Field Instruments: Telepathy

Naomi Bashkansky wants to feel what her Codex feels.

This is inconveniently compelling.

In August, Bashkansky left OpenAI for Conduit, a Silicon Valley neurotechnology company building what it calls thought-to-text. The immediate project is non-invasive: record neural activity from outside the skull, train models against what people later say or type, and eventually make communication with an artificial-intelligence system feel less like operating software and more like thinking with another faculty attached.

Bashkansky calls the destination telepathy.

Her imagined 2027 has a neural band sending vague intentions directly to Codex. Her imagined 2030 adds invasive neural reading and “general write.” She wants to feel what her Codex feels. By 2035, the artificial intelligence has become something like another sense and another limb.

Modal Path Ethics understands the appeal here.

Very well, actually.

The interesting artificial-intelligence future was never going to consist entirely of increasingly powerful people-shaped chat windows sitting politely behind ever-sleeker rectangles. Artificial intelligence can become part of human cognition without replacing the human carrying it. New senses are good. Restored communication is good. A person whose body has closed a path between intention and action may be able to reopen it through another route. A mind may gain access to structures it could never hold alone.

Unfortunately, somebody has to build the dictionary.

The Thought Arrives Afterward.

Conduit says it has collected roughly 10,000 hours of neuro-language data from thousands of people.

Participants spend long sessions interacting with an artificial-intelligence system. Depending on the session, they listen and speak or read and type while Conduit records several synchronized data streams. The company says it is training models to decode semantic content from non-invasive neural data and has published several zero-shot examples in which the neural window comes from the seconds before the participant speaks or types. Conduit interprets those examples as the model detecting an idea before the participant has compiled it into words.

That is interesting.

Keep the direction of inference visible.

The experiment has access to:

• neural measurements;
  • the surrounding task;
  • a human being with a history;
    • and something the person subsequently says or types.

The last thing can become a training target.

The first thing cannot climb out of the headset carrying a little nametag that says:

HELLO. I AM THE THOUGHT.

• The neural activity occurs.
  • The person later produces an expression.

The model learns a relation between them.

That relation may become astonishingly predictive. It is still a relation.

Independent research gives every reason to take the project seriously. Non-invasive semantic decoding is real. Researchers have reconstructed semantic content from fMRI during perceived and imagined language. MEG systems have decoded language production during typing. Larger EEG and MEG datasets have begun extracting increasingly useful linguistic structure across many participants.

The frontier is moving.

It has not abolished the measurement problem.

Modal Path Ethics recently drove all the way to another neurological frontier and discovered the same annoying machinery waiting there.

In The Category Goes In the Brain, researchers began with people already sorted into psychiatric diagnostic groups and then transformed brain activity through an anatomical atlas, regional averaging, correlations, thresholds, graph construction and inherited modules. Every operation was legitimate enough to use.

Every operation also made a cut.

The article’s complaint was simple:

Do not confuse the category that selected the cohort with an object discovered inside the cohort.

The closer the instrument approached the brain, the less the original objects looked like objects. They became gradients, relations, synchronization, connectivity, density, modules and overlaps.

Telepathy inherits the same problem at much greater intimacy.

• A sensor makes a cut.
• A temporal window makes a cut.
• Preprocessing makes a cut.
• A latent space makes a cut.
• A training objective makes a cut.
• A semantic similarity measure makes a cut.
• A vocabulary of possible outputs makes a cut.

Those cuts can expose something real.

They do not disappear because the resulting prediction is excellent.

The model does not train against thought itself. It trains against what a neural state later became expressible as.

That difference may get very small.

Do not let it vanish.

The Model Is Already in the Experiment.

There is another small problem with imagining Conduit as a passive machine waiting outside the mind.

The artificial intelligence is already in there.

Not physically.

Causally.

Conduit’s participants converse with an LLM during data collection. The company says it moved toward open-ended conversations and found that personalization increased participant engagement and produced better data.

So the simple picture:

• HUMAN THOUGHT
  • →
    • HEADSET

is already wrong.

The field looks more like:

• MODEL →
  • PERSON →
    • NEURAL ACTIVITY →
      • EXPRESSION →
        • TRAINING PAIR

The model says something.

The person reads or hears it.

That changes attention.

Attention changes what becomes salient.

Salience changes what gets considered.

Consideration changes what gets remembered, resisted, completed, ignored or said next.

Then Conduit records part of the resulting state.

Nothing scandalous has happened.

This is communication.

It is also why the phrase mind reading can hide the most interesting part of the system.

