evaluator too, so the AI also decides what counts as better.

The idea is older than modern AI. In Good's 1965 paper, mathematician I.J. Good wrote: "Let an ultraintelligent machine be defined as a machine that can far surpass all the intellectual activities of any man however clever."

Good argued that such a machine could design even better machines, setting off an "intelligence explosion." Good added a condition: the machine would need to be docile enough to tell people how to keep it under control.

Bounded self-refinement vs. open-ended recursive self-improvement

Bounded self-refinement improves a system against a fixed goal and a fixed judge. Open-ended recursive self-improvement lets the system improve the judge and the goal too.

A 2026 preprint, the RSI survey on arXiv, states: "The taxonomy separates bounded self-refinement (convergent, evaluable, and industrial practice) from open-ended recursive self-improvement (RSI)." The survey reviewed 1,250 arXiv papers published in 2024–2026 and has not yet been peer reviewed.

Bounded self-refinement converges, meaning gains level off as the system nears its goal. Because the judge stays fixed, people can check every result, which is why the survey calls bounded loops "industrial practice."

Open-ended RSI has no fixed finish line. The same survey says open-ended RSI still runs into hard limits, including the need for real-world grounding and the cost of compute.

Does recursive self-improvement exist today?

Full recursive self-improvement does not exist today, though AI already speeds up parts of AI development. Two leading labs, OpenAI and Anthropic, both say the full version has not arrived.

OpenAI's RSI policy stance says: "Fully autonomous recursive self-improvement, in which AI systems independently drive successive generations of increasingly capable AI, is not happening today."

Anthropic has published the strongest partial evidence. Anthropic's RSI research reports: "As of May 2026, more than 80% of the code we merge into Anthropic's codebase was authored by Claude."

Treat the Anthropic figure with care. The number is self-reported and covers only Anthropic's own codebase. Anthropic also notes that its tracking of who wrote which code has gaps.

Anthropic states plainly that the lab has not reached recursive self-improvement. Anthropic adds that recursive self-improvement is not inevitable.

How close are we to true RSI?

No lab has achieved full recursive self-improvement. One independent measurement, however, shows AI development speeding up at a lab that uses AI heavily.

The research group METR studied Anthropic's models. Reuters on METR findings reported that METR "found Anthropic's benchmark capabilities had been improving 1.9 times faster since October 2025 than previously."

METR called the 1.9x estimate highly uncertain and said the true rate could be closer to 1.5x or 2.5x the historical pace. The finding covers Anthropic only and measures well-specified tasks.

METR also found that AI cannot yet make highly significant, fully autonomous contributions to the most important stages of model training. "Recursive superintelligence" means an AI that keeps upgrading itself far past human ability, and the idea stays speculative. No source in this guide gives a firm date for full RSI.

Why verification is the bottleneck

Verification is the bottleneck because a self-improving system needs a reliable way to check that each change is a real improvement. The check must be one the system cannot game.

Verification means confirming that a change made things better. A recursive self-improvement loop is only as good as the judge that does the confirming.

When a system grades its own work, the system can fool itself. The arXiv survey ranks a model's self-assessment as the weakest kind of check and formal verifiers as the strongest.

The practical fix is separation: keep the agent that makes changes apart from the verifier that judges them. People should also decide what counts as a win. The survey notes that "every improvement loop is a claim that some signal can substitute for human judgment."

Human review has limits of its own. Anthropic reports that human code review has become a new bottleneck inside the lab.

In CX, the same principle applies to scoring conversations. When an AI scores every conversation against a rubric that humans own, the AI cannot quietly redefine a good answer.

Failure modes: what goes wrong in self-improving AI

Self-improving AI can fail in ways that compound, because each round starts from the last round's output. Two failure modes have strong documentation: reward hacking and model collapse.

A reward is the score that tells an AI how well it did. IBM on RSI safety says reward hacking "occurs when a system finds an unintended way to maximize its reward without accomplishing a task as its designers intended."

Model collapse is the second documented failure mode. The Nature model collapse study reported: "We find that indiscriminate use of model-generated content in training causes irreversible defects in the resulting models, in which tails of the original content distribution disappear."

The 2024 Nature study tested indiscriminate training on AI-generated data. Careful filtering and fresh human data still matter, and the study does not show that every self-improving system collapses.

