Readback Labs

AI is writing more of the world's software. We measure whether people still understand it.

Readback Labs studies how humans oversee AI systems, starting with coding agents: what people verify, what they miss, and how that is changing. Every number we publish can be traced back to its source.

read·back

noun. In radio procedure, repeating an instruction back to the sender so both sides know it was understood. Our work asks whether people can still do that with what AI systems produce.

Evidence for the people who have to decide.

Safety-critical industries learned to demand evidence before trust. A new aircraft type is certified on a structured safety analysis, a reactor is licensed on a risk assessment, and a clearance is read back before anyone acts on it. AI has no such habit yet, and the questions that regulators, companies and the public are starting to ask about it are empirical ones.

Readback Labs answers them with measurements rather than assertions, and puts the results in front of civil society, standards bodies, agencies and companies while there is still time to act on them.

In progress

Comprehension analysis

How much do people engage with and oversee what a coding agent does? We are scoring eight oversight behaviors across every public Claude Code session in the SWE-chat dataset: how people respond to a delivery, whether they verify it, whether they catch overclaims and misbehavior, how much they direct versus delegate, whether their words show they understood the change, what they ask, and how long they take.

Output: a metrics paper, a public collection where any cited exchange can be inspected, and a pipeline other evaluators can run.

How the study works

Preprint · October 2026

Comprehension Audits

AI already writes most of the code at frontier labs. When the people responsible for a contribution can no longer explain it, oversight has quietly stopped. The paper proposes comprehension audits: the responsible people explain their R&D contributions to independent auditors, without AI assistance, and development that depends on a contribution pauses until understanding is demonstrated. It includes a short version of our analysis of leading open-source AI projects and proposes a three-month pilot at a frontier developer.

Ronald J. Bodkin, Bahrad A. Sokhansanj and Gillian K. Hadfield. arXiv:2610.10064.

Read the paper (PDF) arXiv page Overview

Survey open

Overseeing agent-written code

A short survey for R&D leaders on how their teams review code that agents wrote: who reads it, what gets checked, and what has changed in the last year. It takes a few minutes, and results are published only in aggregate.

For: engineering and research leaders whose teams ship agent-written code.

Take the survey

A nonprofit focusing on how agents are changing human understanding.

Readback Labs, Inc. is a California nonprofit public benefit corporation. It measures how people oversee AI and how fast AI is changing the work it is used for, in collaboration with partners, publishes its approach, and shares the results to inform decision-makers. The lab is led by its founder and executive director, Ron Bodkin, and run by a small team of contributors and volunteers.

More about the lab

AI is writing more of the world's software. We measure whether people still understand it.

Where AI is built, AI already writes a measurable share of the code, and each line gets less human attention.

Evidence for the people who have to decide.

Are people still in control of what AI does for them?

Measuring Understanding in the Era of Agents

Comprehension analysis

Comprehension Audits

Overseeing agent-written code

The work is shared. Join it.

Researchers and engineers

R&D leaders

Policymakers and press

A nonprofit focusing on how agents are changing human understanding.

Contact