Brain-based memory detection and the new science of mind reading

Chapter 81 in Kahana and Wagner (eds.), The Oxford Handbook of Human Memory (Oxford University Press, 2024), pp. 2329–2350, doi 10.1093/oxfordhb/9780190917982.013.81. Portions are adapted from Murphy and Rissman (2020), Journal of Law and the Biosciences 7(1):lsaa078, which was not read. Read from the publisher PDF posted on the first author's lab website (rissmanlab.psych.ucla.edu). Read for research direction R2, 4 October 2026. Provenance: papers/amadeus/rissman2024_memory_detection.provenance.json.

What was read

• Read in full: all 25 pages, converted with pdftotext -layout: abstract, every section, conclusion, note and reference list.
• Extraction caveat: the PDF's fi, ff, fl and ffi ligatures have no Unicode mapping, so the text shows a gap where they were ("pro les" for profiles). Read with that in mind; no content is missing.
• This is a review. The primary studies it describes (Rissman 2010 and 2016, Uncapher 2015, Peth 2015, Chow 2018, Meixner and Rosenfeld 2014 and others) were not read here. Their numbers below are as this chapter reports them.

Scope

Memory detection asks whether a stimulus is recognised. Lie detection looks for the arousal and conflict of producing a false answer. The chapter covers memory detection only, and points to Farah et al. (2014) for fMRI lie detection.

Main content (verified as the chapter's statements)

EEG / P300 concealed-information test (CIT)

• The test. Rare crime-relevant "probes" appear among frequent "irrelevants" and instructed "targets". A recognised probe evokes a P300 like a target's. Individuals are classified by bootstrapping the averaged ERPs against a set cutoff, with an "indeterminate" outcome possible.
• Reported accuracies. Autonomic measures (skin conductance, heart rate, respiration) give 80–90% in discriminating guilty from innocent. P300 versions report more than 85%, occasionally 100%.
• Countermeasures.
  • The original protocol was beaten by covert actions such as wiggling a toe on irrelevants.
  • The Complex Trial Protocol (CTP) reports more than 90% sensitivity and specificity even under countermeasures.
  • Mental countermeasures, such as silently saying one's name, can be "lumped" with the motor response so that reaction-time telltales disappear. Even then more than 80% of guilty subjects were detected.
  • Attempts to suppress the memory gave mixed results.
• Who did the work. The field is dominated by two groups: Rosenfeld (academic) and Farwell (commercial "brain fingerprinting", which claims no errors ever and is criticised for undisclosed methods and selective reporting). Only one independent replication of the CTP exists. It largely replicated the findings but found countermeasure vulnerabilities.
• Ecological validity problems.
  • Foils. Probe and foil choice can create false positives: a bat may stand out as the only wooden object, or for someone who played baseball as a child. They can also create false negatives: a thief with other undiscovered crimes may recognise the "irrelevant" jewellery.
  • Leaked details. "Innocent but informed" people who knew the crime details were indistinguishable from the guilty.
  • Weak encoding. Incidental details are often barely encoded (intoxication, darkness, haste). The absence of a P300 is therefore not evidence of innocence or of a false alibi.
  • Delay. One study found no loss of detection after a month, but with a single, instructed, highly salient probe.
• A real-world test. In Meixner and Rosenfeld (2014) people wore a camera for a day; probes were words for their own activities. 12 knowledgeable and 12 non-knowledgeable subjects were separated perfectly.
• Single trials. Single-trial EEG pattern classification of memory status reached only 56–61% (50% is chance). A subject-level classifier reached about 92%, but on 6 faces.
• In the courts. Brain fingerprinting was admitted once in a US post-conviction hearing (Harrington v. Iowa, 2001) but did not decide the case. India's BEOS EEG procedure has been used in hundreds of investigations, once as evidence in a 2008 murder conviction that was later overturned. Singapore has licensed it.

