How is it that the brain governs every thought? Every action? Every facet of our identity? How can it be that the entirety of what we perceive emerges from the soft, gelatinous organ nestled within our skulls?

The free will debate is a resonant one, echoing through millennia. Indeed, we still have yet to reach a decisive conclusion. Below, I have compiled a small collection of my favorite arguments.

Are we really in control?

No.

Neural Evidence.

We must begin, of course, with the Classic Libet Experiment.

Participants faced a clock designed to measure time with high precision. Subjects moved at freely chosen times, reporting either “time of willing” or “time of the sense that the movement occurred”. Researchers simultaneously measured participants’ EEG activity (a means of recording the brain’s electrical activity via electrodes attached to the scalp), focusing specifically on the Readiness Potential — a gradual electrical buildup, represented within EEG data, that reflects neural activity preceding a self-initiated movement.

Astonishingly, Libet’s team discovered, across all subjects, a Readiness Potential that took place approximately 700 milliseconds prior to each participant’s conscious decision to initiate movement. In other words, the brain began preparing for self-initiated movement before the individual even became conscious of their voluntary decision to act.

In one extension of Libet’s experiment, researchers studied patients with implanted depth electrodes. As opposed to EEGs, depth electrodes (thin, flexible probes implanted directly into brain tissue, recording deep electrical activity) yield unparalleled temporal resolution. Within the study, a progressive recruitment of neurons associated with voluntary movement, decision-making, and cognitive control was identified, that altered their firing pattern across an interval surpassing 1,000 milliseconds prior to subjects’ awareness of the urge to act. Impressively, based solely on data procured from those neurons, the researchers were able to predict, with accuracy exceeding 80%, a participant’s impending decision to move 700 milliseconds prior to the subject’s awareness.

Of note, much of our brain’s activity is unconscious. We navigate routine, menial tasks predominantly on autopilot. Traversing a familiar route. Completing household chores. Certainly, complex decisions may bring particular brain processes into consciousness. But who’s to say consciousness acts upon the choice?

Perhaps our experience of autonomy is merely an awareness of decisions, out of our control, that have already been formed by our biology.

Causal Determinism.

Within the free will debate, the loudest voices opposing autonomy’s existence emerge from the causal determinism camp.

“Turtles all the way down,” Robert Sapolsky wrote. He borrows the idea from a well-known anecdote in which a woman declares the earth must rest on a turtle’s shell. When asked what that turtle then rests on, she responds, “it’s turtles all the way down.” In Determined: A Science of Life Without Free Will, the acclaimed Stanford neuroscientist delivered a powerful assertion: “we are nothing more or less than the cumulative biological and environmental luck, over which we had no control, that brought us to any moment.”

Each decision is wholly influenced by present circumstances and neural activity. The aforementioned are influenced entirely by (occurring hours before said decision) sleep, hormone levels, recent experiences, and so forth. These are influenced completely by (occurring years before said decision) childhood upbringing, social environment, education, and so forth. Terminating the incessant chain, the aforementioned are thoroughly influenced by genetic inheritance, culture, and evolutionary history. As Sapolsky expressed, there never once existed an opening for free will to emerge.

Childhood stress, for one, produces increased glucocorticoids. This yields a less developed frontal cortex, manifesting as decreased impulse control. Hormonal fluctuations (a product, to some extent, of environmental factors) can influence aggression, social behavior, and risk-taking. Physiological states bias judgements. In one striking study, researchers examined 1,112 parole decisions made by judges. Immediately after food breaks, favorable rulings were more common; this substantially declined as each decision session progressed.

“In order to prove there is free will,” Sapolsky asserts, “you have to show that some behavior just happened out of thin air.”

Are we really in control?

Yes.

Countering Neural Evidence.

A vast number of experiments probing free will, including those described within this article, examine proximal intentions (an intention to perform an action in the immediate future), as opposed to decisions concerning long-term plans or goals (these are typically more meaningful or deliberate), also known as distal intentions.

In one study, participants selected which of two nonprofit organizations should receive a $1,000 donation (here, distal intentions are in play). Those in the control group were informed that each nonprofit organization would receive $500 regardless of their choice (distal intentions are not in play). The control group exhibited Readiness Potential, whereas the meaningful decision group did not.

Perhaps elementary decisions are truly devoid of free will. But maybe those far greater in complexity (these are the ones that shape our lives) emerge, indeed, from autonomy.

Countering Causal Determinism.

Quantum physics asserts the impossibility of definitively ascertaining the location of particles due to their wave-like nature. Scientists, therefore, are fundamentally incapable of making perfect predictions about any event in nature. This presents the conceptual foundation by which choice, creativity, variability, and hence autonomy can arise. “Random variations in behavior are revealed in even the most highly controlled laboratory conditions,” Jackson Marr, a Professor Emeritus of Psychology at the Georgia Institute of Technology, commented in an article discussing determinism.

Numerous evolutionary scientists acknowledge, in tandem, the arbitrary nature of evolution. Which variations are selected and retained is, to some degree, unpredictable. Mutations (an essential driving force of evolution) occur at random, arising from DNA replication errors, chemical damage, or radiation. According to Stephen Gould, a renowned evolutionary biologist, “if we were to replay the tape of life’s evolution on earth, we would be unlikely to arrive at a world like the one we know.”

The foundation of causal determinism rests on the argument that “every event, action, and state of affairs in the universe is completely necessitated by preceding events and the laws of nature”. Yet the aforementioned examples provide evidence for events that seemingly cannot be explained by the episodes preceding it.

One Final Take.

Below is a breathtaking passage penned by Bernardo Kastrup, journalist for Scientific American.

“Why do we think that metaphysical materialism (the notion that our choices are determined by neurophysiological activity in our own brains) contradicts free will? Because, try as we might, we don’t experientially identify with neurophysiology; not even our own. As far as our conscious life is concerned, the neurophysiological activity in our brain is merely an abstraction. All we are directly and concretely acquainted with are our fears, desires, inclinations, etc., as experienced — that is, our felt volitional states. So, we identify with these, not with networks of firing neurons inside our skull.”

The brain may begin preparing for self-initiated movement before the individual becomes conscious of a voluntary decision to act. But are we not simply our brains? A separation seems to exist between our perceived identity and our inherent biology. “We don’t experientially identify with neurophysiology; not even our own.” This line of reasoning renders all arguments invoking neural evidence to reject autonomy’s existence as fundamentally worthless.

Here is another profound passage penned by Ted Chiang, a renowned science fiction writer.

“There have always been arguments showing that free will is an illusion, some based on hard physics, others based on pure logic. Most people agree these arguments are irrefutable, but no one ever really accepts the conclusion. The experience of having free will is too powerful for an argument to overrule.”

Conclusions.

A conglomeration of voices exist, each picking sides within the gargantuan, sprawling debate of the existence of free will. This article offers only a small glimpse into its depths.

Are we really in control?

We lack, evidently, a definitive answer. Ultimately, the choice is yours.

Sources.

Fried, Itzhak, et al. “Internally Generated Preactivation of Single Neurons in Human Medial Frontal Cortex Predicts Volition.” Neuron, vol. 69, no. 3, Elsevier BV, Feb. 2011, pp. 548–62, https://doi.org/10.1016/j.neuron.2010.11.045. Accessed 31 Oct. 2026.

Hallett, M. (2016). Physiology of free will. Annals of Neurology, 80(1), 5–12. https://doi.org/10.1002/ana.24657

Hoefer, Carl. “Causal Determinism (Stanford Encyclopedia of Philosophy).” Stanford.Edu, 23 Jan. 2003, https://plato.stanford.edu/entries/determinism-causal/. Accessed 4 Nov. 2026.

Jamali, Mohammad, et al. How the Human Mind Affects Its Related Brain? A Mechanism for This Influence, Using an Extended Bohmian QM in Avicenna’s Monotheistic Perspective. MDPI AG, May 2019, https://doi.org/10.20944/preprints201905.0223.v1. Accessed 10 Nov. 2026.

Kastrup, Bernardo. “Yes, Free Will Exists.” Scientific American, 5 Feb. 2020, https://www.scientificamerican.com/blog/observations/yes-free-will-exists/. Accessed 31 Oct. 2026.

Mahto, Neil. “Science Supports the Existence of Free Will.” The Johns Hopkins News-Letter, 24 Apr. 2024, https://www.jhunewsletter.com/article/2024/04/science-supports-the-existence-of-free-will. Accessed 31 Oct. 2026.

Rakos, Richard F., et al. “The Free Will: Determinism Debate in Contemporary Science and Society (or: Why Are You Reading This Article?).” Perspectives on Behavior Science, vol. 48, no. 2, Springer Science and Business Media LLC, Apr. 2025, pp. 475–97, https://doi.org/10.1007/s40614-025-00447-3.

Sapolsky, Robert M. Determined. Penguin, 2023.

This article was written by Rou Zhao , an undergraduate student majoring in Neuroscience and Data Science at UC Berkeley. This article was edited by Will Sperling and Bridgette Lai.

Preface: I want to begin by saying I have no affiliation with Meta, I am writing this in my personal capacity as a neuroscience and data science student, and I have nothing but respect for the researchers who built this model. The work is genuinely impressive. What follows is written by a fan of the science who is worried about what happens once the science leaves the lab.

I am getting worried about the world we are about to enter. To not overuse the cliché, I think we are walking into the Brave New World Aldous Huxley laid out in 1932 in the book that gave the phrase its meaning.

I know, I know. Another punchy intro for a Medium article, another hook, another reach for the apocalypse. But I am not trying to be hyperbolic when I say this: we might actually be screwed.

I don’t think enough people, especially the lay public, have a clear sense of what is currently on the table technologically. For me, as someone studying neuroscience and data science at UC Berkeley, the recent pace of the field has been everything but scary. If anything, it has been a dream. I fell into this field by accident in my freshman year of high school after watching the “Secrets of the Mind” documentary by V.S. Ramachandran, and at the time neuroscience still felt niche. Half a decade later, neurotech is one of the loudest stories in both research and pop culture. Anyone can name a player now: Neuralink, OpenBCI, Openwater, and the parade of consumer EEG wearables that have moved from clunky headsets to beanies to earbuds to glasses, all claiming to interface with the brain to optimize focus, sleep, recovery, studying, meditation. Name a thing you want optimized and there is, somewhere, a stealth startup already trying to sell it to you.

I read a lot. I read scientific journal articles compulsively, and even though I would never call myself an expert in any one corner of this field, I know enough that when something I read stops me in my tracks, it is generally not a buzzy headline. It is the real thing.

So what stopped me in my tracks?

TRIBE V2.

For those of you unaware: TRIBE V2 is a foundation model Meta released that predicts fMRI brain responses to naturalistic stimuli (video, audio, and text) in real time, mapped onto the cortical surface. The “TRI” refers to the model’s ability to handle text, video, and audio as inputs simultaneously, with each modality producing its own predicted map of the brain regions that would activate in response. You feed it a movie clip, and it gives you back a second-by-second prediction of what your visual cortex, language network, and auditory regions are doing.

The training pipeline: a group of human subjects watched many hours of naturalistic media inside an fMRI machine, which records blood-flow correlates of brain activity across the cortex and deeper structures (the amygdala, hippocampus, brainstem, and so on), and also captures whether large-scale networks like the default mode network are switched on. Meta took the paired data, what the subjects saw and heard plus what their brains did in response, and trained a transformer-based model on top of it. The result is something that, given a brand-new piece of media it has never seen, can predict with surprising accuracy where activity will show up across the cortex. The model won the 2025 Algonauts Challenge and was released publicly shortly afterward.

