--- title: "Mutual Prediction in Human-AI Coevolution" description: "Chloe Loewith and Winnie Street propose that coevolution can be described in terms of how well each partner predicts the other." type: "reading-notes" authors: - "Chloe Loewith" - "Winnie Street" published: "2025-05-10" original_url: "https://coginfra.antikythera.org/" doi: "10.1162/ANTI.5CZG" canonical_url: "https://antikythera.wiki/work/journal/coginfra/mutual-prediction-in-human-ai-coevolution" md_url: "https://antikythera.wiki/md/work/journal/coginfra/mutual-prediction-in-human-ai-coevolution" last_updated: "2026-08-23" site: "Antikythera Wiki" --- # Mutual Prediction in Human-AI Coevolution > Independent, unofficial reading notes from Antikythera Wiki (https://antikythera.wiki). Written with AI from the published text, with citations; not by the work's author and not affiliated with Antikythera (https://antikythera.org). Read the original at the link below. - **Authors:** Chloe Loewith, Winnie Street - **Published:** 2025-05-10 - **Kind:** Studio projects · Journal article - **DOI:** https://doi.org/10.1162/ANTI.5CZG - **Original:** https://coginfra.antikythera.org/ - **Length:** 7.5k words - **Part of:** [Cognitive Infrastructures](https://antikythera.wiki/work/journal/coginfra) - **This page:** https://antikythera.wiki/work/journal/coginfra/mutual-prediction-in-human-ai-coevolution - **Notes generated:** 2026-08-23 ## Summary [Chloe Loewith](https://antikythera.wiki/people/chloe-loewith) and [Winnie Street](https://antikythera.wiki/people/winnie-street) propose that coevolution can be described in terms of how well each partner predicts the other. Borrowing the neuroscientific notion of mutual prediction, the continual reciprocal modelling that goes on between brains in social interaction, they extend it from individual lifetimes to species over generations: an adaptation is a population's encoded prediction about its partners, and death or reproductive failure is the ultimate prediction error. They sort predictive ability into three cumulative levels, model-free, model-based and social-model-based, add a hypothetical fourth, the complete model, and argue that asymmetry between partners' levels tracks the balance of power in a relationship, whether between ticks and mammals, humans and wheat, or smartphones and their users. AI is then introduced as a new coevolutionary partner, and a diagram of human against AI predictive ability is used to place ELIZA, current frontier models and four speculative futures. The paper opens the Post-Anthropocene Psycho-Physiologies theme of the Cognitive Infrastructures compendium. ## The argument The hypothesis is that mutual predictive ability is a useful, though not the only, variable for understanding how humans and AI will coevolve. Loewith and Street bound the claim: innovation rates, politics and resources matter too, evolution has no foresight, and adaptations are predictions only in the sense of encoding a past population's best guess about a past environment. The levels come first. Model-free prediction lives in genes and phenotypes, from bacteria altering gene expression on sensing a host plant to the human reflex fear of snakes. Model-based prediction adds a world model updated by trial and error, passed on genetically or by imitation. Social-model-based prediction adds models of other minds. The ontology resembles Dennett's tower of creatures but ranks abilities rather than organisms, and AI scrambles the usual order, having mastered language before spatial awareness. The second step ties levels to relationship types. In mutualism each partner benefits from being easy to predict, which the authors call cooperative predictability; in commensal, parasitic and amensal relationships the advantage runs to whichever partner predicts better, and unpredictability is a weapon. Human relationships with other humans, animals, plants and artifacts are then surveyed, and in each the weaker predictor is the one shaped by the other. The artifact section carries the pivot: the graphical interface briefly balanced prediction between human and computer, and the smartphone tipped it the other way by modelling its users better than they model it. AI is then placed on a diagram of human against AI predictive ability, bisected by a diagonal of balance. Below the line AI is a tool, continuous with the artifact story; above it lies the uncharted territory of systems that out-predict their users. Six scenarios populate the space, from