Test found Anthropic Claude did not create scorpion antivenom

In short: A test with an unreleased version of Anthropic’s Claude showed the AI could suggest protein ideas for scorpion venom, but it did not create a real antivenom drug.

What happened

An investment research firm called Citrini Research worked with AI company Anthropic to see if AI could help make scorpion antivenom. Antivenom is a treatment used after a venomous sting or bite. Citrini said its team did not have deep expertise in how AI applies to biology, but it still ran the experiment and published a blog post about it.

The group chose scorpion antivenom because it can be expensive and hard to make. Traditional antivenom often involves using animals, like horses, to produce antibodies. The goal here was to have Claude design proteins that could attach to scorpion toxins, like making custom “keys” that fit specific “locks” in the venom.

Anthropic said Claude generated about 320,000 possible protein candidates and narrowed them to 211 designs. Anthropic covered the AI computing costs, which were close to $50,000, according to the report. Citrini then paid outside labs, including Adaptyv Bio and Twist Biosciences, roughly $150,000 to test whether any of the proteins actually attached to targeted toxins.

Seventeen proteins attached to four of the five toxins the team hoped to target. However, Citrini said none of the proteins attached to more than one toxin. Experts told The Verge that attaching to a toxin is only an early step and does not show the protein can neutralize the toxin or work safely as a medicine.

Why it matters

This is a useful reality check for AI in drug research. AI can help generate ideas faster, but turning those ideas into safe treatments still takes long testing, real labs, and careful medical trials.

Source: The Verge AI