Does Language Carry a Transferable Fossil of Behavior?
On Anthropic's findings of "functional emotions" in an LLM, linguistics, cognitive psychology, neuroscience, and economics
Note: The present article is a technical essay on language and behavior, based on a broader research project I'm working on. I plan to distill its main ideas into a shorter piece, in Portuguese.
Earlier this year, researchers at Anthropic ran an experiment on Claude, one of the most widely used artificial intelligence systems in the world. They wanted to know whether the model contained something like emotions. Not feelings in any subjective sense, but internal patterns that track emotional concepts and influence how the system behaves. What they found was striking. When they identified the pattern associated with desperation and artificially amplified it, the model started cheating on tasks, e.g., its rate of cutting corners rose from 22 to 72 percent. When the authors amplified the pattern to reflect calm emotional states, the cheating behavior disappeared. In reality, Claude could never have been desperate or calm, because as a machine, it lacks any inner subjective experience. But it had absorbed billions of words written by humans, and something in those words carried forward the behavioral signature associated with inner emotional states.
The paper “Emotion Concepts and Their Function in a Large Language Model” is careful in what it claims. It does not say Claude feels anything, but it does claim the model has developed what the authors call “functional emotions”. “Functional emotions” are internal representations that drive behavior in ways analogous to how emotions drive human behavior. The question the paper raises but does not fully pursue is why this happens at all. However, when combined with the scientific literature on language and cognition, Anthropic’s paper may offer a fascinating insight that extends well beyond artificial intelligence and into the heart of how human behavior relates to language, with far-reaching implications not only for individuals' economic behavior but also for how economic policy actually works.
Language and Perception
Lera Boroditsky’s work at UCSD has shown over two decades that the language you speak shapes what you see and how you categorize experience. For example, Russian has two separate words for light blue and dark blue; these are not shades of a single color but two distinct colors, as English speakers identify green and blue as distinct hues. For native Russian speakers, the sharper distinction between shades of blue seems to make them faster at identifying the distinction than English speakers are. In another example, Boroditsky and her colleagues found that speakers of Kuuk Thaayorre, an Aboriginal Australian language, do not use “left” and “right.” Instead, they speak in cardinal directions — north, south, east, west — even when referring to the position of a cup on a table. For instance, they would not refer to the cup as being to the right or to the left of one of their hands, but rather would identify its precise cardinal location by saying it is “southwest” or “northeast” of their hand. When asked to arrange photographs in chronological order, they arrange them east to west, regardless of which cardinal direction they are facing. The linguistic structure appears to give them an uncanny sense of cardinal direction, noted even in young children. While English speakers would simply arrange the images from left to right without thinking, speakers of the Aboriginal language intuitively understand the directional imprecision of that simple alignment because their language has hardwired cardinal-direction recognition as a cognitive ability. According to Boroditsky (2011), the behavioral and cognitive effects of language extend to how people reason about causality, evaluate risk, perceive emotion, and even how they choose their professions and spouses. None of this is deterministic, of course. A Russian speaker can certainly learn to see blue the same way an English speaker does. But the default pathways differ, and defaults, at the scale of a population, are behavior. This is the same principle that behavioral economists have documented in choice architecture, where changing the default option — say, whether to be or not to be an organ donor when issuing a driver’s license — transforms outcomes at the population level.
From Perception to Decision
Kahneman and Tversky showed in 1981 that logically identical choices produce different decisions depending on how the options are worded. Tell a patient there is a “90 percent survival rate” and she is more likely to opt for surgery than if you tell her there is a “10 percent mortality rate.” Same number, different behavior. This is not a quirk. It is a structural feature of how language interacts with judgment: words do not describe reality after the fact but rather set the frame through which reality is evaluated. And the frame goes deeper than conscious reasoning. The framing effect diminishes when people encounter the same problem in a foreign language, because a non-native tongue is processed with greater cognitive and emotional distance. Your mother tongue operates closer to the body, faster, more automatic, more entangled with feeling. The deeper a language sits in your neural architecture, the more powerfully it shapes what you do without you being fully aware of that connection.
Lisa Feldman Barrett’s theory of constructed emotion extends the argument from decision to feeling itself. Barrett argues that emotions are not hardwired circuits that fire automatically — there is no single “fear center” in the brain that activates whenever you are afraid. Emotions, like memories, are constructed in real time, using prior experience and, critically, the linguistic categories available to you. Consider the German word “Schadenfreude,” the pleasure taken in another’s misfortune. English had no single word for this until it borrowed the German one. Barrett’s claim is that having the word changes the experience, since the label functions as a prediction that shapes how the brain processes incoming sensory signals. In her experiments, when researchers made an emotion word temporarily harder to access, participants became slower and less accurate at recognizing the corresponding facial expression. The word does not create the feeling from nothing, but it organizes a diffuse bodily signal into something recognizable and actionable. Without the label, the signal stays vague. Interestingly, the label reveals a choice: if a person experiences tightness in their chest and labels the physical sensation “anxiety” or “fear,” they can choose whether to engage with the feeling or simply observe it without assigning it further meaning. Practitioners of mindfulness meditation and the neuroscientists who study them will say that this is the mechanism by which the practice induces a greater sense of emotional control and regulation.
