MIT-led team gets up to $53 million in federal money to use AI to design new weed killers as weeds outgrow old ones

A research team led by the Massachusetts Institute of Technology will receive up to $53 million from the federal government to use artificial intelligence to invent new weed killers, MIT announced Nov. 10.

The money comes from the Advanced Research Projects Agency for Health, or ARPA-H, a federal agency that funds high-risk health research.  More than 1 billion pounds of herbicide are used in the U.S. each year, MIT says.  Herbicides protect crop yields, but weeds are developing genetic mutations that let them survive the most common ones, glyphosate, the main ingredient in Roundup, and atrazine.  Exposure to herbicides may raise the risk of cancer and hormone and metabolic problems, MIT says; most Americans get small amounts through food and drinking water.  Glyphosate is also made with phosphorus, which the U.S. imports.

“We could find ourselves without these core agrochemicals,” said Christopher Voigt, head of MIT’s Department of Biological Engineering and the project’s lead researcher.

How the project will work

Glyphosate kills plants by blocking an enzyme, a protein that speeds up a chemical reaction, that plants need to make certain building blocks of protein.  The project, called PLOW, short for Precision Learning for On-target Weed control, has three teams.  The first will build generative AI models, the kind that create new things rather than just sort data, to propose molecules that block the same enzymes in new ways.  Some will be chemicals, and some will be small interfering RNA, short strands of genetic material that can switch off a specific plant gene.  The models are meant to favor molecules that are cheap to make, unlikely to trigger resistance and low-risk to people and wildlife.

The models are expected to suggest billions of candidates.  A second team will test them quickly in the lab and feed the results back to improve the AI.  A third will test the best ones in greenhouses, with the first RNA candidates expected to finish greenhouse tests by the end of the first year.  The team will also engineer bacteria and fungi that break down farm chemicals before they reach food, soil or water.

Could the AI design something more dangerous?

Yes, the same kind of tool can be pointed the wrong way, and researchers have shown it. In 2022, scientists at Collaborations Pharmaceuticals, a North Carolina drug-discovery company, reported in the journal Nature Machine Intelligence that they took an AI program built to find medicines and avoid toxic side effects, and simply reversed its goal to reward toxicity instead. In less than six hours, it proposed 40,000 molecules, including the nerve agent VX, one of the deadliest chemical weapons ever made, other known chemical weapons and new molecules it predicted would be even more toxic. The researchers did not make any of the molecules, but they warned that chemical-making services are widely available and poorly regulated.

Nothing in MIT’s announcement suggests PLOW is aimed at anything but weeds. To guard against accidental harm, MIT says the project’s AI models will be built to favor molecules with a low risk of harming people or causing damage to the environment, and the best candidates must pass lab and greenhouse tests. The announcement does not describe specific safeguards against misuse of the AI tools themselves. Before any new weed killer can be sold in the U.S., the company making it must give the Environmental Protection Agency test data, and the agency assesses the risk of harm to people, wildlife, fish and other plants, and of water contamination, according to the EPA.

Partners include the Broad Institute of MIT and Harvard, the University of North Carolina at Chapel Hill, the University of Tennessee, the University of Illinois and the company Robigo.

Source: MIT News announcement, Nov. 10, 2026; Urbina et al., “Dual use of artificial-intelligence-powered drug discovery,” Nature Machine Intelligence, March 2022; EPA: About pesticide registration.