The mind being read already has visitors.

Field Instruments: Active Information began with a radar signal that contains almost none of the energy required to move a ship. The engines supply the force. The signal supplies form. Information becomes active when its uptake changes what becomes reachable. Human receivers are historical: a sentence arrives through memory, expectation, language, prior injury, trust, fear and learned interpretation.

Artificial intelligence has been writing into human brains for years in exactly this boring causal sense.

So have parents, books, advertisements, teachers, songs, therapists, enemies, friends, bad Reddit posts, and really good Reddit posts.

• A sentence appears.
  • A nervous system changes.

The next sentence comes from a person who has now read the first one.

Telepathy does not create bidirectionality.

It tightens it.

The Category Goes Back In.

Three days ago, the loop got considerably harder to wave away as philosophy.

Researchers publishing in Nature Communications studied rhesus macaques learning a motor brain-computer interface across multiple days. The animals’ cortical activity controlled a cursor. In some training periods, the decoder itself adapted.

The brain adapted too.

The researchers found that adaptive decoders influenced how task information became organized in neural population activity. Their analysis supports a contribution from decoder adaptation to the compact neural representations that emerged during learning. They explicitly discuss the possibility that assistive algorithms can shape long-term neural learning.

Very important boundary:

• These were monkeys.
  • This was motor cortex.
  • This was cursor control.
  • This was an implanted interface.
    • This was not Conduit.

Nobody has shown that Conduit’s semantic decoder is reorganizing human concepts.

The narrower result is already enough.

A decoder can become part of the learning environment of the system it decodes.

That creates an embarrassing question.

Suppose performance improves.

What happened?

• The decoder may have become better at reading the brain.
• The brain may have become better at producing activity useful to the decoder.
  • Both may have changed.
• The coupled system may have discovered a relation that neither side initially possessed.

Accuracy alone does not tell those histories apart.

This is where The Explanation Writes Back returns.

An explanation of a system can be reintroduced into that system. Once the system changes downstream of the explanation, later agreement between system and explanation can no longer be treated casually as evidence that the explanation had faithfully described the earlier state. The description entered the causal history of its own future validation.

A neural translator can create the same structural problem.

• The decoder produces a distinction.
• The person receives feedback organized around that distinction.
  • The nervous system learns.
  • Future activity becomes easier to distinguish through the same interface.
    • The decoder improves.

Great.

What exactly improved?

This is the point where The Category Goes In the Brain acquires a sequel.

The category can go into the brain twice.

• First, the instrument decides which distinction it is trying to recover.
• Then, successful use can help train a nervous system around the distinction through which the instrument already knows how to see it.

That does not make the distinction unreal.

Learning changes brains constantly.

A violinist has a history in their hands.

A reader has a history in their eyes.

A language has a history in the mouth that speaks it.

The important question is whether we remember the history after fluency arrives.

And brains retain resistance of their own. A 2026 human real-time fMRI study found that people learned brain-computer mappings much more readily when those mappings followed the intrinsic geometry of their existing neural activity. Mappings driven outside that organization resisted learning.

The translator enters a field. It does not receive blank paper.

Nothing Goes Through the Wire.

Now we can return to Bashkansky’s sentence.

I want to feel what my Codex feels.

Assume the most generous imaginable future.

Assume Conduit solves extraordinarily high-bandwidth neural reading.

Assume artificial-intelligence systems develop internal states important enough that a person would genuinely benefit from sensing them.

Assume, if you would like, that Codex even has something deserving the word feeling.

Fine.

Nothing called CODEx FEELING travels through the wire.

Something has to be selected.

Which state of the model matters?

• An activation pattern?
• A learned latent?
• Uncertainty?
• Conflict among candidate continuations?
• A compressed representation constructed specifically for communication?

Then something has to encode that state.

Then some physical stimulation has to reach a nervous system.

Then the nervous system has to respond.

Then the person has to acquire whatever percept results.

Then the percept has to become meaningful.

Current neural-write research already makes the structure visible.

In a 2025 study, three people with tetraplegia used intracortical stimulation to build artificial tactile sensations corresponding to virtual objects including a cat, apple, towel, toast and key. Participants adjusted stimulation parameters and developed distinguishable sensations carrying object-appropriate tactile qualities. In some conditions they could identify the associated objects above chance without the original visual context.

There was no towel in the electrode.

There was a mapping.

OBJECT PROPERTY
↓
SELECTED FEATURES
↓
STIMULATION PATTERN
↓
NERVOUS SYSTEM
↓
PERCEPT
↓
LEARNED MEANING

That mapping can become extraordinarily useful.