(And please excuse my subsequent personification of TRIBE as if it were a person with opinions. It is more readable to write “TRIBE found X” than “the output of this transformer when fed through this preprocessing pipeline yielded values that map to fMRI activity in regions Y, Z.” In lieu of writing for the approval of Reviewer #2, I am trying to make this blog legible.)

–

Before I get into any alarmist commentary, I want to give the model genuine credit, because the thing that impressed me most, from a data-science perspective, was the level of anatomical specificity it recovered.

The brain is an absurdly specialized organ. There are few things in the universe that match the level of structural specificity you find in a few millimeters of cortex or subcortex. The fact that we have, as a field, mapped as much of this jiggly, fuzzy, water-logged thing as we have is no small miracle, and it owes everything to the painstaking work of medical students grinding through Anki, radiologists squinting at fuzzy MRI lesion scans, and neurosurgeons whose lifetimes of cadaver work and intraoperative mapping have slowly given us the atlas.

I know that networks matter. I know connectivity and plasticity and whole-brain fluctuations and astrocytes matter, and that the brain is not a Lego set of independent regions. But despite all of that, we have localized so many small substrates with such precision that we can point to, say, Broca’s area in the left inferior frontal gyrus and Wernicke’s area in the left posterior superior temporal gyrus and say, with high confidence, this one handles speech production, this one handles speech comprehension. And those claims are testable. Damage one, lose one function. Damage the other, lose the other. You can watch it happen live in this classic VSauce video where a non-invasive transcranial stimulator is applied to Broca’s area and the subject (Mike) begins to stutter mid-sentence.

[see minute 10:32!]

The point is: localization is real, hard-won, and represents tens of thousands of human-lifetimes of work. So when a model arrives that can do something like the same thing, just by ingesting video and audio, it is not a small feat.

When you give TRIBE V2 a stimulus and ask it to predict differential activity, it cleanly recovers regions that visual neuroscience has spent decades isolating:

• the fusiform face area (FFA) for faces
• the parahippocampal place area (PPA) for places
• the extrastriate body area (EBA) for bodies
• the visual word-form area (VWFA) for written characters

TRIBE V2’s predicted activations cleanly recover the canonical category-selective regions of visual cortex (FFA, PPA, EBA, VWFA), each emerging from a subtraction against a phase-scrambled control of the same image.

What is being shown in those panels is technically a subtraction design. The original stimulus (face, place, body, text) is contrasted against a phase-scrambled version of itself. Phase scrambling, for the uninitiated, means: take the image, run a Fourier transform, randomize the phase information, and reconstruct. The output preserves all the low-level statistics of the original (spatial frequency, contrast, luminance) but destroys all the high-level semantic content. The result is the grey, fuzzy nothing-image you see as the baseline (background), which is mathematically identical to your visual cortex at the level of edge detection but no longer contains any meaningful structure.

By subtracting the brain response to the meaningless image from the response to the meaningful one, you isolate the activity driven by categorical content (faces, words, etc.) rather than raw pixel statistics. So when you see those crisp FFA and PPA localizations, you can trust that the model is picking up category-level processing and not just brightness or edge density.

The same subtraction logic applied to audio: speech-selective regions in the superior temporal sulcus emerge cleanly when speech responses are contrasted against matched natural sounds.

The same logic carries over for the speech versus natural-sounds panel.

When the authors ran independent component analysis on TRIBE V2’s predictions, the top five components recovered five canonical functional networks: primary auditory cortex, language network, motion-detection area, default mode network, and visual system (Damoiseaux et al., 2006).

The other thing the model recovered, in my opinion the most striking single result in the paper, is what fell out of an independent component analysis (ICA) on TRIBE V2’s predictions. The top five components closely matched five well-studied functional networks: the primary auditory cortex, the language network, the motion detection area, the default mode network, and the visual system (Damoiseaux et al., 2006). Nobody told the model to look for these, but still the model recovered them from naturalistic stimulus alone.

The default mode network is the network most active when you are not focused on a task: daydreaming, mind-wandering, ruminating about your life choices at 2 a.m. It also tends to be hyperactive in depression and in introspective personalities (Sheline et al., 2009). The fact that this network falls out of TRIBE V2’s predictions, unsupervised, in my opinion, means the model is capturing something real about the brain.

–

I want to flag one finding the model produced that I tried to verify carefully, because I think it is a useful example of how to read these papers responsibly.

According to TRIBE V2, emotional versus physical pain was distinguished by activity in the temporo-parietal junction (TPJ) and the middle temporal gyrus (MTG).

The TPJ part is grounded in real data: Woo et al. (2014), using multivariate pattern analysis in the Wager lab, found that social rejection compared to physical pain was predicted by increased activity in the right TPJ, dorsomedial prefrontal cortex, and precuneus. But the correct interpretation is not that the TPJ is a general emotional-pain detector. It is that the TPJ tracks the mentalizing and social-cognition component that tags social pain as specifically social. Generalizing this to non-social emotional pain like grief or fear is not a clean move.

The MTG part is weaker. The MTG does show up in pain contexts. The Wager lab’s Neurologic Pain Signature includes MTG voxels as predictors of pain intensity (Wager et al., 2013). There is also work showing larger after-pain MTG activity in chronic pain patients correlating with rumination scores. But there is less strong evidence that MTG is a clean dissociator between emotional and physical pain.

The more robust dissociation in the actual literature is different: physical pain preferentially engages the sensory-discriminative system (S1, S2, posterior insula), while social and emotional pain preferentially engage affective regions (dorsal ACC, anterior insula) and dorsomedial PFC (Eisenberger, 2012).

I bring this up not to dunk on the model. The model is doing what models do: surfacing real patterns, occasionally over-confidently. The reminder is that when an AI tool starts outputting neuroscience claims that look clean, the actual story is usually messier (Murphs Law?), and you still need humans who have read the pain literature to interpret what the model is showing you. The tool is powerful, but it is not an oracle.

–

So why does any of this matter outside a paper-discussion seminar?

Let’s follow this TRIBE line to its conclusion.

Yes, you can use TRIBE V2 to localize substrates for things like whether a sentence reads as painful or neutral, threatening or benign, arousing or boring. That is, technically, a standalone research finding. But let’s follow that line all the way to where the fish bites the hook, and then let’s follow it back to the man holding the rod.

Who is the man holding the rod?

Meta.

Who is the fish?

Us.

I know that sounds dramatic, but someone else has to also be reading this the way I am. The model is open source. That means Meta, and anyone with a passable command of GitHub, Hugging Face, and the patience to debug a few dependency issues (seriously Meta, this is one of your latest releases and the dependency misalignments took hours to figure out, although I am sure your engineers have larger problems to deal with), can now run experiments asking: which version of this video clip produces the strongest predicted amygdala response? Which caption variant maximally activates the default mode network and locks the viewer into rumination? Which thumbnail predicts the largest visual cortex bias for our brand?

This is an A/B test with a brain readout instead of a click-through rate. Papers on arXiv are already exploring this direction, using brain-prediction models to optimize stimuli for emotional response, attentional capture, or persuasive effect. The accuracy is going to keep climbing. And once you have an in-house team that can iterate on engagement-maximizing content with a neural-prediction model in the loop, the boost in efficiency over the current setup, where you guess and check via clicks, is non-trivial.

–

A small side tangent.

A few days ago I posted a story on Instagram. The video was an eerie, fluorescent-lit shot of the Boston T’s Silver Line at night, the kind of frame that looks like a screen capture from a horror game. When I tapped to add music, every song Instagram suggested was already from the Backrooms soundtrack.

Mind you, this was a close friends story. So somewhere, in the seconds between me hitting upload and the music drawer opening, an on-device ML classifier ran on my video, tagged it as something like creepy / liminal / dark, and fed that tag into both the music recommender and, presumably, the ad-targeting pipeline. The story had not even posted yet.

The point is not that Instagram does ML classification on user content. We knew that, even if we don’t think about it often. The point is how fast and how granular it has gotten. It is not just metadata anymore. It is interpreting the content of the frame in real time. And this is what is shipping today, on a close friends story, the supposedly private one.

Now imagine the same kind of inference pointed not at the content I make but at how my brain is predicted to respond to the content others make.

That is the world TRIBE V2 makes one degree closer.

–

This is not, by itself, a new field. Neuroeconomics, neuromarketing, and attention research have been around for two decades. Companies have been EEG-testing ads since the early 2010s. What is new is that the resolution available has jumped, the model is publicly downloadable, and the gap between research result and production pipeline at a company optimizing for engagement is measured in months, not decades.

Every piece of the surveillance stack is converging at once:

• the cookies we click through and the ones we don’t see
• the location services we forget to disable
• the things we say within earshot of Alexa
• the websites where we typed our phone number once in 2017 and forgot
• the captions we write on close friends stories
• the on-device ML quietly classifying our images before we hit post
• the Flock cameras and license-plate readers accumulating on suburban streets
• and now a public model that predicts what content will land hardest, neurologically, on the average person

I do not want to live in tinfoil-hat paranoia. I also do not want to pretend this is fine. We are being surveilled, profiled, and modeled at a level that is now approaching the molecular. Down to the neuron.

Huxley and Orwell are rolling in their graves.

Model: https://ai.meta.com/blog/tribe-v2-brain-predictive-foundation-model/

Demo: https://aidemos.atmeta.com/tribev2

This article was written by Bhavna Malladi, an undergraduate student at UC Berkeley studying Neuroscience and Data Science.

Imagine having a miniature version of your brain. That is exactly what scientists have been trying to do in recent research. Researchers create new tools to answer old questions, and in doing so, they open up the scientific world to further exploration. One of these new tools created to study neuroscience is helping scientists around the world understand some of the most prevalent brain diseases and disorders.

Researchers have been using cell cultures to understand human biology for years, but recently scientists have been able to transform these cell cultures into models of human organs, creating mini versions of these organs in which they can experiment on. By using stem cells in a 3-Dimensional environment instead of the popular 2D dishes commonly used in labs, scientists have been able to create differentiated mini clusters of cells that later on form a “mini organ” where diseases and disorders can be modeled. Nowadays, organoids are being used to model organs like the intestines and the retina. One type of organoid that is being used more than any other is the brain organoid, known for its ability to model neurological diseases that cannot be seen in animal models.

The brain is one of the most complex organs in the human body. Made up of neurons and glial cells and responsible for controlling the functions of our entire body, modeling the brain was always going to be a challenge for researchers. 2D cultures, which have been commonly used in labs, are made of flat monolayers of cells. While this method allows for easy imaging and understanding of the biology of single cells, it is not complex enough to model the human body. Because of its structure, it does not allow for much cell-to-cell contact or interaction, which are vital in the human body. Organoids are an advanced form of 3D cultures, which allow for more life-like interactions between cells, mimicking human biology. In 2007, researchers started working with 3D cultures of the brain and were able to create a culture that modeled the forebrain, the largest part of the brain. Years later, in 2013, researchers were able to encompass different brain areas within one organoid, effectively creating a mini-brain.

Brain organoids are replicas of the human embryonic cortex, the brain cortex developed in embryos, which allows researchers to understand brain development in the embryo while also being a great model for disorders that only occur later in life. Researchers create organoids with different stem cells and manipulate them to study their desired disorder. Organoids made of stem cells from microcephalic patients, those who were born with smaller brains due to problems in brain development, have disruptions in cell division, causing smaller organoids. Similarly, researchers can model macrocephaly by deleting tumor suppressor genes in the stem cells before they differentiate. Outside of genetic conditions, researchers can also model how external factors can impact the brain. One example of this is the exposure of organoids to the Zika virus, which causes an increase in cell death, drastically reducing the organoids’ size. Although organoids do not form cortical folds yet, researchers are looking for ways to adapt these cell cultures to be able to model disorders that cause less cortical folding in humans.

