A working paper finds that AI's self improvement feedback loops exist but are not yet self sustaining, though they are tightening.
Feedback loops inside AI research are real but are not yet strong enough to be self-sustaining, and they are tightening. That is the calibrated snapshot in a working paper dated July 13, 2026 that tries to put numbers on a question most AI coverage leaves vague: whether AI is starting to make AI get smarter faster.
If AI compounds its own progress, capability gains can in principle accelerate on their own. If it does not, progress stays a function of human researchers, compute, and capital, with AI as a tool rather than a participant. The new paper, titled "The Economics of Recursive Self-Improvement", treats that as an empirical question rather than a metaphysical one.
The lead author is Tom Cunningham of METR (Model Evaluation and Threat Research), an organization that builds evaluations for frontier AI systems. Co-authors come from multiple institutions, and the work is published under the umbrella of the Elasticity Institute, with administrative support from METR and hosting from Constellation. The model the paper builds is closer to economics than to computer science. It borrows the idea of an elasticity, the way an economist measures how a 1% change in one input moves an output, and applies it to feedback loops inside AI research.
The model is a directed graph. Each node is something that can change AI capabilities: the compute spent on training, the quality of algorithmic ideas, the share of research work done by AI systems, the speed of chip progress. Each edge is a feedback loop, and the size of the effect along that edge is the elasticity. Net acceleration in AI capabilities, the paper argues, depends on the product of elasticities across each loop. A loop that is small today, multiplied by a loop that is also small, is still small. Several loops tightening at once can compound in ways that look gradual and then become non-gradual.
The paper draws a hard line between "narrow" and "broad" AI capabilities. A system can improve on AI R&D benchmarks (writing better code for a chip, finding a smarter optimization trick, debugging a training run) without getting broadly better at economically valuable tasks. Most of what gets measured publicly is narrow. Most of what would matter to the broader economy is broad. The paper does not assume the two move together, and that assumption is what makes its "not yet" finding a snapshot rather than a verdict.
The paper ends with a wish list of empirical objects: measurements that AI companies already collect internally and could feasibly publish without giving away competitive secrets. Examples include the share of code in a training pipeline that was written or reviewed by AI, the share of ML research papers at the company that list an AI co-author, the fraction of bug reports triaged by an AI agent, and similar ratios. None are exotic. The product of elasticities is only as good as the inputs, and right now the inputs are mostly estimates.
The paper is a working draft, not a peer-reviewed publication, and the authors flag their own caveats: the calibration is back-of-envelope, the elasticities are illustrative and drawn from a small evidence base, and the snapshot could date quickly because feedback loops appear to be strengthening. The framing cites Favaro and Clark (2026) on a feedback loop in which AI is accelerating AI research. A reader who treats the paper as a verdict is misreading it. A reader who treats it as a framework for tracking the question is reading it the way its authors wrote it.
Ask the labs to publish the ratios on the wish list. Watch whether they move together or in opposite directions. Update the elasticity product when they do. The latest version of the paper is hosted at github.com/elasticity-ai/elasticity, with corresponding author contact at tom.cunningham@metr.org. Whether the feedback loops cross from "tightening" to "self-sustaining" is the question the paper is built to answer. On the authors' own math, the answer right now is not yet.