Communicating uncertainty without misleading

Year 3 / Module 18 Human-AI interaction and machine learning interfaces / Unit 5 / Communicating uncertainty without misleading

HCI3102 · Unit 5 of 10

Uncertainty displays change behaviour only when the representation matches the inference the decision requires, so a confidence figure that is well calibrated in the model's own terms can be uninterpretable at the interface, and the same uncertainty statement can raise or lower trust depending on how it is framed.

Communicating uncertainty without misleading

A model rarely outputs a decision. It outputs a score, and the interface turns that into a number that looks like a probability. This fifth of the ten units takes the display, where the argument about uncertainty is lost or won. A figure well calibrated in the model's own terms can be uninterpretable at the interface. The inference a person performs from it is not the one the model performed. The same statement can raise or lower trust according to its framing. The question is which inference the display must support, and for whom.

A calibrated score is a claim about the model, not a message to a person

Guo, Pleiss, Sun and colleagues (2017) found the networks they examined poorly calibrated, with predicted probabilities overstating the true correctness likelihood. Temperature scaling was surprisingly effective on most datasets. The score a model emits and the probability a person reads are two objects, and only one is on the screen.

Ovadia, Fertig, Ren and colleagues (2019) found temperature scaling achieving low expected calibration error at low dataset shift. The error increased significantly as shift increased, so in-distribution calibration does not guarantee calibration under shift.

Angelopoulos and Bates (2023) describe the strongest version of the quantity-first case. Conformal prediction wraps any pre-trained model and produces sets guaranteed to contain the truth with a user-specified probability, such as 90 per cent. A prediction set is stronger than a score, but it still claims coverage of a label, not what the reader infers.

The format decides what a probability means

Gigerenzer and Hoffrage (1995) made the representational claim first. They argue that no claim about the mind's capacity for Bayesian inference can be evaluated without specifying the format. Naive participants derived up to 50 per cent of their inferences by Bayesian algorithms in frequency form.

Duke (2023) shows how narrowly such a gain lands. In two experiments, with 275 and 796 participants, a numerical range in the text or a probability lexicon improved interpretations of probability phrases over verbal phrases alone. Neither improved interpretations of confidence, which participants could not read against the precision of their estimate intervals.

Design the display around the decision, not around the quantity

Fernandes, Walls, Munson and colleagues (2018) measured the outcome that matters. Quantile dotplots with 50 outcomes produced expected payoffs 97 per cent of optimal, five percentage points above the control. Their within-subject standard deviation was three percentage points, four points below the control. The same uncertainty in words performed worst.

Kay, Kola, Hullman and colleagues (2016) derived their display from the decision. Quantile dotplots reduced the variance of participants' probabilistic estimates by about 1.15 times compared with density plots, a gain in precision rather than preference.

Individual draws help, and the evidence stops at aggregation

Hullman, Resnick and Adar (2015) tested the more provocative version of the claim. Their animated plots of individual draws beat error bars and violin plots for judgements about the reliability of orderings of two and three quantities. Ancker, Senathirajah, Kukafka and colleagues (2006) add that the features improving quantitative reasoning differ from those most likely to alter behaviour. Greis and colleagues (2018) gave participants two sensor readings of the same quantity with disagreeing values, such as a phone tracker and a wheel sensor reporting different distances. Showing the uncertainty improved the estimates, most where the sensors differed in reliability, and richer displays helped more. The sensors are synthetic, so the case shows how a display changes aggregation. It does not show what a product display does to a decision.

A confidence display changes the person, not only the decision

Cao, Liu and Huang (2024) found in a skin cancer screening task that showing calibrated model uncertainty alone was inadequate. Calibrating it and presenting it in a frequency format let users adjust their reliance and reduced the effect of confirmation bias.

Li, Yang, Liao and colleagues (2025) show what else a confidence figure moves. With 270 participants, self-confidence aligned with the AI's confidence. The alignment persisted after the AI was withdrawn, and real-time correctness feedback reduced it. For most participants it impaired their self-confidence calibration.

