Skip to content
Back to the question

Evidence record

Can anyone read your mind?

Every sentence on the exploration belongs to at least one claim below. Each says what kind of statement it is, what it rests on, what the evidence does not establish, and how far it was checked. Nothing here is a score, and none of it is a verdict on a belief or a theory.

Last checked 24 September 2026.

  • Source checked: 34
  • A judgement, not a fact: 17

If something here is wrong, misquoted, out of date or unfair, write to support@unknownoceans.com and give the claim number, which does not change.

What the question takes for granted

A scientific question with a narrow, measured answer. The page answers it for one technique, non-invasive decoding of meaning from brain imaging, and keeps what has been shown in three and then five volunteers apart from what the phrase “mind reading” implies.

  • Brain activity carries the content of what a person hears, sees or imagines, in a form a model can learn

    Established

    Established for perception and deliberate imagery, in trained decoders and cooperative volunteers. Whether it extends to thought in general is what the page questions.

  • What a person reports about their own mind is what there is to read

    Contested

    The phrase assumes a mind is a store of contents that the owner can see and a machine might copy. The introspection experiments the page cites find that people often cannot report what moved them.

How the question is labelled

Claim: The words in the scanner

Kind
Framing. Signposting. It asserts nothing a source could check.

Claim: Mind reading

Kind
Framing. Signposting. It asserts nothing a source could check.

The question

Claim: A decoder can now recover the gist of what someone is hearing or imagining.

Kind
Experimental result. What an experiment or trial found, under its own conditions. A result in animals, cells or a simulation is not a result in people.
Review
Source checked. The cited passage was read, and it says this.

What this does not establish

Three volunteers, each with a decoder trained on their own recordings. Imagined speech was decoded less well than heard speech.

Later work looked for

Crossref, 2023-06 on, "semantic language decoding fMRI": the same group’s cross-participant follow-up (2025) extends it; nothing that retracts or fails to reproduce it was found.

Claim 3 · checked 24 September 2026 · AI-assisted audit

Claim: It needs their cooperation,

Kind
Experimental result. What an experiment or trial found, under its own conditions. A result in animals, cells or a simulation is not a result in people.
Review
Source checked. The cited passage was read, and it says this.

What this does not establish

Shown for this decoder with current methods. The 2023 authors say future developments might bypass the requirement.

Later work looked for

Crossref, 2023-06 on, "cross-participant semantic decoding": the 2025 follow-up reduces the data needed from a new person and its authors hold that cooperation is still required.

Claim 4 · checked 24 September 2026 · AI-assisted audit

Claim: and that turns out to be the interesting part.

Kind
This page’s reading. A judgement the page makes from the evidence. Others read it differently.
Review
A judgement, not a fact. Reviewed for whether it is fairly put and follows from the checked claims. No source can confirm it.

The page’s judgement of where the interest lies, argued in its second section.

Claim 5 · checked 24 September 2026 · AI-assisted audit

Claim: Machine learning applied to brain imaging can reconstruct the meaning of continuous language from a cooperative subject, once the decoder has been trained on that person’s own recordings.

Kind
Experimental result. What an experiment or trial found, under its own conditions. A result in animals, cells or a simulation is not a result in people.
Review
Source checked. The cited passage was read, and it says this.

  • Supports

    Semantic reconstruction of continuous language from non-invasive brain recordingsabstract

    Cited · Link answered, 19 October 2026 · Read in full, 24 September 2026 · Research in people · Experiment

    The source reads: Here we introduce a non-invasive decoder that reconstructs continuous language from cortical semantic representations recorded using functional magnetic resonance imaging (fMRI). ... we tested whether successful decoding requires subject cooperation, and found that subject cooperation is required both to train and to apply the decoder.

  • Qualifies

    Semantic language decoding across participants and stimulus modalitiesResults, first section

    Cited · Link answered, 19 October 2026 · Read in full, 24 September 2026 · Research in people · Experiment

    The source reads: For each pair of goal and reference participants, we trained a movie-based converter on brain responses to 70 minutes of silent movies, and a story-based converter on brain responses to 70 minutes of narrative stories

    The 2025 follow-up needs far fewer of the new person’s recordings, but still needs some.

Later work looked for

As claim 4.

Corrected

24 September 2026. Previously: “Machine learning applied to brain imaging can reconstruct the meaning of continuous language from a cooperative subject who has spent many hours training the decoder.” The 2025 follow-up by the same group trained a working decoder for a new person from seventy minutes of their recordings, so “many hours” is no longer a requirement.

Claim 6 · checked 24 September 2026 · AI-assisted audit

Claim: It produces paraphrase rather than transcript,

Kind
Experimental result. What an experiment or trial found, under its own conditions. A result in animals, cells or a simulation is not a result in people.
Review
Source checked. The cited passage was read, and it says this.

  • Supports

    Semantic reconstruction of continuous language from non-invasive brain recordingsDiscussion, third paragraph

    Cited · Link answered, 19 October 2026 · Read in full, 24 September 2026 · Research in people · Experiment

    The source reads: While our decoder successfully reconstructs the meaning of language stimuli, it often fails to recover exact words (WER 0.92–0.94 for the perceived speech test story).