Another aspect of organoid research that has been steadily growing is the modeling of neurodegenerative diseases. Because organoids are most similar to the human embryonic cortex, they are not yet the most useful tools to understand diseases like Alzheimer’s and Parkinson’s that develop late in life. Researchers, however, have not given up on using organoids to model these diseases. One study, for example, chose to surround the organoids in a serum that mimics the breakdown of the blood-brain barrier that happens in Alzheimer’s patients. Labs in UC Berkeley, like the Bateup and Kaufer labs, are using similar techniques to model neurodegenerative diseases in an attempt to understand how we can treat them.

We are in an era of technological and scientific breakthroughs, where it seems every other month a novel way of researching is developed. Imaging techniques are becoming clearer and faster every day, allowing us to view high-resolution images of the tiniest brain structures, and computer models seem to be taking over the study of cognition. This obviously begs the question: Why use brain organoids at all? They require immense precision and time to be created and even longer to be experimented on, without taking into account how much money they cost to produce. Why would researchers use up so many resources to model the brain like this?

One of the main ways researchers study neurological problems is through animal models, mainly mice. The problem with this, however, is that the human brain is much more complex than the mouse brain. Not only does this limit the research that can be done in terms of cognitive abilities, but it also prevents in-depth research in neurodegenerative diseases. Although genetically modified mice are commonly used in labs to study diseases like Alzheimer’s and Parkinson’s, they lack the complexity of the human brain and do not act as perfect models for age-related degeneration. Because of this lack of similarity to the human brain, drug testing in mice for degenerative diseases often fails to translate to human patients. Organoids, on the other hand, are made from human cells and are practically a replica of the embryonic brain. This allows scientists to see how drugs and environmental exposures would impact the human brain directly without having to translate the findings from another species.

Another common research method for neuroscientists is using humans. Brain imaging allows scientists to observe what parts of the brain become active when a patient is doing a certain task or how the brain deteriorates after a patient is diagnosed with a neurodegenerative disease. The main issues with human research are the ethical considerations and the limitations of the observable data, neither of which is a concern when using organoids. When studying a drug that might cure a brain disorder, a control group that does not take the drug is necessary to compare the results of the group that does take the drug. Without a control group, scientists would not know if the drug is actually curing the patients or if it is an outside factor. This obviously causes a problem for the control group, who are not allowed to take the drug even if it might benefit their health. Brain organoids are not living beings with life spans, so there are no ethical concerns when it comes to having a control group.

Although it is impossible to say if in the future, the technology will become so advanced that these mini-brains gain consciousness; for the time being there are no ethical concerns in the use of these cells. Organoids also allow researchers to understand how a disorder progresses on a cellular level, something which is not yet possible in living human subjects. Brain disorder research on human brains is limited to when the patients are diagnosed, which leaves many questions about the initial stages of the disorder and how it might grow. With organoids, however, a researcher can suppress or increase transcription of genes that will cause these disorders. Before there is any manifestation of the disorder on the organoid, researchers already know the disorder is present and are able to map its growth and impact from the very beginning.

While organoids may seem to be a miracle tool used to study the brain, they still have many limitations that prevent researchers from retiring other methodologies. One of the main shortcomings of organoids is their lack of vascularization. The brain needs a constant oxygen supply, which is provided through its vascularization. The brain’s vasculature also provides the needed nutrients for brain function and development, along with being the source of many disorders in the brain, like strokes, aneurysms, and dementia. Organoids’ inadequate vascularization prevents the cell culture from growing past a couple of millimeters; it also limits the organoids’ long-term functions. The lack of nutrients and metabolic processes through blood vessels leads to shorter survival times for organoids, which is a great limitation for researchers as they prevent them from studying long-term effects on the brain, as well as forces them to keep continuously producing more organoids, something that is not only time-consuming but also expensive.

Furthermore, organoids are still only replicating embryonic forms of human organs. They are unable to fully differentiate and mature into grown human organs, which limits the studies that can be conducted and the results that can be seen. Researchers have been finding ways around this by gene editing the organoids or simulating environments that cause the cells to act as more mature ones; however, this research is still novel and does not perfectly replicate the mature human brain. Another issue with using organoids as a method of research is their variability. Because organoids are made from differentiated stem cells, it is incredibly hard to control every single variable that causes the maturation of the cell culture, which causes a lot of variation in organoid batches. Not only is variability in size and structure observed, but also in functional aspects. This provides a challenge for researchers when testing drugs or other treatments in organoids. The differences in each batch of organoids can cause varying results even if all other variables are controlled in the experiment. The lack of precision in results poses a challenge to researchers when they have to generalize the treatments to all human brains. This also causes a problem with replicability. For a drug or treatment to be legalized, the initial study using the drug or treatment has to be replicated with similar results. The great variability in organoids makes it harder for these replicated studies to come up with results that match the original study. Studies are being conducted every day to minimize these limitations of organoids and improve their function in hopes of making organoids the symbol of the future of biological research.

Scientists are constantly looking for ways to improve organoid research and its applications. One of the more optimistic and revolutionary research areas of organoids is that of personalized medicine. Labs around the world are trying to adapt organoids to be more mature brain models by adding vascularization in the hope of one day being able to create small models of individual human brains. If this becomes possible, then drugs can be tested on the mini-brain before being consumed by the patients themselves. This would allow researchers to observe individualized side effects of drugs and treatments on each patient before testing on the person themselves. The biggest hurdle that scientists are trying to overcome when it comes to organoid research is the lack of maturation in the cell culture, and while some labs are trying to use vascularization to grow the organoids, other labs are using different methods. Researchers at UC San Diego have come up with a way of stimulating neurons in brain organoids and maturing the cell culture by using Graphene-Mediated Optical Stimulation, a technique that uses graphene’s ability to transform light into electrical energy to activate neurons in the organoid. When graphene is implanted in the organoid and light is shone on it, more neurons activate and mature. This could open many doors for future studies of neurodegenerative diseases using organoids, as the cells would more closely resemble those of the adult human brain.

With the ever-growing world of artificial intelligence and all its applications in our lives, it is no surprise that researchers are finding ways to incorporate the technology into organoids. Cognitive scientists are constantly trying to improve computer algorithms to mimic brain cognition and function, and a new area of research called Organoid Intelligence is pioneering a method of integrating those computational systems with biology. Although the studies are still in their early stages, labs have been trying to integrate organoids with AI systems to improve their processing capabilities and make their algorithms work more similarly to human cognition. Researchers at Johns Hopkins University are attempting to power computers and cognitive models using organoids in an attempt to compete with the ever-growing Artificial Intelligence industry.

Brain organoid research is opening many doors for labs around the world to fully understand the human brain and its capabilities. This model of the human brain will allow researchers to manipulate human brain cells in ways that have never been done before and that might lead to treatments and even cures for the most deadly brain disorders. Although the technology is still new and has many limitations, scientists are working on ways to improve and adapt organoids to fit their every need. From personalized medicine to the future of artificial intelligence, there is no limit to what organoid research might lead to. In the next few years, we will undoubtedly see an uptick in research with organoids and will be able to answer questions about the human brain that have been unanswered for decades.

Bibliography

Chen, Xianwei, et al. “Modeling Sporadic Alzheimer’s Disease in Human Brain Organoids under Serum Exposure.” Advanced Science, vol. 8, no. 18, Aug. 2021, p. 2101462. PubMed Central, https://doi.org/10.1002/advs.202101462.

Contact Us via LiveChat! https://secure.livechatinc.com/customer/action/open_chat?license_id=12180372&group=1&embedded=1&widget_version=3&unique_groups=0&organization_id=8ded8051-43c5-40a8-b076-d6ebb33d7280&use_parent_storage=1&x-region=us-south1. Accessed 13 Mar. 2026.

Corrò, Claudia, et al. “A Brief History of Organoids.” American Journal of Physiology — Cell Physiology, vol. 319, no. 1, July 2020, pp. C151–65. PubMed Central, https://doi.org/10.1152/ajpcell.00120.2020.

Dawson, Ted M., et al. “Animal Models of Neurodegenerative Diseases.” Nature Neuroscience, vol. 21, no. 10, Oct. 2018, pp. 1370–79. PubMed Central, https://doi.org/10.1038/s41593-018-0236-8.

Deb, Bipan K., and Helen S. Bateup. “Modeling Somatic Mutations Associated With Neurodevelopmental Disorders in Human Brain Organoids.” Frontiers in Molecular Neuroscience, vol. 14, Jan. 2022, p. 787243. PubMed Central, https://doi.org/10.3389/fnmol.2021.787243.

Fan, Xulong, et al. “Strategies to Overcome the Limitations of Current Organoid Technology — Engineered Organoids.” Journal of Tissue Engineering, vol. 16, Apr. 2025, p. 20417314251319475. PubMed Central, https://doi.org/10.1177/20417314251319475.

Forget AI; Organoid Intelligence May Soon Power Our Computers. https://www.forbes.com/sites/williamhaseltine/2023/04/28/forget-ai-organoid-intelligence-may-soon-power-our-computers/. Accessed 13 Mar. 2026.

Kim, Jihoon, et al. “Human Organoids: Model Systems for Human Biology and Medicine.” Nature Reviews Molecular Cell Biology, vol. 21, no. 10, Oct. 2020, pp. 571–84. www.nature.com, https://doi.org/10.1038/s41580-020-0259-3.

Konan, Landry M., et al. “Neuroanatomy, Cerebral Blood Supply.” StatPearls, StatPearls Publishing, 2026. PubMed, http://www.ncbi.nlm.nih.gov/books/NBK532297/.

Marinho, Luciana Simões Rafagnin, et al. “The Impact of Antidepressants on Human Neurodevelopment: Brain Organoids as Experimental Tools.” Seminars in Cell & Developmental Biology, Special Issue: Deconstructing Organs: Decellularized organs, organoids, stem cells niche and engineering new organs, vol. 144, July 2023, pp. 67–76. ScienceDirect, https://doi.org/10.1016/j.semcdb.2022.09.007.

Qian, Xuyu, et al. “Brain Organoids: Advances, Applications and Challenges.” Development (Cambridge, England), vol. 146, no. 8, Apr. 2019, p. dev166074. PubMed Central, https://doi.org/10.1242/dev.166074.

This article was written by Juliana Amaral, an undergraduate student majoring in Neuroscience at UC Berkeley. This article was edited by Meenakshi Yarlagadda.

Imagine a childhood characterized by the blaring sounds of drones, war planes, and explosions. Imagine the warm glow of a cartoon playing on the Disney Channel in your living room replaced by flashes of white phosphorus strikes, illuminating the sky above your village and community. Millions of children around the globe do not have to imagine this, nor is this scenario one of their nightmares; it is their daily life. While we, in the United States, talk about war in terms of politics, economic instability, or territorial disputes, the reality of the lived experience of war extends far beyond the scope of our understanding. For children growing up in conflict zones, the constant presence of danger and trauma, rooted in uncertainty of survival, drastically alters how their brains develop.

Childhood is a vital part of human development, especially brain development. 90% of brain development occurs before the age of five. In the early years, the mind forms neural connections and sets the stage for how we think, feel, and interpret the world around us. Another important aspect of proper brain development is that early development shapes long-term mental health and stress response throughout life. Extensive global research proves that when children are repeatedly exposed to violence and traumatic experiences, the neural pathways that are in charge of stress, memory, and emotional regulation are negatively altered, leaving long-term impacts on human behavior.

Today, over 460 million children live in areas that are actively affected by conflict; these children are exposed to violence and chronic stress characterized by (but not limited to) loss of loved ones, physical abuse, displacement, and mental trauma. Repeated exposure to traumatic events further leads to children in conflict zones facing significantly higher risks of developing anxiety disorders, post-traumatic stress disorder (PTSD), and, often, eventually, depression.

