Human experience often involves a struggle with self-control. We might tell ourselves that just one more piece of chocolate, one more glass of wine, or one more episode of a captivating series before bed is acceptable, only to find ourselves continuing regardless. But who, or indeed what, is this “self” that engages in this internal conflict before succumbing to temptation? The cells within our gut collaborate with those in our brain and hands, leading us to reach for that chocolate bar, uncork the wine bottle, or click the “next episode” button. At some point, and with ever-increasing complexity, a threshold is crossed, and the collective entity becomes more than the sum of its individual parts. This emergent entity, the “self,” is the being that acts in the world in ways that align with your goals and desires.
However, consider the possibility that “selves” might exist within these fundamental cellular units, predating their merger into a larger, unified whole. While this concept may sound unconventional, biological simulations suggest that these minute life forms, typically perceived as passive mechanisms—cogs blindly governed by physical laws—possess their own objectives and exhibit autonomy. Surprisingly, even elementary networks of biomolecules appear to demonstrate a degree of selfhood, a discovery that could pave the way for novel therapeutic approaches to health conditions with significantly fewer adverse effects.
Furthermore, some biologists propose that this evolving understanding of selfhood may illuminate what distinguishes life itself and its initial emergence. Cognitive scientist Tom Froese, from the Okinawa Institute of Science and Technology in Japan, states, “The origins of agency coincide with the origins of life.”
Intelligent Agents
More technically, biologists and neuroscientists define selves as “agents” possessing goals and acting in ways to achieve them. Unlike entities simply carried along by their environment, agents actively modify themselves and their surroundings in purposeful ways. In essence, they exert causal influence over their own state and their environment.
To exhibit agency, an entity must absorb information, utilize that information to solve problems, and then learn from the outcomes of those actions. Neuroscientists generally refer to this as “cognition” and employ brain imaging techniques and behavioral experiments to investigate this interconnected set of processes. Traditionally, cognition has been attributed only to beings with brains. Theoretical biologist Emily Dolson at Michigan State University notes, “It’s easy to get caught up in the idea that brains are our first example of cognition, and a lot of people therefore think that brains must be special [in this respect].” This perspective, however, is increasingly being challenged.
A growing number of researchers are exploring other domains where these capabilities manifest. They are applying similar methodologies to much simpler organisms that lack brains in any conventional sense. In recent years, studies examining the behavior and signaling patterns, both electrical and chemical, of slime molds, plants, and even single-celled organisms have revealed remarkable aptitudes, including learning, memory formation, and adaptive decision-making in response to incoming information. The scope of cognition has even been extended to smaller systems within the human body. For instance, the immune system develops its own memory of which proteins are effective against harmful invaders, and cell clusters coordinate their actions to grow and repair the body autonomously. This indicates that both the immune system and these cellular collectives operate with varying degrees of agency.
The question then arises: how far down the complexity ladder can this concept be extended? Theoretical biologist Michael Levin and his colleagues at Tufts University in Massachusetts have recently applied cognitive frameworks to systems far simpler than even basic single-celled organisms—systems that most would classify as inanimate. Levin emphasizes, “You can’t just assume things have a certain level of agency. You have to do experiments and then you get surprises.”
Levin’s team investigated the gene regulatory networks (GRNs) found within every cell, which are crucial for determining the timing, location, and intensity of gene expression. These networks comprise genes, proteins, RNA, and other biomolecules that interact across numerous “nodes.” In the human body, a malfunction in a GRN, such as improper regulation of an essential protein, might prompt an intervention like gene therapy to alter its structure, akin to adding a new transistor to a faulty electrical circuit. This conventional approach treats these networks as passive machines requiring rewiring to alter their function.
Pavlov’s Dogs and Learning Networks
Levin and his collaborators explored an alternative method to modify GRN behavior: investigating whether it could actively “learn” features of its environment. Their inspiration came from a now-classic cognitive experiment pioneered by physiologist Ivan Pavlov in the 1890s. Pavlov repeatedly presented dogs with food following the sound of a ticking metronome. After several repetitions, the dogs learned to associate the ticking sound with the impending meal and began salivating at the metronome’s sound alone. This demonstrated that dogs process environmental information and use it to make predictions, a phenomenon known as associative learning.
Instead of dogs and metronomes, Levin and his team developed computer simulations of 29 different GRNs, data derived from biological sources. They trained each GRN to associate the presence of a neutral drug, which elicits no inherent response, with a functional drug that does affect it. This was achieved by repeatedly stimulating nodes within the network simultaneously. Through this process, they achieved the desired behavioral change in each GRN without the presence of the functional drug—analogous to a dog salivating at a metronome’s sound without food. Thus, their experiment established that GRNs possess learning capabilities, adapting their behavior in a manner that necessitates a form of memory. Levin commented, “These are examples of cognition, for sure. You’re not going to have a scintillating conversation with a GRN, but it’s something, it’s not zero.”
These findings hold potential for reducing the detrimental side effects associated with many medications, according to Levin. For example, opioids like morphine effectively manage chronic pain, but users often develop tolerance, necessitating higher doses that can lead to addiction and subsequent withdrawal risks. However, by manipulating the memory within biomolecular pathways, as Levin’s team did with GRNs, the build-up of tolerance could potentially be slowed or prevented. It might even become feasible to trigger the effects of potent medications with severe side effects, such as chemotherapy drugs, using a less harmful biomolecule instead. Despite these possibilities, no real-world medical treatments have yet been developed based on these computer model findings.
Beyond healthcare applications, the demonstration that computer models of GRNs can learn, much like Pavlov’s dogs, carries significant implications for our understanding of molecular network agency. Each GRN appears to function as an agent that directs the behavior of its chemical components towards collective objectives.
