Judge
Confidence & Uncertainty
How certain is the system?
Signals how sure the system really is about an output, in a form non-technical users can read at a glance.
- Evidence: research-based (Kay et al. 2016, Guo et al. 2017)
- Tested: concept only, mock data
Overview
Why it matters
Precise-looking numbers invite too much trust.
A risk score of 73 or a confidence of 94% looks exact. Most AI outputs are not. When people see a clean number with no sign of uncertainty, they trust it more than they should: clinicians accept flags they should question, caseworkers approve recommendations they should override.
For hallucination in generated text, see Grounding & Hallucination Indicators. It’s a different problem with a different design response.
What does it look like?
Caseworker handling a borderline case
Caseworker handling a borderline case
- Caseworker opens case
- Score shows: 71, with a confidence band of 64 to 78.
- A label shows: “Moderate confidence: employment data is incomplete.”
- Caseworker is prompted: “Consider requesting updated employment records before making a final decision.”
- Caseworker requests documents: decision deferred.
Applicant reading their score
Applicant reading their score
- Applicant sees: “Risk score: Moderate” with a plain-language description.
- Below the score: “The system is less certain than usual about this result because some of your information could not be verified.”
- Applicant understands why and is prompted to provide verification documents.
Out-of-distribution alert
Out-of-distribution alert
- System encounters a self-employed applicant with non-standard income structure.
- Score is generated but flagged: “This application type is less common in our dataset. The estimate may be less accurate than usual.”
- Case is automatically routed to a specialist caseworker.
When does it fit?
Use it when
- The system produces scores, risk levels or predictions
- Decisions based on them have real consequences
- Confidence varies from case to case, and that variation matters
Skip it when
- The output is a verified fact, or the system is deterministic
- The model doesn’t produce a reliable confidence measure
- Speed matters more than nuance, as in emergency interfaces
Design
How do I apply it?
-
Show it at the decision point.
Low confidence must be visible when someone is about to act, not hidden in a detail view.
-
Say where it comes from.
Incomplete data and an unfamiliar case call for different actions.
-
Pair it with an action.
Uncertainty without a next step only creates doubt.
-
Show only what the model can back up.
An uncalibrated range is worse than none. Check with your technical team first.
-
Don’t use it as a disclaimer.
Uncertainty helps people decide. It is not a way to shift responsibility.
For your technical team
Raw softmax scores from modern neural networks are systematically overconfident (Guo et al., 2017). Confirm that post-hoc calibration, such as temperature scaling, has been applied before surfacing confidence values. Where feasible, distinguish aleatoric uncertainty (randomness in the world) from epistemic uncertainty (gaps in the model’s knowledge). At minimum, make clear whether the uncertainty is in the data, the model, or both.
How can I show it?
Verbal label
High, moderate or low, next to the output.
A plain-language label alongside the numeric output: “High confidence,” “Moderate confidence,” “Low confidence: additional review recommended.” More accessible than a numeric range for non-technical users.
Confidence band
A range instead of a single number.
Rather than showing a single number (score: 73), show a range (score: 68 to 78). The width of the band communicates uncertainty. Visually: a central value with a shaded range extending on either side.
Confidence colors
Color shows how sure, not how high.
Calibrated color scales tied to confidence levels, not just to the output value itself. A high score with low confidence should look different from a high score with high confidence. Make the levels differ in lightness, not hue alone.
Quantile dotplot
Countable dots for possible outcomes.
A discrete representation of a probability distribution: countable dots rather than a smooth density curve, where each dot is one equally-probable outcome. Research shows this reduces the variance of users’ probability estimates versus density strips. Prefer it for non-technical users, and for small screens where a density curve compresses into an unreadable sliver.
Distribution view
The full range of outcomes, for experts.
For users who need more detail: a probability distribution showing the full range of possible outputs, not just the central estimate. Appropriate for analyst or expert views. Can be placed in a secondary layer.
Warning icon
Appears only when confidence drops below a threshold.
A visual signal (a warning icon, a shaded zone, a distinct color) that appears when confidence is below a defined threshold. Draws attention to uncertain cases without requiring the user to interpret a number.
Out-of-distribution warning
Flags cases unlike anything the model has seen often.
A specific warning that appears when the current input falls outside the model’s training distribution: “This case has characteristics the system hasn’t seen frequently. Results may be less reliable.” Separate from uncertainty in the output: this is uncertainty in the model’s applicability.
Risk slider
Reviewers adjust a threshold for their case, and the change is logged.
A user-adjustable control allowing a reviewer to move a displayed point estimate toward a more conservative or permissive threshold to match their situation-specific risk tolerance. The adjustment is made explicit and logged rather than silently absorbed into the reviewer’s reasoning. Connect to the override mechanism in Human-in-the-Loop so that threshold adjustments are recorded alongside the final decision.
Further left for non-technical readers and routine cases, further right for experts and high stakes.
What does it cost?
Signal vs. fatigue
Too many warnings and people stop reading them. Save prominent signals for genuinely low confidence.
Precision vs. clarity
“73.4% ± 4.1%” is exact but hard to act on. “Moderate confidence” is easier. Match the reader.
Honesty vs. credibility
Frequent low confidence can erode trust in the whole system. Explain what it means and what to do.
Connections
Which patterns work with it?
Grounding & Hallucination Indicators
Can I trust what the system generated?
Uncertainty covers statistical confidence in predictions. Grounding covers factual reliability in generative AI output. These are distinct failure modes that require different design treatment.
Model Scope & Limits
What can this system reliably do?
Out-of-distribution uncertainty is a specific form of scope violation. The two patterns work together to communicate where the system’s reliable operating range ends.
Human-in-the-Loop
Who can review or override this decision?
Low confidence is one of the primary triggers for human review. Uncertainty signals and the HITL mechanism should be directly linked.
Attribution
Why did the system make this decision?
Uncertainty applies not just to the final output but to individual attribution factors. A factor’s contribution may itself be uncertain: especially in noisy or incomplete data.
Sources (4)
Sources (4)
Doshi-Velez et al. (2017) — Accountability of AI Under the Law: The Role of Explanation
argues that societal and legal accountability requires AI systems whose outputs can be explained, and examines when explanation should be demanded. Foundational for the normative case underlying this pattern
Kay et al. (2016) — When (ish) Is My Bus? User-Centered Visualizations of Uncertainty
empirical study comparing uncertainty visualization formats. Finds that quantile dotplots outperform continuous density strips on decision precision because they leverage subitizing, the ability to instantly recognise small group quantities, reducing estimation variance by approximately 1.15× compared to density plots. Directly motivates the quantile dotplot primitive in this pattern
Yaniv & Kleinberger (2000) — Advice Taking in Decision Making: Egocentric Discounting and Reputation Formation
shows that decision-makers systematically discount advice relative to their own opinion and adjust the weight they give it based on the advisor’s perceived quality and track record. The psychological basis for why the way a system communicates its confidence changes how much weight users give its output. Published in Organizational Behavior and Human Decision Processes, 83(2)
Guo et al. (2017) — On Calibration of Modern Neural Networks
demonstrates that modern deep neural networks are systematically overconfident: their softmax confidence scores are not reliable probabilities, particularly after training on large datasets. Introduces temperature scaling as a post-hoc remedy. Essential context for designing honest uncertainty displays: you cannot simply surface the model’s own confidence score as trustworthy
Your system shows a clean score, and it drives real decisions.
Has anyone checked how reliable that score is across different kinds of cases? That’s where this pattern starts. Let’s look at your interface together.
Work with me