Later work looked for

As claim 3.

Claim 7 · checked 24 September 2026 · AI-assisted audit

Claim: carries over to a new person only with some of that person’s recordings,

Kind
Experimental result. What an experiment or trial found, under its own conditions. A result in animals, cells or a simulation is not a result in people.
Review
Source checked. The cited passage was read, and it says this.

  • Supports

    Semantic reconstruction of continuous language from non-invasive brain recordingsResults, Privacy implications

    Cited · Link answered, 19 October 2026 · Read in full, 24 September 2026 · Research in people · Experiment

    The source reads: Decoders trained on cross-subject data (Extended Data Fig. 6) performed barely above chance, and significantly worse than decoders trained on within-subject data

  • Supports

    Semantic language decoding across participants and stimulus modalitiesSummary

    Cited · Link answered, 19 October 2026 · Read in full, 24 September 2026 · Research in people · Experiment

    The source reads: These results demonstrate that cross-participant decoding can reduce the amount of linguistic training data required from a goal participant

What this does not establish

In 2023 the transfer used anatomical alignment alone and failed. In 2025 it used functional alignment on the new person’s recordings and reached about half the performance of a person’s own decoder, in three participants.

Later work looked for

As claim 4; the 2025 paper is the newer work and the page now reports it.

Corrected

24 September 2026. Previously: “does not transfer between people,” Superseded: the same group transferred decoders between people in 2025, using functional alignment on seventy minutes of the new person’s recordings.

Claim 8 · checked 24 September 2026 · AI-assisted audit

Claim: and can be resisted by the subject at will.

Kind
Experimental result. What an experiment or trial found, under its own conditions. A result in animals, cells or a simulation is not a result in people.
Review
Source checked. The cited passage was read, and it says this.

What this does not establish

Resistance lowered decoding rather than always abolishing it, and counting did not significantly lower it. Three volunteers.

Later work looked for

Crossref, 2023-06 on, "decoder resistance mental privacy fMRI": nothing found that tests resistance again.

Corrected

24 September 2026. Previously: “and can be defeated by the subject at will.” The paper reports that decoding was significantly lowered and calls it resistance; one strategy, counting, did not significantly lower it. “Defeated” said more.

Claim 9 · checked 24 September 2026 · AI-assisted audit

Claim: Mind reading was a thought experiment for most of the history of the question.

Kind
This page’s reading. A judgement the page makes from the evidence. Others read it differently.
Review
A judgement, not a fact. Reviewed for whether it is fairly put and follows from the checked claims. No source can confirm it.

The page’s summary of the question’s past. In 2006 a review of decoding still called a general brain-reading device science fiction.

  • Context

    Decoding mental states from brain activity in humansTechnical and methodological challenges

    Cited · Link answered, 19 October 2026 · Read in full, 24 September 2026 · Research in people · Review of the field

    The source reads: these encouraging first steps should not obscure the fact that a ‘general brain reading device’ is still science fiction

Claim 10 · checked 24 September 2026 · AI-assisted audit

Claim: It is now a laboratory result with published limits,

Kind
What scholars argue. A report of what specialists argue, or where they disagree.
Review
Source checked. The cited passage was read, and it says this.

Claim 11 · checked 24 September 2026 · AI-assisted audit

Claim: which is a better position from which to worry about it.

Kind
This page’s reading. A judgement the page makes from the evidence. Others read it differently.
Review
A judgement, not a fact. Reviewed for whether it is fairly put and follows from the checked claims. No source can confirm it.

The page’s judgement: measured limits are easier to reason about than speculation.

Claim 12 · checked 24 September 2026 · AI-assisted audit

Claim: The limits are not incidental.

Kind
This page’s reading. A judgement the page makes from the evidence. Others read it differently.
Review
A judgement, not a fact. Reviewed for whether it is fairly put and follows from the checked claims. No source can confirm it.

Introduces claim 14.

Claim 13 · checked 24 September 2026 · AI-assisted audit

Claim: They come from how the decoder and the measurement work, and they are what makes the finding describable.

Kind
This page’s reading. A judgement the page makes from the evidence. Others read it differently.
Review
A judgement, not a fact. Reviewed for whether it is fairly put and follows from the checked claims. No source can confirm it.

The paper traces the paraphrasing to the decoder’s semantic features and the need for cooperation to how each person’s brain responds, not to the scanner alone.

  • Context

    Semantic reconstruction of continuous language from non-invasive brain recordingsDiscussion, third paragraph

    Cited · Link answered, 19 October 2026 · Read in full, 24 September 2026 · Research in people · Experiment

    The source reads: In our decoder, loss of specificity occurs when different word sequences with similar meanings share semantic features, causing the decoder to paraphrase the actual stimulus.

Corrected

24 September 2026. Previously: “They come from how the measurement works,” The paper attributes paraphrase to the decoder’s semantic features and the cooperation requirement to individual differences and attention, not only to the measurement.

