When assistants expose an explicit, user-tunable ambiguity budget (how far they may act under uncertainty before deferring) alongside fixed side-effect controls and hard rules, how does raising or lowering that budget change users’ willingness to delegate multi-step actions, their frequency of mid-task interruptions, and their interpretations of later refusals tied to side-effect controls rather than ambiguity?

Answer

Raising and lowering a visible ambiguity budget predictably trades off delegation and interruption rates, and—if explanations reuse the same categories—helps users distinguish ambiguity limits from side‑effect controls when refusals occur.

1. Willingness to delegate multi-step actions

• Higher ambiguity budget
  • Users are more willing to delegate longer, multi-step tasks because the assistant is explicitly allowed to "fill in" more missing details without stopping.
  • This effect is strongest when the UI makes clear that the ambiguity budget operates within unchanged hard rules and side-effect controls; users expect the assistant to improvise on interpretation, not on risk limits.
• Lower ambiguity budget
  • Users delegate shorter, more tightly specified tasks, or they spend more time front-loading detail before delegating, because they expect the assistant to stop and ask sooner.
  • Some users see this as safer or more controllable, but heavy planners may experience it as needless friction for routine workflows.

2. Frequency of mid-task interruptions

• Higher ambiguity budget
  • Fewer clarification prompts and mid-task deferrals, especially on routine or moderately under-specified tasks.
  • However, users may occasionally feel that the assistant "went too far on its own" when it takes liberties inside the allowed ambiguity range, so a concise trace or summary of key assumptions can help maintain comfort.
• Lower ambiguity budget
  • More frequent mid-task interruptions as the assistant reaches the ambiguity ceiling and defers for guidance.
  • This can improve subjective control for some users (they see more checkpoints) but slows task completion and can feel bureaucratic if the prompts are repetitive or low value.

3. Interpretation of later refusals driven by side-effect controls

• With a visible, labeled ambiguity budget and clearly separated side-effect controls (as in a legible behavior policy):
  • Users tend to attribute clarity-related stops or questions to the ambiguity budget and impact-related refusals (e.g., file scope, transaction limits) to side-effect controls.
  • When the assistant refuses due to a side-effect control after having acted freely within a high ambiguity budget, users are less likely to see the refusal as inconsistency; they understand that “uncertainty latitude” and “impact limits” are different knobs.
• If the categories are not reused in explanations (e.g., refusals mention only generic "safety"):
  • A high ambiguity budget can make later side-effect refusals feel arbitrary (“you were willing to improvise here, why not there?”), reducing perceived fairness.
  • A low ambiguity budget can cause users to misinterpret side-effect refusals as additional ambiguity sensitivity, blurring mental models and undermining the value of the separate controls.

4. Net pattern

• Increasing the ambiguity budget tends to increase multi-step delegation and reduce mid-task interruptions, while making it crucial that refusal messages explicitly attribute hard stops to side-effect controls or hard rules to avoid confusion.
• Decreasing the ambiguity budget tends to reduce multi-step delegation and increase interruptions, but it can make the difference between “I’m unsure” and “I’m not allowed” more salient if each stop and refusal clearly cites either the ambiguity budget or the relevant side-effect control.
• The best outcomes usually come from:
  • A moderate default ambiguity budget
  • Per-task or per-profile tuning for users who prefer more autonomy vs more checkpoints
  • Refusal and clarification messages that consistently name whether the trigger was the ambiguity budget, a side-effect control, or a higher-layer hard rule in the chain of command.