Trauma and stress that is induced by conflict-driven outcomes typically impact key brain regions, especially those that process fear and regulate emotions. Neuroscience initiatives show that traumatic stress morphs activity in the amygdala, hippocampus, and prefrontal cortex; together, these regions regulate threat detection, memory processing, and decision making. The amygdala acts as the brain’s internal alarm, specifically when a human is in danger, and their fight-or-flight response is activated. In children who are consistently exposed to violence, we see the amygdala become hyperactive, which causes the brain to remain in a heightened state of alertness and alarm for long periods of time.

Simultaneously, chronic stress also impacts the hippocampus, which is the brain region involved in learning and memory. This is when cortisol comes into play: when stress hormones are elevated within the hippocampus of a developing child, their hippocampal function is impacted, leading to impaired spatial navigation, wayfinding, and memory retention. These consequences are lifelong; for instance, everyday tasks such as navigating to a familiar location, remembering directions, or recalling a simple grocery list can become more difficult. It is equally important to zoom into the prefrontal cortex, which is the region also responsible for regulating emotions and, more specifically, controlling impulses. As these reflexes are disrupted by prolonged exposure to stress and violence, it becomes more and more difficult for children to regulate fear, control physical impulses, and respond appropriately to emotional situations. Early factors in a child’s life that are inevitable due to their geographic circumstance can lead to permanent consequences, compared to children in non-war-ridden climates.

Modern neuroscientists have begun using brain imaging techniques to better understand how exposure to violence influences underlying neural activity and behavior. A prominent study examining children exposed to violence early on were found to show increased activation in brain regions associated with fear processing, which is a pathway highly involved with the amygdala and hippocampus. These recent findings suggest that repeated exposure to environments that feel threatening and hostile result in strengthening neural circuits involved in detecting danger. Although that is a necessary adaptation for children in these areas, it unfortunately causes them to perceive threats even in safe situations throughout their life. This habituation leads to dangerously higher levels of stress, anxiety, depression, and further mental health challenges throughout the person’s life.

The most astonishing and recognized aspect of the brain is its plasticity — its unique ability to adapt and reorganize itself. However, plasticity is a two way street. Some say that during a traumatizing childhood, neural circuits are especially flexible, allowing these children to learn from environmental cues and adapt to their situation in a way that feels normal to their own brains and bodies. Contrary to this belief, plasticity within these children actually means that chronic exposure to stress or trauma shapes neural pathways in ways that persist into adulthood.

Furthermore, comparative studies that examine childhood trauma find that individuals who are exposed to early trauma show major developmental differences in brain regions correlated with memory, stress response, and emotional control, including reduced hippocampal volume and heightened amygdala reactivity. Individuals who grew up in non-war zones do not show these developmental differences in key brain regions. The neurological differences observed explain why individuals who grew up in conflict zones often experience higher rates of PTSD, anxiety, and depression somewhere down the road in their lives.

War is inherently traumatic, especially for children, whose developing brains are particularly vulnerable to its effects. However, I must note that brain research implores and highlights the importance of protection and support in a few domains: social support, stable caregiving environments, and access to mental health resources are a few buffers that can support healthier brain development within children in war-ridden areas and ease the impact of constant trauma. Research suggests that supportive environments can directly influence brain plasticity by regulating stress-response systems. Strong caregiving and social connections can reduce overactivation of the amygdala and support healthier development of the prefrontal cortex, allowing children to better regulate fear and emotional responses despite exposure to trauma.

War not only reshapes borders, nor is its biggest consequence elevated prices of gasoline: war profoundly affects the developing brain. For children growing up surrounded by conflict, the never-ending sounds and detection of danger lead to constant threat detection, fear, and emotional override. These factors create long lasting effects on emotional regulation, development, and mental health. Understanding these neurological impacts is essential for developing interventions that support children affected by war. As war persists in the modern day, it falls on our shoulders to mitigate long-term consequences of the trauma in early brains that are just barely beginning to engage with the world around it.

Bremner, J. D. (2006). Traumatic stress: Effects on the brain. Dialogues in Clinical Neuroscience, 8(4), 445–461.
https://pmc.ncbi.nlm.nih.gov/articles/PMC3181836/

McLaughlin, K. A., Sheridan, M. A., & Lambert, H. K. (2014). Childhood adversity and neural development: The impact of violence exposure on the brain. Annual Review of Developmental Psychology.
https://pmc.ncbi.nlm.nih.gov/articles/PMC7145742/

Teicher, M. H., & Samson, J. A. (2016). Annual research review: Enduring neurobiological effects of childhood abuse and neglect. Journal of Child Psychology and Psychiatry.
https://pmc.ncbi.nlm.nih.gov/articles/PMC6428430/

Sharma, N., & colleagues. (2024). Long-term effects of childhood exposure to violence in fragile settings. Global Mental Health.
https://pmc.ncbi.nlm.nih.gov/articles/PMC12930244/

Wymbs, N. F., et al. (2020). Social supports moderate the effects of child adversity on neural correlates of threat processing.
https://pmc.ncbi.nlm.nih.gov/articles/PMC8060780/

This article was written by Toros Kehribarian , an undergraduate student majoring in Neuroscience and Public Health at UC Berkeley. This article was edited by Meenakshi Yarlagadda.

I was the type of child that tried every sport you could think of and never really stuck to one. My parents always made me practice a sport because they said it was good for my body and brain. I, however, liked learning new things more than I did improving a skill. In university I wanted to keep moving my body because even if I had no idea how exercise could help my brain, I could see the impact it had on my body. I had less time for team sports and pivoted my attention to strength training in the gym. During the holidays I wanted to go back to practicing a sport with other people but didn’t want to stop all the progress I had made in the gym. That is how I got introduced to Brazilian Jiu Jitsu (BJJ).

BJJ is a martial art which combines traditional techniques from Judo and Jiu Jitsu with self defense. It requires athletes to constantly exert great amounts of strength to immobilize their opponents with various parts of their body. My curious mind wouldn’t allow me to just practice the sport without researching how it actually impacted my body and mind. That was how I stumbled upon a study that showed that training BJJ could improve symptoms of PTSD, depression and anxiety.

Post Traumatic Stress Disorder (PTSD) is a well known condition that can occur after a patient experiences a traumatic moment in their life. In the United States approximately 6% of the population will have PTSD at some point in their lives. Currently there are three main treatments for PTSD. Prolonged exposure teaches individuals to slowly go through traumatic memories and feelings which leads to reduced avoidance and decreases PTSD symptoms. Cognitive processing therapy teaches patients to challenge and change beliefs they have about their trauma which can help them understand it more and have less symptoms. The third treatment is eye movement desensitization and reprocessing, which makes the patient recount the traumatic experiences while having to move their eyes. This is intended to reduce the impact of traumatic memories on patients.

Exercise’s impact on the brain has been widely researched. Studies show that subjects who exercise regularly have a better overall mood and decreased symptoms of depressive and anxiety related disorders. Most studies show the impact of aerobic exercises like running. After a period of low intensity running, one study showed that there are lower cortisol levels in the brain causing subjects to feel more calm than before their run. Strength training and weight lifting are becoming increasingly popular modes of exercise and they have also been shown to have similar effects on the brain as aerobic exercises. A significant drop in anxiety levels were shown after different modes of strength training with more drastic differences after concurrent training. Scientists have shown that exercise has similar positive impacts on PTSD symptoms. The type of exercise that produced the most significant improvement on patients with PTSD symptoms was a combination of aerobic, strength and flexibility training.

This mixture of exercise techniques is one of BJJ’s most important qualities. It requires incredible amounts of strength, stamina and flexibility to be able to take down and immobilize a bigger opponent without striking them. In 2019 a study with United States veterans sought to find out the impact that this martial art and its multimodal approach could have on PTSD symptoms and if it could be used as a treatment for the disorder.

The current treatments for PTSD require expensive therapists and techniques, so to minimize cost and retain efficiency and results the scientists decided to study the impact that BJJ had on PTSD symptoms. In the past a study had concluded that self defense training similar to BJJ helped women with military sexual trauma, PTSD and depression.

The participants, men ages 18 and above, went through a 5 month long BJJ training program, keeping a record of their symptoms and behaviors before, during and after the training program. Results showed that not only were PTSD symptoms significantly decreased, but so were depressive disorder symptoms and alcohol and substance use. The graph below measures the average symptoms of PTSD on the PTSD checklist (PCL-5) before the 5 months of training, halfway through the training program and after the participants finished the program. There is a clear and significant decrease in the PCL-5 score determining a lessening of symptoms.

The scientists in charge of this experiment pose two possible explanations for this decrease in PTSD and depression symptoms in the participants. The first is the impact of physical activity on the brain that has been explored in this article. A second explanation is one of resilience. BJJ teaches athletes to problem solve in the moment and get themselves out of tough situations. The study explores whether this trained resilience is something that not only impacts BJJ but is applied to life causing the participants to have a more resilient mind. The perception of being capable and resilient is one that can decrease symptoms of PTSD. BJJ potentially rewires the brain, teaching athletes how to react in uncomfortable situations. These lessons act as a form of exposure therapy to those with PTSD.

Exercise is commonly known to improve both our physical and mental health but it is rare that we can see the specific impacts that a singular type of exercise can have on the brain. BJJ’s impact on PTSD symptoms bridges different types of therapies to improve the mind. The constant physical exercise with aspects of aerobic, strength and flexibility training combined with the exposure therapy aspects of the martial art allows BJJ to have a greater impact on the brain ultimately decreasing symptoms of PTSD and other psychological disorders.

What is Jiu Jitsu | BJJ Heroes. (2011, June 26). https://www.bjjheroes.com/bjj-news/brazilian-jiu-jitsu

Bäärnhielm, S., Ramel, B., Theunis, E., Mijaljica, G., Dyster-Aas, J., & K Arnberg, F. (2024). [Post-traumatic stress disorder (PTSD) and complex PTSD (CPTSD) — A clinical update of knowledge]. Lakartidningen, 121, 23090.

Merians, A. N., Spiller, T., Harpaz-Rotem, I., Krystal, J. H., & Pietrzak, R. H. (2023). Post-traumatic Stress Disorder. Medical Clinics of North America, Clinical Psychiatry, 107(1), 85–99. https://doi.org/10.1016/j.mcna.2022.04.003

Peluso, M. A. M., & Andrade, L. H. S. G. de. (2005). PHYSICAL ACTIVITY AND MENTAL HEALTH: THE ASSOCIATION BETWEEN EXERCISE AND MOOD. Clinics, 60(1), 61–70. https://doi.org/10.1590/S1807-59322005000100012

Schneider, S., Askew, C. D., Diehl, J., Mierau, A., Kleinert, J., Abel, T., Carnahan, H., & Strüder, H. K. (2009). EEG activity and mood in health orientated runners after different exercise intensities. Physiology & Behavior, 96(4), 709–716. https://doi.org/10.1016/j.physbeh.2009.01.007

Barahona-Fuentes, G., Huerta Ojeda, Á., & Chirosa-Ríos, L. (2021). Effects of Training with Different Modes of Strength Intervention on Psychosocial Disorders in Adolescents: A Systematic Review and Meta-Analysis. International Journal of Environmental Research and Public Health, 18(18), 9477. https://doi.org/10.3390/ijerph18189477

Jadhakhan, F., Lambert, N., Middlebrook, N., Evans, D. W., & Falla, D. (2022). Is exercise/physical activity effective at reducing symptoms of post-traumatic stress disorder in adults — A systematic review. Frontiers in Psychology, 13, 943479. https://doi.org/10.3389/fpsyg.2022.943479

Willing, A. E., Girling, S. A., Deichert, R., Wood-Deichert, R., Gonzalez, J., Hernandez, D., Foran, E., Sanberg, P. R., & Kip, K. E. (2019). Brazilian Jiu Jitsu Training for US Service Members and Veterans with Symptoms of PTSD. Military Medicine, 184(11–12), e626–e631. https://doi.org/10.1093/milmed/usz074

This article was written by Juliana Amaral, an undergraduate student majoring in Neuroscience at UC Berkeley. This article was edited by Meenakshi Yarlagadda.