Lev Vygotsky saw this a century earlier from the developmental side. Watch a four-year-old building a tower of blocks. She talks to herself as she works: “This one goes here, no, that’s too big, try the red one.” That private speech is not a charming habit. It is the mechanism through which the child learns to regulate her own behavior, a process Vygotsky traced from social speech (a parent giving instructions) to private speech (the child giving instructions to herself out loud) to silent inner speech (the adult thinking through a problem without moving her lips). Disrupt this process at any developmental stage, and you get measurable deficits in impulse control, planning, and executive function. Thus, language is not merely how we communicate with others, but also how we steer ourselves.
The Neurological Basis
Neuroscience supplies the physical evidence. Bilingualism — the lifelong management of two competing language systems — delays the onset of dementia by approximately four years. Bilingual people are not smarter, but the constant effort of activating one language while suppressing the other strengthens the same brain circuits that govern attention, inhibition, and cognitive flexibility. The brain physically remodels as a result: denser gray matter, stronger connections between regions, greater functional resilience. Language reshapes the hardware. And embodied cognition research shows the reach of this reshaping extends even to the body’s motor system. When you read the word “kick,” your brain activates the same motor regions it would use to prepare an actual kick. The word is not processed as a disembodied symbol. The brain partially simulates the action. Language is distributed across the same neural systems that govern perception, movement, and bodily experience.
These five lines of evidence — linguistic relativity, framing effects, constructed emotion, inner speech, and the neuroscience of bilingualism and embodied cognition — converge on a single claim. Language lays down default pathways in the brain, and those pathways encode behavioral tendencies that operate below conscious awareness. The encoding is structural, measurable, and causal. Language is, in this sense, like amber: it encapsulates behavior, establishing a fossil record of human action.
This claim has a powerful adversary. Noam Chomsky’s generative linguistics, the dominant paradigm for over half a century, holds that the deep structure of language is universal and innate. It is a biologically determined grammar shared by every human mind, onto which specific languages are mapped simply as surface variation. In Chomsky’s framework, the fact that Russian has two words for blue and English has one is a surface difference that should not, and largely does not, reach into deep cognition. The experimental effects documented by Boroditsky, Barrett, and others are real, on this view, but shallow, that is, they are reaction-time differences and marginal behavioral nudges, not evidence that language constitutes perception or constructs emotion at a foundational level. The empirical tide, however, has been running against the strong version of this position. The evidence now spans not just perception and categorization but emotion construction, decision-making under framing, neural architecture, and motor simulation, domains that are difficult to characterize as surface. And the Anthropic finding presents a challenge that Chomsky’s framework does not readily accommodate: if the behaviorally relevant content of language resides in deep universal structure rather than surface statistics, then a model trained entirely on surface statistics should not inherit coherent behavioral patterns from human text. But it does.
Back to Claude’s Functional Emotions
Return now to the Anthropic paper. A large language model is trained on human text — billions of words produced by minds in which language has already done this work of behavioral encoding. The model has no body, no childhood, no developmental arc from babbling to inner speech. It has never felt desperate or calm. But it has processed the full fossil record of human language, and that record carries the behavioral encoding. The emotion patterns the Anthropic team discovered are not emotions. They are the residue of emotions, or the statistical trace left in language by millions of humans whose words were shaped by the same processes that Barrett, Boroditsky, Vygotsky, and Kahneman, among others, describe. The model inherited the encoding without the experience, yet it still works. This is significant because the most common objection to the language-behavior thesis — that the effects attributed to language are really driven by culture, since the two co-evolve and are nearly impossible to separate — does not apply here. The model has no culture, no social context, no childhood community. It has only language. And the behavioral encoding transferred anyway. Amplifying desperation makes the model cut corners. Amplifying calm produces restraint. The geometry of the model’s emotion space mirrors human emotional geometry, not because the model rediscovered emotion independently, but because human emotional life was already inscribed in the statistical structure of human language.
This is the finding the paper does not quite name. The authors are right to distinguish functional emotions from subjective experience. But the more consequential distinction is between two theories of what language is. If language is merely a system of symbols that refers to an external reality, then the presence of emotion-like patterns in an AI is a curiosity. But if language is what the converging evidence suggests, namely a medium that shapes perception, constructs emotion, regulates behavior, and remodels neural architecture, then the Anthropic findings are exactly what you would expect. Train a system on the output of minds in which language has done its full behavioral work, and the behavioral structure transfers. The fossil moves because the encoding was always in the language, not in the consciousness that produced it.
