In the detective television series Homicide: Life on the Street, interrogating suspects may often include the question, “Have you seen this person before?”  This being a 1-hour television show, the suspect shown is normally guilty and often attempts to evade prosecution by lying, “No, never seen him before…” whereby the detectives then need to use the remaining time to build their case.  They need to walk the streets and find direct or indirect evidence that supports their hypothesis that the suspect has indeed seen the person and has in fact been involved in a capital crime.

Imagine if the same scene can be played again, this time within a massive fMRI scanner.  “Have you seen this person before?  Wait, don’t bother answering.  According to this fMRI scan, you have.  Guilty.  Done.  Now what do we do for the next 59 minutes?”  Technically, a real life fMRI can do this.  This is not science fiction.  This is today.  How would it work and why is it not in everyday use?

Nothing is ever done in isolation and it pays to briefly scan the background with similar technology and with similar labor-reducing goals.  A customer service desk needs to hire customer service employees that await incoming calls.  When a caller connects with the service employee, they exchange greetings (“Hi, how can we help you?”) and get down to the business (“Hi, I need help checking on my order…”).  It is a mundane phone call that allows two people to communicate at a distance. 

To reduce costs, enhance service employee productivity, and provide 24-hour coverage at greatly reduced operating costs, the customer service desk might add machine operators.  In this case, a caller connects to the machine operator and hears a recording (“For order status say ‘status.’  For new orders say ‘new order.’”)  To get a mechanical view of what the machine operator hears back and needs to interpret, figure 1 shows how the speech waveform might appear.

Figure 1.  A sample view of how the speech waveform for “Status” and “New Order” might appear.

The machine operator needs to interpret the speech waveform pattern and translate it back into a menu path - “Status” or “New Order.”  To a human operator, this is a trivial task.  Considering that this is a two-alternative-forced-choice task, non-humans could do this.  But for a machine, it needs to consider linguistics, natural language processing, speech sample time warping, and a heavy dose of machine learning pattern recognition.  This despite the fact that “Status” has two syllables while “New Order” has three.  It bears stating that this is a hard task for a machine operator. 

Reading a brain fMRI uses the much of the same tools, albeit on both spatial and temporal scales.  Figure 2 shows a sample fMRI image.   

Figure 2.  Sample brain fMRI scan image.

fMRI scans record oxygen use in brain regions at any given point in time.  More oxygen use in a region implies more activity.  Correlating the brain regions with a brain function atlas could indicate which functions just occurred.  The colors in a typical fMRI image indicate changes in oxygen use.  This forms a spatio-temporal pattern of brain activity.  Norman, et al. (2006) provides a brief overview of pattern analysis in fMRI data.   If visually recognizing a known person activates a specific pattern of brain activity detectable via fMRI, then seeing that specific fMRI pattern again implies visual recognition of the same person.

In a crime interrogation scenario, if the detectives could see the suspect’s fMRI scan patterns while posing the question, “Have you seen this person before?” they can get their answer.  While dimensionally more complex than the speech waveform pattern recognition task, the sample warping, normalization, and pattern recognition algorithms essentially remain the same.  Detecting whether a suspect recognizes a photo or not follows the same procedure as determining whether a caller said, “Status,” or “New Order.” 

However, there are key functional differences that render this simplistic technical process translation moot.  Speech is voluntary; a caller can say, “Status” or “New Order” or possibly remain silent.  The machine operator only needs to be programmed to detect one of two potential choices in a direct answer detection mode.  A brain is involuntarily constantly active; the suspect could recognize, partially recognize, be distracted, be confused, not recognize, and still need to breathe, plan, feel hungry, feel sleepy, tap the foot, feel the seat underneath, and more, all of which gets picked up in an fMRI scan.  Speech can be isolated into a static sample.  Brain fMRI scans may need to be placed in context over time and space. 

Speech into an automated machine operator is deliberate and transmitted as a deliberate message intended to be unambiguous.  Brain activity is internal.  Reading brain activity inside the skull while it quietly transmits neurotransmitters in the dark is far more intrusive than interpreting audio signals intended for public consumption.  Using an fMRI to read a brain is like placing hidden cameras in someone’s dressing room.  Privacy concerns aside, most of the time nothing of note happens.   When something does happen, it has no context unless anchored to a known external event.  Inducing a known external event (e.g. posing a question) biases the internal activity and introduces non-stationary behavior.  Comparison to past baseline patterns becomes moot.  

Speech is fundamentally one-dimensional.  A speaker may think and plan several different utterances concurrently, but is physically constrained to produce only a single intentional vocalized output via a single larynx.  The difference between thoughts and speech may not be solely technical differences that can be bridged with better data mining, noise cancellation, and pattern recognition techniques.  Finding a single coherent, un-vocalized, internalized thought in an active brain is not merely an issue of finding a needle in a haystack.  It is an issue of determining the intent of a person before that person knows their own intent.  And doing so without introducing the observer’s intent.  In legal terms, reading unspoken thoughts in the brain with an fMRI machine picks up so many divergent and immature thoughts that  can be directed and interpreted so many ways it should be considered an objectionable leading question.