fMRI

• True against false memories. Across studies the networks for true and false memories overlap heavily. The sensory-reactivation account (Slotnick and Schacter 2004) has mixed support. There is no "spot" that marks a memory as true.
• Rissman et al. 2010 (faces, MVPA).
  • Behavioural accuracy was about 70%.
  • Hits against correct rejections: 83%, rising to 95% for the classifier's most confident guesses.
  • True against false recognition, holding confidence constant: 59%.
  • Studied against unstudied faces among those judged new: chance.
  • In an implicit task (gender judgment), old against new was chance.
  • Classifiers transferred across people about as well as within.
  • Conclusion: the patterns track the subjective memory state, not experiential history.
• Uncapher et al. 2015. Simple mental countermeasures, such as attending to superficial features of familiar faces or thinking of a lookalike for new ones, brought classification to chance. The memories were weak: faces seen twice for 2 s.
• Rissman et al. 2016 (3 weeks of a wearable camera).
  • Self/other judgments were about 80% correct. That left too few false memories to test true against false.
  • Own-life hits against others'-life correct rejections: 91%, with no subject below 80%.
  • Recollection against familiarity: 72%.
  • Events 3 weeks old were decoded as well as 1-week-old ones. Cross-person classifiers worked, and lab-trained classifiers transferred to life photos above two thirds.
• Peth et al. 2015 (mock crime, fMRI). Knowledge of crime details was detectable. But people who did the crime, people who only planned it, and people merely informed of the details could not be told apart (chance).
• Chow et al. 2018. Patterns in an autobiographical network told own experiences from others'. Patterns in a "laboratory memory" network told whether photos had been previewed the day before. Firsthand experience and secondhand familiarity draw on partly separate processes.
• Content. Content decoding during recall works at the level of category (face, scene, object), not reliably for which specific item is recalled on a single trial.

The authors' bottom line. Even with the best tools there may be no reliable way to:

• (a) tell a false but believed memory from a true one;
• (b) detect simple mental countermeasures on single trials;
• (c) tell participation from mere knowledge of, or intention towards, an event.

They attribute this to the constructive nature of memory more than to the instruments. Retrieval fills gaps with schematically plausible detail, so "it is unlikely that fMRI will ever ... provide a veridical readout of what a person experienced in the past". Individual differences in autobiographical memory may further complicate general-purpose detectors.

Limits of the chapter

• A narrative review by authors of several of the fMRI studies it covers. No systematic search, no pooled accuracy estimates.
• Fully forensic in framing (guilt, witnesses). It does not address cooperative settings where the person wants to be read accurately.
• Several reported EEG accuracies come from one group or from a commercial source with undisclosed methods. The chapter flags this.

What it means for Amadeus (inference)

• Neural memory signals read the person's belief, not the past. The strongest results decode subjective recognition, recollection and strength, which the person can also report. Brain patterns fail where belief and history diverge (59%, or chance). So the "brain" channel cannot audit self-deception about one's own memories: a false but believed memory looks like a true one. For P6 this is the most important qualification. The brain can corroborate that the person holds a memory; it cannot show the memory is true. Historical truth needs contemporaneous records ("did"), not neural data.
• This supersedes the hope in the architecture doc that deliberative recall under recording could show "whether a memory is genuine or reconstructed (Slotnick and Schacter)". That rested on a group-level contrast. The trial-level evidence, from studies built to give neural decoding its best chance, says no, or at best weakly.
• Detecting lies to others needs conditions Amadeus will not have. It requires a known ground truth (the probe), carefully matched foils, details that have not leaked, and a subject not using countermeasures. In a cooperative month of recording these conditions largely fail or are unnecessary. A person who wants to hide something from their own replica can beat fMRI recognition decoding with a simple mental strategy.
• Firsthand against secondhand knowledge is the most promising neural lead (Chow 2018). It is exactly the provenance question of brief 5.3 ("did I live this or hear about it?"). But the result is at group or network level with camera ground truth, and it was not tested on old memories.
• Wearable cameras (and the EAR audio of Sun and Vazire) are the realistic ground-truth source. Every positive real-world result here used camera records to know what happened. A month-long Amadeus protocol should log contemporaneous records for that reason. They are what later lets recall be scored, by the person and by any neural decoder.