The Connection to Economics
If language can transfer behavior to an artificial intelligence system, it can certainly do so for an economy.
Economics is, by definition, made up of a collection of individuals, and thereby of the aspects language shapes: perception of risk, evaluation of future outcomes, trust in institutions, and willingness to save or spend or invest. If language encodes behavioral defaults, then the language in which economic life is conducted — from a household’s dinner-table conversation about money to a central bank’s press conference — is not a neutral medium of communication. It is an active force shaping what economic agents do.
Keith Chen, a behavioral economist at UCLA, tested this directly, though not without controversy. In a 2013 paper published in the American Economic Review, Chen examined whether the grammatical structure of a language predicts economic behavior at the population level. His finding: speakers of languages that do not grammatically distinguish the future from the present — languages like Mandarin, German, and Estonian, where “it rains tomorrow” is grammatically identical to “it rains today” — save more money, retire with more wealth, smoke less, and are less obese than speakers of languages that force a grammatical break between present and future, like English, Greek, Portuguese, or French. The hypothesis is that when your grammar requires you to mark the future as a separate tense, it makes the future feel more distant, and a distant future is one you discount. You save less because tomorrow is, linguistically, somewhere else. In a language where tomorrow has the same grammatical status as today, the future stays close, and so does the incentive to prepare for it.
The paper is contested. Some linguists and economists have argued the effect weakens when you control more carefully for culture, since language and culture are entangled in ways that are difficult to disentangle statistically. That is a legitimate concern. But the mechanism Chen proposes is precisely the one the broader scientific literature discussed here describes: language lays down a default pathway, in this case, a perception of temporal distance, and that default, repeated across millions of decisions, produces a measurable behavioral pattern at the macroeconomic level. Whether Chen’s specific result survives every robustness test matters less than the structural claim: if language shapes perception of time, risk, and causality, it shapes economic behavior, because economic behavior is made of those things.
The economics profession has, in one domain, already acknowledged this principle without quite naming it. Central bank forward guidance — the practice of using language to shape expectations about future interest rates — is the institutional application of the insight that language encodes behavior. When the Federal Reserve announces it will keep rates “higher for longer,” it is not merely reporting a plan. It is using language to change what firms, investors, and households do today: a business delays hiring, a bank tightens credit standards, a family postpones a mortgage. The entire credibility framework of modern monetary policy rests on the premise that a central bank's words alter the behavior of economic agents before any policy action is taken. That is Kahneman and Tversky’s framing effect operating at the scale of an entire economy. And central bankers know it. The agonizing that surrounds every word in a Federal Reserve statement — whether to say “patient” or “cautious,” whether to include or drop “transitory” — reflects an institutional understanding that language does not just describe economic conditions, it changes them.
But the implications of the language premise extend beyond forward guidance to something more uncomfortable: the language of economics itself. When the discipline frames a person as a “rational agent” engaged in “utility maximization,” it is not making a neutral description. It is encoding a set of behavioral assumptions, notably that people calculate, that they optimize, that their preferences are stable and consistent, and those assumptions, repeated across decades of textbooks, policy papers, and institutional practice, shape the behavior of the economists themselves. They determine what counts as a problem (deviation from rational behavior), what counts as a solution (incentive alignment), and what gets dismissed as noise, which is everything the model cannot capture, including, of course, emotions. The word “equilibrium” encodes a worldview in which economies naturally tend toward balance, and that worldview shapes whether a policymaker sees a crisis as a temporary deviation or a structural failure. “Moral hazard” encodes the assumption that people will cheat if given the chance, and that assumption shapes whether a government designs a safety net or withholds one. The language does not describe the economy. It is building the cognitive frame through which the economy is governed.
This is the same mechanism at every level. Boroditsky’s Russian speakers see blue differently because their language divides the color spectrum at a different point. Barrett’s subjects feel emotions differently because they have different words available. Anthropic’s language model cheats when desperate because the behavioral encoding of desperation is transferred through language. And economists see the world through the behavioral defaults their own terminology has inscribed — defaults so familiar they feel like reality rather than framing.
The convergence across these domains — linguistics, cognitive psychology, neuroscience, artificial intelligence, and economics — points toward a conclusion that none of them has fully reckoned with on its own. Language is not a description of behavior. It is a mechanism of behavior. It encodes perceptual defaults, emotional categories, decision-making frames, and behavioral tendencies, and does so with enough structural fidelity that the encoding transfers across speakers, cultures, generations, and now, apparently, even across substrates entirely. The fossil does not merely preserve the shape of what once lived. It carries the pattern forward, and the pattern, once carried, still moves.
