A neural decoder's sentence is an estimate, not a verified intention

An AI system can produce fluent text from a weak signal. Fluency does not establish that the text preserves what someone meant.

That matters in neural communication interfaces, where an error can put words into a person's mouth rather than merely produce a poor search result.

This is a proposed software-design checklist, not a tested clinical system or evidence that general thought reading is possible.

Keep the evidence boundary visible

In a 2023 Nature study, Willett and colleagues decoded attempted speech from implanted electrodes at 62 words per minute. Reported word error rates were 9.1% with a 50-word vocabulary and 23.8% with a 125,000-word vocabulary.

Those are results for that research setting. They are not a universal accuracy figure for all people, sensors, thoughts or conversations. The paper describes the demonstration as a proof of concept and says further work is needed before a complete clinically viable system.

A product interface should preserve that distinction. It should not label every generated sentence as a confirmed statement from the user.

Separate prediction from permission to send

A possible message flow is:

record only within an authorised session
    -> produce a candidate interpretation
    -> expose uncertainty or abstain
    -> allow correction through an accessible control
    -> obtain the configured send approval
    -> deliver to the intended recipient
    -> support clarification and correction

This is a design proposal. The accessible control and approval method would need validation with the intended users, including people who cannot use an ordinary keyboard or voice command.

A candidate interpretation and an authorised message should be different application states. Receiving a model output should not silently move the message into the sent state.

Test meaning, not only readable words

For a prototype, define what would count as failure before testing it. Consider a sentence such as:

I agree with the plan, but not until next week.

Dropping the condition changes the decision. A word-level score can help evaluate a decoder, but it does not replace checking whether the recipient understood the intended agreement and timing.

Useful evaluation questions include:

• Was the intended meaning retained, including conditions and negation?
• How long did the complete exchange take, including corrections?
• Could the user reject a wrong interpretation without unreasonable effort?
• Did the recipient understand that an unconfirmed output was uncertain?
• Was performance checked on held-out data and across separate sessions?

Compare the complete workflow with an appropriate communication alternative. A fast model response alone is not proof of a faster human-to-human exchange.

Make stopping and deletion part of the design

Define when recording begins, when it ends, who receives the output and which data must be retained. A stop control should be tested as carefully as a send control. Retaining every raw neural recording by default is not a substitute for a justified data policy.

Clinical, neural-recording and stimulation work requires appropriate expertise, permissions and safety review. A web application checklist does not provide those.

Further reading and disclosure

I wrote the IT Impress review and proposed research framework Beyond Speech and Writing: Can AI Let Humans Exchange Thoughts at High Speed?. It brings together published research and proposed evaluation criteria. It does not report new IT Impress experimental results or a working thought-transfer product.

The English version on our second site is also available. This post does not claim that a multilingual rollout is complete.

By Keshavan Thiagaraja, IT Impress