Skip to main content

Memory on Trial: fMRI Pattern Analysis and Mental Privacy

By Erin Morrow

Image courtesy of Jamie Allen on Flickr
The cost of wrongful incarceration manifests with devastating consequence – over 2,500 prisoners currently reside on death row in the United States (Death Penalty Information Center, 2020), with 20 individuals having been exonerated via biological evidence as of mid-2019 with an average of over a decade spent in prison (Innocence Project, 2020). The potential to reduce the number of erroneous convictions in addition to accurately identifying guilty individuals is immediately relevant to the well-being of the innocent and the well-functioning of the legal system. At times fraught with contention and faced with public demands for reconsideration of evidence, it is easy to see how this system might benefit from a clarifying tool that generalizes to crimes outside of those implicating bodily fluids (and therefore DNA evidence). In particular, the domain of memory becomes extremely important. The consequences of assessing memory accurately in a legal context are paramount despite the fallibility of the current toolbox, which includes examinations like the polygraph that can be readily influenced by physiological countermeasures used to ‘mislead’ the machine (Bles & Haynes, 2008).

Thus, the more direct tool of neurotechnology has been considered to evaluate memory and/or affect how memories are accessed. For instance, some have recognized transcranial magnetic stimulation (TMS) – a method that involves using a coil positioned on the exterior of the skull to temporarily inhibit the activity a targeted brain region – as a way to potentially ‘enhance’ memory retrieval. This would occur by weakening false memories, presumably improving the quality of evidence provided during an investigation (for review, see Vedder & Klaming, 2010). However, when the testimony of the defendant is in question, it becomes important to be able to verify the content of a specific memory, especially when the presence of other witnesses is limited. Techniques developed over the past decade and a half in functional magnetic resonance imaging (fMRI) – which measures relative blood flow in the brain as a surrogate for neuronal activity – may assist in doing so.

Image courtesy of Free SVG
Multivoxel pattern analysis (MVPA) takes the hemodynamic information collected in fMRI over time in each voxel (one of thousands of three-dimensional volumes dividing up the brain) and attempts to identify a common pattern spanning many voxels that is present during an experimental condition, such as viewing images of faces. A classifier, which can operate using linear or nonlinear parameters, then trains on many iterations of differing conditions (e.g., various faces versus various houses) to learn the activity pattern associated with each condition. Finally, this classifier can be tested on novel stimuli or situations – like new faces and houses – to identify which of these previous conditions an individual is perceiving or thinking based on their brain activity. Accuracy can be subsequently evaluated and improved, typically within a single individual (Norman et al., 2006).

Researchers have used MVPA in multiple domains, perhaps most widely in perception to identify which category of visual stimuli an individual is viewing, as described in the example above (for review, see Norman et al., 2006). However, MVPA has also been used to train classifiers on brain activity typical of truth-telling versus lying (Davatzikos et al., 2005), with at least two businesses offering fMRI services to clients eager to demonstrate their innocence (see No Lie MRI, CEPHOS; see Brain Fingerprinting Laboratories for use of event-related potentials, or ERP) (Greely & Illes, 2007). This assessment has obvious legal implications and has even been formally offered as evidence in court, although rejected at least once (Miller, 2010). The movement for fMRI lie detection has proceeded despite objections, for example, to the heterogeneity of lying and the translation of laboratory paradigms to ‘real-world’ scenarios (Meegan, 2008; Miller, 2010).

However, perhaps more astonishingly, researchers have also used MVPA to target memory processes. In 2010, Rissman and colleagues examined specific memories by successfully classifying both individuals’ degree of recognition for a given face (i.e., having a level of ‘familiarity’ with versus ‘recalling’ the face with context) and whether they subjectively identify a given face as ‘old’ or ‘new.’ Although the findings of this study are limited in that the classifier performed more poorly when assessing whether the faces were objectively ‘old’ or ‘new,’ this has important implications for the future of testimony (Rissman et al., 2010).

Imagine a scenario in which an image of the murder weapon – within the context of the original crime scene – is shown to the defendant during an fMRI scan. Based on classifiers not unlike those developed so far, brain activity could one day have the power to exonerate or condemn based on a memory only the offender would have; this is analogous to an investigation a decade ago in India (Giridharadas, 2008). Indeed, it is not difficult to envision the significance of such an analysis in an era of #MeToo and #BlackLivesMatter. As the world engages with highly visible trials such as that of Harvey Weinstein, conversations surrounding the reliability of memory are prominent – perhaps fMRI pattern analysis could offer a new perspective.