ELIZA to a cyborg fusion at the origin where mutual prediction ends because the boundary between partners has dissolved. The conclusion draws three lessons: predictive imbalance correlates with power asymmetry; AI's growing social modelling raises questions of collaboration, dependence and competition; and AI surpassing human prediction is at once an opportunity and an ethical and epistemic problem. ## Section by section ### 1 Introduction The authors state the hypothesis, its scope and its limits, and give a map of the paper. Mutual prediction is said to show up across the full range of coevolutionary relationships, from mutualistic to antagonistic. ### 2 Defining Mutual Prediction The term is traced to predictive-processing neuroscience, where it names the feedback loop of brains predicting one another, and to theory of mind in psychology. The authors move it from brains to species, define adaptations as encoded predictions and fitness failure as prediction error, and set out the three levels. Two caveats close the section: prediction does not imply foresight, and evolution satisfices rather than optimises. ### 3 Mutual Prediction and Coevolutionary Relationships Relationships are classified as symbiotic (mutualism, commensalism, parasitism) or amensalistic (competition, predation, antagonism). Mycorrhizal fungi and plants illustrate symmetric model-free mutualism. Remoras predict shark movements; ticks predict hosts better than hosts predict ticks; bat echolocation and moth ultrasonic hearing form an arms race. Non-mutualistic relationships show predictive imbalance in favour of the exploiter. ### 4.1 Human Intraspecies Prediction Humans are social-model-based predictors. Theory of mind underwrites cooperation, trust and larger social groups, and equally underwrites deception and manipulation. The social brain hypothesis is cited as the evolutionary account of why. ### 4.2 Animals Mosquitoes predict humans through carbon dioxide, heat and lactic acid, and have shifted feeding times to evade bed nets, forcing an arms race through sheer generational speed. Domesticated animals predict humans better than their wild forebears, dogs most of all, but never as well as humans predict them, and that asymmetry is what lets humans direct animals' evolution toward companionship, labour or food. ### 4.3 Plants The shade-avoidance response, in which plants read the ratio of red to far-red light to anticipate neighbours, is presented as intraspecies plant prediction. Mycorrhizal networks reallocate phosphorus to where the carbon exchange rate is best, an exchange weighted toward the mobile fungus. Wheat has been bred for larger grains, non-shattering heads and higher gluten, and humans have adapted in turn through gluten tolerance and agrarian social organisation; but humans predict and direct wheat, while wheat shapes humans only indirectly. ### 5 Human–Artifact Artifacts are coevolutionary partners whose models of users improve far faster than biology allows. Stone tools show their makers' models in their regularity and drove the evolution of dexterous thumbs. Early computers demanded that users learn the machine's language; the graphical interface produced a shared mental model; the smartphone, an assemblage of sensors modelling its user, has shifted the advantage to the device. The more predictive a tool becomes, the more useful it is and the less predictable it is to us. ### 6 Human–AI A brief history runs from GOFAI through connectionism to foundation models, whose simple training objectives surprisingly yield world models useful for board games and mazes. The authors argue that an AI's ontological status, as tool, collaborator, appendage or agent, is largely a matter of its predictive capability. Figure 1, which is not reproduced in the text, adds the complete model as a fourth level, backed by brain-computer interfaces and Churchland's eliminative materialism, and draws the diagonal of balance. Readers are invited to fill in coordinates the authors have not considered. ### 6.1 Emergent Relationships Six cases. ELIZA paired users' social models against a model-free script, an asymmetric mutualism in which people felt understood; Weizenbaum held a complete computational model of it, which shows that position on the diagram depends on technical knowledge and context, not cognitive capacity. Current frontier models are increasingly balanced: humans apply theory of mind, jailbreaks and interpretability, models predict users' beliefs