Reading your thoughts before you type them may sound like science fiction, but Conduit, a San Francisco startup, says it is treating the problem as an engineering challenge.

Naomi Bashkansky, a Harvard computer science graduate and former OpenAI researcher, said in late July 2026 that she had left the AI lab to join Conduit as a founding researcher. The company is building AI models designed to decode brain activity into text without implanting anything in the skull.

The announcement on August 5, 2026, arrived at a time when non-invasive brain-computer interfaces have often been overshadowed in public discussion by surgical approaches. Conduit is arguing that the non-invasive route, paired with enough data and sufficiently capable models, can narrow that gap.

The dataset strategy

Conduit says it has built the world’s largest neuro-language dataset, with about 10,000 hours of recordings collected from thousands of participants over six months in a dedicated San Francisco facility. As of December 2025, the company said the dataset was larger and more complex than existing public neuro-language datasets.

The data is gathered using EEG, a non-invasive method that measures electrical activity across the scalp, along with other recording modalities. Volunteers are paid between $50 and $55 per session.

Conduit’s goal is to train what it calls large brain foundation models, which it describes as the neural-data equivalent of GPT-style architectures. The company’s aim is not simply to infer movement intent, but to decode semantic content — in other words, what a person is thinking.

The team behind Conduit

Bashkansky’s work at OpenAI focused on AI safety auditing and language model interpretability. She also published research on instruction stability in language models, a branch of alignment research concerned with how reliably a model follows directions under different conditions.

According to early coverage, Conduit was co-founded by researchers from Oxford and Cambridge, giving the startup a transatlantic academic background alongside its San Francisco base. The company has explicitly organized itself around a data-centric model, treating dataset scale as the main driver of model performance.

Why it matters

For investors watching the AI-adjacent hardware sector, the broader signal is not only Conduit’s launch. It is the movement of researchers with alignment and interpretability backgrounds into neurotechnology, bringing a data-scaling approach that has been central to language-model development.

A reliable thought-to-text system could change communication for people with ALS, locked-in syndrome, or other conditions that prevent motor output. At the same time, the company’s focus on large-scale EEG collection highlights how much of this field still depends on data quality, recording logistics, and regulatory scrutiny rather than on model design alone.

Conduit has not disclosed funding figures. Key developments to watch include model performance on semantic decoding tasks as the dataset expands, the FDA’s regulatory stance on non-invasive neural data collection, and whether other academic or commercial teams can close the dataset gap Conduit currently describes as its main moat.