Some of the most fascinating questions in the world are the ones that seem impossible to answer: is there other sentient life in our universe? What makes us conscious human beings? The latter question has gnawed at humans for centuries; after all, what is more human than analyzing ourselves? Different schools of thought have attempted to offer theories and explanations for this question. In the 1600s, philosopher René Descartes proposed that consciousness, or one’s existence, is the only thing in life a person can believe because it is immaterial, and can not be an illusion: cogito ergo sum, or “I think, therefore I exist.” Religion attributes consciousness to a divine spark. Science looks to neurotransmitters — chemicals in the brain that allow communication between neurons — and neuronal networks to explain the basis of behavior that signifies consciousness. However, despite its extensive findings of brain activity behind observable human behavior, science has not yet been able to explain conscious experience itself. For example, if we mapped all human brain functions to a computer, would it gain sentience? Probably not, so what are we missing?

To start tackling this problem, we must first define consciousness. To be conscious is to be aware of one’s existence and recognize the qualitative experience of something, also known as “qualia.” In his landmark 1974 paper “What Is It Like to Be a Bat?” philosopher Thomas Nagel summarizes this concept as “There is something that it is like to be in a mental state”; there is a qualitative feeling to experience. A conscious state is classified by the “type” of consciousness, or state of awareness, one is experiencing, from sleep to alertness. As a result, most consciousness studies compare different states of consciousness and the neural underpinnings of the observable differences between states, such as the brain on sleep, various substances, or while alert. These studies do not uncover the root of conscious experience itself, which is considered the “hard problem” of consciousness.

Neuroscience studies have attempted to explain the nature of consciousness itself, even if explaining how consciousness itself emerges was a gargantuan task. In 1983, neuroscientist Benjamin Libet ran an experiment to test human free will and consciousness. Test subjects were instructed to simply flex their wrist whenever they wanted to and mark when they first felt aware of their intention to act. However, readiness potentials from the brain, which are neural activity spikes in the motor cortex, preceded the movement and the subject’s awareness of their intention. In other words, the brain initiated an action potential to flex the wrist before the subject was even aware of it. This result implies we do not possess free will to initiate actions, and thus are not conscious in the moment, even if we feel the illusion of control. So, when are we actually conscious?

This experiment unfortunately suffers from a few experimental design and data interpretation errors: for example, the recorded readiness potential may not be correlated to intention to act, the experiment does not resemble real world conditions, and its claim to be measuring free will is dubious. An important aspect in our search for consciousness is ensuring that we are measuring the same thing: a tough feat, as there is no universally agreed-upon definition of consciousness. However, the questions Libet posed in his experiment are crucial in future experimentation and our search for answers.

Neuroscientists Roger Sperry and Michael Gazzaniga conducted a series of split-brain experiments in the 1960s to observe the behavior of humans whose corpus callosum — a bundle of neurons that connects the brain’s two hemispheres — was split as an epilepsy treatment. In these experiments, because the hemispheres could no longer communicate, a person claimed not to have seen an object placed in front of them: the left side of the brain, responsible for language production, could not receive input from the right brain, which processed the visual object in the subject’s left visual field. As a result, the subject could not process or articulate that the object had been perceived. This confusing behavior calls into question: are there now multiple distinct, conscious minds in the brain? If so, does the brain with an intact corpus callosum have multiple loci of consciousness that integrate together well for regular functioning? The general modern-day consensus is that more information is needed to answer these questions; however, some theories of consciousness rely on the assumption that consciousness is split in a split-brain.

All of these experiments culminate in the elusive mind-body problem: this problem attempts to define the relationship between the mind and body. The body exists in the physical world, and is thus subject to the laws of physics. Old schools of thought believed that the mind was a distinct entity and unrelated, but the current consensus is that the mind is a product of the brain, and is thus also grounded in the laws of the physical world. Cognitive and behavioral neuroscience attempts to understand the precise relationship between the mind and brain by finding the neurobiological basis of observable mental phenomena but we still do not know how conscious experience emerges. We can correlate neurobiological brain states to mental states, but we can not correlate brain states to our qualitative experience of mental states.

Neural theories of consciousness have emerged to understand how conscious experience is encoded in the brain, which is distinct from studies about different states of consciousness. Each theory seeks to identify a neural correlate of consciousness: the minimum activation of neurobiological mechanisms necessary to generate conscious experience (although some believe there is no specific neural correlate of consciousness, and it is instead bound to visual processing mechanisms).

One such theory is the memory theory of consciousness, which suggests that the qualitative experience of consciousness that we feel might be a memory that was processed extremely quickly. The theory also proposes that instead of one unified neural correlate across the brain, different cortical areas contribute different aspects to consciousness, and thus have different correlates. This theory would accurately account for the phenomenon of conditions following a stimulus affecting our conscious perception of it.

Alternately, higher-order theories of consciousness suggest that an experience can be deemed conscious only after a “higher area” has created a brain state that correlates to a thought about the experience. For example, in visual processing, visual stimuli are passed to the primary visual cortex, which has neurons that respond to the stimuli’s basic, simple features. As areas of higher complexity process the stimuli, the initially selected basic features are combined to create more complex features, until the highest-level area can represent an entire object like a face. Similarly, the higher-order theory claims that in order for conscious perception to arise, a higher-order network in the brain must process stimuli up this hierarchy and have an associated brain state correlated to thinking about the experience. For example, for a color to be consciously perceived, the visual system must represent the color, and there must be a representation of the thought of perceiving the color — quite layered!

The global workspace theory posits that stimuli we process are only registered as conscious once they are broadcasted to different regions that separately process the information; it then is synthesized into one stream of consciousness.

However, both higher-order theories and the global workspace theories are computational models that explain how consciousness functions, not what provides us with the feeling of conscious experience. Understanding how physical states enable the experience of subjective states may require more biological models, but scientists still do not know how to close this explanatory gap.

In order to understand the precise relationship between the mind and brain, we need to understand what mechanisms enable the qualitative experience of consciousness (the “hard problem” of consciousness), and understanding the neural correlates of consciousness may be the key to unlocking this current knowledge gap. We can not stop asking these tough questions, even if they feel too broad to tackle all at once.

Bibliography

Aru J., Suzuki M., Rutiku R., Larkum M. E., Bachmann T. (2019). Coupling the State and Contents of Consciousness. Frontiers in Systems Neuroscience, 13. https://doi.org/10.3389%2Ffnsys.2019.00043.

Block, N. (2009). Comparing the major theories of consciousness. In Michael Gazzaniga (ed.), The Cognitive Neurosciences IV. (pp. 1111–1123). MIT Press.

Chalmers, D. J. (2010). What Is a Neural Correlate of Consciousness? The Character of Consciousness (pp. 59–90). Oxford Academic. https://doi.org/10.1093/acprof:oso/9780195311105.003.0003.

de Haan, E.H.F., Corballis, P.M., Hillyard, S.A. et al. (2020). Split-Brain: What We Know Now and Why This is Important for Understanding Consciousness. Neuropsychology Review, 30, 224–233. https://doi.org/10.1007/s11065-020-09439-3.

Dehaene, S., Lau, H., Kouider, S. (2017). What is consciousness, and could machines have it? Science, 358(6362), 486–492. https://doi.org/10.1126/science.aan8871.

Friedman, G., Turk, K.W. & Budson, A.E. (2023). The Current of Consciousness: Neural Correlates and Clinical Aspects. Current Neurology and Neuroscience Reports 23, 345–352. https://doi.org/10.1007/s11910-023-01276-0.

Libet, B., Gleason, C. A., Wright, E. W., Pearl, D. K. (1983). Time of conscious intention to act in relation to onset of cerebral activity (readiness-potential): the unconscious initiation of a freely voluntary act. Brain, 106(3), 623–642. https://doi.org/10.1093/brain/106.3.623.

Nagel, T. (1971). Synethese, 22(3/4), 396–413. https://www.jstor.org/stable/20114764.

Northoff, G., Lamme, V. (2020). Neural signs and mechanisms of consciousness: Is there a potential convergence of theories of consciousness in sight? Neuroscience and Biobehavioral Reviews, 118, 568–587. https://doi.org/10.1016/j.neubiorev.2020.07.019.

Seth, A. K., Bayne, T. (2022). Theories of consciousness. Nature Reviews Neuroscience, 23, 439–452. https://doi.org/10.1038/s41583-022-00587-4.

Valencia, A. L., Froese, T. (2020). What binds us? Inter-brain neural synchronization and its implications for theories of human consciousness. Neuroscience of Consciousness, 2020(1), https://doi.org/10.1093/nc/niaa010.

This article was written by Arhana Aatresh, an undergraduate student at UC Berkeley majoring in Neuroscience and minoring in Science and Technology Studies. This article was edited by Aileen Xia and Jade Harrell.

When you’re by yourself, content and at peace with whatever task you’re doing at the moment, are you ever interrupted by the intrusive feeling that someone is watching you? Or, do you feel silently watched all the time? Perhaps you’re reading this and are actually very concerned by these questions — “What is this person talking about?” you think to yourself; “Are they experiencing onset schizophrenia?” If you truly do not understand what I mean, though, then you are simply ignorant to the reality occurring around you.

In fact, you are this mysterious “someone” I keep referring to: you are watching yourself. Maybe you are not watching yourself all the time, but rather, in short moments littered all throughout your day. In order to determine whether or not you were asking yourself the questions I proposed for you earlier, you had to enter your mind for a second and silently watch — waiting to see if these sorts of questions came up, or had come up previous to me suggesting them. It is a fact that you are being watched often, although comfortingly, you are likely only being watched by yourself.

This phenomenon of watching yourself, or more specifically, being aware of and thinking about your thinking, is called metacognition in psychological literature. The concept of metacognition has been of interest to many in psychological and technological fields, as it is the basis of what many believe manifests mental growth and everyday functioning. If our models of metacognition are correct, then it can serve as a very useful tool to uncover how we are able to learn more and what this implies about our neurocognitive function. This information can also be applied towards computational modeling in self-learning devices (artificial intelligence).

Metacognition

Metacognition can be described as the higher-order thinking that you can utilize to investigate a first order/lower order thought. To break this down, imagine you are walking around your neighborhood. While walking, you see an obnoxious black sports car speed right past you. This is a first-order thought — essentially an observation you make. In this first-order thought, you have identified whether or not a stimulus is present. If you were to be asked by another person on the street whether or not an obnoxious black sports car had sped past you a few seconds ago, you can then make a claim about it depending on how confident you are — “Yes, an obnoxious black sports car just sped past me” or “I don’t know, I wasn’t paying attention.” This would be a second-order thought, as you are entering your mind and retrieving back to your memory to see if this had happened, and you are also implicitly reporting how you feel about the information you retrieved in your response (saying “yes” implies you are sure of yourself, saying “I don’t know” implies you are not). The metacognitive part of this story lies in how you are actively engaging with and thinking about another thought that you already had. You are climbing up the ladder of thoughts and looking down at the preceding ones in your new thoughts. An interesting case of metacognitive impairment is illustrated in a study dubbed “blindsight,” in which patients with lesions to the primary visual cortex are asked to report whether or not they see a stimulus, similar to whether or not you saw the obnoxious black sports car. Despite being able to respond to the stimulus (e.g., being startled by the presence of something only detectable by sight, etc.), and therefore confirming that the patients are able to actually see the stimulus, the patients will report that they did not see anything. The patients do not lack the ability to see the object — they are able to have the first-order thought of observation — but rather, they lack the awareness of having seen it, demonstrating that their metacognition of sight is poor because they cannot process a higher-order thought about what they have seen.

