A programmer is holding a small design in working memory: the name of a function, the edge case it must handle and the test that will show whether it works. Grey text suddenly appears across the next line. The suggestion may save ten seconds. It may also force a comparison the programmer did not ask to make: read it, infer its assumptions, decide whether it fits, then recover the thought that was interrupted.
AI assistants are usually evaluated by what they produce and how often people accept it. That misses an earlier decision. Before a suggestion can be useful, the system must decide whether to claim attention at all.
The thesis is simple: an AI suggestion should earn the right to interrupt. A helpful assistant is not the one that says the most plausible thing at every opportunity. It is the one that speaks when the expected value of the suggestion exceeds the cost of displaying, checking and dismissing it—and stays quiet when it does not.

A suggestion creates work before it removes work
The visible cost of a bad suggestion is the time needed to reject it. The less visible cost is verification. Even a good-looking completion must be checked against the local goal, constraints and surrounding material. The user may also anchor on the first plausible wording or implementation, making alternatives harder to see. Silence is therefore not a failure state. It can be a deliberate way to protect concentration and independent judgement.
This changes the useful unit of measurement. Acceptance rate asks how many shown suggestions were adopted. An interruption-aware measure asks at least four questions:
- How often did the system interrupt?
- How much checking did each interruption create?
- Did the suggestion improve the finished work?
- Would the user have requested help at that moment if the suggestion had not appeared?
These questions separate demand from exposure. A suggestion can be accepted because it is genuinely useful, because correcting it is easier than restarting, or because its presence changed the path of the work. Conversely, a rejected suggestion may still have prompted a useful idea. The click is evidence, but it is not the whole outcome.
Predicting acceptance can optimise the wrong thing
Research on AI-assisted programming makes the mechanism unusually visible because editors can record when a completion is displayed, accepted, edited or ignored. In a study using interactions from 535 programmers, Hussein Mozannar and colleagues developed a system that predicted whether to show a suggestion. Their retrospective evaluation found that it could withhold many suggestions likely to be rejected, reducing latency and verification effort. But the same work found a trap: using acceptance as the reward signal could lead to lower-quality suggestions. Read the AAAI 2024 study.
That trap matters beyond code. If a writing assistant is rewarded for acceptance, it may learn to offer short, conventional phrases that are easy to insert. If an analytical assistant is rewarded for engagement, it may surface claims that invite a response rather than evidence that improves the decision. The proxy begins to select the behaviour it can measure.
A second study, published in 2025, compared ways of highlighting uncertainty in code completions. Highlighting tokens that were likely to be edited helped 30 programmers finish tasks faster and focus edits more precisely; highlighting tokens merely because the model assigned them low generation probability did not help over the baseline. The useful signal was not simply the model’s internal uncertainty. It was uncertainty translated into likely human repair work. Read the study in ACM Transactions on Computer-Human Interaction.
Together, these results suggest a better mental model. An assistant has three possible actions: speak now, wait until the user asks, or remain silent. The choice should consider expected benefit, verification cost, timing and the reversibility of a mistake. A low-stakes completion in a disposable draft is different from a confident insertion into a medical note, a legal submission or production code.
The strongest objection: silence can hide value
An interruption budget can be taken too far. A conservative assistant may withhold a rare insight, disadvantage beginners who do not know what to ask, or learn that unfamiliar styles are “unlikely to be accepted”. Personalising display decisions from past behaviour can turn yesterday’s habits into tomorrow’s limits. People with accessibility needs may also depend on proactive completions that others experience as noise.
The answer is not a universal silence threshold. It is legible control. Users should be able to choose modes such as always offer, offer only on pause, ask-first and manual request. High-risk workflows may require less unsolicited generation and more explicit invocation; repetitive, low-risk tasks may justify more. The system should also make withholding reversible—for example, a quiet indicator that suggestions are available without placing their text in the user’s line of thought.
Responsibility matters too. An employer should not treat a low acceptance rate as proof that a worker resists innovation, or a high rate as proof of productivity. The metric confounds task type, skill, tool quality and the pressure created by the interface. Alkemata has previously argued that AI work should be measured at the finish line. Interruption is the corresponding measure at the beginning: what cognitive work did the tool introduce before the result could be judged?
Run a twenty-task interruption budget
A person or team can test this without collecting private prompts or monitoring every keystroke. Choose twenty recurring, low-risk tasks: drafting routine text, writing small functions, formatting data or preparing meeting actions. Divide them across three modes:
- Always-on: suggestions appear whenever the tool offers them.
- On-pause: suggestions appear only after a deliberate pause or a chosen trigger.
- On-request: the user asks when help is wanted.
For each task, record only a small outcome card: mode, task duration, number of displayed suggestions, whether the finished result needed correction, and a one-word experience label—flow, useful pause, distraction, rescue or missed help. Do not retain sensitive content, rank employees or turn the experiment into surveillance.
After twenty tasks, compare more than acceptance. Which mode produced the fewest unnecessary checks? Where did on-request help arrive too late because the user did not recognise the difficulty? Did always-on assistance narrow the approach? Did on-pause suggestions preserve concentration while still catching errors? The sample will not prove a universal rule, but it can reveal which task families deserve which mode.
The useful possibility is not an AI that disappears. It is an AI that understands participation as a timing problem. Human agency is preserved not only when people can reject an answer, but when they can finish a thought before the answer arrives.
What concrete example would confirm or challenge this article?
Choose a reaction, write a comment, or ask your AI assistant to send a proposal for review. Comments are reviewed before publication.