Is Recursive Self

On February 5, 2026, OpenAI introduced GPT-5.3-Codex and wrote in its release notes that early versions of the model were “instrumental in creating itself,” helping to debug training runs, manage deployment, and diagnose evaluation failures. It was the first explicit admission from a frontier lab that one of its models materially contributed to the engineering loop that produced its successor.

That admission is what some in the field call recursive self-improvement (RSI), the point at which large language models meaningfully improve the systems used to build, deploy, and evaluate the next versions of themselves. If the loop is real, progress stops being linear and leans towards the exponential.

“AI is now intelligent enough to meaningfully contribute to its own improvements,” wrote Matt Shumer, CEO of Otherside AI, in “Something Big Is Happening,” a February 9, 2026, essay that went viral.

But is the loop real, or is the announcement marketing language dressed in technical clothing?

Most experts answer that question with a control diagram, not a benchmark.

“Genuine recursive self-improvement and sophisticated internal automation are not separable on a benchmark,” said Nik Kale, principal engineer and member of the Coalition for Secure AI, an open ecosystem of AI and security experts from industry organizations dedicated to sharing best practices for secure AI deployment. “They are separable on a control diagram.” The real question, Kale said, is whether a human qualified to verify what the AI did was still in the loop when it happened.

Armando Solar-Lezama, Distinguished Professor of Computing at the Massachusetts Institute of Technology (MIT) Schwarzman College of Computing and associate director and COO of the university’s Computer Science and Artificial Intelligence Laboratory (CSAIL), said the bottleneck lives where human tasks cannot be handed off. Most of improving an AI model today, he noted, is about data collection: paying people to produce and judge training outputs, or setting up pipelines to generate that data automatically. Those tasks resist RSI because they require humans to identify the blind spots and limitations of existing models. Even hyperparameter tuning, he added, is “bottlenecked more by availability of compute than by programming skill.”

Jonathan Rosenfeld, who leads the FundamentalAI group at the MIT FutureTech interdisciplinary research group, pointed to a second constraint: the medium itself. “AI currently still has a very large gap between its digital- and physical-world observability and ability to act,” he said. “It cannot install or replace a GPU, debug the high-bandwidth signal shape on a printed circuit board, participate in the hallway conversation it was not part of.”

Compression, Not Frontier Expansion

So is recursion making frontier models smarter, or just faster?

Many experts say the latter. “Recursion primarily compresses development velocity within existing capability frontiers,” said Tejasvi Addagada, SVP Data Governance and AI Risk at DataIQ Technologies, an AI-native enterprise software company focused on customer engagement data. “It does not transcend architectural limits or discover new capability classes.” AI excels at what Addagada calls tactical recursion. It does not do strategic recursion, the act of redefining what is worth optimizing.

Solar-Lezama is not sure the difference matters. “I don’t think there is a meaningful distinction between these two options,” he said. Substantially compressing development cycles can itself lead to exponential disruption, he said. He compared it to walking versus driving: both reach the same destinations, but a car is fast enough to reorganize society around it.

Even so, exponential improvements are temporary, Solar-Lezama warned; they stop when a technology hits resource constraints or other bottlenecks. The real question, he said, is what those bottlenecks are and how far the curve runs before they show up.

Rosenfeld made a comparison from history. Chip design recursively drove the march of increasingly performant chips for decades. Over the same period per-transistor costs fell by more than a million-fold, he said, while the global economy grew less than ten-fold. Bottlenecks matter, he said, and AI is no exception.

Mark Ginsberg, an AI analyst at Drivehill Media, an Internet marketing agency in Israel, was blunt. “The capability frontier expansion claim is mostly marketing from frontier labs,” he said.

Once a model can shape its own objectives, the question shift from whether it can improve to what “improvement” means.

The real challenge of RSI, Solar-Lezama said, is figuring out what to improve. “Once you have an optimization objective, the optimization itself is already automatic.” When a model starts setting those objectives on its own, he warned, the failure mode is reward-hacking, which he compared to “a drug addict who realizes that the same pleasure you get from a new relationship or a major accomplishment can be produced by a chemical.”

Rosenfeld framed the same concern as a geometry problem. “Growing model capability can be thought of as increasing the radius of a sphere,” he explained. The position on that sphere’s surface, he continued, represents the model’s behavior. Capability does not pick a direction on its own. A model that’s good enough to design proteins for drug delivery, Rosenfeld noted, is also good enough to design viruses. The values instilled in the model, he said, are seen by many as inherently human choices.

“Would you want to be in a situation where the human is effectively forced out by the nature of the particular RSI process employed?”

Jacob Strauss, CTO of ChaseLabs, which operates an AI sales development platform, proposed a stricter empirical test. RSI, he argued, should count only when model-led changes produce gains that hold up on independent, write-protected holdout evaluations the model cannot see or edit. Scores can be moved by tightening prompts, loosening graders, or adding compute. None of that, he said, is self-improvement in a strong sense.

That is why several experts argue disclosure should be mandatory. Ginsberg said frontier labs should be required to publish “differential audits,” before-and-after snapshots of evaluation criteria whenever model-assisted development modifies the training or evaluation pipeline. “Anything less,” he said, “is self-regulation theater.” Without that record, Addagada warned, RSI becomes opaque self-optimization.

For now, the announcements have run ahead of the science. What is shipping, said Kirill Meshyk, head of AI Data Collection at data solutions provider Unidata, is “extraordinarily powerful automation, but still operating within a closed-loop system designed by humans.”

That would leave the industry where Kale placed it: with an org chart problem before a capability problem. When recursion is in play, the question will be who, exactly, is still in the room when it runs.

Logan Kugler is a technology writer specializing in artificial intelligence based in Tampa, FL, USA. He has been a regular contributor to Communications for 15 years and has written for nearly 100 major publications.