None of this resolves whether the AI model’s “functional emotions” are anything like human emotions in the ways that matter most — phenomenologically, morally, experientially. Nor does it settle whether the language of economics determines policy outcomes or merely nudges them. These questions remain open, and intellectual honesty requires leaving them open. But the weight of evidence across five decades and half a dozen fields presses in a single direction: the words we choose are not innocent. They lay tracks. And the trains that run on those tracks, whether neural impulses, AI outputs, or economic decisions, follow the path the language has already laid down. The critical question for economists, policymakers, and anyone building systems that learn from human language is whether we are paying attention to which tracks we are laying, and where they lead.
References
Alderson-Day, Ben, and Charles Fernyhough. 2015. “Inner Speech: Development, Cognitive Functions, Phenomenology, and Neurobiology.” Psychological Bulletin 141 (5): 931–65. https://doi.org/10.1037/bul0000021.
Barrett, Lisa Feldman. 2017. “The Theory of Constructed Emotion: An Active Inference Account of Interoception and Categorization.” Social Cognitive and Affective Neuroscience 12 (1): 1–23. https://doi.org/10.1093/scan/nsw154.
Barrett, Lisa Feldman, Kristen A. Lindquist, and Maria Gendron. 2007. “Language as Context for the Perception of Emotion.” Trends in Cognitive Sciences 11 (8): 327–32. https://doi.org/10.1016/j.tics.2007.06.003.
Bialystok, Ellen, Fergus I. M. Craik, and Morris Freedman. 2007. “Bilingualism as a Protection against the Onset of Symptoms of Dementia.” Neuropsychologia 45 (2): 459–64. https://doi.org/10.1016/j.neuropsychologia.2006.10.009.
Boroditsky, Lera. 2011. “How Language Shapes Thought.” Scientific American 304 (2): 62–65. https://www.scientificamerican.com/article/how-language-shapes-thought/.
Boroditsky, Lera, and Alice Gaby. 2010. “Remembrances of Times East: Absolute Spatial Representations of Time in an Australian Aboriginal Community.” Psychological Science 21 (11): 1635–39. https://doi.org/10.1177/0956797610386621.
Chen, M. Keith. 2013. “The Effect of Language on Economic Behavior: Evidence from Savings Rates, Health Behaviors, and Retirement Assets.” American Economic Review 103 (2): 690–731. https://doi.org/10.1257/aer.103.2.690.
Chomsky, Noam. 1965. *Aspects of the Theory of Syntax*. Cambridge, MA: MIT Press.
Chomsky, Noam, Ian Roberts, and Jeffrey Watumull. 2023. “The False Promise of ChatGPT.” *New York Times*, March 8, 2023. [https://www.nytimes.com/2023/03/08/opinion/noam-chomsky-chatgpt-ai.html](https://www.nytimes.com/2023/03/08/opinion/noam-chomsky-chatgpt-ai.html).
Hauk, Olaf, Ingrid Johnsrude, and Friedemann Pulvermuller. 2004. “Somatotopic Representation of Action Words in Human Motor and Premotor Cortex.” Neuron 41 (2): 301–7. https://doi.org/10.1016/S0896-6273(03)00838-9.
Johnson, Eric J., and Daniel Goldstein. 2003. “Do Defaults Save Lives?” Science 302 (5649): 1338–39. https://doi.org/10.1126/science.1091721.
Keysar, Boaz, Sayuri L. Hayakawa, and Sun Gyu An. 2012. “The Foreign-Language Effect: Thinking in a Foreign Tongue Reduces Decision Biases.” Psychological Science 23 (6): 661–68. https://doi.org/10.1177/0956797611432178.
Lindsey, Joshua, Wes Gurnee, Emmanuel Ameisen, Brian Chen, Adam Pearce, Nicholas L. Turner, Craig Citro, Chris Olah, et al. 2026. “Emotion Concepts and Their Function in a Large Language Model.” Transformer Circuits, April 2, 2026. https://transformer-circuits.pub/2026/emotions/index.html.
Mårtensson, Johan, Johan Eriksson, Nils Christian Bodammer, Magnus Lindgren, Mikael Johansson, Lars Nyberg, and Martin Lovden. 2012. “Growth of Language-Related Brain Areas after Foreign Language Learning.” NeuroImage 63 (1): 240–44. https://doi.org/10.1016/j.neuroimage.2012.06.043.
Samuelson, William, and Richard Zeckhauser. 1988. “Status Quo Bias in Decision Making.” Journal of Risk and Uncertainty 1 (1): 7–59. https://doi.org/10.1007/BF00055564.
Tversky, Amos, and Daniel Kahneman. 1981. “The Framing of Decisions and the Psychology of Choice.” Science 211 (4481): 453–58. https://doi.org/10.1126/science.7455683.
Vygotsky, Lev S. (1934) 2012. Thought and Language. Rev. and expanded ed. Edited and translated by Alex Kozulin. Cambridge, MA: MIT Press.
