When there is no more data left to collect, Artificial Intelligence will start reading your thoughts.

In Minority Report, the 2002 movie directed by Steven Spielberg, Tom Cruise plays an assiduous police detective in charge of a covert program called “pre-crime” that gives officers insight into someone’s innermost, private thoughts. Using that information, Cruise goes after regular, unassuming citizens, stopping them just before they commit their bad deeds.

Brain-Computer Interface (BCI) technologies powered by artificial intelligence cannot (yet) read your deepest thoughts and condemn you to pre-crime jail. However, recent clinical breakthroughs, along with the emergence of wearable consumer BCI devices should give us pause as it raises numerous ethical and privacy challenges.

In the not-too-distant future, who will control our neural data? How will such data be collected, stored, retained and used? How can we stop these technologies from enabling a pre-crime surveillance state?

We have reached peak data

Anthropic made headlines this summer when it was disclosed that it was scanning and unceremoniously destroying millions of printed books to train their advanced models. At that same time, Stanford University’s 2026 AI Index Report noted that the availability of high-quality human text and web data for training large language models is close to being exhausted.^[1]

Given that data is a key ingredient for improving model performance, researchers are searching for new and unconventional sources of data. It’s not difficult then to imagine how our private thoughts may one day become a rich repository for training powerful AI systems.

Brain-Computer Interfaces - Decoding Meaning

BCI technologies have been used for decades to restore movement and communication in individuals left paralyzed by neurological conditions (e.g., ALS, locked-in syndrome, spinal cord injuries, stroke). They work by decoding brain signals associated with physical movement to enable patients’ control over robotic arms or other prosthetic devices.^[2]

Recent advances aided by artificial intelligence have enabled BCI’s to decode both attempted speech (which relies on muscle motor movements) and inner speech (your silent inner monologue) of paralyzed individuals unable to communicate.^[3]

A study by Stanford University found a critical insight – neural signals associated with attempted speech and inner speech overlap in the same brain region – the motor cortex, and both signals are “remarkably similar.” The second critical insight was that a BCI device using a recurrent neural network (RNN) was able to successfully decode inner speech signals with “up to 74 percent accuracy.”^[4]

AI models are trained on brain signal data to recognize patterns associated with speech, language and movement. Brain data is collected from electrodes implanted in the brain, or through non-invasive methods like EEGs (electroencephalography), MEGs (magnetoencephalography) or functional magnetic resonance imaging (fMRI).

When a paralyzed individual attempts to speak in real-time or imagines a word or phrase, electrical signals are generated by neurons firing in regions of the brain corresponding to language and speech. BCI implants capture these signals and decode them using deep learning algorithms into sequences of phonemes (e.g., smallest unit of sound in a spoken word) or semantic word representations, which are then translated into words, phrases or sentences that are relayed to external devices enabling keyboard clicks, cursor movements or audio outputs.^[5]

Thought Leakage

Critically, the Stanford study found instances of thought leakage where a BCI was able to decode aspects of private inner speech when it was not explicitly programmed to do so.

Participants were asked to count the number of times they saw a particular shape and colour combination in a 10 by 10 grid. Participants were given no explicit instruction to use inner speech to complete the task, however researchers assumed that they would.^[6] A BCI device trained only on cued inner speech data (e.g., words that participants were explicitly instructed to say on command) was able to decode sequences of numbers from the participant’s brain activity data.

Commercializing your inner thoughts

Neuralink may be the most well-known player in the BCI market, but other firms are emerging with next-generation, non-invasive and wearable BCIs aimed at industries like gaming, employment, education, smart cities, industrial automation and health and wellness. For the latter, firms are targeting the able-bodied, promising improvements in cognitive health and functionality - including in memory, sleep quality and emotional regulation.

Although BCI technologies remain experimental and commercially unproven (especially due to the high cost of research and development), venture capitalists have invested over $3 billion in BCI startups since 2024.^[7] Companies are treating BCI-captured neural data as critical assets with high monetization value.

Risks to privacy and autonomy

The legal landscape controlling the flow and use of neural data is limited and fragmented. In Canada, federal privacy law doesn’t specifically address neural data, while in the U.S., only Colorado, California, and Montana have added neural data usage provisions to their existing privacy laws. Internationally, Chile amended its constitution to protect “neurorights,” and finally UNESCO adopted a recommendation on neurotechnology ethics in 2025.

BCI devices used in consumer-facing applications and domains pose novel privacy risks that are not appropriately addressed by existing privacy laws and data security frameworks:

1. Challenges in neural data anonymization: Neural signals produced by the brain have signatures corresponding to an individual’s unique brain structure and neural response patterns. Thus, unlike traditional data sources, neural signal data cannot be completely de-identified and are at risk of being re-linked to a specific individual/user.^[8]

2. Creation of user “Neural Profiles”: Neural data can facilitate the creation of granular “neural profiles” on individuals that offer expanded insight into individual attributes or characteristics. Advances in signal decoding aided by AI algorithms enable what’s called “inference leakage” in which neural signals – once processed by an AI model, reveal sensitive personal attributes that were embedded in the original training data, including behavioural, emotional or disease status. Such “profiles” can be sold by organizations to third parties like advertisers.^[9]

3. AI models themselves are major sources of privacy risk: An AI model’s weights, parameters, embeddings, decision-boundaries and inference outputs can encode attributes of the people whose neural data was used for training. Risks arise through all stages of AI model training and inference, including risks from unauthorized exposure of raw neural data, re-linking neural data for identification purposes, inferring sensitive traits from decoded neural data, and reverse-engineering models to recover neural training data.^[10]

4. Difficulty in consent and disclosure: An individual may participate in a BCI-enabled online game or register in a health and wellness initiative that tracks neural activity through a wearable BCI device. In either case, the individual may be unaware of the extent to which their neural data is being tracked, collected, processed and used. Thus, neural data changes the nature of consent and disclosure and requires controls that go beyond traditional “opt-ins” and “opt-outs.” We cannot control the neural signals generated in our brains and have less control over how such signal data are captured and extracted for commercial purposes.^[11]

We are no longer in the realm of science-fiction. BCIs may not be able to read our minds just yet, but the fragmented legal and regulatory landscape leaves an opening for technology firms to assert ownership rights over our neural data and control how it can be used.