Design for uncertainty
User-Centered Design
Design Systems
Human-Centered AI Design

Design for uncertainty

Expose uncertainty so users understand confidence, risk, and limitations of AI outputs.

Transparency
Human-Centered AI Design

AI systems operate with inherent uncertainty. They make predictions, classifications, and recommendations based on probabilities, not certainties.

Designing for uncertainty means explicitly acknowledging that AI systems can be wrong, incomplete, or ambiguous, and shaping interactions so users are not misled into false confidence.

Uncertainty is not a flaw to hide, but a property to surface and manage.

A human-centered approach avoids binary outcomes (“right/wrong”) and instead communicates confidence levels, assumptions, and limits, enabling users to make informed decisions and apply appropriate judgment.

Users should understand:

  • how confident the AI is in its outputs,
  • what factors might affect accuracy,
  • and what the consequences of relying on uncertain information might be.

Hiding uncertainty may make the system feel more confident, but it leads to:

  • over-reliance on potentially incorrect outputs,
  • poor decision-making when confidence is low,
  • and loss of trust when users discover the system was uncertain.

By exposing uncertainty appropriately, I help users:

  • calibrate their trust in the system,
  • make better decisions based on confidence levels,
  • and understand when human judgment is needed.

Communicate Assumptions and Data Gaps

Explain what data the AI used, what assumptions it made, and what information might be missing that affects its outputs.

Usage scenarios:

  • When the outcome influences operational decisions
  • When uncertainty is significant
  • When the output is not exploratory
off on
Explain what data sources the AI used to generate outputs. Present outputs without explaining underlying assumptions or data limitations. Highlight when important information is missing or unavailable. Make recommendations appear more certain than the data supports.

Design for Ambiguous or Multiple Outcomes

When the AI is uncertain or multiple interpretations are possible, present all reasonable options rather than forcing a single answer.

Usage scenarios:

  • Recommendations (movies, music, products)
  • Weather forecasts with scenarios
  • Recipe suggestions based on ingredients
  • Traffic estimation with alternative routes
off on
Present multiple reasonable interpretations when the AI is uncertain. Force a single answer when multiple outcomes are equally plausible. Show confidence levels for each possible outcome or interpretation. Hide ambiguity by presenting only the most likely outcome.

Encourage Verification and Human Judgment

When uncertainty is high, explicitly prompt users to verify outputs and use their judgment rather than blindly accepting AI recommendations.

Usage scenarios:

  • Hiring or firing
  • Granting credit
  • Access to public services
  • Diagnosis or healthcare priorities
  • Sanctions or controls

Here, human verification cannot be optional; it is a governance measure.

Prompt users to verify outputs when low or moderate confidence directly impacts people. Present uncertain outputs as if they're definitive and don't need verification. Provide easy ways to request human review or additional verification. Make it difficult for users to question or override AI outputs.

Make Uncertainty Visible

Show confidence levels, probability scores, or uncertainty indicators so users understand how certain the AI is about its outputs.

Usage scenarios:

  • When the output is probabilistic
  • When the error has significant consequences
  • When the model works on noisy data
  • When the system is perceived as reliable
Display confidence scores or probability indicators for AI outputs. Present uncertain outputs with the same visual weight as high-confidence ones. Use visual indicators (colors, badges, progress bars) to communicate confidence levels. Hide uncertainty behind technical jargon or bury it in fine print.

Why this principle matters

Uncertainty is a fundamental characteristic of AI systems. Pretending it doesn't exist creates false confidence and poor decision-making.

When users understand uncertainty:

  • they can weigh AI recommendations appropriately,
  • they know when to seek additional information or human input,
  • and they can make informed choices about risk tolerance.

Without transparency about uncertainty, users may:

  • blindly trust outputs that are actually uncertain,
  • make critical decisions based on low-confidence predictions,
  • or lose trust when they discover the system was uncertain about something important.
Design for trust and accountability Enable meaningful user control