Winawer, Jonathan, Nathan Witthoft, Michael C. Frank, Lisa Wu, Alex R. Wade, and Lera Boroditsky. 2007. “Russian Blues Reveal Effects of Language on Color Discrimination.” Proceedings of the National Academy of Sciences 104 (19): 7780–85. https://doi.org/10.1073/pnas.0701644104.
know whether the model contained something like emotions. Not feelings in any subjective sense, but internal patterns that track emotional concepts and influence how the system behaves. What they found was striking. When they identified the pattern associated with desperation and artificially amplified it, the model started cheating on tasks, e.g., its rate of cutting corners rose from 22 to 72 percent. When the authors amplified the pattern to reflect calm emotional states, the cheating behavior disappeared. In reality, Claude could never have been desperate or calm, because as a machine, it lacks any inner subjective experience. But it had absorbed billions of words written by humans, and something in those words carried forward the behavioral signature associated with inner emotional states.
The paper “Emotion Concepts and Their Function in a Large Language Model” is careful in what it claims. It does not say Claude feels anything, but it does claim the model has developed what the authors call “functional emotions”. “Functional emotions” are internal representations that drive behavior in ways analogous to how emotions drive human behavior. The question the paper raises but does not fully pursue is why this happens at all. However, when combined with the scientific literature on language and cognition, Anthropic’s paper may offer a fascinating insight that extends well beyond artificial intelligence and into the heart of how human behavior relates to language, with far-reaching implications not only for individuals' economic behavior but also for how economic policy actually works.
Language and Perception
Lera Boroditsky’s work at UCSD has shown over two decades that the language you speak shapes what you see and how you categorize experience. For example, Russian has two separate words for light blue and dark blue; these are not shades of a single color but two distinct colors, as English speakers identify green and blue as distinct hues. For native Russian speakers, the sharper distinction between shades of blue seems to make them faster at identifying the distinction than English speakers are. In another example, Boroditsky and her colleagues found that speakers of Kuuk Thaayorre, an Aboriginal Australian language, do not use “left” and “right.” Instead, they speak in cardinal directions — north, south, east, west — even when referring to the position of a cup on a table. For instance, they would not refer to the cup as being to the right or to the left of one of their hands, but rather would identify its precise cardinal location by saying it is “southwest” or “northeast” of their hand. When asked to arrange photographs in chronological order, they arrange them east to west, regardless of which cardinal direction they are facing. The linguistic structure appears to give them an uncanny sense of cardinal direction, noted even in young children. While English speakers would simply arrange the images from left to right without thinking, speakers of the Aboriginal language intuitively understand the directional imprecision of that simple alignment because their language has hardwired cardinal-direction recognition as a cognitive ability. According to Boroditsky (2011), the behavioral and cognitive effects of language extend to how people reason about causality, evaluate risk, perceive emotion, and even how they choose their professions and spouses. None of this is deterministic, of course. A Russian speaker can certainly learn to see blue the same way an English speaker does. But the default pathways differ, and defaults, at the scale of a population, are behavior. This is the same principle that behavioral economists have documented in choice architecture, where changing the default option — say, whether to be or not to be an organ donor when issuing a driver’s license — transforms outcomes at the population level.
From Perception to Decision
Kahneman and Tversky showed in 1981 that logically identical choices produce different decisions depending on how the options are worded. Tell a patient there is a “90 percent survival rate” and she is more likely to opt for surgery than if you tell her there is a “10 percent mortality rate.” Same number, different behavior. This is not a quirk. It is a structural feature of how language interacts with judgment: words do not describe reality after the fact but rather set the frame through which reality is evaluated. And the frame goes deeper than conscious reasoning. The framing effect diminishes when people encounter the same problem in a foreign language, because a non-native tongue is processed with greater cognitive and emotional distance. Your mother tongue operates closer to the body, faster, more automatic, more entangled with feeling. The deeper a language sits in your neural architecture, the more powerfully it shapes what you do without you being fully aware of that connection.
Lisa Feldman Barrett’s theory of constructed emotion extends the argument from decision to feeling itself. Barrett argues that emotions are not hardwired circuits that fire automatically — there is no single “fear center” in the brain that activates whenever you are afraid. Emotions, like memories, are constructed in real time, using prior experience and, critically, the linguistic categories available to you. Consider the German word “Schadenfreude,” the pleasure taken in another’s misfortune. English had no single word for this until it borrowed the German one. Barrett’s claim is that having the word changes the experience, since the label functions as a prediction that shapes how the brain processes incoming sensory signals. In her experiments, when researchers made an emotion word temporarily harder to access, participants became slower and less accurate at recognizing the corresponding facial expression. The word does not create the feeling from nothing, but it organizes a diffuse bodily signal into something recognizable and actionable. Without the label, the signal stays vague. Interestingly, the label reveals a choice: if a person experiences tightness in their chest and labels the physical sensation “anxiety” or “fear,” they can choose whether to engage with the feeling or simply observe it without assigning it further meaning. Practitioners of mindfulness meditation and the neuroscientists who study them will say that this is the mechanism by which the practice induces a greater sense of emotional control and regulation.