Image courtesy of Public Domain Pictures
But does this assessment infringe a fundamental privacy, a sole ownership of one’s own thoughts? MVPA has drawn comparisons to ‘mind-reading,’ given its capacity to distinguish between categories of seemingly nuanced mental states with over 90% accuracy in some cases (e.g., Davatzikos et al., 2005).  However, this outlook may be overconfident (Farah, 2005). University of Guelph’s Daniel Meegan (2008) argues that the memory processes in question may be unconscious; therefore, the process is akin to collecting a physical trace, like a strand of hair, from a crime scene. He claims that obtaining this kind of neural evidence is hardly more private than DNA extraction; any brain signal would require several layers of interpretation – none of which directly elucidate a memory as we understand it – and would need to be prompted by the same stimuli as present during encoding. He suggests that, although future neurotechnology may get closer to representing memory more clearly and would require renewed ethical consideration, fMRI analyses do not yet have the capacity to divulge this information. Limitations include length of time since original encoding (i.e., how long ago the initial crime occurred), generally good but not always very-high accuracy (introducing the potential for both false-positives and false-negatives), and a few possible countermeasures. For example, a potential countermeasure could involve a guilty defendant imagining a new event when shown ‘old’ stimuli, triggering a false-negative. This scenario assumes that the retrieval is driven instead by intentional processes and therefore able to be manipulated (Meegan, 2008).

Mart Bles and John-Dylan Haynes (2008) offer a similar argument in that fMRI pattern analysis is intrinsically limited; “decoding techniques perform pattern recognition, not pattern interpretation” (p. 88). As such, MVPA relies on experimenters identifying subjective mental states in order to then categorize them. These experimenters would also have to collect an infinitely large data set over an extended period of time to properly train the classifier on a given individual’s brain activity. Furthermore, given the incredible sensitivity of MRI to motion artifacts, individuals could easily distort data by moving during the scan (although, if trained beforehand, MVPA may be able to detect these and other non- motion-based countermeasures) (Bles & Haynes, 2008).

Aside from the analysis process itself, Farah and colleagues (2014) note that consent and data privacy also remain an issue. Although no analysis method yields results that are 100% certain, individuals must be comfortable with a variable margin of error and the potential for third-party interpretation of fMRI data that may be less than accurate. Parties must also be aware of “the indirect coercion that results when refusal to take the test is seen as indicative of guilt” (p.129). Additionally, legal limits should be considered in the future handling and storage of fMRI data, notwithstanding its potential societal benefit to assist law enforcement in identifying the guilty and – perhaps more importantly – exonerating the innocent (Farah et al., 2014).

With new classifiers and applications of MVPA being developed, it is becoming increasingly important to consider and monitor the impact of fMRI pattern analysis on the legal system. Although findings indicate that these analyses are not yet capable of ‘mind-reading’ to the extent popularized in science fiction, we ought to consider now the impact of memory assessment through neuroimaging.


  1. Bles, M., & Haynes, J. D. (2008). Detecting concealed information using brain-imaging technology. Neurocase, 14(1), 82-92.
  2. Davatzikos, C., Ruparel, K., Fan, Y., Shen, D. G., Acharyya, M., Loughead, J. W., … & Langleben, D. D. (2005). Classifying spatial patterns of brain activity with machine learning methods: Application to lie detection. Neuroimage, 28(3), 663-668.
  3. Death Penalty Information Center. (2020). Death row. Retrieved March 19, 2020, from
  4. Farah, M. J. (2005). Neuroethics: The practical and the philosophical. Trends in Cognitive Sciences9(1), 34-40.
  5. Farah, M. J., Hutchinson, J. B., Phelps, E. A., & Wagner, A. D. (2014). Functional MRI-based lie detection: Scientific and societal challenges. Nature Reviews Neuroscience15(2), 123-131.
  6. Giridharadas, A. (2008, September 14). India’s Novel Use of Brain Scans in Courts Is Debated. The New York Times, p. 10.
  7. Greely, H. T., & Illes, J. (2007). Neuroscience-based lie detection: The urgent need for regulation. American Journal of Law & Medicine, 33(2-3), 377-431.
  8. Innocence Project. (2020). Exoneration statistics and databases. Retrieved March 19, 2020, from
  9. Meegan, D. V. (2008). Neuroimaging techniques for memory detection: Scientific, ethical, and legal issues. The American Journal of Bioethics, 8(1), 9-20.
  10. Miller, G. (2010). fMRI lie detection fails a legal test. Science, 328(5984), 1336-1337.
  11. Norman, K. A., Polyn, S. M., Detre, G. J., & Haxby, J. V. (2006). Beyond mind-reading: Multi-voxel pattern analysis of fMRI data. Trends in Cognitive Sciences, 10(9), 424-430.
  12. Rissman, J., Greely, H. T., & Wagner, A. D. (2010). Detecting individual memories through the neural decoding of memory states and past experience. Proceedings of the National Academy of Sciences, 107(21), 9849-9854.
  13. Vedder, A., & Klaming, L. (2010). Human enhancement for the common good—Using neurotechnologies to improve eyewitness memory. AJOB Neuroscience, 1(3), 22-33. 

Erin Morrow is an undergraduate student at Emory University majoring in Neuroscience and Behavioral Biology. Along with having a particular interest in the ethical ramifications of altering and accessing memories and the impact of neurotechnology, Erin assists with neuroimaging analysis and memory research at the Hamann Cognitive Neuroscience Laboratory. She hopes to integrate her pursuits in neuroethics with her engagement in volunteerism and her future academic aspirations in research.

Want to cite this post?

Morrow, E. (2020). Memory on Trial: fMRI Pattern Analysis and Mental Privacy. The Neuroethics Blog. Retrieved on , from


  1. This is fascinating . What an enlightening read!


Post a Comment

Emory Neuroethics on Facebook