The ability to understand and respond to one’s own thinking can occur in a multitude of ways — not just through visual observation and memory retrieval. Metacognition is broken up into three different “kinds” of metacognition based on the subject to which the judgment of thought pertains: anoetic, noetic, and autonoetic. Anoetic metacognition deals with reflecting on objects within the world, like the car and blindsight examples from before. Noetic metacognition concerns the mental representations that a person holds, as metacognition can also occur towards the various schema/models of things that we have in our minds. For example, it would be involved in determining if your personal idea of “obnoxious” is accurate when compared to many other definitions of “obnoxious.” This is a very important process in learning, as in order to learn, we must reflect on our current representation/understanding of a subject and see how it holds up to new information. Lastly, autonoetic metacognition describes judgments of oneself, similar to the concept of introspection. Unlike noetic and anoetic metacognition, where a fixed piece of information is processed in our mind, autonoetic metacognition focuses more on being able to narrow in intuitively on fluid features of the self and our ability to take agency on account of those findings. Making a calculated decision about something troubling you would require the use of autonoetic metacognition.

As a result of these definitions, it has been sometimes assumed that metacognition might merely mean introspection or consciousness, although this is not true. The two concepts of consciousness and metacognition are fairly comparable, as consciousness describes the state of us being aware of ourselves and metacognition similarly describes us utilizing our awareness of ourselves. One might think that metacognition is a manifestation of consciousness — and this would be true — but then we must remind ourselves that nearly all cognitive processes start with consciousness. To think of metacognition and consciousness as the same neglects the uniqueness of metacognition being able to manipulate the mind. Consciousness is a state of awareness whereas metacognition further uses this awareness to spur on future actions and thoughts. In a sense, consciousness allows for the possibility of metacognition to exist, but not vice versa, so they do not describe the same phenomenon but are instead adjacent to each other. Similarly, introspection is also sometimes conflated to mean the same thing as metacognition. While properties of introspection require the use of metacognition (in order to reflect on one’s actions, one must have mental representations of themselves to be analyzed by higher order thought), metacognition is not just introspection. This can be best seen in the blindsight example from earlier, in which patients lacked the metacognition to identify that they were seeing something, but they did not necessarily lack introspection in the subjective sense. They could still, theoretically, identify features of themselves like their personality — or they could even be subjectively aware of their cognitive deficit — but this doesn’t change their inability to use metacognition in regards to visual perception. They can be self-aware, but not in cognitive control of their minds.

Metacognition is intimately tied to the idea of cognitive control and monitoring. In order to use metacognition, one must be mentally monitoring, consciously or subconsciously. Further, by the use of metacognitive monitoring, people have more cognitive control and the capacity to better deliberate over their thoughts, reasoning, and overarching mental skills.. Though, as always, it is important not to conflate these ideas together. In an experiment done on the cognitive control/monitoring of typists, it was found that they could intuitively account for errors in moments of lack of awareness — as in, they instinctively recognized that they made an error in their typing even while not fully present and corrected them — demonstrating their enhanced monitoring and control. However, when experimenters added in random errors of their own, the typists noticed them and took accountability for these errors that they themselves did not make. Despite being aware of when they actually did make mistakes via monitoring, they lacked the metacognition to realize when they did not make mistakes — suggesting that cognitive monitoring/control and metacognition are not the same inherently, although related.

Digging deeper into the meaning of metacognition, it becomes apparent that the nature of metacognition is rather subjective and difficult to quantify. Its definition overlaps with many other definitions/phenomena, and one cannot make a clear, distinct statement about what a metacognitive, higher order thought must look like, as the thought can, quite frankly, be about anything. Further, higher order thinking can go on almost continuously — building and building to look upon a previous thought indefinitely — so fixing it in order to make clear statements on it would be neglecting the full nature of metacognition. Therefore, it is important to be delicate and nuanced when operationalizing metacognition for scientific studies. The definition must be strict and simple enough so that it is quantifiable, but abstract enough so that the true nature of metacognition is captured within the study. Because of this, to compute metacognition we often restrict it to measuring feelings that can only exist if metacognition did/didn’t take place.

For example, a key indicator of the presence of metacognition is confidence in regards to performing a task. In this case, the task that a person performs is the object-level, first order thought. The task can be anything that requires the use of varying individual knowledge and skills, such as answering a history or math question. After completing the task, the person’s metacognition can be measured by having them rate how confident they are in the accuracy of their task/answer. Defining how confident someone is in their performance requires them to look within themselves for their skills and knowledge — thereby thinking about past thinking, comparing that with the object-level thought, and “activating” their metacognition. Depending on whether or not their performance was accurate, this can be compared with how well they thought they did when asked to report their confidence in their performance. If they perform correctly and are confident in their performance, they would be considered to have good metacognition, as their reflection of their skills/cognition aligned with their actual skills/cognition as measured by the task. However, if they perform incorrectly but were confident that they had performed correctly, then they would have poor metacognition as their interpreted skills and actual skills did not align. Confidence and question-answering like this is easily self-reported and measured, so it is a popular way of operationalizing metacognition in research. This way of measurement does not come without flaws, though, because sometimes not having an aligned confidence and performance does not mean the metacognition is poor. For example, accurately having low confidence on answering a question because the knowledge is not strong in one’s mind, and then getting the answer right anyway does not mean that one’s metacognition is bad, but rather, that they got the answer right by chance or through extreme difficulty. An adjacent method of measurement is to have subjects make a decision, and then reflect on why they made that decision or the difficulty of making that decision. This method is good for looking more closely at how metacognition works case-by-case between people, as opposed to just measuring the accuracy or strength of metacognition. These kinds of operationalization are referred to as retrospective metacognition, as it is the present use of metacognition regarding something of the past.

Metacognition is also broken up into different, specific “acts,” including “judgements of learning” and “feelings of knowing”. Judgements of learning are a kind of metacognition, as they require reflection on cognitive thought — recognizing how well or if one has learned a subject adequately requires the analysis of one’s own cognition of the subject. On the other hand, feelings of knowing are essentially “tip-of-the-tongue” states, in which one recalls that they know of something, like the name of an actor, although they cannot remember it at the present moment. This demonstrates use of metacognition, as it again requires awareness of one’s cognitive state (memory), and a report back about the state. It occurs to me here whether or not feelings of knowing imply the existence of unconscious metacognition, as technically, the piece of knowledge that one is recalling that they know is not directly conscious to the person. Feelings of knowing confirm that information is constantly stored at the unconscious level, and that it can be deliberately accessed at any time (although with delays and sometimes unsuccessfully). Perhaps, if metacognition is not unconscious in and of itself, it could be that it is the gate into accessing unconscious mental processes. Not only does metacognition go up the metaphorical ladder of thoughts by pressing down on the thought directly beneath it — it also can climb down the ladder rapidly, going underground, and shoot back up while skipping steps along the way for efficiency. Feelings of knowing and judgments of learning are called prospective applications of metacognition.

Relating Metacognition to Neurobiology

Being as fluid and subjective as it is, it would be any neuroscientist’s fantasy to be able to objectively pinpoint where and how metacognition occurs in the brain. Unsurprisingly, metacognition is a multi-step process that involves many parts of the brain. Different “types” of metacognition are also processed in different regions, as confirmed by people who have brain lesions and only specific kinds or steps of metacognition impaired. Despite being multifaceted, metacognition has predominantly been localized to the prefrontal cortex (PFC), the anterior cingulate cortex (ACC), and the hippocampus.

Activity in the ACC is mainly seen working in tandem with the lateral PFC and anterior insula when an individual experiences response conflict. Response conflict describes when an individual is receiving contradictory or noisy information from a stimuli that they must sort through mentally before giving an answer or “response” for. The noise/contradiction causes a conflict in the response that the individual is inclined to give. For example, in the Stroop Test a person is given a series of color words (e.g. “purple,” “green”) that are themselves colored as a different color than what the word says, and are told to report the color of the letter, not the word. This process causes a response conflict as the individual being tested is inclined to say the word and not the color of the word. While experiencing response conflict, there is increased activity in the ACC, anterior insula, and lateral PFC. As we will discuss in a moment, the lateral PFC is correlated with behavior adjustment, increased caution, and cognitive control. Using this fact, it is theorized that the ACC detects when there is response conflict or a lack of confidence and then signals for the lateral PFC to be activated in order to heighten cognitive control and adjustment.

Figure 1. The Stroop Test: One must say aloud the color of the word, not the word itself. For example, the first “Orange” should be read “pink.”

Within the PFC, metacognitive processes are further split between the lateral PFC and the medial PFC. The lateral PFC has been seen to play a role in retrospective metacognitive acts. Dorsolateral PFC activation is associated with individuals experiencing any decreased confidence in performance or a thought, thus suggesting that it’s responsible for holding mental representations to be crunched by metacognitive processes — similar to working memory. In a study where subjects with rostrolateral PFC lesions were tested for their metacognitive ability, it was found that while their metacognition was intact, they required a greater threshold for decreased/increased confidence before being able to report the difficulty/ease of the metacognitive tasks they were doing. Therefore, the rostrolateral PFC likely plays a part in the ability to communicate confidence and in the detection of confidence within metacognition. Supporting this further, varying metacognitive confidence is positively related to gray matter volume in the rostrolateral PFC. The rostrolateral PFC is also more noticeably correlated with increased confidence (greater activation with greater confidence, and vice versa) and stronger white matter integrity in this area was associated with stronger metacognitive accuracy. In other cases, the rostrolateral PFC is seen to integrate information from the ACC and anterior insula. Further, it was found that there is a separation between the functions of the ACC and lateral PFC, as subjects with lesions to the lateral PFC could not report in what way the Stroop Test was difficult, but they were able to perform the test accurately.

With all of this information in mind, I think it is plausible that, if the ACC signals a lack of confidence to all of the lateral PFC (which includes the rostrolateral PFC), then the rostrolateral PFC (being associated with increased confidence) tries to make sense of the information it is being given and find as much confidence that it can from the puddle of uncertainty and doubt gifted by the ACC. To picture this more intuitively, imagine you are trying to figure out whether or not a girl you’re interested in is also interested in you. This girl is sending you mixed signals — she texts you all the time but ignores you and is rather apathetic with you in person. Having both of these conflicting ideas is noticed and then correctly sorted by your ACC: “Texting me all the time means she’s interested, but being apathetic means she’s disinterested, and she cannot be both interested and disinterested in me at the same time.” Then, this information gets sent over to your rostrolateral PFC for confidence detection and communication: “Well, I know that she texts me all the time, so I am confident that she probably likes me at least a little, but maybe she’s on the fence about me. This is what I will tell my friend if they ask me about the situation with her.” At some point when the information is being sent to the ACC, it is also sent to the dorsolateral PFC: “I am not confident about whether or not she likes me.” This relationship between the ACC, dorsolateral PFC, and rostrolateral PFC is only a theory of mine, but it is a fact that the ACC and dorsolateral PFC are associated with decreased confidence and the rostrolateral PFC with increased confidence.

