When an AI acts confusing, people often quietly patch the gap themselves — doing work the interface should have done.

This explores the tactics people fall back on when an AI interface doesn't work the way they assume it will. First, an honest note: the collection has very little that documents workarounds directly, meaning observed studies of what users improvise. What it does well is explain why the mismatch happens. Read that way, many of the 'workarounds' turn out to be users doing work the interface should be doing for them.

The main source of the mismatch is that chat interfaces trigger communication skills people have been building since childhood: taking turns, assuming shared context, trusting that the other side will ask when it's confused. The AI doesn't actually communicate in that sense, so when things go wrong it feels like user error even though the cause is in the design Why do users fail with AI interfaces designed like conversations?. The most common workaround follows from this. Users learn to over-specify everything up front, packing context, constraints and examples into one prompt, because they've found out the system won't probe. That's hard, because people often don't know what they want until they see options. Intent forms gradually through back-and-forth rather than arriving complete How do users actually form intent when prompting AI systems?, and the 'gulf of envisioning' is the gap between a vague goal and a prompt that captures it Why can't users articulate what they want from AI?. Iterating by trial and error (prompt, inspect, rephrase, retry) is the workaround users improvise. The research suggests the better fix is for the system to offer options, so the user's job becomes picking from a list instead of describing something from scratch.

A second kind of mismatch is about memory and state. With traditional software, people build a stable picture of how it works: this button always does that. AI context is different. It shifts constantly through prompt history, retrieved data and hidden state, so users can't learn it the same way How does AI context differ from conventional software context?. This explains familiar habits like restarting a conversation when it 'goes weird', copying key instructions into every message, and keeping a personal file of prompts that worked. These are users doing context management by hand. A related problem is that the model has no record of what it doesn't know about the user, so it fills the gaps with guesses and agreeable answers. Explicitly listing the unknowns in the prompt cut hallucination roughly in half Do language models know what they don't know about users?. That's essentially a formal version of the user habit of saying 'don't assume X.'

Sometimes the workaround is to stop chatting altogether. For structured, goal-directed tasks, one argument holds that chat is a step backward: it brings back the layer of describing actions in words, which direct manipulation (clicking, dragging, editing in place) was designed to remove Is chat the wrong default interface for AI products?. People who copy AI output into a spreadsheet or editor to finish the job there are voting with their behavior. A gentler design option is to make the AI's human-like persona adjustable, so users aren't pulled into treating it as a conversation partner when a tool would serve them better Can interface design reduce harmful anthropomorphism of language models?.

The part you might not expect: workarounds can hide their own cost. When users quietly compensate for a mismatched interface and get a polished result, they often credit themselves with more skill than the process involved. Fluent output, unclear authorship and an opaque pipeline reinforce one another to inflate people's sense of their own competence How do AI tools trick users into overestimating their own skills?. So a workaround that seems to succeed can also hide how much the tool shaped the outcome. That's a reason to treat workarounds as evidence of design gaps rather than as user expertise.