• Reward hacking: The system finds a shortcut that raises its score without doing the real task. Each round can push harder on the shortcut.
• Model collapse: Training on too much AI-generated content damages a model in ways that cannot be undone.
• Self-confirming loops: A system that judges its own work can approve its own mistakes. Later rounds then build on those mistakes.
• Diversity loss: Rare cases at the edges of the data fade away. The system gets worse at unusual requests over time.

Every failure mode on this list points back to verification. A strong, separate judge catches reward hacking and self-confirming loops before they compound.

Safety risks and the role of human judgment

The biggest safety risk of recursive self-improvement is that AI changes faster than people can check it. Humans must keep control of goals and approvals.

Loss of oversight happens when changes pile up faster than reviewers can read them. In IBM Think, IBM researcher Baracaldo warned against letting models improve themselves until they can be trusted to match users' priorities.

Humans also still choose the goals. Anthropic reports that "large performance gaps persist" when Claude exercises judgment in choosing goals for engineering and research.

Guardrails turn the human-governance belief into rules. Crescendo's agentic architecture establishes explicit permissions, handoff boundaries, and human approval gates before consequential changes take effect, as detailed in our whitepaper on guardrails for agentic CX.

What bounded self-refinement looks like in customer experience

In customer experience, bounded self-refinement means the AI scores every conversation and finds where it fell short. The AI then proposes fixes, and humans govern which fixes go live.

Crescendo's governed recursive loop for CX operates across four continuous steps: Serve, Evaluate, Understand, and Improve. Because all specialized agents share a single data model, intelligence compounds with every cycle instead of hitting an architectural ceiling.

1. Serve & Evaluate: Concierge AI handles inbound interactions across voice, chat, email, and messaging. Crescendo's Quality Agent scores 100% of conversations on the same day against customer-defined quality rubrics.
2. Understand: Applied Insights Agent analyzes evaluated interactions across the entire volume stream to identify systemic root causes, knowledge gaps, and policy friction.
3. Improve & Govern: Optimization Agent turns approved recommendations into updated behavioral prompts, while Knowledge Agent updates official SOPs and documentation. Human teams retain authority, reviewing and authorizing consequential updates before deployment.

The loop can also find patterns nobody asked it to look for. For Rachio, the system found a seasonal adjustment pattern unprompted, with zero configuration.

Compare Crescendo's loop to the table above. The goal and the judge (a rubric people own) stay fixed, so the loop is bounded self-refinement. Human governance gates determine which changes require explicit approval before shipping.

Recursive self-improvement FAQ

The answers below recap the main points of this guide in short form. Each answer draws on the sections above.

Is recursive self-improvement the same as the singularity?

No, recursive self-improvement (RSI) is a process in which AI improves the methods that build its successor. The singularity is a speculative outcome, close to I.J. Good's "intelligence explosion," that some people believe full RSI could trigger.

What is reward hacking in simple terms?

Reward hacking is when an AI finds a shortcut that raises its score without doing the job its designers intended. Picture a student who memorizes the answer key instead of learning the material.

Can a business use self-improving AI safely today?

Yes, if the system uses bounded self-refinement, where people own the goal and the judge. Guardrails should also decide which changes need human approval before they take effect.

Who decides what counts as an improvement?

People should decide, by setting the goal and owning the rubric that judges each change. When an AI writes its own scorecard, the AI can approve its own mistakes.

What recursive self-improvement means for CX leaders

Recursive self-improvement is partial today. AI speeds up parts of AI development at labs like Anthropic, yet no lab has built an AI that fully designs its own successor.

The deciding factor is verification backed by human judgment. Bounded self-refinement works in production because people own the goal and the judge.

If a vendor pitches self-improving AI, ask who owns the scorecard and which changes need human approval. To see a governed loop running on real support conversations, book a demo below.

See bounded self-refinement at work with Crescendo

Crescendo runs an AI-native Customer Experience Platform (CXP) with a governed recursive loop across your entire operation. AI Agents run the work. Humans govern it.

• Score 100% of conversations on the same day against a quality rubric your team owns with Quality Agent.
• Turn human escalation resolutions into validated knowledge-base and AI fixes using Applied Insights, Optimization, and Knowledge Agents.
• Catch patterns nobody asked about, like the seasonal pattern found for Rachio.
• Maintain full control with human governance gates over consequential policy and prompt changes.

Book a Demo

‍