On the other hand, the medial PFC has shown to play a relevant role in prospective uses of metacognition. Damage to the ventromedial PFC leads to a decrease in feelings of knowing accuracy, and activity in the ventromedial PFC increases during accurate feelings of knowing. Activity in the ventromedial PFC also increases during accurate judgments of learning, thus implicating this region in contributing to the accuracy of all prospective metacognition. Outside of prospective metacognition, the anterior medial PFC has been shown to have increased activity (and myelination) during moments of self-reflection/introspection. An interesting finding to note here is that, while anterior medial PFC activity is increased in healthy individuals during self-reflection, this is not true of those with mental health conditions such as schizophrenia while trying to introspect. While this could be because those who are schizophrenic do not self-reflect, and thus there is nothing to be activated, this opens up another interpretation — those with schizophrenia lack the ability to engage with their anterior medial PFC, and thus cannot produce accurate accounts of self-reflection. This new interpretation does not really tell us much that we did not already know, but it does expand the idea of schizophrenia being more of a cognitive issue, and it suggests that there is a process which must occur before the self-reflection process does (perhaps something needs to send a signal to the anterior medial PFC before introspection can start). This step-by-step thinking is supported by Paul Lysaker’s work in schizophrenia treatment, in which he breaks down the self-reflection process to its roots in object-level (first order) cognitive awareness. This treatment has been supposedly successful if approached carefully in a step-by-step process of building cognitive awareness and subjective introspective ability!

Lastly, metacognitive ability has been linked to magnetization transfer and myelination in the retrosplenial cortex and hippocampus. In the retrosplenial cortex, stronger metacognition was correlated with less magnetization transfer. In the left hippocampus, stronger metacognition was correlated with having weaker myelination in this area. Both the retrosplenial cortex and hippocampus are important components of the neural memory network, so having a correlation with metacognition suggests the existence of metamemory and memory working in conjunction with metacognition. This comes as no surprise, as in order to make retrospective judgments of metacognition one must remember the event in which they are judging. What is shocking about this finding is that there is a negative correlation between these parts and processes. A negative correlation suggests that more sparse connectivity in the hippocampus leads to a strengthening of accurate encoding/recall of decision making “evidence” or pieces of information to be used in metacognition. I believe that it is important to consider why this may be occurring, so, thinking back on the idea of response conflict and the ACC, it could be possible that is the case because of noise. Perhaps having sparser connections ensures that redundant information is not being encoded or recalled in memories during metacognitive processes, which could potentially cause conflict in the cognition and lead to an unproductive/unnecessary lessening of confidence that would extend more mental energy than desired. For example, someone attempting to believe that they are a good person will not recall neutral things (not bad nor good) that they have done in order to avoid redundancy, and this is a natural process because the brain does not want to exert more energy than necessary to a given task. By not exerting energy on redundant ideas, more energy can be given to the few relevant mental connections made — therefore, stronger, accurate metacognition occurs.

Metacognition in Artificial Intelligence

Considering the formulaic nature of metacognition, it is of interest if metacognition can be described in computer algorithms. If we were to, for example, give AI the capacity for introspection, they could take in information with more ease and efficiency as they could recognize when new information contradicts the old information that they are aware of having. At this point, we could give them a command to abandon all old information which contradicts the new information — in a sense, creating a self-updating program that requires less work from the programmer. Also, by giving AI metacognition, we could theoretically give the AI a “goal” and context of the “problem” that the goal is resolving, and it would be self-aware enough to know how to carry out that goal using its knowledge of its own programming limitations. An example of this would be having an AI translate between programming languages. If we give the AI knowledge of a set of languages and what various commands in the language mean, with metacognition, the AI could identify the overall goals and limitations of a piece of code in one language, and then match that goal in another language and translate the code while adhering to the new language’s own limitations.

To begin giving AI these abilities, one would need to deconstruct the various abstractions that define metacognition, such as “knowing of oneself” subjectively. One way of doing so is by giving the AI a double-meta model of itself, in which the programmer gives the AI the knowledge/data of how itself is programmed and what it looks like for itself to know how itself is programmed. This may sound abstract, but this is a fairly intuitive concept. Take yourself, for example. You have an idea of yourself, and who you are, broadly speaking. Along with this, you also have an idea of what it looks like when you are analyzing — you are aware of and see the reasoning that occurs when you are thinking of yourself, someone else, or a concept. By giving AI a model of itself, it would understand who it is and how it functions, but it would still lack the ability to make claims or have reasoning about itself. Hence, if you give the AI a model of a model of itself, it can make claims about the way it thinks of itself. This double model accounts for most of the metacognitive ability that an AI would need to know to complete the tasks that we want it to do, but theoretically we could continue this process over and over again. Each new model of itself represents a higher order of metacognition thinking. To describe this in a more computational way, we can also think of higher orders of thought as “dimensions” within the “space” of AI, which makes our goals of AI metacognition more mathematically tangible. This way of creating metacognition for AI is called model based reasoning.

Another important aspect of metacognition relevant to AI is the idea of decision-making and rationality. Metacognition is terribly important for decision making, as one must engage with higher order thinking to make more complex decisions. Generally, one wants to make a rational decision as opposed to an irrational one — requiring even more metacognition to decide what is irrational. Rationality is subjective, so oftentimes rational decision-making is operationalized to mean the decision which has the least cost and the most benefit. This operationalization can then be computed into an anytime algorithm. An anytime algorithm is an algorithm that ensures an output still occurs even if the algorithm is interrupted by additional inputs during the time that it takes to run. This way, all new inputs are accounted for which is imperative for rational decision-making, as what defines the most rational decision can change in an instant. Metacognitively, we are good at adapting to new information and accounting for it right away while making decisions because we constantly have the information readily available to us by means of cognitive monitoring. For computers, this is not the case as commands are fixed and not adaptable — therefore requiring that we make an algorithm that forces adaptation when necessary.

To do so, what first must be done is collecting our meta-data. Meta-data, in this case, means essentially the span of all possible decisions and decision making outcomes. This span is fluid and changing though, as more data gets added with each second of running, so we must make this an anytime algorithm from the very start of the process. Building off of our meta-data, we must implement our decision making process and commands which define what, in our data, has the least cost and most reward. This process could, theoretically, go on indefinitely as the system weighs out the pros and cons of each decision over and over and as the data continuously changes. Because of this, we must consider when enough “thinking” or running time has occurred before terminating the run of the algorithm (would it be beneficial or a waste of time to continue thinking about which choice is best?) — which would require another algorithm that monitors this decision making process if we want this process to be automatically functioning in the AI. A decision making monitor system would also require another computation, where we compute what it means when it is important to continue debating a decision as opposed to it being more important to abandon it for efficiency. In humans, this process is natural and swift as we intuitively know when we have “finished” the time required to make a decision — taking about 5 minutes to decide what is best to make for dinner versus taking 5 hours to decide what kind of car you should buy.

Another way of implementing decision making in AI would be to adopt Pavlovian learning strategies, and to condition the AI to make a certain kind of decision that has worked in the past. This way, one would only need to program the AI to detect similar features of problems it has solved in the past, the problem-and-goal method described earlier, and a means by which to apply past solutions to new situations. Humans do this constantly via experiential learning. This strategy is called case-based reasoning.

Failure driven learning is also being attempted in the programming of AI. For example, the project Meta-AQUA showcases an AI that is capable of generating a story, explaining what is going on in the story/why things occurred as they did, and find faults in their explanation of the story. Meta-AQUA uses model based reasoning and introspection to exhibit this metacognition, as it first has a model of itself, then uses introspection to gather potential reasons why its explanation is incorrect or not up-to-par yet. The project is still ongoing, and its rigorous methods of machine learning are likely to be fruitful in the discussion of technological metacognition for the future.

Conclusion

Metacognition is the process by which we use higher order thinking to make a judgment on lower order thoughts. This can be in a number of ways, and about a number of things. Metacognition can take the form of making a statement about an outside object that one perceives. It can be seen in introspection of oneself that requires deep mental analysis, thinking about choices in order to make a rational decision, or determining if one should be confident about their performance. Common examples of metacognition used in studies are feelings of knowing and judgments of learning. Cognitive control and monitoring is oftentimes mediated by the process of metacognition. Neurologically, the ACC is thought to signal a lack of confidence to the lateral PFC, in which the lateral PFC can then communicate and “bring to life” the presence of the uncertainty via metacognition. The medial PFC is relevant for prospective metacognitive judgments and introspection, and the hippocampus is relevant for more accurate metacognitive recall. Metacognition is being attempted to be replicated in AI for the purposes of efficiency, because if AI can use metacognition, they may become self-functioning and self-fixing. There are a number of metacognitive strategies being worked on in AI, such as rational decision making and failure-driven learning, that are performed on the basis of model and case based reasoning.

Why is any of this important? I’m sure that most of this has felt like being pummeled with a lot of abstract information, and while I described many of these concepts in practical applications, I understand it can still be hard to internalize as something relevant. After all, this is just base-level information — how can we even move forward with applying metacognition? Metacognition seems like merely a fact of life. However, metacognition is sneakily very important for everyday health and functioning. It would be impossible to move throughout your day (safely, at least) without the ability to spontaneously “check” your mind and analyze the best next action. What’s more is that metacognition is varying for many people. Some have really good metacognition, while others do not — and those who are struggling with, let’s say autonoetic metacognition, are typically experiencing other problems that impact their quality of life (for example, people who have schizophrenia). This is why studying metacognition, how it works in the brain, and trying to replicate metacognition in technology is so important. By understanding how metacognition works, we can find who struggles with metacognition and in what ways. With our neurological understanding, we can target these places in the brain and stimulate it for those struggling. To stimulate it, we can use neurotechnology, or we can implant a “metacognitive device” which simulates the experience of metacognition if the brain is completely deficient in this ability. Or, we can try a non-invasive approach such as that of the introspective exercises done with schizophrenic patients, as described earlier. Metacognition holds the key to improving the lives of people with a range of cognitive health issues. So, if we were to ignore metacognition in further research for the sake of its nuance or abstractness, we would simultaneously be dismissing the needs of large populations of people who are facing the consequences of impaired metacognition and could desperately use an improved one.

Bibliography

Fleming, S. M., Dolan, R. J., & Frith, C. D. (2012). Metacognition: computation, biology and function. Philosophical Transactions of the Royal Society B: Biological Sciences, 367(1594), 1280–1286.

Allen, M., Glen, J. C., Müllensiefen, D., Schwarzkopf, D. S., Fardo, F., Frank, D., … & Rees, G. (2017). Metacognitive ability correlates with hippocampal and prefrontal microstructure. NeuroImage, 149, 415–423.

Fleming, S. M., & Dolan, R. J. (2012). The neural basis of metacognitive ability. Philosophical Transactions of the Royal Society B: Biological Sciences, 367(1594), 1338–1349.

Lysaker, P. H., Buck, K. D., Carcione, A., Procacci, M., Salvatore, G., Nicolò, G., & Dimaggio, G. (2011). Addressing metacognitive capacity for self reflection in the psychotherapy for schizophrenia: a conceptual model of the key tasks and processes. Psychology and Psychotherapy: Theory, Research and Practice, 84(1), 58–69.

Cox, M. T. (2005). Metacognition in computation: A selected research review. Artificial intelligence, 169(2), 104–141.

This article was written by Jade Harrell, an undergraduate student at UC Berkeley studying Molecular and Cell Biology with a concentration in Neurobiology, as well as Psychology. This article was edited by Shobhin Logani and Jacob Marks.

For the most part, I have a pretty standard TikTok feed: I get a few DIY videos here and there, a fair share of movie and show reviews, and a weird amount of instant ramen hacks. I never imagined, though, that I would ever watch a video of a talking dog. Alright, Bunny the dog doesn’t really talk — she uses buttons with pre-programed phrases to communicate with her owner, Alexis Devine. Still, that’s pretty impressive. I found myself watching video after video, fascinated by Bunny’s apparent ability to express her desires and thoughts through the press of a button. At first, I took these videos at face value, and simply assumed that Bunny was just an extremely intelligent, communicative animal. However, after some more research into animal cognition, I’ve come to believe that Bunny’s behavior may have more to do with her training than her ability to actually ‘communicate’ in the way humans do.