Levin and his team then investigated whether this induced associative learning in GRNs would enhance their collective behavior, effectively measuring their degree of “selfhood.” To further explore this, they employed a mathematical tool known as causal emergence. This concept was initially developed by neuroscientist Erik Hoel, also at Tufts University, in conjunction with the integrated information theory (IIT) of consciousness. IIT posits that the extent of the brain’s functioning as a unified whole can be quantified by a measure called phi, which also serves as an indicator of conscious awareness. If researchers can predict future brain states more accurately by viewing the brain holistically rather than focusing on its individual components, it is said to have a higher phi and exhibit greater causal emergence.
Setting aside the complexities of consciousness, causal emergence has evolved into a general method for determining when a complex system operates as an agent rather than a collection of disparate parts. Broadly, if the components act independently, phi is low; if they coalesce into collective patterns, phi is higher. Applying causal emergence metrics to GRNs, Levin and his team observed that after a GRN successfully learned to associate a neutral drug with a functional one, its phi value increased. The more the GRN learned, the greater these phi gains became, collectively suggesting the emergence of a new level of agency. Levin noted, “A lot of people will say, ‘Ah, you’ve taken these tools past their domain of applicability.’ But if you like the tools, let the science tell you where they work. If the tools are crap, you will find out pretty quickly.”
Evolution and the Origins of Life
Kevin Mitchell, a geneticist and neuroscientist at Trinity College Dublin in Ireland, finds such results noteworthy because agency is considered “a defining characteristic of life.” When a group of cells merges and acquires a new cognitive capacity, this enhanced skill allows it to exert influence from a higher level, compelling individual cells to subordinate their own interests for the sake of collective goals. Mitchell characterizes this as a form of “meta-control,” enabling agents to actively respond to their environments.
These findings not only shape our perception of who or what constitutes an agent but also propose that agency itself could be a driving force behind evolution. Dolson explains, “In the history of life, there are these major evolutionary transitions where what it means to be an individual changes.” For instance, simple prokaryotic cells were engulfed by others to form more complex eukaryotes, which then combined to create multicellular organisms. This inclination for parts to unite and form new levels of agency, according to Dolson, may be a significant mechanism contributing to the tendency of life to evolve towards greater complexity.
This concept is further supported by a subsequent study where Levin and his colleagues trained GRNs to learn and then “unlearn” a behavior by imposing new Pavlovian associations. This is akin to teaching dogs to associate a metronome with food and then teaching them to associate a bright light with food, leading them to forget the metronome association. The expectation was that once a learned behavior became obsolete, the agent would “release” that information, and the causal emergence increase attributed to that behavior would diminish. However, upon measuring the causal emergence of each GRN, Levin’s team discovered that it continued to rise, even after the original behavior had been forgotten. “If you are forced to lose that memory, you don’t lose your phi gains, which is astounding because it means there is an asymmetry to this, it becomes an intelligence ratchet,” Levin stated.
Rather than simply discarding information to forget a behavior, it appears that GRNs forget by learning the inverse of the original concept. Richard Watson, a complexity researcher at the University of Southampton in the UK, observes, “Now, instead of knowing nothing, you know that concept and its inverse.” Counterintuitively, teaching a GRN to forget results in a more sophisticated cognitive model, and its levels of agency and causal emergence continue to climb.
Neuroscientist Nikolay Kukushkin at New York University cautions against overemphasizing results derived from computer models of biological systems. “You can prove that something is possible in silico, but you can’t prove that is how it works [in real-world cells],” he remarks. Nevertheless, he finds the findings intriguing, suggesting that even though simulations are less complex than actual cells, valuable lessons can still be extracted.
Additionally, simulations of even simpler systems offer a more precise reflection of the real world and align with Levin’s ideas on agency and evolution. In 2022, complexity scientist Stuart Bartlett at the California Institute of Technology and David Louapre at Ubisoft Entertainment in Paris, France, found that basic “autocatalytic” chemical systems, which react with each other to self-replicate, could also learn through association. In autocatalysis, one chemical serves as fuel, while another is produced by consuming that fuel. The pair discovered that the reaction rate between these two chemicals is influenced by prior patterns in the concentration of available fuel—a behavior Bartlett characterizes as a “primitive form of learning.” This suggests that cognitive abilities may exist at an even lower level of molecular complexity than GRNs.
Bartlett chose to study autocatalysis because these simple chemical reactions mimic behaviors like self-replication in living systems. Self-replication and evolution are widely recognized as fundamental characteristics of life, leading some researchers to believe that autocatalysis could even shed light on the origin of life. However, Froese suggests that to fully comprehend this possibility, these chemical systems must be considered as agents acting with a degree of purpose, rather than mere collections of inanimate particles.
In this perspective, agency and cognition are best understood as a continuum, rather than as exclusive properties of highly complex life forms. Simple agents, at one end of this spectrum, learn from their environments, progressively acquiring more elaborate forms of agency—along with the capacity to control themselves, their components, and their surroundings.
However, Watson argues that while GRNs and autocatalytic chemicals may exhibit goals and rudimentary “thinking” abilities, concluding that they possess any form of inner mental world represents a leap too far. He states, “You don’t necessarily need to describe the parts as having beliefs, intentions or desires.” Levin, meanwhile, suggests that the strangeness of attributing characteristics typically associated with complex life forms like ourselves to simple systems should not deter exploration. Levin concludes, “All I’m saying is here is this bag of tools [from cognitive and behavioural science] I’m going to bring. I’m not interested in arguing with philosophers about this stuff. If you’ve got a better worldview that gets you to better discoveries, great, let’s roll.”