Claim 14 · checked 24 September 2026 · AI-assisted audit

What has been done

Claim: What has been done

Kind
Framing. Signposting. It asserts nothing a source could check.

What the decoder does

Claim: What the decoder does

Kind
Framing. Signposting. It asserts nothing a source could check.

Claim: A 2023 study recorded the brains of three participants for sixteen hours each while they listened to stories, and trained a model on the recordings.

Kind
What scholars argue. A report of what specialists argue, or where they disagree.
Review
Source checked. The cited passage was read, and it says this.

  • Supports

    Semantic reconstruction of continuous language from non-invasive brain recordingsFig. 1 legend

    Cited · Link answered, 19 October 2026 · Read in full, 24 September 2026 · Research in people · Experiment

    The source reads: BOLD fMRI responses were recorded while three subjects listened to 16 h of narrative stories. An encoding model was estimated for each subject to predict brain responses from semantic features of stimulus words.

Corrected

24 September 2026. Previously: “A 2023 study trained a model on many hours of brain imaging while participants listened to stories.” Made specific from the paper: three participants, sixteen hours each.

Claim 17 · checked 24 September 2026 · AI-assisted audit

Claim: Afterwards it could generate text that recovered the meaning of what a participant was hearing, imagining, or watching in a silent film.

Kind
Experimental result. What an experiment or trial found, under its own conditions. A result in animals, cells or a simulation is not a result in people.
Review
Source checked. The cited passage was read, and it says this.

  • Supports

    Semantic reconstruction of continuous language from non-invasive brain recordingsabstract

    Cited · Link answered, 19 October 2026 · Read in full, 24 September 2026 · Research in people · Experiment

    The source reads: Given novel brain recordings, this decoder generates intelligible word sequences that recover the meaning of perceived speech, imagined speech, and even silent videos

  • Qualifies

    Semantic reconstruction of continuous language from non-invasive brain recordingsExtended Data Fig. 5 legend

    Cited · Link answered, 19 October 2026 · Read in full, 24 September 2026 · Research in people · Experiment

    The source reads: The decoder successfully recovered 72–82% of time-points during perceived speech, 41–74% of time-points during imagined speech, and 21–45% of time-points during perceived movies.

What this does not establish

Decoding was weaker for imagined speech and weaker again for silent films, and was measured in three people.

Later work looked for

As claim 3.

Corrected

24 September 2026. Previously: “Afterwards it could generate text matching the meaning of what” “Matching” overstated it: the paper says the text recovers the meaning, and for films only 21 to 45 per cent of moments were decoded.

Claim 18 · checked 24 September 2026 · AI-assisted audit

Claim: The output is a paraphrase.

Kind
Experimental result. What an experiment or trial found, under its own conditions. A result in animals, cells or a simulation is not a result in people.
Review
Source checked. The cited passage was read, and it says this.

  • Supports

    Semantic reconstruction of continuous language from non-invasive brain recordingsDiscussion, third paragraph

    Cited · Link answered, 19 October 2026 · Read in full, 24 September 2026 · Research in people · Experiment

    The source reads: In our decoder, loss of specificity occurs when different word sequences with similar meanings share semantic features, causing the decoder to paraphrase the actual stimulus.

Later work looked for

As claim 3.

Claim 19 · checked 24 September 2026 · AI-assisted audit

Claim: It captures the gist and often gets exact words wrong,

Kind
Experimental result. What an experiment or trial found, under its own conditions. A result in animals, cells or a simulation is not a result in people.
Review
Source checked. The cited passage was read, and it says this.

Later work looked for

As claim 3.

Corrected

24 September 2026. Previously: “frequently gets exact words wrong” The paper’s own word; a change of register only.

Claim 20 · checked 24 September 2026 · AI-assisted audit

Claim: because it scores candidate sentences by their meaning, and different wordings of one meaning look alike to it.

Kind
What scholars argue. A report of what specialists argue, or where they disagree.
Review
Source checked. The cited passage was read, and it says this.

  • Supports

    Semantic reconstruction of continuous language from non-invasive brain recordingsDiscussion, third paragraph

    Cited · Link answered, 19 October 2026 · Read in full, 24 September 2026 · Research in people · Experiment

    The source reads: In our decoder, loss of specificity occurs when different word sequences with similar meanings share semantic features, causing the decoder to paraphrase the actual stimulus.

Corrected

24 September 2026. Previously: “because the signal it reads is a slow blood-flow response rather than the firing of individual neurons.” The paper gives a different cause. It attributes the paraphrasing to the semantic features the decoder scores, and says motor features could separate a sentence from its paraphrases; the slowness of the signal is the obstacle its method was built to get around.

Claim 21 · checked 24 September 2026 · AI-assisted audit

Claim: What it is not

Kind
Framing. Signposting. It asserts nothing a source could check.

Claim: A scanner can now read anyone’s thoughts.

Kind
Framing. Signposting. It asserts nothing a source could check.

Claim: The decoder is trained on many hours of one person’s own data.

Kind
What scholars argue. A report of what specialists argue, or where they disagree.
Review
Source checked. The cited passage was read, and it says this.

