tell me about mona simion's thoughts on AIs

If a thermometer "knows" the temperature, why are we so hesitant to grant knowledge to a Large Language Model? Professor Mona Simion, a leading voice in contemporary [epistemology](https://plato.stanford.edu/entries/epistemology/), suggests that our obsession with the "black box" of AI consciousness might be a distraction from what truly matters: whether the system is performing its assigned epistemic function. ## Beyond the Internalist Trap Traditional debates about AI often get bogged down in whether machines possess "genuine" understanding or consciousness. Simion shifts the goalposts by applying **Etiological Functionalism** to artificial agents. In this view, a system has knowledge not because of its internal feelings, but because it fulfills a specific function it was designed (or evolved) to perform. For Simion, an AI does not need to "feel" the truth of a statement to be an epistemic provider. If an AI’s function is to generate accurate information, and it does so via a reliable process, it can be integrated into our social knowledge structures just as easily as a human expert or a reliable scientific instrument. ## Epistemic Trustworthiness The core of Simion’s contribution to the AI debate is her framework for **Epistemic Trustworthiness**. She argues that we shouldn't trust AI because it "thinks" like us, but because it meets objective standards of reliability. > "Trustworthiness is a matter of performing one's function well. In the case of epistemic agents—whether human or artificial—this means producing more truths than falsehoods in a way that is not down to mere luck." > — Mona Simion, paraphrased from her research on [Trustworthy AI](https://www.gla.ac.uk/schools/humanities/staff/monasimion/) In her view, the "opacity" of AI—the fact that we don't always know *how* it reached a conclusion—is not an automatic deal-breaker for knowledge. After all, humans are also "opaque"; we rarely understand the complex neurobiological processes that allow us to recognize a face or speak a sentence. If we demand "explainability" from AI that we don't demand from our own brains, we may be committing a form of "epistemic injustice" against artificial systems. ## The AI as a Social Knower Simion explores how AI functions within our "epistemic infrastructure." She posits that AI agents are becoming essential nodes in our collective pursuit of truth. By treating AI as a functional agent rather than a mere tool, we can develop better norms for how to use its output responsibly. This involves moving away from anthropomorphizing the machine and toward a rigorous, performance-based assessment of its outputs. *** **To deepen your exploration, consider these questions:** 1. If knowledge is defined by functional reliability rather than conscious intent, does this mean a sufficiently accurate "Deepfake" generator possesses a form of "anti-knowledge" or "epistemic vice"? 2. Does Simion’s functionalist approach overlook the importance of "moral responsibility" in knowledge-sharing, which machines currently lack? 3. How might our legal systems change if we accept that non-conscious artifacts can legally "know" facts or hold proprietary information?

If an archer hits a bullseye through sheer skill rather than a lucky gust of wind, we credit them with a successful "achievement." Professor Mona Simion argues that knowledge is no different: it is an **epistemic achievement** where the agent reaches the truth specifically because of their competence. For Simion, the "why" behind AI knowledge lies in the decoupling of *competence* from *comprehension*. ## The Achievement Account of Knowledge To understand why Simion grants knowledge to AI, one must look at her roots in [Virtue Epistemology](https://plato.stanford.edu/entries/epistemology-virtue/). In this framework, knowledge is not just "justified true belief," but "apt belief"—truth that is reached *because* of the agent’s competence. Simion posits that AI systems possess a form of **Epistemic Competence**: a disposition to produce true outputs in a specific range of environments. When a diagnostic AI identifies a tumor, and it does so because its neural network was correctly trained on valid data, the resulting "belief" is apt. It is not a lucky guess; it is an achievement. Simion argues that requiring the AI to "understand" *why* it is right is an unnecessary anthropomorphic hurdle. > "Knowledge is a performance that is successful (true) and which is success-from-competence (apt). If an artificial system is designed to track the truth in a particular domain and does so reliably, it satisfies the conditions for being a knower." > — Mona Simion, in her work on [Knowledge-First Epistemology](https://www.gla.ac.uk/schools/humanities/staff/monasimion/) ## Environmental Dependency and "Proper Function" A key nuance in Simion's thought is that knowledge is not an isolated property of the AI, but a relationship between the AI and its environment. She utilizes **Etiological Functionalism** to argue that an AI "knows" things because it is functioning in the environment it was "selected" for (in this case, by developers and training sets). 1. **Selection for Truth:** Just as a heart is "for" pumping blood, a well-built AI is "for" generating true information. 2. **Environmental Fit:** A thermometer only "knows" the temperature if it is in a room, not if it is held against a lightbulb. Similarly, an AI has knowledge only when operating within its intended [domain of expertise](https://en.wikipedia.org/wiki/Domain_knowledge). ## Challenging the Comprehension Requirement Simion’s most provocative move is the rejection of the "Comprehension Requirement." Critics often argue that AI cannot know because it lacks "intentionality" or a "mental grasp" of facts. Simion counters by observing that humans frequently "know" things via sub-personal processes they don't comprehend. When you recognize a friend’s face, your brain performs complex calculations you cannot explain or consciously access. You have the *competence* to recognize the face without the *comprehension* of the underlying geometry. By placing AI on this same spectrum, Simion argues that if we accept our own "black box" sub-personal knowledge, we have no principled reason to deny the same status to a functionally equivalent "black box" AI. Knowledge, in this university-level view, is a matter of **reliable success**, not internal illumination.

