Rep. Ro Khanna is preparing to introduce artificial intelligence safety legislation that would bar the development or deployment of so-called “recursive” AI systems until adequate safeguards are in place, according to a report from CNBC.
The California Democrat, whose district includes much of Silicon Valley, has become one of the more closely watched voices in Congress on technology policy â in part because he represents the companies most affected by it. His proposal would mark one of the more aggressive federal attempts yet to place conditions on a specific class of AI capability rather than regulating the industry broadly.
What ‘recursive’ AI means
The term generally refers to AI systems capable of improving themselves â models that can modify, retrain, or redesign their own architecture or successor systems without meaningful human involvement in each step. Researchers sometimes describe the concept as “recursive self-improvement,” and it sits at the center of long-running debates about whether advanced AI could improve at a pace that outstrips human oversight.
For years, the idea was largely confined to academic papers and speculative writing. As frontier labs have built models that can write and debug code, run automated experiments, and orchestrate other AI agents, the question has moved closer to a practical policy concern. Critics of self-improving systems argue that once a model can meaningfully accelerate its own development, the window for humans to evaluate, test, or halt it narrows considerably.
A ban framed as conditional â in place “until safeguards exist” â suggests the legislation would be structured as a pause rather than a permanent prohibition, with some mechanism for lifting the restriction once safety standards, testing regimes, or oversight bodies are established.
A difficult definition problem
Any bill of this kind faces an immediate technical challenge: writing a definition of “recursive” AI precise enough to be enforceable but not so broad that it sweeps in routine machine learning practice. Automated hyperparameter tuning, neural architecture search, and models generating synthetic training data for future models all involve degrees of automation in the development loop. Lawyers and engineers on both sides of the debate would likely spend considerable time arguing over where the line falls.
Enforcement is another open question. Frontier AI development is global, and restrictions applied only to U.S. developers would not bind labs operating elsewhere â an argument industry groups have used against domestic rules in the past.
An uncertain path in Congress
The proposal arrives against a backdrop of stalled federal AI legislation. Congress has held numerous hearings on artificial intelligence over the past several years without passing comprehensive rules, leaving much of the regulatory activity to state legislatures and executive action. A single member’s bill, particularly one from the minority or without bipartisan co-sponsorship, faces long odds of reaching a floor vote.
Still, bills like this often serve a purpose beyond immediate passage. They stake out positions, force industry and researchers to respond to specific definitions, and can supply language that later gets folded into larger packages.
Khanna has previously pushed for tech accountability measures while maintaining close ties to the industry, a balance that has drawn both praise and skepticism. Whether this bill advances or not, its introduction signals that the debate over self-improving AI is moving from research conferences into legislative text. Read More
Leave a Reply