Ex-OpenAI researcher joins Conduit to build non-invasive mind-reading AI

The migration of talent from large language model alignment to neurotechnology reflects a broader pattern where researchers seek new domains for scaling law applications after reaching perceived plateaus or ethical constraints in pure AI safety. Historically, BCI development was siloed within neuroscience and medical device regulation, characterized by slow clinical validation cycles and invasive hardware requirements. The entry of AI-native researchers introduces a paradigm shift, treating neural signals as high-dimensional data amenable to transformer architectures rather than purely biological phenomena. This mirrors earlier transitions where computer vision and NLP researchers moved into genomics and protein folding, applying deep learning to biological problems previously dominated by domain-specific heuristics. Such cross-pollination often accelerates technical benchmarks but frequently outpaces the development of domain-specific governance structures. The precedent suggests that when AI scaling methodologies are applied to sensitive biological substrates, technical capability tends to advance faster than societal consensus on acceptable use cases, creating recurring friction between innovation velocity and normative safety frameworks.

On August 6, 2026, Naomi Bashkansky publicly announced her departure from OpenAI’s alignment research team to join Conduit, a neurotechnology startup focused on non-invasive brain-computer interfaces (BCI). According to social media disclosures and blog summaries, Bashkansky resigned from OpenAI on July 23, 2026, and began her role at Conduit the following day. Her transition marks a significant personnel shift, moving expertise from large language model safety to neural decoding. Bashkansky has framed this career pivot not as an abandonment of safety, but as an application of scaling laws to a new domain where she believes human-in-the-loop safeguards can be effectively implemented.

Conduit’s stated objective is to achieve "thought-to-text" capabilities by 2027 and latent communication by 2030 using non-invasive hardware. According to Bashkansky’s public roadmap, the near-term product involves a headband that decodes user reactions while reviewing code, automatically transmitting these signals to AI coding agents in ten-second intervals. The longer-term vision extends to 2035, where AI integration is described as becoming a "sixth sense," allowing users to access information through mental imagery without verbal intermediaries. This approach relies on applying LLM scaling methodologies to neural data, positing that imperfect brain signals combined with strong language priors can yield functional decoding without surgical intervention.

The company claims to have already amassed over 10,000 hours of neural language data, which proponents describe as the largest known dataset of its kind for non-invasive BCI. The data collection methodology reportedly involves participants sitting in a booth wearing a headset while interacting with an LLM for extended periods. Bashkansky argues that this greenfield problem requires a dedicated focus distinct from general AI alignment, suggesting that the entity solving general neural read/write capabilities could achieve significant market valuation. She maintains that the technology remains human-in-the-loop by design, countering potential concerns about passive surveillance or unauthorized cognitive access.

Critics and observers have noted the implications of applying aggressive scaling paradigms to biological data. While OpenAI has not publicly commented on Bashkansky’s departure or Conduit’s specific technical approach, the move highlights a growing divergence in how AI researchers view the intersection of machine learning and neuroscience. The controversy centers less on alleged misconduct and more on the strategic and ethical validity of treating neural decoding as a tractable scaling problem. Skeptics question whether non-invasive signal fidelity can support the ambitious timelines proposed, while supporters argue that the convergence of AR/VR hardware and neural priors makes this a logical next step in human-computer interaction.

Bashkansky’s public statements emphasize that interim solutions will likely involve BCI paired with augmented reality, making it easier to send discernible brain signals to wearable devices before achieving full telepathic integration. This phased approach attempts to bridge the gap between current technological limitations and future aspirations. The narrative presented by Conduit and Bashkansky positions their work as a natural evolution of AI development, leveraging established scaling principles to unlock direct neural communication. However, the rapid timeline and the involvement of former alignment personnel in building cognitive readout technologies have sparked industry discussion regarding the adequacy of existing safety frameworks for non-invasive neurotech.

The primary stakeholders are future users of non-invasive BCI systems and the broader public whose cognitive data may become a training substrate. If Conduit’s scaling hypothesis proves valid, the ability to decode thoughts via consumer headbands by 2027 could outpace regulatory frameworks designed for medical devices, potentially exposing users to unprecedented cognitive privacy risks. Conversely, failure to achieve promised fidelity could result in misallocated research capital and erosion of trust in neurotech. The magnitude hinges on whether 10,000 hours of neural data proves sufficient for semantic decoding; success would validate a new AI-neuro convergence paradigm, while failure would reinforce skepticism about applying LLM methodologies to biological signals. OpenAI’s silence creates ambiguity about whether this talent migration represents a sanctioned expansion of AI scope or an uncontrolled diffusion of alignment expertise into less-regulated domains.