7
$\begingroup$

I'm intrigued by (and want to write a story on) the social implications of research in neurology that is improving our understanding of the brain. I've seen studies that have done everything from used fMRI machines to identify the correlation of brain activity to thoughts of certain objects to sending brain signals across the world

I want to write a short story that is at least within the plausible penumbra of outcomes of the development of this technology. Sometime in the future, a combination of human ingenuity, technological development, vastly improved AI, and a lot of luck has resulted in an extraordinarily powerful brain-scanning technology—that can, slightly (but acceptably) implausibly, detect individual neurons firing— that can be (unethically) deployed covertly. My starting point is that the government covertly scans and observes a political opponent's surroundings and their internal brain activity over a period of six months, and all this time plugs the data of both the brain and the person's words and actions into a ridiculously powerful AI that machine "learns" the idiosyncratic correlations of that person's brain.

However, where I'm stuck, is figuring out the plausible limits of what this best possible brain scan and computer can figure out. In my story, I can justify that the AI figures out that the target is, for example, inappropriately secretly sexually aroused by something (good for blackmail), or is secretly worried or angry. But can any scanning technology make the leap from observing brain patterns to anything even close to what we would think of as "thought reading", as a telepath in a fantasy or superhero would? This is where my not being a neurologist hampers me: I don't understand to what extent my internal monologue is an empirically identifiable scientific state or a complete black box? I'd like to have the AI be able to use the correlations achieved by the suspect talking out loud to figure out when someone was thinking of particular words, but I am really not sure if being able to read someone's thoughts—even by use of technology—is just magic disguised as being on the edge of scientific plausibility.

$\endgroup$
6
  • 1
    $\begingroup$ If it were, I bet it would be invasive. We can't even probe machines without filling them full of sensor stuff. $\endgroup$
    – DKNguyen
    Jun 21, 2021 at 2:33
  • 2
    $\begingroup$ The big hurdle is not detecting brain activity, and reading memories, but understanding them. We have absolutely no clue how actual though works, nor memory. Some think that every human brain has its own, chaos-fractal-based encoding scheme for memories! $\endgroup$
    – PcMan
    Jun 21, 2021 at 7:32
  • $\begingroup$ I think basically, if you want to write a story where it's possible, then it's possible in your story and nobody can definitively say it isn't realistic. $\endgroup$
    – kaya3
    Jun 21, 2021 at 12:00
  • $\begingroup$ I sense a fellow avid JRE listener. :) $\endgroup$
    – AnoE
    Jun 21, 2021 at 14:38
  • 1
    $\begingroup$ This sounds like a case of complexity addiction: Why do you need brain scans? If you have half a year of full-time high-quality video recordings and human-picked educated guesses of the targets state of mind as a training set, a good AI should be able to infer anger, arousal, and similar from gestures. Moreover, your video recordings would probably give sufficient material for blackmail anyway. $\endgroup$
    – Wrzlprmft
    Jun 21, 2021 at 14:58

7 Answers 7

14
$\begingroup$

We do not know the answer to this question. When it comes to the psyche (inner thoughts are part of it) current technological and theoretical levels are not enough to make predictions or even educated guesses.

Whether it is possible to extrapolate (what is essentially) 'mind-reading' from existing technologies depends not on technologies but on your philosophical position:

In its extreme form, this position states that if we know everything about the brain we will know everything about the mind.

You would have to adopt a position of strong reductionism if you want some technology that allows reading specific thoughts. You will have to assume that all thoughts can be fully decoded from neuron activity. It is a possibility but we are not anywhere near being able to prove it.

If you decide to go with reductionism, you will have to read on determinism and free will. Some extreme versions of reductionism logically lead to total determinism and the complete absence of free will. Philosophy of mind and to some extent epistemology are concerned with these topics and their relationship to each other.

  • holism or emergentism (here is a more detailed overview of emergentism, including discussions of the conscious mind [section 5.1])

This position can be summarised as 'the whole is greater than the sum of its parts'.

If you follow this view, the mind cannot be reduced to neuron activity and it will not be possible to build 'mind-reading' technologies based on advances in neurology and/or limited only to the study of the brain. The mind requires its own, mind specific instruments to be studied and understood. (Whether these instruments are possible or not we do not know, but nothing stops you from inventing them for your story.)


You can choose either of these two philosophical approaches or any combination of them. You can limit your technology to rough emotional states (for example, stress levels are associated with changes in hormonal status) or you can go with detailed mind-reading. The latter would require a much more advanced overall technological level than the former.