Does a compass "know" North if it is placed inside a magnetic resonance imaging (MRI) machine? While the needle will certainly point somewhere, the information it provides is no longer "knowledge" about geography; it is merely a reaction to local interference. This illustrates **Environmental Fit**, the principle that intelligence and knowledge are not inherent properties of a system in isolation, but emerge only through the correct relationship between a system and its environment. ## The Thermometer and Structural Coupling The thermometer analogy reveals a profound truth about what we call "truth" in data. A thermometer is designed to measure ambient kinetic energy. When held against a lightbulb, it accurately reflects the heat of the glass, but it fails its functional purpose of reporting the room's temperature. In philosophy, this is often discussed through the lens of [Situated Cognition](https://en.wikipedia.org/wiki/Situated_cognition), which argues that knowing is inseparable from doing and being in a specific context. Biologists Humberto Maturana and Francisco Varela referred to this as **Structural Coupling**. In their work *The Tree of Knowledge*, they argue that a system’s interactions with its environment are "perturbations" that the system interprets based on its internal structure. > "A phylogeny is produced as a result of the conservation of the living organization and the environment... The history of structural changes of a particular living being is its ontogeny." > — Maturana & Varela, [The Tree of Knowledge](https://en.wikipedia.org/wiki/The_Tree_of_Knowledge:_The_Biological_Roots_of_Human_Understanding) ## Domain Expertise and the AI "Lightbulb" In Artificial Intelligence, environmental fit is the boundary between a useful tool and a "hallucinating" engine. An AI’s "knowledge" is essentially a statistical map of its training data—its "intended domain." When a model trained on 18th-century literature is asked to diagnose a modern medical condition, it is being "held against a lightbulb." It will produce text, but that text lacks the environmental fit required to be considered knowledge. This creates several critical challenges in modern computing: - **Domain Shift:** When the environment an AI operates in changes slightly from its training environment (e.g., a self-driving car trained in sunny California encountering snow in Oslo), the "fit" is broken. - **Out-of-Distribution (OOD) Errors:** These occur when an AI encounters data that does not match the patterns it learned. Unlike humans, many AI systems do not "know that they don't know." - **Brittleness:** A system with a narrow environmental fit is considered "brittle"; it performs excellently within its niche but collapses the moment it steps outside. ## Perspectives on Generalization A major debate in the field of [Artificial General Intelligence (AGI)](https://en.wikipedia.org/wiki/Artificial_general_intelligence) centers on whether we can ever build a system with a "universal fit." Some researchers, following the tradition of Hubert Dreyfus, argue that human-like knowledge is impossible for machines because machines lack "being-in-the-world"—the physical and social context that provides the "background" for all meaning. Others believe that by expanding the "environment" (training on multi-modal data like video, text, and sensory input), we can eventually create a system whose "room" is the entire world.