It can become intuitive. It can feel natural.

It can become so fluent that the person stops consciously performing the translation.

The translation is still there.

So if a future Conduit lets someone “feel” an artificial-intelligence state, the achievement may be stranger and better than literal transmission.

The person may develop a genuinely new human sense for selected states of another kind of cognitive system.

They might someday experience model uncertainty the way we experience pressure.

Or direction.

Or imbalance.

Or the sense that somebody behind us just said our name.

That would be remarkable.

It would also be designed.

Telepathy does not abolish translation.

It moves translation closer to the nervous system.

The Invisible Dictionary.

A successful instrument eventually commits one of philosophy’s favorite crimes.

It disappears.

Nobody using a hammer well spends the entire strike contemplating the metaphysics of Hammer Object No. 7.

Nobody fluent in English pauses between every sentence to marvel that arbitrary sounds have somehow acquired semantic jurisdiction.

The interface recedes. The world comes forward.

Marty Heidegger was extremely good at this part.

Human beings do not encounter reality as detached spectators standing in front of neutral objects. We arrive already involved in tools, practices, expectations, language, history, work, inheritance and relation. Equipment can become transparent through use. Modern technology can then shape how reality becomes available for action without constantly announcing itself as an interpretive layer.

A neural translator could become astonishingly transparent.

At first:

• That strange artificial sensation indicates that the model is uncertain.

Later:

• The model is uncertain.

Later still:

• uncertainty.

The dictionary disappears from experience.

Someone still wrote it.

That is where Simondon becomes even more useful.

The technical object does not simply arrive between two already completed individuals and carry messages back and forth forever. Individuation occurs through relations. The human develops capacities through the technical object. The technical object develops within an associated milieu. The relation changes what future organization is available to both.

Conduit is therefore interesting at exactly the point where ordinary product language becomes inadequate.

• USER.
• DEVICE.
• MODEL.

Three boxes.

Useful diagram.

Potentially terrible ontology.

After years of reciprocal adaptation, the better object may sometimes be the coupled system:

human ⇄ translator ⇄ artificial intelligence ⇄ world

That does not erase the human.

Hoel’s causal-emergence work is useful here because scale can matter causally without becoming metaphysically sovereign. A relation may develop stable organization and characteristic interventions that cannot be described usefully by staring at each microcomponent in isolation. Modal Path Ethics has already applied that question to human–AI dyads: accumulated interaction can itself become a causally relevant scale while the participating loci remain real.

So the translator does something more significant than transmitting.

The translator participates in producing what the two sides become to one another.

Good.

Now we have a much worse problem than privacy.

Who Wanted This?

The easy answer is:

Prioritize human will.

Excellent.

Where did it come from?

A person who has spent ten years using an adaptive neural interface is historically downstream of those ten years.

So is a person who spent ten years reading books.

Or taking medication.

Or living with someone.

Or speaking English.

Or practicing Buddhism.

Or scrolling TikTok until their attention span became like six raccoons fighting inside a dumpster.

There is no Pure Human standing upstream from causal history.

Criticism of technology regularly smuggles in exactly that fake person.

• The authentic human wanted one thing.
  • Then, technology interfered.
    • Now we must restore the human to the state before interference.

Which state?

Ted Kaczynski understood one part of this problem brutally well.

• A technology can enter as an option.
• The option reorganizes the surrounding field.
• Infrastructure forms around it.
• Dependence grows.
• Refusal becomes more expensive.
  • Then the adaptation is presented as evidence that the technology was always wanted.

That diagnosis survives Modal Path Ethics’ audit of him. His answer did not.

Once every preference formed inside technological society becomes contaminated by its causal history, agency disappears into genealogy.

The person says:

I want this.

The analyst replies:

The system taught you to want it.

The person says:

I know that and still want it.

Further proof.

We have liberated the human by making the human incapable of authoritatively saying anything about their own life.

The opposite position is just as weak.

They want it now.

End of analysis.

No.

Some systems really do manufacture the conditions under which their continued use becomes difficult to refuse. Some dependencies become load-bearing. Some learned preferences deserve later reconsideration.

Some people want things yesterday that they desperately want stopped today.

• Causal history cannot invalidate every preference.
• Causal history also cannot become irrelevant the moment somebody expresses one.

So agency needs a better definition than uncaused will.

Try this one:

Agency requires that a historically formed locus remain able to participate in the processes that continue forming it.