The only important thing is consistency. If your world is based on the philosophy of reductionism, everything should be reducible to the same extent. If you see the world in a more holistic manner, you should define what is reducible and what is not.

$\endgroup$
5
  • $\begingroup$ "current technological and theoretical levels are not enough to make predictions or even educated guesses" Well it depends on what is meant by "stranger". I know that the French intelligence (along with that of the CIA) can do it with nearly 100% accuracy if enough data can be gathered about the stranger beforehand without interacting directly with him (here data would be something like triples (what's seen ; what's heard ; thermal imaging) gathered over probably a long period). Now one might argue this person isn't a "stranger" anymore, but research shows extrapolation seems feasible. $\endgroup$
    – Evariste
    Jun 21, 2021 at 20:28
  • $\begingroup$ @Evariste Are you talking about reading thoughts? $\endgroup$
    – Otkin
    Jun 22, 2021 at 1:03
  • $\begingroup$ Yes. Such as reproducing what one sees in one's mind $\endgroup$
    – Evariste
    Jun 22, 2021 at 3:40
  • $\begingroup$ @Evariste Sources? Perhaps I am a bit behind the latest technologies. $\endgroup$
    – Otkin
    Jun 22, 2021 at 8:31
  • $\begingroup$ @Evariste: Yes, humans (and chimpanzees, and gorillas and dogs) have theory of mind; that is, if we live in close association with a person for some time we get to understand them and intuit what that person will do next in a given situation. We all can do it, and we all do it instinctively all the time; French spies are not exceptional in any way. But this doesn't mean that we can tell what they are thinking: die Gedanken sind frei. $\endgroup$
    – AlexP
    Jun 22, 2021 at 12:14
6
$\begingroup$

AI problems

On reading minds

Strange as it sounds, we don't understand how the brain works. We have decades of data, with increasingly better machines, but our theories are still full with assumptions. MRI is in many ways a method with a high brain resolution on a fast timescale. But what is measures is iron in the red blood cells (hence the huge super magnets). It detects when a part of the brain is requesting more blood. But requesting more blood happens after the initial start of the activity. Continued activity will just show that. An activity continues occuring. ElectroencEfaloGram (EEG) measures electro activity, but in a more general way. Positron Emission Topography (PET) is in a way much the same as MRI. Inject radioactive materials to track different kinds of brain activities, depending on what radioactive substance you add.

The problem here is that we extrapolate meaning out of secondary processes or bundle a ton of data. In addition, correlation is not causation. It is unlikely, but still possible that tertiary processes cause the brain to function and take decisions. A thought might cause a brain activity, or brain activity causes thought.

What we extrapolate from the data is that neurons fire in sequences in certain structures. It can be compared to morse code on a pattern of neurons. This is efficient, as the brain structure can be used for many cross purposes. Take the identification of faces. Many faces look alike, showing activity in certain structures. The slight changes in the neurons that light up and the morse code make us identify a different face. The morse code practically never is a single pulse, as neurons need to fire every so often when idle.

Problem here is that we skip over 'administration'. Between the neurons there is a synapse. This modifies information, changing how neurons fire. If you're incredibly hungry, you'll first notice the sandwich before you notice the snake.

Still, it seems possible to know from neurons firing alone to extrapolate what someone is roughly thinking.

On AI and muddied waters

The AI learns by getting cases. The more cases the better. This might be easy for some things, like the seeing of a face. We know the stimuli, we know the area where it should be processed. We can say a certain pattern+code is his brother. We see a ton of faces a day, making it easy to get a large portfolio of faces.

Now a thought. We can get words and meaning from the area of Broca and Wernicke. But thought and meaning is certainly possible without words. Figuring out how gears work, or imagining the engine are largely visual based and can be seen active in the visual cortex. But what about understanding of a difficult math equation? Or planning a trip tomorrow? You might get glimpses in the language centers, but they can be fully internally initiated without the AI knowing why it started and what it is.

From there it'll get harder. Spontaneous thought needs to be separated, but we can both cycle through traffic and plan a trip tomorrow. The AI, even a smart one, will start trying to fit it with planning traffic and not with planning a trip tomorrow. Adding wrongful cases has an impact that quickly gets bigger the more there are. This muddies the water for the AI, making even already learned things uncertain.

