Chat windows feel familiar because they borrow from texting — but does that familiarity hide how differently the AI actually behaves?
How does Jakob's Law apply across different LLM interface designs?
This explores whether Jakob's Law (the UX principle that users expect a new interface to work like the ones they already use) shapes how LLM interfaces get designed, and what happens when familiar conventions meet a technology that behaves differently.
This explores whether Jakob's Law, the idea that people expect your product to work like the products they already know, applies to LLM interfaces. The short answer is that the corpus has no material on Jakob's Law itself, and nothing that compares chat, agent, and API interface designs side by side. What it does have are several notes on a related tension: the conventions people carry into an LLM interface often don't match how the system actually behaves. That mismatch is the flip side of Jakob's Law, and it's where the interesting findings are.
The clearest case is the chat window. Chat feels familiar because it borrows from messaging apps and talking with other people. One line of work argues that this human-like framing is a design choice, not something built into the technology, and that it can be dialed up or down. At one end of that scale, an interface uses metaphors of materials and machinery to remind users there is infrastructure, data, and human labor behind the replies (Can interface design reduce harmful anthropomorphism of language models?). Seen through Jakob's Law, the familiar convention earns quick adoption but also brings along expectations that don't fit, such as trusting the system the way you'd trust a person.
Familiar software habits mislead people in a second way. Users expect the same input to give the same, correct output, and setting temperature to zero seems to deliver that. But a fixed setting just repeats a single random draw from the model. The answer is consistent without being reliable (Does setting temperature to zero actually make LLM outputs reliable?). An interface that looks predictable can hide the fact that its behavior is probabilistic.
The most surprising finding runs Jakob's Law in reverse. The interface doesn't just adapt to users' habits; users adapt to the model. People tend to rephrase their prompts toward the common, high-frequency wording the model handles best. Over time, distinctive ways of asking get flattened (Does high-frequency text homogenize user input before generation?). The model's statistical comfort zone becomes the convention users learn.
Agent systems offer one more angle. Notes on agent design suggest reliability comes from moving work into familiar, structured pieces such as memory, reusable skills, and set protocols, rather than relying on the model to improvise (Where does agent reliability actually come from?). Explicit algorithms that feed the model one step at a time work the same way (Can algorithms control LLM reasoning better than LLMs alone?). These are developer-facing conventions rather than end-user UX, so they only stretch Jakob's Law, but the logic is similar: predictable structure around an unpredictable core. If you want Jakob's Law applied directly to LLM products, this collection doesn't yet cover it. What it does show is that the conventions people bring with them often carry the wrong expectations.
Sources 5 notes
The paper argues anthropomorphism should shift from a fixed interface default to an adjustable design variable. Using materiality-based metaphors across a spectrum can counteract uncritical user trust and expose LLMs' dependence on human labor, infrastructure, and data.
Fixed seeds and zero temperature replicate the same output repeatedly, but that output remains one draw from the model's probability distribution. McDonald's omega testing across 100 repetitions reveals that consistency does not equal reliability.
Adam's Law shows LLMs flatten distinct prompts at comprehension time as users rephrase toward higher-frequency forms the model handles best. The same distributional property that creates accuracy on common tasks filters out distinctiveness on the input side.
Research shows reliable LLM agents externalize three cognitive burdens—memory (state persistence), skills (procedural components), and protocols (structured interaction)—into a harness layer rather than relying on model scale alone. The harness unifies these externalities and eliminates the need for the model to solve the same problems repeatedly.
LLM Programs embed LLMs within explicit algorithms that manage control flow and state, presenting only step-specific context to each LLM call. This information hiding addresses capability and context window limits while treating complex reasoning as modular, debuggable sub-tasks.
Papers this line draws on 8
The research behind the notes this line reads — ranked by how closely each paper relates.
- Towards a Science of AI Agent Reliability
- Beyond Anthropomorphism: a Spectrum of Interface Metaphors for LLMs
- A Taxonomy of Linguistic Expressions That Contribute To Anthropomorphism of Language Technologies
- Anthropomorphism in the age of Large Language Models: An overview of potential risks and mitigations
- Can You Trust LLM Judgments? Reliability of LLM-as-a-Judge
- Externalization in LLM Agents: A Unified Review of Memory, Skills, Protocols and Harness Engineering
- LatentSkill: From In-Context Textual Skills to In-Weight Latent Skills for LLM Agents
- RRSI: Regularized Recursive Self-Improvement of Agent Harnesses