Let’s take a closer look at how Bunny actually seems to express her thoughts. On her incredibly popular TikTok page, the black and white sheepadoodle mix can be seen discussing a variety of topics with her owner, including everyday observations about the weather outside, her mood, and the people she lives with. For example, in one video, Bunny lets her owner know that she wants to be let outside by pressing the buttons ‘Now’, ‘Want’, and ‘Outside’. Her descriptions get much more interesting and complicated than this: in another video, she hears a dog barking outside and presses the buttons ‘Stranger’ and ‘Dog’. When she wants to go on a walk on the beach, she presses ‘Please’, ‘Water’, and ‘Walk’. Her most interesting video may also be her most existential: Bunny looks into a mirror, pauses, and then walks up to her buttons to ask, ‘Who This?’1 This example demonstrates the possibility of deeper, more complex thinking not commonly observed in dogs. While her use of these buttons is undoubtedly impressive, most experts in animal cognition are skeptical that Bunny is actually communicating in the way that humans do. On a fundamental level, dogs communicate in a much different manner than humans — they mainly use pheromones, body language such as tail-wagging or tucking, and vocalizations such as barking and whimpering.

Surprisingly enough, the concept of a ‘talking’ animal isn’t completely new. In the early 1900s, a horse named Clever Hans emerged as an animal-celebrity of sorts in Berlin, Germany. His owner, Wilhelm von Osten, claimed that Hans could perform arithmetic, tell time, read, and even understand German. For example, when his owner asked him a certain question, Hans would provide an answer by tapping his foot. Hans traveled all over Germany, amazing crowds and drawing national attention.2

However, after Hans was put to the test in an environment where an official psychologist was present, he could no longer communicate in the way his owner claimed he could. The reason was this: rather than truly understanding and processing actual language, Hans was simply responding to cues from his owner. Instead of actually performing the mental tasks, the horse was watching the reactions of his owner and responding directly to involuntary cues in the owner’s body language. Hans didn’t actually know how to tell time — he was just acting in response to a certain stimulus, such as his owner’s facial expressions. As a result, Hans got the right answer 89% of the time when Von Osten knew the correct answer and the horse could see him, and when Von Osten didn’t know the answers, Hans got the correct answer only 6% of the time.5 The misinterpretation of an animal’s behavior as higher-level cognition is now aptly named the ‘Clever Hans Effect’. While Bunny may be a different species, her button-pressing behavior may be slightly similar to Hans’s hoof-tapping.

Animal communication researcher Federico Rossano acknowledges that it’s very possible Bunny could be pushing buttons by accident, leading to random presses that occasionally make sense to us. It’s also possible she may be memorizing certain combinations that produce positive responses from Devine, while not actually understanding what the sounds mean.4 There is an important difference between understanding what a word means in English versus simply having an association between a word and an action or reaction. As canine cognition expert Alexandra Horowitz puts it, “A lyrebird mimicking someone yelling ‘timber’ because they’ve heard someone yelling ‘timber’ is not themselves meaning to say ‘timber!’”3 Bunny has incredible perceptive abilities, but may not truly understand the actual meaning of the words she presses.

Even Bunny’s owner, Devine, says that she too is unsure of whether or not Bunny can truly ‘speak’:

“It’s really an emotional roller coaster for me…There are some days where I am so frustrated and I believe it’s all random and the skepticism in me overrides everything else. Then there are days when almost every single utterance is clean and concise and contextually appropriate and I’m like, ‘There’s definitely something going on here and it’s not random.”4

To say that Bunny understands English at a human-level is a stretch, and not a claim we can make with certainty just yet. While it would be interesting to continue studying the advanced abilities of animals like Bunny, I believe that there is also value in simply trying to understand the way dogs and other animals communicate naturally, rather than trying to make them adapt to our way of spoken language.

Sources:

1. @Whataboutbunny (https://www.tiktok.com/@whataboutbunny?lang=en)
2. https://www.ncbi.nlm.nih.gov/pmc/articles/PMC3921203/#:~:text=In%20the%20first%20decade%20of,in%20order%20to%20answer%20questions.
3. https://www.sciencefocus.com/nature/bunny-how-to-speak-dog/
4. https://www.vice.com/en/article/4ad4dm/can-bunny-the-talking-dog-really-talk
5. https://www.horsejournals.com/popular/history-heritage/clever-hans#:~:text=Clever%20Hans%2C%20an%20Orlov%20Trotter,%2C%20read%2C%20and%20understand%20German.

This article was written by Paru Nair, a freshman undergraduate student at UC Berkeley studying cognitive science and was edited by Jacob Marks, a senior undergraduate student at UC Berkeley and NT@B Publications Co-Lead.

What is there to find in your eyes?

Imagine a pair of glasses that could detect if you were looking at a far away object and would automatically zoom in for you; a keyboard that types as your eyes glance from word to word; or even a search tool that tells you about that thing you were looking at just a little too long. Futuristic devices like these that can monitor eye movements have a variety of applications ranging from diagnosing eye disorders to providing interactive gaming experiences. The key to designing such systems is gaze-tracking: measuring the eye movement in real-time with precision.

The most popular approach to gaze-tracking today is computer vision: where cameras capture pictures of the eye, and algorithms process the data in real time to determine how the eye is moving. Modern computer vision programs incorporate machine learning algorithms to detect patterns in images and efficiently classify objects. Despite the breakthroughs in this field, these algorithms suffer from substantial computational complexity. Training machine learning algorithms and deciphering images quickly requires high-performance chips which consume a significant amount of power. While feasible in a self-driving car or a smartphone, embedding a computer vision system in a pair of eyeglasses is very challenging.

In the Devices division of Neurotech@Berkeley, we focus on developing hardware to measure and analyze electrical signals generated by our bodies. Inspired by the potential of gaze-tracking systems, we set out to explore a technique with substantially lower computational cost than computer vision: electrooculography (EOG). Unlike computer vision that relies on cameras, EOG relies on intrinsic electrical signals generated in our eye.

What is EOG?

In 1951, Elwin Marg, an American neuroscientist and optometrist at UC Berkeley, discovered and defined electrooculography. EOG is a physiologic test that uses several carefully placed electrodes to measure the standing electrical potentials between Bruch’s membrane (at the back of the eye) and the cornea. Since a dipole is formed between the cornea and the retina, where the cornea has a positive charge relative to the retina, eye movement would change the electrical signal detected by the EOG electrodes. Now, imagine a scenario of two electrodes where one is placed just above the eyebrow and one is placed underneath the eye. As the subject looks up, the cornea moves closer to the top electrode and the Bruch’s membrane moves closer to the bottom electrode — the potential difference between the two electrodes increases! Likewise, as the subject glances down, the potential then decreases. Left and right eye movement can also be detected by placing electrodes to the left and right of the eye. Through such a simple mechanism, EOG can be leveraged to detect blinks, winks, and several other forms of eye movement.

Shortly after its discovery, EOG was primarily applied in the medical field in diagnosis of mental and sleep disorders, assistive technology for those who have sustained injuries, and more. More recently, it is being integrated into more accessible technologies such as applications in virtual reality (VR) and other consummerables for the general public. Notably, EOG data can be used for gaze-tracking. In fact, researchers have been able to detect eye movements as small as 1.5 degrees. Suddenly the gaze-tracking problem we described earlier just became significantly simpler. No longer do we require sophisticated computer vision algorithms.

Among the countless applications for an EOG-based eye-tracker, we think three stand out.

Looking at the Applications

An EOG headset would make it possible to measure signals in real time with high precision, allowing for a number of different applications in a variety of settings.

1. Tracking Consumer Behavior:

There are a number of possible applications for EOG tracking within the realm of ads and product marketing. Currently, a lot of product and marketing research is either done through focus groups or online surveys (sometimes served through ads that are targeted at consumers) but these research methods don’t provide immediate feedback on customers’ reactions to a product. For example, ad designers may want to see what part of an ad copy a potential customer’s eyes are drawn to first. They may want to test whether the design that they’ve created is coming across to consumers as intended. Through the use of an EOG headset they would be able to track in minute detail where customers’ eyes are and what they are drawn to at first glance. They can measure second by second what parts of an ad a consumer is focusing on. In this way, EOG headsets are potentially a cost-effective and scalable way to track where consumers’ visual attention is.

2. Typing Assistance:

On a consumer level, EOG signals can be applied to support keyboard use on a computer. An EOG aid that can bluetooth pair with a keyboard to predict what users are typing based on the textual context and their gaze could be useful to increase typing speed and productivity. The proof of concept for devices like this already exists, products like the Tobii Eye Tracker are already being used by gamers to increase click rate and to perform in-game actions faster than possible with a joystick. The applications extend beyond gaming and productivity. For those unable to type on a standard keyboard and who cannot utilize speech-to-text, on-screen keyboards paired with an EOG tracker can allow for faster communication.

3. Stepping into The Virtual Reality:

EOG could also play a crucial role in increasing the mass appeal of VR and the Metaverse. Today, VR remains a novelty. Clunky headsets, stiff controls, and countless other barriers to immersion have trapped the technology in an uncanny valley where VR environments do not feel convincingly real. By extension, these barriers to VR’s immersion also threaten the viability of the Metaverse as a convincing replacement to real 3D spaces. However, with the help of EOG eye-tracking in VR headsets, these virtual environments could begin to seem a whole lot more real. In today’s VR headsets, users have to move their heads to look around their virtual environments. Since eye movements go unregistered, natural saccadic eye scanning is forcibly replaced by often awkward head movements, contributing greatly to the clunkiness of VR. However, with EOG integrated headsets, participants could look naturally about their virtual environments by moving their eyes. Thus, EOG could both increase the ease

with which users can navigate VR, while also making it more immersive.

Designing Our Next Steps

When acquiring biomedical signals like EOG, there are two main challenges for hardware designers. First, these signals are extremely small in magnitude (on the order of microvolts) which makes them difficult to detect. Second, these signals are so small, these signals are susceptible to sources of electrical noise and interference. For instance, lights and appliances in a room are powered by 60 Hz 120V AC power. These strong power supply signals can “capacitively couple” into our circuit and overwhelm the EOG signal, making it impossible to decipher eye movements from noise.

The first part of our circuit consists of an instrumentation amplifier, which is responsible for rejecting sources of interference like the 60 Hz power supply. The next step is filtering. EOG signals are known to exist between about 0 and 50 Hz. Thus, we pass our EOG signal through filters that reject frequencies outside of this range. Finally, we pass our signal through an operational amplifier circuit. This circuit is responsible for providing a large gain to the signal, amplifying it to a level where it can be detected by a computer.

Additionally, we’ve created a 3D printed headset to mount the circuit and electrodes. We are now in the process of developing efficient classification techniques in Python to distinguish between blinks and other eye movements in our signals. With all of this progress, we are on our way to making our gaze-tracking device a reality.

References:

1. “The Computational Limits of Deep Learning” — Thompson, Greenewald, et al. 2020
2. https://www.automate.org/blogs/eye-tracking-technology-the-future-of-human-computer-interaction
3. https://venturebeat.com/2020/10/20/microsoft-researchers-design-software-based-eye-tracking-ai-that-works-on-any-device/
4. https://eyewiki.aao.org/Electrooculogram
5. https://www.nature.com/articles/s41699-018-0064-4#:~:text=EOG%20has%20various%20applications%20including,mental%20disorder%20diagnosis%2C%20and%20HMIs.&text=EOG%20along%20with%20electroencephalogram%20(EEG,progressive%20neuro%2Dmotor%20degenerative%20diseases.
6. https://www.cnet.com/tech/computing/watching-me-watching-you-how-eye-tracking-is-coming-to-vr-and-beyond/

This article was written by the Devices division of Neurotech@Berkeley. This article was edited by Publications Lead Jacob Marks and VP of Projects Ashwin Rammohan.