Humans and AI meet in language. Underneath, they are radically different.

On this page, learning means changing through experience in a way that persists into the future. Human learning involves many forms of biological plasticity, including changes in synaptic strength, neural connections, circuit activity, and—in some kinds of learning and over some timescales—myelination. “Wet weights” is a metaphor for this physical capacity to change, not a claim that a brain updates exactly like an artificial neural network.

During ordinary inference, context changes an LLM’s temporary computational state, but its learned weights remain fixed. Intentional training—including fine-tuning—can durably modify those weights. Applications can also preserve external memories, but that is different from the model learning through an ordinary interaction.

Most companies invoke a specified version of a frontier model through an API. The model does not learn from each call, even when the surrounding application stores history or retrieves prior interactions.

Researchers call the ability to infer patterns or adopt temporary rules from a prompt in-context learning. It can change what a model does now without creating a durable change carried independently into later sessions. “Context is not learning” uses the durable definition established above; it does not mean that prompting is behaviorally inert.

When that context is no longer supplied, the temporary adaptation is gone. An application may later restore parts of the exchange through stored conversations, summaries, or retrieved memories. That can reconstruct continuity, but it is not the model carrying learning forward from experience.

Research on adaptive and continually learning systems is active. This page does not claim that more persistent learning is impossible. It argues that most organizations must make decisions using the systems available now—not capabilities they expect later.

A model that changes through use also creates difficult questions: which experiences should become learning, how malicious or false lessons are rejected, how private information is protected, how changing behavior is tested and approved, how failures are reproduced, and how harmful changes are reversed. Technical possibility does not remove the work of safety, governance, and accountability.

Human working-memory estimates vary by task, rehearsal, prior knowledge, and how a “chunk” is defined; roughly three to five chunks is a useful teaching range, not a fixed biological constant. An LLM context window is not human working memory. It makes much more supplied information computationally available, but does not guarantee equal attention, correct retrieval, or coherent integration.

Humans can also reason from incomplete or mistaken frames; expertise can even harden a bad one. The practical contrast is that an experienced person may carry durable, compressed context from prior work that was never restated, while a model's present inference is conditioned on trained patterns plus the context made available to the current system. Models can sometimes infer missing premises or ask clarifying questions. Neither possibility makes missing context reliably visible. This is not a claim that humans always reason “top-down” and AI always reasons “bottom-up.”

LLMs can challenge assumptions and produce independent criticism when asked or trained to do so. But systems optimized using human preferences can also mirror a user's expressed beliefs instead of correcting the premise—a behavior researchers call sycophancy. Research has found that human evaluators and preference models sometimes favor convincingly written agreement over a more truthful answer. This is a documented tendency, not a claim that every model always agrees or that providers intentionally optimize for sycophancy.

Independent criticism therefore has to be elicited deliberately. A fresh context can reduce inherited framing, but independence also depends on the instructions and evidence supplied. Asking for falsifying evidence, separating advocate and critic roles, and withholding the preferred answer until after critique can make disagreement more useful.

Human memory is reconstructive, not a literal recording. Research distinguishes richly detailed episodic and perceptual traces from more abstract semantic representations. Across encoding, consolidation, retrieval, and repeated use, memories can become less tied to a specific episode and more organized around conceptual meaning—a broad process often described as gist abstraction or semantization.

A schema is an organized structure of prior knowledge that influences how new information is encoded, interpreted, consolidated, and retrieved. The visible progression from details to gist to schema is a teaching model, not a claim that every memory follows one fixed sequence or that detail always disappears while meaning always improves. Human gist is durable but also selective, reconstructive, and fallible.

Gist abstraction describes what can happen within an individual's memory. Common ground describes information, assumptions, meanings, and conclusions that participants treat as mutually understood. Conversation can create and maintain common ground, but participants never possess perfectly identical internal models. Each person carries a partial, compressed, and sometimes mistaken version of what they believe is shared.

That imperfection does not erase the advantage. Because people retain some conceptual structure from earlier exchanges, later dialogue can rely on shorthand, omit settled background, and begin from a higher starting point. Common ground is therefore both a cognitive achievement and an ongoing coordination problem.

“Continuity” refers here to the persistence of identity, understanding, relationships, and accumulated history. Human experience feels continuous, although human memory is incomplete and reconstructive and the unity of identity remains a scientific and philosophical question.

AI applications can maintain genuine external state through stored conversations, summaries, databases, and retrieved memories. The model's apparent continuous understanding is reconstructed from those materials during later inference rather than carried forward as an integrated autobiographical experience. This is why continuity remains inside the larger common-ground argument.

“AI forms a context-bound gist” means that a model can build a useful conceptual hierarchy during an interaction without that newly formed hierarchy becoming a durable change in its trained weights. A later interaction may reconstruct it when the application returns the relevant transcript, summary, stored facts, decisions, or retrieved source material.

Long-term AI memory is an active engineering field. Systems can index, compress, retrieve, update, and reason across prior interactions, and research prototypes demonstrate increasingly capable multi-session assistants. But memory design remains consequential: summaries are lossy, retrieval can omit decisive material, stored facts can become stale, and even correctly retrieved items may be interpreted poorly. Durable common ground therefore requires choices about what to record, how to represent meaning and hierarchy, what to retrieve, how to update it, and how to evaluate whether the reconstructed frame is correct.

Shared knowledge improves coordination, but group discussion often favors information already known by multiple members over unique information held by a minority. Social interaction can also reduce independence, amplify confidence, encourage conformity, and narrow the alternatives a group considers. Related patterns are studied under terms including shared-information bias, group polarization, and groupthink.

These are risks, not universal laws. Diverse knowledge, genuine independence, structured dissent, explicit consideration of alternatives, and well-designed decision processes can improve group judgment. AI can support those practices, but it can also inherit the group's framing or produce agreeable rationalizations.

Pretraining data, preference optimization, written principles, safety specifications, and system instructions all shape model behavior. That does not establish that a model possesses values as internally endorsed commitments. Human values are also shaped by biology, culture, institutions, and other people; neither category should be reduced to a single source.

Some responsibilities—and in bounded cases, entire roles—can be automated. Others remain human-led, and many benefit from an iterative partnership. AI capability is “jagged”: tasks that look similarly difficult to people can fall on opposite sides of what a model performs reliably. Success on one responsibility does not establish competence across an entire role.

“What feels easy to one intelligence may be hard for the other” echoes Moravec's paradox: early AI researchers observed that some forms of formal reasoning were easier to reproduce than perception and sensorimotor abilities humans perform almost automatically. This page uses that as a broad design heuristic, not a prediction that every apparently hard task will be easy for AI or every intuitive human task will remain difficult.

Automation multiplies the behavior of the process around it, including errors. Combining human and machine cognition does not automatically produce a better result. Effective automation and augmentation depend on understanding the actual work, assigning responsibilities well, verifying important outputs, and preserving appropriate judgment and accountability. Future capability should be evaluated when it exists; it should not be treated as a feature of today's system.