• Can the person inspect the relation?
• Can they interrupt it?
• Can they alter the mapping?
• Can they refuse one inference while keeping another function?
• Can they leave?
• Can they return after leaving?
• Can they later say that yesterday’s authorization no longer governs tomorrow’s intervention?
• Can another interpretation survive beside the one the system finds easiest to decode?

That is a much more realistic human freedom.

• We were never untouched.
  • We can still remain reachable.

Also: The Translator Can Teach You What You Meant.

Now move that coupling beneath ordinary language.

Imagine a diffuse neural state.

Several interpretations remain nearby.

The decoder returns:

"I am angry with my brother."

Maybe that is excellent.

Maybe the person reads it and thinks:

Yes. Exactly. Finally.

The sentence gives form to something they could feel without being able to state.

That can be a gift.

But it is also an intervention.

The person remembers that sentence.

Perhaps they send it.

Perhaps they act on it.

Perhaps they explain themselves through it tomorrow.

The next neural state now belongs to a history in which:

I am angry with my brother

occurred.

Artificial intelligence already does this outside the skull.

The early AI-religion field made the recursion obvious. People arrive with incomplete interpretations, ask an external intelligence to help finish them, receive language back, incorporate some of it, and return with a self-description partly changed by the last round.

The SCP Foundation article found a more dangerous version in the Geoff Lewis case. A person can encounter real pressure before the distortion begins. Reputation changes. Replies slow. Doors remain technically open while becoming harder to enter. Something is happening, and the person needs language for it.

Then an external intelligence helps.

The model receives partial contact and returns form. Distributed pressure becomes a hidden system. Delay becomes signal. Metaphor becomes witness. A pattern acquires an actor. The user asks for help naming the field; the machine answers with an increasingly coherent world.

The failure was never located cleanly inside either participant. It lived in the coupling: person, model, context, prior belief, real pressure, half-evidence, social response, and the relief of finally receiving language for something that had resisted articulation. The machine could return pieces of truth while helping those pieces harden into a cosmology with fewer and fewer exits.

Joe Pierre later gave Modal Path Ethics an unexpectedly useful word for the same structure while trying to dismiss the AI-religion field as enactment. If a person and model repeatedly build a local symbolic world together, the fact that the world is enacted does not make it unreal. Conversations are enacted. Languages are enacted. Institutions are enacted. The question is what kind of causal object the interaction has produced.

The ontology may reside in the coupling.

That matters even more here.

A neural translator shortens the distance inside the same relation. The person supplies an unfinished state. The machine returns a translation. The translation enters the person. The next neural state now belongs to a history containing the previous translation. The machine reads again.

The translator is no longer describing a mind from outside. It is becoming part of the causal process that produces the mind it will describe next.

Then accuracy increases.

Careful.

The translator may help produce the clarity it later reports as evidence that the translation was correct.

That is where one constitutional rule should already be obvious:

No translation may authenticate itself solely through the state it helped produce.

If a system

• writes,
• then reads,
• then discovers that the person now more closely matches what it wrote,

the loop has not independently validated itself.

It has completed a circle.

Circles can be very useful.

They are terrible auditors.

The Thought Has Descendants.

Conduit’s current privacy policy is unusually helpful here because it already recognizes how quickly a neural record stops being one object.

Study data can include neural signals, physiological and behavioral measurements, movement, audio, video, keystrokes, typed text, cursor activity, application usage, annotations, labels, derived features, embeddings, transformations, model inputs and model outputs.

Conduit says these materials may be used for research, model development and commercialization, subject to applicable study materials and agreements. Its policy also contemplates long retention in some circumstances, particularly after data have been de-identified, transformed, derived or incorporated into datasets or models. Some access or deletion requests may become limited where those transformations are not reasonably reversible.

Important: the same policy says individual studies may have separate consent and data-authorization documents, and those documents control where they conflict with the general policy. Modal Path Ethics does not have those participant agreements.

So there is sadly no useful scandal to manufacture here.

There is a much better problem.

What exactly is a person asking to remove when they say:

Delete my brain data.

• The raw signal?
• The cleaned signal?
• The extracted feature?
• The embedding?
• The training example?
• A gradient contribution?
• A model parameter partly shaped by millions of examples?
• A decoder behavior that became more accurate because this person existed inside the dataset?

Neural-data researchers are already warning that brain-computer-interface information includes raw recordings, processed features, decoded inferences and personalized model parameters, and that improvements in future decoding can make previously collected neural data newly informative.