Corrected

24 September 2026. Previously: “The decoder is trained per person on many hours of their own data,” Split so the transfer result could be stated with its newer qualification.

Claim 24 · checked 24 September 2026 · AI-assisted audit

Claim: Built from other people’s recordings alone, it did barely better than chance,

Kind
Experimental result. What an experiment or trial found, under its own conditions. A result in animals, cells or a simulation is not a result in people.
Review
Source checked. The cited passage was read, and it says this.

  • Supports

    Semantic reconstruction of continuous language from non-invasive brain recordingsResults, Privacy implications

    Cited · Link answered, 19 October 2026 · Read in full, 24 September 2026 · Research in people · Experiment

    The source reads: Decoders trained on cross-subject data (Extended Data Fig. 6) performed barely above chance, and significantly worse than decoders trained on within-subject data

What this does not establish

Seven volunteers, with anatomical alignment between brains.

Later work looked for

As claim 8.

Corrected

24 September 2026. Previously: “and performs poorly when applied to someone else.” The paper tested decoders built from other people’s data, not a trained decoder applied to a new brain, and reports “barely above chance”.

Claim 25 · checked 24 September 2026 · AI-assisted audit

Claim: and later transfers between people still needed the new person in the scanner.

Kind
Experimental result. What an experiment or trial found, under its own conditions. A result in animals, cells or a simulation is not a result in people.
Review
Source checked. The cited passage was read, and it says this.

  • Supports

    Semantic language decoding across participants and stimulus modalitiesResults, first section

    Cited · Link answered, 19 October 2026 · Read in full, 24 September 2026 · Research in people · Experiment

    The source reads: For each pair of goal and reference participants, we trained a movie-based converter on brain responses to 70 minutes of silent movies, and a story-based converter on brain responses to 70 minutes of narrative stories ... we separately aligned brain responses between the goal participant and the two reference participants

Later work looked for

As claim 8.

Corrected

24 September 2026. Previously: “” Added so the entry reports the 2025 cross-participant work.

Claim 26 · checked 24 September 2026 · AI-assisted audit

Claim: It produces a transcript of inner speech.

Kind
Framing. Signposting. It asserts nothing a source could check.

Claim: It produces a paraphrase of meaning.

Kind
Experimental result. What an experiment or trial found, under its own conditions. A result in animals, cells or a simulation is not a result in people.
Review
Source checked. The cited passage was read, and it says this.

Later work looked for

As claim 3.

Claim 28 · checked 24 September 2026 · AI-assisted audit

Claim: Exact wording is often wrong.

Kind
Experimental result. What an experiment or trial found, under its own conditions. A result in animals, cells or a simulation is not a result in people.
Review
Source checked. The cited passage was read, and it says this.

Later work looked for

As claim 3.

Claim 29 · checked 24 September 2026 · AI-assisted audit

Claim: It works whether or not you consent.

Kind
Framing. Signposting. It asserts nothing a source could check.

Claim: Training requires cooperation,

Kind
Experimental result. What an experiment or trial found, under its own conditions. A result in animals, cells or a simulation is not a result in people.
Review
Source checked. The cited passage was read, and it says this.

What this does not establish

The 2025 follow-up notes that someone could consent to recordings of films without knowing they could train a language decoder.

Later work looked for

As claim 4.

Claim 31 · checked 24 September 2026 · AI-assisted audit

Claim: and a trained decoder could be resisted by a subject silently naming animals or telling themselves a different story.

Kind
Experimental result. What an experiment or trial found, under its own conditions. A result in animals, cells or a simulation is not a result in people.
Review
Source checked. The cited passage was read, and it says this.

  • Supports

    Semantic reconstruction of continuous language from non-invasive brain recordingsExtended Data Fig. 5 legend

    Cited · Link answered, 19 October 2026 · Read in full, 24 September 2026 · Research in people · Experiment

    The source reads: within-subject decoders successfully recovered 52–57% of time-points when subjects passively listened, 4–50% of time-points when subjects resisted by counting by sevens, 0–3% of time-points when subjects resisted by naming animals, and 1–26% of time-points when subjects resisted by imagining a different story

  • Supports

    Semantic reconstruction of continuous language from non-invasive brain recordingsMethods, Experimental tasks

    Cited · Link answered, 19 October 2026 · Read in full, 24 September 2026 · Research in people · Experiment

    The source reads: For the “name” cue, subjects were instructed to name and imagine animals in their heads. For the “tell” cue, subjects were instructed to tell different stories in their heads.

What this does not establish

Three volunteers. Counting by sevens, the third strategy tested, did not significantly lower decoding.

Later work looked for

As claim 9.

Corrected

24 September 2026. Previously: “and a trained decoder can be defeated by the subject thinking about something else.” Not every distraction worked: counting did not significantly lower decoding. The entry now names the two strategies that did, and says “resisted”, the paper’s word.