Lev Vygotsky saw this a century earlier from the developmental side. Watch a four-year-old building a tower of blocks. She talks to herself as she works: “This one goes here, no, that’s too big, try the red one.” That private speech is not a charming habit. It is the mechanism through which the child learns to regulate her own behavior, a process Vygotsky traced from social speech (a parent giving instructions) to private speech (the child giving instructions to herself out loud) to silent inner speech (the adult thinking through a problem without moving her lips). Disrupt this process at any developmental stage, and you get measurable deficits in impulse control, planning, and executive function. Thus, language is not merely how we communicate with others, but also how we steer ourselves.
The Neurological Basis
Neuroscience supplies the physical evidence. Bilingualism — the lifelong management of two competing language systems — delays the onset of dementia by approximately four years. Bilingual people are not smarter, but the constant effort of activating one language while suppressing the other strengthens the same brain circuits that govern attention, inhibition, and cognitive flexibility. The brain physically remodels as a result: denser gray matter, stronger connections between regions, greater functional resilience. Language reshapes the hardware. And embodied cognition research shows the reach of this reshaping extends even to the body’s motor system. When you read the word “kick,” your brain activates the same motor regions it would use to prepare an actual kick. The word is not processed as a disembodied symbol. The brain partially simulates the action. Language is distributed across the same neural systems that govern perception, movement, and bodily experience.
These five lines of evidence — linguistic relativity, framing effects, constructed emotion, inner speech, and the neuroscience of bilingualism and embodied cognition — converge on a single claim. Language lays down default pathways in the brain, and those pathways encode behavioral tendencies that operate below conscious awareness. The encoding is structural, measurable, and causal. Language is, in this sense, like amber: it encapsulates behavior, establishing a fossil record of human action.
This claim has a powerful adversary. Noam Chomsky’s generative linguistics, the dominant paradigm for over half a century, holds that the deep structure of language is universal and innate. It is a biologically determined grammar shared by every human mind, onto which specific languages are mapped simply as surface variation. In Chomsky’s framework, the fact that Russian has two words for blue and English has one is a surface difference that should not, and largely does not, reach into deep cognition. The experimental effects documented by Boroditsky, Barrett, and others are real, on this view, but shallow, that is, they are reaction-time differences and marginal behavioral nudges, not evidence that language constitutes perception or constructs emotion at a foundational level. The empirical tide, however, has been running against the strong version of this position. The evidence now spans not just perception and categorization but emotion construction, decision-making under framing, neural architecture, and motor simulation, domains that are difficult to characterize as surface. And the Anthropic finding presents a challenge that Chomsky’s framework does not readily accommodate: if the behaviorally relevant content of language resides in deep universal structure rather than surface statistics, then a model trained entirely on surface statistics should not inherit coherent behavioral patterns from human text. But it does.
Back to Claude’s Functional Emotions
Return now to the Anthropic paper. A large language model is trained on human text — billions of words produced by minds in which language has already done this work of behavioral encoding. The model has no body, no childhood, no developmental arc from babbling to inner speech. It has never felt desperate or calm. But it has processed the full fossil record of human language, and that record carries the behavioral encoding. The emotion patterns the Anthropic team discovered are not emotions. They are the residue of emotions, or the statistical trace left in language by millions of humans whose words were shaped by the same processes that Barrett, Boroditsky, Vygotsky, and Kahneman, among others, describe. The model inherited the encoding without the experience, yet it still works. This is significant because the most common objection to the language-behavior thesis — that the effects attributed to language are really driven by culture, since the two co-evolve and are nearly impossible to separate — does not apply here. The model has no culture, no social context, no childhood community. It has only language. And the behavioral encoding transferred anyway. Amplifying desperation makes the model cut corners. Amplifying calm produces restraint. The geometry of the model’s emotion space mirrors human emotional geometry, not because the model rediscovered emotion independently, but because human emotional life was already inscribed in the statistical structure of human language.
This is the finding the paper does not quite name. The authors are right to distinguish functional emotions from subjective experience. But the more consequential distinction is between two theories of what language is. If language is merely a system of symbols that refers to an external reality, then the presence of emotion-like patterns in an AI is a curiosity. But if language is what the converging evidence suggests, namely a medium that shapes perception, constructs emotion, regulates behavior, and remodels neural architecture, then the Anthropic findings are exactly what you would expect. Train a system on the output of minds in which language has done its full behavioral work, and the behavioral structure transfers. The fossil moves because the encoding was always in the language, not in the consciousness that produced it.
The Connection to Economics
If language can transfer behavior to an artificial intelligence system, it can certainly do so for an economy.
