Google DeepMind and MIT jointly develop AI agent CoDaS: capable of autonomous scientific research, writing papers in just 8 hours

AI doesn’t just chat anymore—it now even does its own research and writes papers! The AI scientist CoDaS, jointly developed by Google DeepMind and MIT, has recently shocked the academic world. It can autonomously analyze intelligence-wearable data from thousands of people, not only automatically identifying that “late-night doomscrolling” is a potential indicator of depression, but also verifying and writing scientific papers on its own. What originally required experts more than a month to complete, CoDaS can finish in just 6 to 8 hours.
(Background: He Yi’s talk—Boost efficiency by 10 times with AI; let’s serve 3 billion users worldwide)
(Additional context: An open-source AI tool that no one was paying attention to warned about Kelp DAO’s $292 million vulnerability 12 days ago)

Table of Contents

Toggle

• Without human guidance, AI discovers “late-night doomscrolling” triggers depression
• Built-in “adversarial validation” automatically checks facts and prevents getting things wrong
• Reduces 37 days of work to 8 hours; blind paper testing earns expert approval

With artificial intelligence technology advancing by leaps and bounds, AI’s role is evolving from a simple “supporting tool” into an independent “scientific researcher” that can operate on its own.

Recently, a major study jointly released by Google Research, Google DeepMind, and the Massachusetts Institute of Technology (MIT) has demonstrated a breakthrough called CoDaS (AI Co-Data-Scientist)—a multi-agent AI system that successfully achieves a fully autonomous scientific discovery workflow. Renowned tech community opinion leaders Wes Roth and Samuel Schmidgall also widely reshared this highly pathbreaking academic achievement on the X platform.

A joint team from Google Research, Google DeepMind, and MIT has introduced CoDaS, a multi-agent AI system designed to autonomously run the entire biomarker discovery lifecycle from analyzing raw wearable sensor data and generating hypotheses to conducting statistical analysis and… https://t.co/KLgxFT4OSq pic.twitter.com/4ursWqeo7l

— Wes Roth (@WesRoth) April 20, 2026

Without human guidance, AI discovers “late-night doomscrolling” triggers depression

CoDaS is a system specifically designed to autonomously discover health biomarkers from raw data from wearable devices (“wearable sensors”). Its operating process includes: hypothesis generation, statistical analysis, adversarial validation, and literature-based reasoning—and ultimately it can even produce a complete draft of a scientific paper.

In testing, the research team fed CoDaS a large-scale wearable dataset covering nearly 10,000 participants (including sleep, activity, heart rate, phone-use habits, and more). Without any human prompts, the AI discovered multiple meaningful health features, and the most striking was a psychological health indicator:

AI discovered that excessive nighttime browsing of social communities or negative news is significantly positively correlated with depression severity (correlation coefficient ρ = 0.177, p < 0.001, sample size n = 7,497).

What’s truly astonishing is that the AI even independently named this behavior “late-night doomscrolling.” Besides mental health, it also successfully found a negative correlation between the ratio of daily step count to resting heart rate and metabolic diseases (insulin resistance).

Built-in “adversarial validation” automatically checks facts and prevents getting things wrong

To prevent AI from producing common “scientific hallucinations” or making meaningless statistical inferences, CoDaS includes a powerful adversarial validation mechanism (Adversarial Validation).

For example, when looking for metabolic health features, the system once proposed using “the square of glucose” to predict insulin resistance. Although this formula appears to have extremely high statistical correlation, CoDaS’s validation mechanism immediately detected that it was a scientifically meaningless tautology, and decisively rejected this feature. This mechanism greatly improves the scientific reliability and clinical potential of AI outputs.

Reduces 37 days of work to 8 hours; blind paper testing earns expert approval

CoDaS’s work efficiency and output quality completely upend the traditional research model. According to the paper data, a large-scale data analysis and writing task that would normally require 37 person-days from human experts can now be completed by CoDaS in just 6 to 8 hours.

More convincingly, in blind testing reviews conducted by domain experts:

• Papers generated by CoDaS achieved as high as an 86% “non-rejection rate” (i.e., accepted, minor revisions, or major revisions).
• By contrast, the rejection rate of papers from other benchmark AI scientific agents is as high as 85% to 100%.

This study proves how multi-agent AI systems can efficiently transform passive, consumer-grade wearable data into insights with clinical value. As a representative advancement of “agentic AI” in the digital health space, CoDaS signals a new era of scientific discovery jointly led by humans and AI—perhaps it’s already here.