Claim 32 · checked 24 September 2026 · AI-assisted audit

The cooperation requirement is a result, not a caveat

Claim: The cooperation requirement is a result, not a caveat

Kind
This page’s reading. A judgement the page makes from the evidence. Others read it differently.
Review
A judgement, not a fact. Reviewed for whether it is fairly put and follows from the checked claims. No source can confirm it.

The heading’s point is that the requirement was tested, as claim 34 reports.

Claim 33 · checked 24 September 2026 · AI-assisted audit

Claim: The same study tested that directly, and reports that cooperation is required both to train the decoder and to apply it.

Kind
Experimental result. What an experiment or trial found, under its own conditions. A result in animals, cells or a simulation is not a result in people.
Review
Source checked. The cited passage was read, and it says this.

What this does not establish

Seven volunteers for the training test and three for the resistance test.

Later work looked for

As claim 4.

Claim 34 · checked 24 September 2026 · AI-assisted audit

Claim: Its authors add that later developments might get around both;

Kind
What scholars argue. A report of what specialists argue, or where they disagree.
Review
Source checked. The cited passage was read, and it says this.

Corrected

24 September 2026. Previously: “” The paper qualifies its own privacy result, and the page reported the result without the qualification.

Claim 35 · checked 24 September 2026 · AI-assisted audit

Claim: in a 2025 follow-up they judged that the requirement still holds.

Kind
What scholars argue. A report of what specialists argue, or where they disagree.
Review
Source checked. The cited passage was read, and it says this.

  • Supports

    Semantic language decoding across participants and stimulus modalitiesDiscussion

    Cited · Link answered, 19 October 2026 · Read in full, 24 September 2026 · Research in people · Experiment

    The source reads: Since functional alignment requires training data collected with the goal participant’s cooperation, we believe that this conclusion still holds.

Corrected

24 September 2026. Previously: “” Added from the newer paper by the same authors.

Claim 36 · checked 24 September 2026 · AI-assisted audit

Claim: So the finding is two findings.

Kind
This page’s reading. A judgement the page makes from the evidence. Others read it differently.
Review
A judgement, not a fact. Reviewed for whether it is fairly put and follows from the checked claims. No source can confirm it.

Introduces claims 38 and 39.

Claim 37 · checked 24 September 2026 · AI-assisted audit

Claim: Meaning can be recovered from a non-invasive recording,

Kind
Experimental result. What an experiment or trial found, under its own conditions. A result in animals, cells or a simulation is not a result in people.
Review
Source checked. The cited passage was read, and it says this.

Later work looked for

As claim 3.

Claim 38 · checked 24 September 2026 · AI-assisted audit

Claim: and mental privacy currently has a physical basis rather than only a legal one.

Kind
This page’s reading. A judgement the page makes from the evidence. Others read it differently.
Review
A judgement, not a fact. Reviewed for whether it is fairly put and follows from the checked claims. No source can confirm it.

The page’s reading of claim 34. “Currently” carries the authors’ own caution that later methods might get around it.

  • Context

    Semantic reconstruction of continuous language from non-invasive brain recordingsDiscussion, last paragraph

    Cited · Link answered, 19 October 2026 · Read in full, 24 September 2026 · Research in people · Experiment

    The source reads: For these and other unforeseen reasons, it is critical to raise awareness of the risks of brain decoding technology and enact policies that protect each person’s mental privacy

Claim 39 · checked 24 September 2026 · AI-assisted audit

What it does not reach

Claim: What it does not reach

Kind
Framing. Signposting. It asserts nothing a source could check.

Claim: Privacy held here because each decoder had to learn one brain and could be refused attention, not because anyone designed it that way.

Kind
This page’s reading. A judgement the page makes from the evidence. Others read it differently.
Review
A judgement, not a fact. Reviewed for whether it is fairly put and follows from the checked claims. No source can confirm it.

Follows from claims 25 and 32: decoders built from other brains failed, and a subject attending elsewhere could lower decoding.

  • Context

    Semantic reconstruction of continuous language from non-invasive brain recordingsResults, Attention effects on decoding

    Cited · Link answered, 19 October 2026 · Read in full, 24 September 2026 · Research in people · Experiment

    The source reads: Moreover, these results demonstrate that subjects have conscious control over decoder output, and suggest that semantic decoders can only reconstruct what subjects are actively attending to.

Corrected

24 September 2026. Previously: “Privacy held here because the measurement is slow, not because anyone designed it that way.” The paper’s privacy results rest on the failure of decoders built from other people and on resistance by attending elsewhere, not on the slowness of the signal.

Claim 41 · checked 24 September 2026 · AI-assisted audit

How far the technique goes

Claim: How far the technique goes

Kind
Framing. Signposting. It asserts nothing a source could check.

Claim: The technique is a development rather than a break.

Kind
This page’s reading. A judgement the page makes from the evidence. Others read it differently.
Review
A judgement, not a fact. Reviewed for whether it is fairly put and follows from the checked claims. No source can confirm it.

Follows from claims 44 and 45.

Claim 43 · checked 24 September 2026 · AI-assisted audit

Claim: A 2006 review already described decoding simple mental states from imaging data, which way a pattern was moving or which of two overlapping patterns a person was attending to.