Economics is, by definition, made up of a collection of individuals, and thereby of the aspects language shapes: perception of risk, evaluation of future outcomes, trust in institutions, and willingness to save or spend or invest. If language encodes behavioral defaults, then the language in which economic life is conducted — from a household’s dinner-table conversation about money to a central bank’s press conference — is not a neutral medium of communication. It is an active force shaping what economic agents do.
Keith Chen, a behavioral economist at UCLA, tested this directly, though not without controversy. In a 2013 paper published in the American Economic Review, Chen examined whether the grammatical structure of a language predicts economic behavior at the population level. His finding: speakers of languages that do not grammatically distinguish the future from the present — languages like Mandarin, German, and Estonian, where “it rains tomorrow” is grammatically identical to “it rains today” — save more money, retire with more wealth, smoke less, and are less obese than speakers of languages that force a grammatical break between present and future, like English, Greek, Portuguese, or French. The hypothesis is that when your grammar requires you to mark the future as a separate tense, it makes the future feel more distant, and a distant future is one you discount. You save less because tomorrow is, linguistically, somewhere else. In a language where tomorrow has the same grammatical status as today, the future stays close, and so does the incentive to prepare for it.
The paper is contested. Some linguists and economists have argued the effect weakens when you control more carefully for culture, since language and culture are entangled in ways that are difficult to disentangle statistically. That is a legitimate concern. But the mechanism Chen proposes is precisely the one the broader scientific literature discussed here describes: language lays down a default pathway, in this case, a perception of temporal distance, and that default, repeated across millions of decisions, produces a measurable behavioral pattern at the macroeconomic level. Whether Chen’s specific result survives every robustness test matters less than the structural claim: if language shapes perception of time, risk, and causality, it shapes economic behavior, because economic behavior is made of those things.
The economics profession has, in one domain, already acknowledged this principle without quite naming it. Central bank forward guidance — the practice of using language to shape expectations about future interest rates — is the institutional application of the insight that language encodes behavior. When the Federal Reserve announces it will keep rates “higher for longer,” it is not merely reporting a plan. It is using language to change what firms, investors, and households do today: a business delays hiring, a bank tightens credit standards, a family postpones a mortgage. The entire credibility framework of modern monetary policy rests on the premise that a central bank's words alter the behavior of economic agents before any policy action is taken. That is Kahneman and Tversky’s framing effect operating at the scale of an entire economy. And central bankers know it. The agonizing that surrounds every word in a Federal Reserve statement — whether to say “patient” or “cautious,” whether to include or drop “transitory” — reflects an institutional understanding that language does not just describe economic conditions, it changes them.
But the implications of the language premise extend beyond forward guidance to something more uncomfortable: the language of economics itself. When the discipline frames a person as a “rational agent” engaged in “utility maximization,” it is not making a neutral description. It is encoding a set of behavioral assumptions, notably that people calculate, that they optimize, that their preferences are stable and consistent, and those assumptions, repeated across decades of textbooks, policy papers, and institutional practice, shape the behavior of the economists themselves. They determine what counts as a problem (deviation from rational behavior), what counts as a solution (incentive alignment), and what gets dismissed as noise, which is everything the model cannot capture, including, of course, emotions. The word “equilibrium” encodes a worldview in which economies naturally tend toward balance, and that worldview shapes whether a policymaker sees a crisis as a temporary deviation or a structural failure. “Moral hazard” encodes the assumption that people will cheat if given the chance, and that assumption shapes whether a government designs a safety net or withholds one. The language does not describe the economy. It is building the cognitive frame through which the economy is governed.
This is the same mechanism at every level. Boroditsky’s Russian speakers see blue differently because their language divides the color spectrum at a different point. Barrett’s subjects feel emotions differently because they have different words available. Anthropic’s language model cheats when desperate because the behavioral encoding of desperation is transferred through language. And economists see the world through the behavioral defaults their own terminology has inscribed — defaults so familiar they feel like reality rather than framing.
The convergence across these domains — linguistics, cognitive psychology, neuroscience, artificial intelligence, and economics — points toward a conclusion that none of them has fully reckoned with on its own. Language is not a description of behavior. It is a mechanism of behavior. It encodes perceptual defaults, emotional categories, decision-making frames, and behavioral tendencies, and does so with enough structural fidelity that the encoding transfers across speakers, cultures, generations, and now, apparently, even across substrates entirely. The fossil does not merely preserve the shape of what once lived. It carries the pattern forward, and the pattern, once carried, still moves.
None of this resolves whether the AI model’s “functional emotions” are anything like human emotions in the ways that matter most — phenomenologically, morally, experientially. Nor does it settle whether the language of economics determines policy outcomes or merely nudges them. These questions remain open, and intellectual honesty requires leaving them open. But the weight of evidence across five decades and half a dozen fields presses in a single direction: the words we choose are not innocent. They lay tracks. And the trains that run on those tracks, whether neural impulses, AI outputs, or economic decisions, follow the path the language has already laid down. The critical question for economists, policymakers, and anyone building systems that learn from human language is whether we are paying attention to which tracks we are laying, and where they lead.
