Why the domain fits
Active inference casts any system that maintains its integrity as minimizing variational free energy relative to a generative model of its environment, and linguistic communication maps onto this directly: speakers select utterances expected to change a listener's beliefs, while listeners parse speech to infer hidden intentions and meanings, both processes framed as free-energy minimization. The formalism's four-way partition of internal, external, sensory, and active states via Markov blankets separates an agent's beliefs from the communicative situation, with heard speech as sensory states and articulation as active states. Language's inherent hierarchy — phonemes into syllables, words into phrases, phrases into discourse — also aligns with active inference's nested generative models, where higher levels encode slow abstract causes (intentions, topics) and lower levels encode fast sensory detail (acoustic realization).
State of the literature
The foundational contribution is the 2020 work on generative models, linguistic communication, and active inference (Friston and colleagues), which simulates dialogue between synthetic agents that infer each other's hidden semantic states and select utterances to minimize expected free energy, producing emergent shared lexical conventions. A companion hermeneutic account (Paulus and Friston) frames communication as mutual interpretation, where action fulfills predictions about one's own future speech and its effect on the listener. The literature then branches into speech motor control and an active-inference account of stuttering (disfluency as failure to smoothly minimize free energy across motor and linguistic levels), a preprint arguing natural language syntax itself complies with the free-energy principle, an essay reframing inner speech as a translator between messy sensory states and compact verbal labels, and a Frontiers in Psychology call to bring active inference into neuropragmatics. Running through all of this is an unresolved tension with predictive-coding and Bayesian-brain theories: a 2023 critique argues neural systems may prioritize feature discovery over exhaustive prediction, and a separate critical review challenges whether the brain implements literal Bayesian computation at all.
Key projects and tools
Concrete implementations are still largely research prototypes rather than named public tools: synthetic linguistic agents are built as probabilistic graphical or state-space models with message-passing or gradient-descent inference (typically MATLAB or Python), and speech motor control models are envisioned as dynamic Bayesian networks over articulatory states. Dynamic causal modeling (DCM), already used in the active inference community to model brain connectivity, is proposed as extendable to neuropragmatics experiments alongside standard neuroimaging pipelines (SPM, FSL). On the AI side, three distinct 2025-era projects integrate active inference with LLMs: a review architecture using an outer active-inference controller to manage an LLM's prompts and interactions, a Nature Digital Medicine system that selects prompts via expected free energy to improve response reliability, and an arXiv preprint on a multi-LLM system where a hierarchical controller assigns tasks across specialized LLMs to minimize free energy relative to user goals. Research is loosely associated with Friston-affiliated groups (e.g., University College London) and the LISN Lab, which hosts the inner-speech essay.
Open problems
Direct empirical evidence that human language processing actually follows active inference principles remains limited: many of its predictions (seeking information under uncertainty, acting on preferences) are qualitatively shared with reinforcement learning and other frameworks, so tests need to target distinctive features like the epistemic-pragmatic decomposition of expected free energy. The Bayesian-brain critique specifically warns against reading generic prediction-error signals in neuroimaging as proof of full Bayesian computation in language circuits. Multiscale integration is largely unaddressed in practice — most models cover a narrow band (word-meaning mappings or speech motor control) rather than spanning neural circuits through discourse to language change — and clinical translation (stuttering, apraxia, aphasia, psychosis-linked inner-speech disruption) has only an initial worked example (stuttering) with the rest still hypothetical. For LLM-based systems, open questions include how to formalize tractable generative models of user state, how to define expected free energy functionals capturing alignment and reliability, and how to evaluate interactive agents since standard NLP benchmarks do not capture dynamic active-inference-driven dialogue.
Reference Backbone
Karl J. Friston (2010). The free-energy principle: a unified brain theory? Nature Reviews Neuroscience. DOI: 10.1038/nrn2787. Thomas Parr, Giovanni Pezzulo, Karl J. Friston (2022). Active Inference: The Free Energy Principle in Mind, Brain, and Behavior. MIT Press. Giovanni Pezzulo, Francesco Rigoli, Karl J. Friston (2017). Active Inference, homeostatic regulation and adaptive behavioural control. Progress in Neurobiology. DOI: 10.1016/j.pneurobio.2017.08.001. Maxwell J. D. Ramstead, Karl J. Friston, Axel Constant, Lancelot Da Costa, Casper Hesp, Beren Millidge, Alexander Tschantz (2023). On Bayesian Mechanics: A Physics of and by Beliefs. arXiv.