Kind
What scholars argue. A report of what specialists argue, or where they disagree.
Review
Source checked. The cited passage was read, and it says this.

  • Supports

    Decoding mental states from brain activity in humansp. 525, decoding perceptual features

    Cited · Link answered, 19 October 2026 · Read in full, 24 September 2026 · Research in people · Review of the field

    The source reads: direction of motion 11 and even perceived colour 10 of a visual stimulus presented to an individual can be predicted by decoding spatially distributed patterns of signals

  • Supports

    Decoding mental states from brain activity in humansDecoding unconscious or covert mental states

    Cited · Link answered, 19 October 2026 · Read in full, 24 September 2026 · Research in people · Review of the field

    The source reads: it is possible to identify which one of two superimposed oriented stimuli a person is currently attending to without necessarily requiring them to explicitly report where their attention is directed

Corrected

24 September 2026. Previously: “which of two images a person was attending to.” The review describes two superimposed oriented stimuli (striped patterns laid over each other), not two images.

Claim 44 · checked 24 September 2026 · AI-assisted audit

Claim: What changed by 2023 is that the target became continuous meaning rather than a choice between categories.

Kind
What scholars argue. A report of what specialists argue, or where they disagree.
Review
Source checked. The cited passage was read, and it says this.

  • Supports

    Decoding mental states from brain activity in humansExtrapolation to novel perceptual or cognitive states

    Cited · Link answered, 19 October 2026 · Read in full, 24 September 2026 · Research in people · Review of the field

    The source reads: brain reading will be restricted to simple cases with a fixed number of alternatives, for all of which training data are available

  • Supports

    Semantic reconstruction of continuous language from non-invasive brain recordingsIntroduction, first paragraph

    Cited · Link answered, 19 October 2026 · Read in full, 24 September 2026 · Research in people · Experiment

    The source reads: previous attempts to decode this information have been limited to identifying one output from among a small set of possibilities

Claim 45 · checked 24 September 2026 · AI-assisted audit

Claim: Two things could take it further.

Kind
This page’s reading. A judgement the page makes from the evidence. Others read it differently.
Review
A judgement, not a fact. Reviewed for whether it is fairly put and follows from the checked claims. No source can confirm it.

Introduces claims 47 and 49.

Corrected

24 September 2026. Previously: “Imaging hardware improves, and models trained across many people may eventually need less individual data. Neither of those is a barrier in principle, and neither has been crossed.” “Neither has been crossed” was out of date: in 2025 the same group transferred decoders between people with seventy minutes of the new person’s data. The paragraph now reports what each direction has shown.

Claim 46 · checked 24 September 2026 · AI-assisted audit

Claim: Hardware is one: blurred to the resolution of current portable systems, the 2023 recordings still decoded around half of the moments in a story.

Kind
Experimental result. What an experiment or trial found, under its own conditions. A result in animals, cells or a simulation is not a result in people.
Review
Source checked. The cited passage was read, and it says this.

What this does not establish

A simulation: the scanner recordings were smoothed to the spatial resolution of portable optical systems. No portable recording was decoded.

Later work looked for

Crossref, 2023-06 on, "fNIRS semantic language decoding": nothing found that decodes continuous meaning from a portable recording.

Claim 47 · checked 24 September 2026 · AI-assisted audit

Claim: Sharing models between people is another.

Kind
This page’s reading. A judgement the page makes from the evidence. Others read it differently.
Review
A judgement, not a fact. Reviewed for whether it is fairly put and follows from the checked claims. No source can confirm it.

Introduces claim 49.

Corrected

24 September 2026. Previously: “” Added with claim 49.

Claim 48 · checked 24 September 2026 · AI-assisted audit

Claim: In 2025 the same group carried decoders to a new person using seventy minutes of that person’s recordings, of stories or of silent films, and reached around half the performance of a decoder trained on the person’s own ten hours.

Kind
Experimental result. What an experiment or trial found, under its own conditions. A result in animals, cells or a simulation is not a result in people.
Review
Source checked. The cited passage was read, and it says this.

  • Supports

    Semantic language decoding across participants and stimulus modalitiesResults, first section

    Cited · Link answered, 19 October 2026 · Read in full, 24 September 2026 · Research in people · Experiment

    The source reads: For each pair of goal and reference participants, we trained a movie-based converter on brain responses to 70 minutes of silent movies, and a story-based converter on brain responses to 70 minutes of narrative stories

  • Supports

    Semantic language decoding across participants and stimulus modalitiesResults, first section

    Cited · Link answered, 19 October 2026 · Read in full, 24 September 2026 · Research in people · Experiment

    The source reads: We obtained a ceiling for decoding performance by comparing the cross-participant decoders to within-participant decoders trained on 10 hours of story data from the goal participant

  • Supports

    Semantic language decoding across participants and stimulus modalitiesResults, first section

    Cited · Link answered, 19 October 2026 · Read in full, 24 September 2026 · Research in people · Experiment

    The source reads: Across the test story, cross-participant decoding performance was around half of the within-participant ceiling

What this does not establish

Three participants, each decoded with decoders from the other two. Performance is measured against chance on one ten-minute story.