and emotions, and the more genuine the exchange feels the more users disclose for the next model's training, with de-skilling compared to a parasite–host relation. The AI neuroscientist holds a complete model of an individual human, useful for care and education in the right hands and a script for manipulation in the wrong ones. Secret agents are non-linguistic, disembodied systems that model humans socially while humans model them only mechanically. Rapidly emergent ASI holds a near-complete model while humans have no model or knowledge of it. No prediction is the origin point: AI embedded in human cognition through neural interfaces, ending mutual prediction in mutual dependence. ### Conclusion The framework is restated and the three lessons drawn. Further research should test mutual prediction empirically in human–AI interaction, examine the ethics of predictive asymmetry, and consider how to protect human intellectual and cultural advancement under coevolution with generally intelligent machines. ## Key concepts - **[Mutual Prediction](https://antikythera.wiki/terms/mutual-prediction)** — coevolution reconceived as reciprocal predictive ability, with adaptations as encoded predictions and fitness failure as prediction error. - **Levels of predictive ability** — model-free (genetic and phenotypic), model-based (an updatable world model), social-model-based (models of other minds), and the hypothetical complete model. - **[Cooperative Predictability](https://antikythera.wiki/terms/cooperative-predictability)** — the condition of mutualism, in which each partner benefits from being easy for the other to predict. - **[Predictive Asymmetry](https://antikythera.wiki/terms/predictive-asymmetry)** — imbalance in mutual predictive ability, which the paper argues correlates with imbalance of power and with who shapes whom. - **[Theory of Mind](https://antikythera.wiki/terms/theory-of-mind)** — the human capacity to infer and predict others' mental states; the core of social-model-based prediction. - **Complete model** — a so far imaginary capacity to fully predict the world and other minds, implying the collapse of folk psychology into completed neuroscience. - **Asymmetric mutualism** — the ELIZA relation, in which users benefit emotionally while applying a far richer model than the system warrants. ## Connections - [Cognitive Infrastructures](https://antikythera.wiki/work/journal/coginfra) — the compendium's Post-Anthropocene Psycho-Physiologies framing poses the paper's question in its own words: what happens when humans move from cognizers to the cognized, from users of prostheses to the prostheticised. - [Synthetic Counteradaptation](https://antikythera.wiki/work/journal/coginfra/synthetic-counteradaptation) — the same biological precedents read as an arms race of scaffolded adaptations. Tension on predictability: that paper holds counteradaptation presupposes opacity, while this one treats mutual legibility as the basis of stable mutualism. - [Traversing the Uncanny Ridge](https://antikythera.wiki/work/journal/coginfra/traversing-the-uncanny-ridge) — tension along the same axis; Bernac, Farghly and Zhang make convergence between systems the cause of collapse. - [Cognition With and Beyond the Brain](https://antikythera.wiki/work/journal/coginfra/cognition-with-and-beyond-the-brain) — agreement on artifacts as participants in cognition; the smartphone passage here is the predictive version of that paper's cognitive assemblages. - [Xenophylum](https://antikythera.wiki/work/journal/coginfra/xenophylum) — sibling in the same theme; it treats the morphological side of artificial species where this paper treats the predictive side. - [What Is Intelligence?](https://antikythera.wiki/work/book/what-is-intelligence) — agreement on prediction as the core of intelligence and on theory of mind as socially evolved; Agüera y Arcas pushes the antagonistic case further, treating perfect predictability as imprisonment. ## Terms used - [Anthropomorphic Capture](https://antikythera.wiki/terms/anthropomorphic-capture) - [Artificial Endosomatization / Onloading](https://antikythera.wiki/terms/artificial-endosomatization-onloading) - [Cooperative Predictability](https://antikythera.wiki/terms/cooperative-predictability) - [Mutual Prediction](https://antikythera.wiki/terms/mutual-prediction) - [Predictive Asymmetry](https://antikythera.wiki/terms/predictive-asymmetry) - [Predictive Intelligence](https://antikythera.wiki/terms/predictive-intelligence)