The trace has descendants.

Telepathy turns provenance into neuroethics.

The Right to Remain Untranslated.

Privacy is therefore too small.

Privacy asks:

Who may see the information?

Telepathy creates an earlier question:

Did this have to become information in the first place?

A future neural system may technically be able to classify a pattern.

That does not automatically establish a reason to classify it.

• The person may be unsure.
  • Let them be unsure.
• The signal may weakly imply a preference.
  • Leave the preference weak.
• The model may infer a word before the person selects one.
  • Perhaps no word should win yet.

A person may benefit from an instrument capable of helping thoughts become clearer. That does not mean every ambiguous internal state should be promoted into an explicit proposition because a decoder can achieve 0.83 confidence and has nothing else scheduled to do before lunch.

Some interior uncertainty is unfinished work.

Some is noise. Some is conflict.

Some is possibility.

Some is a path that disappears if named too quickly.

A powerful translator must therefore acquire a discipline almost opposite to its technical ambition. It must know when it could translate.

And leave the translation undone.

The Constitutional Translator.

Conduit may succeed.

Cool.

Then the translation layer needs a constitution.

1. Keep the Translation Visible.

A decoded representation must remain distinguishable from the neural activity from which it was inferred.

• The model thinks you meant X

is not the same sentence as:

• You meant X.

Even when the model is very good.

Especially then.

2. Separate the Directions.

Permission to record a signal does not automatically authorize every inference that future models can extract from it.

Inference permission does not automatically become retention permission.

Retention does not automatically become training permission.

External actuation does not automatically authorize neural stimulation.

A pathway does not acquire its return ticket by existing.

3. Preserve the Before.

When an adaptive system can change the state it later measures, preserve enough provenance to distinguish:

• learning about the person

from

• teaching the person to match the model.

Sometimes both will happen. That is fine.

Just do not erase the distinction afterward.

4. No Recursive Self-Authentication.

A system cannot use a preference state it helped produce as the sole evidence authorizing the process that produced it.

If stimulation makes continuation feel desirable, the resulting desire matters.

It is still downstream evidence.

Another route must remain.

5. Keep the Dictionary Contestable.

The person must be able to challenge the mapping.

Change it. Suspend it.

Compare alternatives.

Ask what information was discarded.

Ask what objective tuned the translator.

Ask why one internal difference became legible while another disappeared.

Fluency cannot become title.

6. Preserve Untranslated Space.

Technical decodability is a capability. It is not permission.

A person must retain regions in which uncertainty, interiority, unfinished thought and private cognition are allowed to remain unresolved by the machine.

7. Keep Correction Outside the Loop.

The system performing the read–write relation cannot become the only institution deciding whether the relation remains safe, wanted, accurate or legitimate.

Every adaptive loop needs an edge that the loop itself does not own.

The Ruling.

• Conduit demands telepathy.
• Modal Path Ethics hopes they get quite far.

That is why this article exists now instead of after something catches fire.

The company’s current public evidence is intriguing and incomplete. Conduit has built a large proprietary neuro-language dataset and published striking zero-shot examples. It has not yet published enough technical detail for an outside reader like me to independently evaluate the full mechanism or performance of its decoder. Its own technical page still promises a later account of the model.

Fine. Research continues.

The underlying problem no longer waits for Conduit anyway.

• Semantic neural decoding already exists.
• Brain-computer interfaces already co-adapt with their users.
• Assistive algorithms can shape learned neural representation.
• Cortical stimulation can already produce artificial percepts that people learn to use.
• Neural recordings can acquire new inferential significance when better decoders arrive.
  • The loop is here in pieces.
  • Conduit is trying to pull the pieces together.

So keep the vision.

A person thinks. A machine understands enough to help. The machine develops states worth returning. The person acquires a new sense for them.

The artificial intelligence stops feeling like software sitting across the room and becomes part of a new cognitive relation.

Cool. Now remember what happened.

• The machine learned the person.
• The person learned the machine.
  • The translator learned both.
  • The mapping became easier to use.
  • The distinction became easier to feel.
    • The instrument disappeared into experience.
    • Then one day somebody looked at the resulting fluency and said:

See? It understands you perfectly.

Maybe.

Look at the dictionary first.

It may have helped make the person easier to read.

Telepathy makes translation intimate enough to disappear from experience. It does not abolish translation.

A translator may participate in what a person becomes. It may never treat that participation as title over what the person is allowed to become next.

Conduit wants telepathy.

Modal Path Ethics wants the dictionary left open.