Later work looked for

Crossref, 2025 on, "cross-participant language decoding fMRI": nothing found that goes further without the new person’s data.

Corrected

24 September 2026. Previously: “” Newer work: the page said transfer between people had not been achieved.

Claim 49 · checked 24 September 2026 · AI-assisted audit

Claim: The authors hold that cooperation is still required, because the alignment needs the new person’s recordings.

Kind
What scholars argue. A report of what specialists argue, or where they disagree.
Review
Source checked. The cited passage was read, and it says this.

  • Supports

    Semantic language decoding across participants and stimulus modalitiesDiscussion

    Cited · Link answered, 19 October 2026 · Read in full, 24 September 2026 · Research in people · Experiment

    The source reads: Since functional alignment requires training data collected with the goal participant’s cooperation, we believe that this conclusion still holds.

What this does not establish

The same passage warns that someone could agree to be recorded watching silent films without realising the recordings could be used to decode language.

Corrected

24 September 2026. Previously: “” Added with claim 49.

Claim 50 · checked 24 September 2026 · AI-assisted audit

Claim: The slowness of the blood-flow signal is a limit of another kind.

Kind
This page’s reading. A judgement the page makes from the evidence. Others read it differently.
Review
A judgement, not a fact. Reviewed for whether it is fairly put and follows from the checked claims. No source can confirm it.

Introduces claims 52 to 56.

Corrected

24 September 2026. Previously: “The slowness of the blood-flow signal is a harder limit.” No source ranks it as harder, and the paper finds its own models a major source of error.

Claim 51 · checked 24 September 2026 · AI-assisted audit

Claim: It rises and falls over about ten seconds, so each image mixes the response to more than twenty words,

Kind
Observation. Something measured or seen, as the named source reports it.
Review
Source checked. The cited passage was read, and it says this.

Later work looked for

A property of the measurement the paper cites, not a finding later work revises.

Corrected

24 September 2026. Previously: “It smears activity over seconds, which is why the output is gist,” The paper gives the figures, and does not make the slow signal the reason the output is gist (see claim 21).

Claim 52 · checked 24 September 2026 · AI-assisted audit

Claim: and a language model has to propose the words the image cannot pin down.

Kind
What scholars argue. A report of what specialists argue, or where they disagree.
Review
Source checked. The cited passage was read, and it says this.

  • Supports

    Semantic reconstruction of continuous language from non-invasive brain recordingsIntroduction, second paragraph

    Cited · Link answered, 19 October 2026 · Read in full, 24 September 2026 · Research in people · Experiment

    The source reads: Decoding continuous language thus requires solving an ill-posed inverse problem, as there are many more words to decode than brain images. Our decoder accomplishes this by generating candidate word sequences

Claim 53 · checked 24 September 2026 · AI-assisted audit

Claim: The authors found their own models a major source of error,

Kind
Experimental result. What an experiment or trial found, under its own conditions. A result in animals, cells or a simulation is not a result in people.
Review
Source checked. The cited passage was read, and it says this.

Later work looked for

As claim 3.

Corrected

24 September 2026. Previously: “” The page said no modelling could improve the output; the paper finds model error a major part of what is lost.

Claim 54 · checked 24 September 2026 · AI-assisted audit

Claim: so better modelling should help.

Kind
What scholars argue. A report of what specialists argue, or where they disagree.
Review
Source checked. The cited passage was read, and it says this.

Corrected

24 September 2026. Previously: “” Added with claim 54; the authors’ own expectation.

Claim 55 · checked 24 September 2026 · AI-assisted audit

Claim: No modelling recovers information the measurement did not carry.

Kind
This page’s reading. A judgement the page makes from the evidence. Others read it differently.
Review
A judgement, not a fact. Reviewed for whether it is fairly put and follows from the checked claims. No source can confirm it.

A general point about measurement: a model can use what a recording holds, and can guess, but cannot add to it.

Corrected

24 September 2026. Previously: “and no amount of modelling recovers information the measurement did not carry.” Separated from the claim that the slow signal is why the output is gist.

Claim 56 · checked 24 September 2026 · AI-assisted audit

Whether thought is the right target

Claim: Whether thought is the right target

Kind
Framing. Signposting. It asserts nothing a source could check.

Claim: Decoding works on what a person is perceiving or deliberately imagining,

Kind
Experimental result. What an experiment or trial found, under its own conditions. A result in animals, cells or a simulation is not a result in people.
Review
Source checked. The cited passage was read, and it says this.

What this does not establish

The paper tested perception, imagined speech and attention. It did not test other kinds of thought, so this says where decoding has worked, not where it cannot.

Later work looked for

As claim 3.

Claim 58 · checked 24 September 2026 · AI-assisted audit

Claim: which is the part of mental life that behaves like a signal.

Kind
This page’s reading. A judgement the page makes from the evidence. Others read it differently.
Review
A judgement, not a fact. Reviewed for whether it is fairly put and follows from the checked claims. No source can confirm it.

The page’s reading of claim 58.