But it gets worse. The 'administration' is still forgotten. The AI is extrapolating data from incomplete data. We don't know what the synapses are doing, or what neurotransmitters are there. As said before, the state of an individual makes things respond differently in the brain. This can be slightly different structures and morse code. This is not what an AI would want. It needs to learn this per mood, which can swing quickly in some cases. In six months it is also difficult to get many of these moods. How often is someone sad, or crying? Some moods can coincide. Horny and sad produces different results than horny and happy. Or sad and happy, like a depressed person laughing at a joke?

The fact is that you can probably infer a lot with an AI knowing each neuron firing. But with so many unexpected thoughts, some more difficult to pinpoint concepts, changing neuron firing patterns and not measuring the full picture it is very difficult to extrapolate all thought.

$\endgroup$
1
  • $\begingroup$ I think it is a great overview of AI-related difficulties. I would only add that brain reactions to the same stimuli are not necessarily identical during every exposure, which adds more uncertainty. For example, if one repeats the same lie several times, their brain and body reactions start to resemble those of telling the truth. With just a bit of training truth and lie are no longer distinguishable. $\endgroup$
    – Otkin
    Jun 21, 2021 at 18:55
2
$\begingroup$

As @Otkin said in another answer, we do not know and we probably won't know for a long, long time.

From my personal estimation, it is highly unlikely that we will be able to read the minds of all people. As I, a non-neurospecialist, see it, brains are giant association engines. Our brains don't actually store memories and such as series of images, but what is stored are references to other experiences, thoughts and images in a giant web of associations, down to the first things we learn as children. That means that to know what a person is thinking, we need to know the associations in the brain, which developed for every person individually based on life experiences.

From that theory, it might be possible to calibrate a perfect neuron-scanning machine to an individual by giving some impulses and see where in the brain the associations lead. That would most likely be a lengthy process that needs to be done on every individual person you're scanning.

That might not be necessary though. If we only want to scan for emotions, it would be enough to monitor areas of the brain which are already understood to some small degree today. We can't monitor them from a distance, but we can see the brain areas responsible for anger, attraction and such. Not perfectly yet, but well enough that we can extrapolate that an emotion-scanner might be possible in the future. This might already be enough to give away secrets to the AI if you follow a person long enough.

$\endgroup$
0
$\begingroup$

There are two questions to consider here. The first is whether or not this is a technology that we think could actually exist in the future, and the second is whether or not readers will find it plausible enough to not break their suspension of disbelief.

The possibility of this technology existing in the future has been covered by other answers, and I fully agree with them that it's unknown.

As for the second question, I'd say this it is reasonably plausible to say you can understand a person's thoughts. Due to the complexity of the human brain it would not be plausible to start scanning someone and immediately know how to read their thoughts, but you've given a training time of six months. Consider this - suppose you had the (lesser) superpower to determine someone's actual emotional state and were able to devote a full six months to trying to completely understand someone. How well do you think you would know them by that point? Even without the brain scanning, devoting that much time to an individual would produce a very good model for how they will act in any given situation.

Another thing to consider is that you're not restricted to scanning the individual. A political opponent can be prodded and provoked - when you bring up certain topics, how do they react? If they have a strange internal reaction to the discussion of someone having an affair, you can start looking into their personal and family history. A bit of detective work could uncover that they are having an affair, or something like their parent having an affair that significantly affected their childhood.

You may not be able to know the exact words of a person's inner dialogue after six months, but that is not necessary to understand the individual. It would be a stretch to lift a password from their thoughts, but as they're typing it you could be able to recognize that the password is related to their dog. To maintain plausibility, you can avoid going into detail about exactly how much you know from the brain scan and how much you know through other means - knowing that a password is based on someone's dog and knowing the dog's name, it would be relatively easy to determine the password. Showing the audience (or telling them about) at least one scene of supplemental information gathering would make it plausible to pull whatever information you wanted to out of the brain scan.

$\endgroup$
0
$\begingroup$

It is hypothetically possible to simulate a mind using current technology, even without knowing how that mind works.

This requires the assumption that we know, in full totality, how to accurately simulate a neuron - as well as the assumption that there are no unknown external factors acting on those neurons - as well as the assumption that there are no unknown factors which arise when very large numbers of neurons act together, as in the brain. Basically there are a lot of assumptions required, but it is hypothetically possible. If you can simulate one unit of the system then you can simulate the system

Nick Bostrom covered this in his theories on superintelligence. It may be currently possible to simulate a single neuron firing with many hundreds of thousands of digital operations; if this is indeed possible and if the current models are accurate then it is theoretically possible to simulate 100 billion neurons interacting in virtual real-time. The practical ceiling to this capability is available computing capacity; in this theory even a computer from 1969 could simulate the interactions of an entire brain, it would just take millions years to compute one second of real-time behaviour and the physical memory infrastructure would probably consume most of the available land mass of Earth.

