Bashkansky is a real person with a verifiable history: a Woman International Master chess title, a Harvard computer science degree, and 1.5 years on OpenAI's alignment team. She resigned July 23, 2026, and started at Conduit the next day. The move was covered by Gizmodo, TechTimes, and multiple other outlets within days of her post.

Conduit was founded in 2024 by Rio Popper (Oxford) and Clem von Stengel (Cambridge, CEO). By December 2025 the team had collected roughly 10,000 hours of non-invasive neural recordings from thousands of participants, which they describe as the largest neuro-language dataset in the world. The operation is charmingly unglamorous: participants recruited largely through Craigslist, paid about $50 per two-hour session, sitting in phone-booth enclosures wearing multi-pound headsets built from disassembled commercial sensors in 3D-printed housings, running 20 hours a day in a San Francisco basement. Forbes profiled the company in December 2025 and quoted both founders.

"Largest dataset" is the company's own characterization. Nobody audits neuro-dataset bragging rights.

This is the part most readers will find hardest to believe, and it is the best supported. Meta's Brain2Qwerty decoded typed sentences from non-invasive MEG recordings at a 32 percent character error rate in early 2025, and the v2 model released in June 2026 reached 61 percent average word accuracy (78 percent in the best participant). Meta's own analysis found a scaling law: more training data yields better decoding, with no plateau detected yet. On the invasive side, a Stanford-led team published in Cell in August 2025 that imagined sentences (inner speech, no movement, no attempted speech) can be decoded in real time from motor cortex, with word error rates of 26 to 54 percent against a 125,000-word vocabulary and as low as 14 percent on a constrained 50-word set. The direction of travel is unambiguous.

Conduit's own reported metric, roughly 45 percent zero-shot semantic match between decoded and intended text, appears only in a Forbes profile and the company's blog. No peer review, no published benchmark, no independent replication. The number is plausible given the field. It is also unverified.

Bashkansky is careful to label her 2027, 2030, and 2035 scenarios as predictions. Credit for honest framing.

Meta's headline results come from MEG, which requires a magnetically shielded room and a machine the size of a hair salon dryer chair from the 1960s. The wearable modality, EEG, performed dramatically worse in the same study (67 percent character error rate versus 32 for MEG). Conduit's bet is that a custom multimodal headset plus orders of magnitude more data closes that gap. Maybe. But the vignette of ambient thought-decoding through a Bluetooth headband while making coffee describes a device that does not exist, running at a fidelity no one has demonstrated outside a shielded room, on a signal type (vague unspoken intentions rather than typed or imagined sentences) that no published system decodes. The Cell authors are explicit on this last point: even with electrodes implanted in motor cortex, it was not possible to accurately decode complete, intelligible sentences during free-form thinking. If invasive arrays cannot yet do it, a headband doing it in three years is not a prediction. It is a wish with a timeline.

A 2024 Neurorights Foundation audit of 30 consumer neurotechnology companies found that 96.7 percent reserved the right to transfer brain data to third parties, fewer than 20 percent mentioned encryption in their privacy policies, and 16.7 percent committed to breach notification. Conduit's privacy policy has not been independently audited. The Cell paper is instructive here. When participants silently counted colored shapes on a grid, the decoder read out a climbing sequence of numbers nobody had asked them to think in words. When participants were prompted to just think, the decoder produced output the authors declined to publish, citing mental privacy and uncertainty about whether the words reflected actual thoughts. So the team engineered a lock: the BCI ignores inner speech until the user mentally speaks an unlock keyword, detected at 98.75 percent accuracy. The keyword was "ChittyChittyBangBang." The researchers building this technology consider accidental thought-capture real enough to defend against with a magic word. The companies collecting the training data mostly do not mention it.