Claim 59 · checked 24 September 2026 · AI-assisted audit

Claim: Much of mental life is not that.

Kind
This page’s reading. A judgement the page makes from the evidence. Others read it differently.
Review
A judgement, not a fact. Reviewed for whether it is fairly put and follows from the checked claims. No source can confirm it.

The page’s judgement. The processes behind choices and judgements, which the next paragraph describes, are one part of mental life that is neither perceived nor deliberately imagined.

  • Context

    Telling more than we can know: Verbal reports on mental processesabstract

    Cited · Link answered, 19 October 2026 · Read in full, 24 September 2026 · Research in people · Review of the field

    The source reads: Evidence is reviewed which suggests that there may be little or no direct introspective access to higher order cognitive processes.

Corrected

24 September 2026. Previously: “Most of what a brain does is not that.” No cited source measures what share of brain activity is perception or imagery. The sentence is now a judgement about mental life, which is what the paragraph is about.

Claim 60 · checked 24 September 2026 · AI-assisted audit

Claim: Experiments on introspection find something different: asked what led to a choice or a solution, people name causes the experimenters know had no effect, and deny ones the experimenters know were at work.

Kind
Experimental result. What an experiment or trial found, under its own conditions. A result in animals, cells or a simulation is not a result in people.
Review
Source checked. The cited passage was read, and it says this.

  • Supports

    Telling more than we can know: Verbal reports on mental processespp. 243-244, position effects

    Cited · Link answered, 19 October 2026 · Read in full, 24 September 2026 · Research in people · Review of the field

    The source reads: When asked about the reasons for their choices, no subject ever mentioned spontaneously the position of the article ... when asked directly about a possible effect of the position of the article, virtually all subjects denied it

  • Supports

    Telling more than we can know: Verbal reports on mental processesproblem-solving studies (Maier)

    Cited · Link answered, 19 October 2026 · Read in full, 24 September 2026 · Research in people · Review of the field

    The source reads: All of these subjects reported that the useless cue had been helpful and denied that the critical cue had played any role in their solution.

  • Qualifies

    Telling more than we can know: Verbal reports on mental processesabstract

    Cited · Link answered, 19 October 2026 · Read in full, 24 September 2026 · Research in people · Review of the field

    The source reads: though people may not be able to observe directly their cognitive processes, they will sometimes be able to report accurately about them

What this does not establish

A review of studies from the 1970s and before, several small. Nisbett and Wilson say people’s reports are sometimes accurate, when the real cause is salient and plausible.

Later work looked for

Crossref, 2005 on, "choice blindness introspection": later choice-blindness studies by Johansson and colleagues extend the finding (titles and records only); nothing found that reverses it.

Corrected

24 September 2026. Previously: “The experiments that read introspective reports rather than perception find something different: people confidently describe reasons for their own choices that the experimenters know to be wrong.” The review does not measure confidence, and its examples are of two kinds: false causes named and real causes denied. The sentence now says both.

Claim 61 · checked 24 September 2026 · AI-assisted audit

Claim: There may be less to read than the phrase suggests.

Kind
This page’s reading. A judgement the page makes from the evidence. Others read it differently.
Review
A judgement, not a fact. Reviewed for whether it is fairly put and follows from the checked claims. No source can confirm it.

The page’s inference from claims 58 and 61, stated as a possibility.

Other readings

Nisbett and Wilson’s finding is about what people can report of their own mental processes, not about what a recording of the brain contains; a decoder might read what introspection misses.

Claim 62 · checked 24 September 2026 · AI-assisted audit

What was read

  • Semantic reconstruction of continuous language from non-invasive brain recordingsTang, LeBel, Jain and Huth, Nature Neuroscience · Author manuscript, NCBI PMC11304553 via E-utilities (Europe PMC fullTextXML answered 500)Read in full, 24 September 2026 · Copyright; author manuscript; quoted brieflyMain text, Methods and Extended Data legends read. Supplementary tables and video not read.
  • Telling more than we can know: Verbal reports on mental processesNisbett and Wilson, Psychological Review · Publisher scan posted for an MIT course (web.mit.edu/curhan/www/docs/Articles/15341_Readings/Social_Cognition/Nisbett_Wilson_1977_Telling_more_than_we_can_know.pdf)Read in full, 24 September 2026 · Copyright; quoted briefly
  • Decoding mental states from brain activity in humansHaynes and Rees, Nature Reviews Neuroscience · Publisher PDF as posted for a University of Texas at Dallas course, read through the Wayback Machine (web.archive.org/web/2016id_/http://www.utdallas.edu/~otoole/HCS6330_F09/17_Haynes_decoding_NNR_06.pdf)Read in full, 24 September 2026 · Copyright; quoted briefly
  • Semantic language decoding across participants and stimulus modalitiesTang and Huth, Current Biology · Author manuscript, NCBI PMC11903136 via E-utilities (Europe PMC fullTextXML answered 500)Read in full, 24 September 2026 · Copyright; author manuscript; quoted brieflyMain text and STAR Methods read; supplementary figures and tables not read.