Crucially, this can be done without knowing why the specific interactions are taking place or why the brain structure is the way it is so a complete understanding of neurobiology is not required - all that is necessary is an accurate scan of the subject and the necessary computing capacity to run the simulation within a reasonable time frame.

With simulated sensory inputs the simulated subject can then be placed in various scenarios in order to extract information, for example it could be placed in a certain environment in order to extract passwords, placed in an interview setting in order to extract information, placed in various scenarios to identify emotional responses, weaknesses, etc. And the simulation could be reset again and again meaning that the simulation could be subjected to virtually unlimited information-extracting scenarios.

If we invoke super-fast computer magic, a hyper advanced AI could run millions of parallel simulations faster than physical real-time, performing a large number of data extraction operations on the subject in a matter of moments.

Current research

Current research in this area is still a very long way off from simulating an actual brain and there are still major deficiencies - for example simulated neural networks nearly all use highly simplified neuron models which may provide good insights for the purposes of a specific piece of research but are not actually representative of real-world neurons. Here's a reasonable list of major real-world limitations: https://www.nature.com/articles/d41586-019-02209-z

$\endgroup$
0
$\begingroup$

The current status is using microwaves to read emotions.

60% accuracy was achieved in 2020.

So, I am reasonably confident - that in 50 years, it might be possible to classify the thoughts in large bins , if not outright "read" such thoughts of a stranger.

$\endgroup$
2
  • $\begingroup$ That's not reading thoughts... (the linked article reads just breath patterns and such, which is as far from reading actual thoughts as you can get and still know about emotions). And 60% accuracy is pretty marginal above the 50% you'd get by throwing dice. $\endgroup$
    – toolforger
    Jun 23, 2021 at 16:40
  • $\begingroup$ indeed, i should have specified that it is only classifying emotions. I did mention classifying "thoughts" in bins tho. I will edit. $\endgroup$
    – Sean
    Jun 23, 2021 at 18:06
0
$\begingroup$

Computer neural net mimicing AI are grown

A computer AI technique today actually models neurons. It simulates a neural net and teaches the neural net how to do certain things by training. The thing is, each neural net is trained individually and thus, each resulting AI system is unique. Two neural net AIs trained on exactly the same input set will not be the same.

If you were to write an analysis program that could read the internal state of one of these neural nets, it would probably be idiosyncratic to that net and not generically applicable to any neural network solving the same problem.

People are likely similar

Otkin's answer is correct: we do not know the specifics of how the mind and/or soul (if such a thing exists) work. The sum total of our knowledge is mostly maddeningly vague generalities. We don't have anything like a system we can hook up to read your thoughts, because we don't know how your thoughts are formed.

We do know people's minds have a lot more pre-baked structure than software AIs (which are a crude imitation). People come out of the (ah) oven with a lot of capabilities pre-installed. These essentially have to be genetically pre-baked into gestation. We have no idea how, but it has to have something to do with the more complex structure of the brain (compared to the simple models AI projects use). People have neocortices and brain stems and these structures always look similar enough that they must have similar neurocircuit base designs.

This must be so because if it weren't, we'd already be able to develop software that could perform at the level of a human. And we can't.

But neural nets still adapt and grow and mold to match their inputs. So each person's 'thought engine' is likely going to be subtly different. Even though they are built off of a pre-baked template, they still develop and grow over time. Any mind-reading device will likely need a significant amount of attunement time to get accustomed to the specific individual it's trying to read**.

So even if the technology does eventually prove workable, it likely won't be like flipping a switch. If it works at all you'll likely need to probe the subject for a significant length of time, anywhere from a couple of hours to several weeks.

** Caveat: Far enough in the future, 'standard models' may arise based on prior configuration databases of millions of known minds, but there will still be outliers needing training. Especially people with brain damage or illness, or significantly above or below average ability.

$\endgroup$
1
  • $\begingroup$ The simulated neurons are vast simplifications of real neurons. (Real neurons can rewrite, simulated cannot - not because it's technically infeasible, but we don't know how to make a useful neural net out of them.) Also, simulated neural networks have a few hundred, maybe a few thousand neurons, a human brain has 100 billion(!) neurons, so you can expect emergent phenomena not present in simulations. $\endgroup$
    – toolforger
    Jun 23, 2021 at 16:38

You must log in to answer this question.

Not the answer you're looking for? Browse other questions tagged .