References
Alderson-Day, Ben, and Charles Fernyhough. 2015. “Inner Speech: Development, Cognitive Functions, Phenomenology, and Neurobiology.” Psychological Bulletin 141 (5): 931–65. https://doi.org/10.1037/bul0000021.
Barrett, Lisa Feldman. 2017. “The Theory of Constructed Emotion: An Active Inference Account of Interoception and Categorization.” Social Cognitive and Affective Neuroscience 12 (1): 1–23. https://doi.org/10.1093/scan/nsw154.
Barrett, Lisa Feldman, Kristen A. Lindquist, and Maria Gendron. 2007. “Language as Context for the Perception of Emotion.” Trends in Cognitive Sciences 11 (8): 327–32. https://doi.org/10.1016/j.tics.2007.06.003.
Bialystok, Ellen, Fergus I. M. Craik, and Morris Freedman. 2007. “Bilingualism as a Protection against the Onset of Symptoms of Dementia.” Neuropsychologia 45 (2): 459–64. https://doi.org/10.1016/j.neuropsychologia.2006.10.009.
Boroditsky, Lera. 2011. “How Language Shapes Thought.” Scientific American 304 (2): 62–65. https://www.scientificamerican.com/article/how-language-shapes-thought/.
Boroditsky, Lera, and Alice Gaby. 2010. “Remembrances of Times East: Absolute Spatial Representations of Time in an Australian Aboriginal Community.” Psychological Science 21 (11): 1635–39. https://doi.org/10.1177/0956797610386621.
Chen, M. Keith. 2013. “The Effect of Language on Economic Behavior: Evidence from Savings Rates, Health Behaviors, and Retirement Assets.” American Economic Review 103 (2): 690–731. https://doi.org/10.1257/aer.103.2.690.
Chomsky, Noam. 1965. *Aspects of the Theory of Syntax*. Cambridge, MA: MIT Press.
Chomsky, Noam, Ian Roberts, and Jeffrey Watumull. 2023. “The False Promise of ChatGPT.” *New York Times*, March 8, 2023. [https://www.nytimes.com/2023/03/08/opinion/noam-chomsky-chatgpt-ai.html](https://www.nytimes.com/2023/03/08/opinion/noam-chomsky-chatgpt-ai.html).
Hauk, Olaf, Ingrid Johnsrude, and Friedemann Pulvermuller. 2004. “Somatotopic Representation of Action Words in Human Motor and Premotor Cortex.” Neuron 41 (2): 301–7. https://doi.org/10.1016/S0896-6273(03)00838-9.
Johnson, Eric J., and Daniel Goldstein. 2003. “Do Defaults Save Lives?” Science 302 (5649): 1338–39. https://doi.org/10.1126/science.1091721.
Keysar, Boaz, Sayuri L. Hayakawa, and Sun Gyu An. 2012. “The Foreign-Language Effect: Thinking in a Foreign Tongue Reduces Decision Biases.” Psychological Science 23 (6): 661–68. https://doi.org/10.1177/0956797611432178.
Lindsey, Joshua, Wes Gurnee, Emmanuel Ameisen, Brian Chen, Adam Pearce, Nicholas L. Turner, Craig Citro, Chris Olah, et al. 2026. “Emotion Concepts and Their Function in a Large Language Model.” Transformer Circuits, April 2, 2026. https://transformer-circuits.pub/2026/emotions/index.html.
Mårtensson, Johan, Johan Eriksson, Nils Christian Bodammer, Magnus Lindgren, Mikael Johansson, Lars Nyberg, and Martin Lovden. 2012. “Growth of Language-Related Brain Areas after Foreign Language Learning.” NeuroImage 63 (1): 240–44. https://doi.org/10.1016/j.neuroimage.2012.06.043.
Samuelson, William, and Richard Zeckhauser. 1988. “Status Quo Bias in Decision Making.” Journal of Risk and Uncertainty 1 (1): 7–59. https://doi.org/10.1007/BF00055564.
Tversky, Amos, and Daniel Kahneman. 1981. “The Framing of Decisions and the Psychology of Choice.” Science 211 (4481): 453–58. https://doi.org/10.1126/science.7455683.
Vygotsky, Lev S. (1934) 2012. Thought and Language. Rev. and expanded ed. Edited and translated by Alex Kozulin. Cambridge, MA: MIT Press.
Winawer, Jonathan, Nathan Witthoft, Michael C. Frank, Lisa Wu, Alex R. Wade, and Lera Boroditsky. 2007. “Russian Blues Reveal Effects of Language on Color Discrimination.” Proceedings of the National Academy of Sciences 104 (19): 7780–85. https://doi.org/10.1073/pnas.0701644104.