A Startup Is Paying Strangers $50 to Help It Build 'Telepathy'

A four-pound EEG headset, studded with electrodes, sits on a table in a San Francisco basement. For the next two hours, the volunteer inside the helmet types without ever looking at their fingers while an AI system records what their brain is doing. At the end, Conduit cuts them a check for $50. The company calls this "telepathy at scale." Inside the basement, the transaction is simpler: a helmet, a typing test, and a check at the end of every two-hour shift.

Conduit, a San Francisco startup founded in 2024, is paying volunteers to sit in electrode-studded helmets and talk to an AI, with one rule that doubles as a screening test: participants must be able to type without ever looking at their fingers. The founders are Rio Popper, an Oxford graduate who has been blind since age five, and CEO Clem von Stengel. The company says the goal is to "decode semantic content from neural data in the seconds before a subject types or speaks," meaning the thought, in their framing, before the words. That is the part that, if it works, would be a step beyond any consumer product on the market: a model that reads the sentence the brain is about to produce, rather than transcribing the keystrokes the body has already made.

In a December 7 blog post, Conduit claimed to have collected roughly 10,000 hours of neuro-language data from thousands of participants over six months, running recordings about 20 hours a day. Ten thousand hours is the figure the company has put at the center of its pitch. Existing public neuro-language datasets sit in the low hundreds of hours, the company says; Conduit's pile, if real, is an order of magnitude larger. The bet is that more labeled brain data will do for non-invasive thought-to-text what more text did for language models. A December Forbes profile of the AI-brain frontier frames the bet the same way: scale of training data is the thesis.

Naomi Bashkansky, a former OpenAI alignment researcher, announced on her personal blog that she is leaving the lab to join Conduit. Bashkansky spent years on OpenAI's alignment team, the group that works on keeping powerful AI systems doing what their operators want. Her stated vision, in her own words, is a 2030 "neural headband" that lets a person send thoughts directly to AI assistants. She also floated, as a personal forecast, that Conduit could eventually be "worth over $1 trillion." Neither the headband nor the valuation is company guidance. Both are predictions, in a personal blog post, and they are dated.

The privacy and funding picture is unclear. Conduit has not disclosed a funding round, and no U.S. agency has decided whether consumer EEG data collected outside a medical setting counts as health data, biometric data, or something new. The closest analogue is medical-device data, which sits under FDA rules, and biometric data, which a handful of states regulate for consumer use. EEG data collected outside a clinic, paid for by check, and used to train a consumer model, falls into the gap between those two. Participants are paid, they sign whatever consent form Conduit uses, and the helmet records. The session cap also varies between sources: the Gizmodo feature cites a 32-session limit per participant, while the Conduit site lists 10 two-hour slots. The discrepancy has not been reconciled publicly.

The competitive context is also narrow. Conduit is not in the same race as Neuralink, which builds implanted brain-computer interfaces. The closest public parallel is Meta's non-invasive Brain2Qwerty project, which uses brain scans to translate thoughts into text, primarily for people with ALS and other neurodegenerative conditions, as the Gizmodo feature on Conduit notes. Brain2Qwerty is published research with a defined medical use case. Conduit's examples are demo videos on a marketing site, and no peer-reviewed validation of the model's zero-shot generalization to new users has been published.

Conduit is selling "telepathy." It is buying training data, $50 at a time. Whether the headset ever reads minds is a research question, and one Conduit has not yet answered in a peer-reviewed paper. Whether $50 is a fair price for an hour of someone's labeled brain activity, in a country that has not decided what category that data belongs to, is a legal one. Bashkansky says her first task is "to get more people in helmets." That is not a forecast. It is a recruiting target, and the program is already running.