Showing uncertainty costs less credibility than the field assumes

van der Bles, van der Linden, Freeman and colleagues (2020) tested the assumption that admitting uncertainty damages credibility. Across five experiments, with 5,633 participants, people perceived greater uncertainty when it was communicated. The set included a preregistered replication with a national sample and a field experiment on the BBC News website. Trust in the numbers and in the source fell only slightly, mostly for verbal communication.

van der Bles and colleagues (2019) supply the vocabulary for comparing displays. Spiegelhalter, Pearson and Short (2011) add the moderator. Limited experimental evidence exists on how formats are processed, although the effectiveness of some graphics depends on the audience's numeracy.

Contested question: the number, the display, or the audience?

The literature disagrees about where the defect lies.

The first position holds that the defect is in the quantity, and that the display is a rendering problem solved afterwards. Its evidence is the calibration programme. Guo and colleagues (2017) found the networks they tested poorly calibrated. Ovadia and colleagues (2019) found methods marginalising over models holding up better under shift. Angelopoulos and Bates (2023) describe prediction sets whose coverage is guaranteed. What would settle it is a study varying only the quantity, with the representation fixed and the joint decision as the outcome.

The second position, which this unit defends, holds that the defect is in the mapping from quantity to representation. Calibrated uncertainty presented alone was inadequate (Cao, Liu and Huang, 2024). A frequency format changed whether a Bayesian inference was performed (Gigerenzer and Hoffrage, 1995). Decision quality moved with the representation (Fernandes et al., 2018). A display carrying less statistical machinery gave better inference (Hullman, Resnick and Adar, 2015). What would settle it is the mirror image: one model, several representations, replicated across tasks.

The third position holds that the defect is the audience, so no representation generalises far. Spiegelhalter, Pearson and Short (2011) report that effectiveness depends on numeracy, and Ancker and colleagues (2006) found preference decoupled from accuracy. Cao, Liu and Huang (2024) found reliance varying with age and statistical familiarity, and Li and colleagues (2025) found the effect depending on the user's initial confidence. What would settle it is preregistered moderation tests reporting the same formats by numeracy and expertise.

None of the three is decidable from the present record. Studies of uncertainty communication commonly conflate the object and the level of uncertainty (van der Bles et al., 2019). Many of them measure preference rather than the outcome the decision turns on (Ancker et al., 2006).

What to do while the dispute is open

The three positions are not equally evidenced. The quantity-first case rests on benchmark measurements of calibration and on a guarantee about label coverage, so it is strong about the model and silent about the reader. The representation-first case rests on a psychological experiment, one incentivised decision experiment and one visualisation comparison, all in laboratory tasks. The audience case rests on two reviews and on one experiment with 270 participants, where the effect depended on the user's initial confidence. The strongest objection to this unit's thesis is that a better display can make a worse number more persuasive. Ovadia and colleagues (2019) found calibration degrading as shift increased, so a clean representation of an uncalibrated score raises a reader's confidence without raising the score's validity. While the dispute is open, a designer should state the object and the level of the uncertainty a display carries, and test it on decisions rather than on preference. A display also has to work on a metered connection and in a screen reader.

Where the argument is heading

The direction is towards displays that state what they are about and are tested on the decision. van der Bles and colleagues (2019) give the taxonomy that makes object and level explicit. van der Bles and colleagues (2020) show the trust cost of disclosure is small and format-dependent. Li and colleagues (2025) set the sharpest requirement for the next studies: report what a display does to the user's calibration after the tool is gone. Ovadia and colleagues (2019) constrain all of it, since such a display inherits the conditions of its calibration.

The sources this unit rests on

Ancker, J. S., Senathirajah, Y., Kukafka, R., et al. (2006). Design Features of Graphs in Health Risk Communication: A Systematic Review. Journal of the American Medical Informatics Association 13(6), 608-618. doi:10.1197/jamia.m2115 The review separates health-risk graph studies by outcome measured. Features improving quantitative reasoning differ from those that alter behaviour, and viewers preferred simple features unrelated to accuracy.
Angelopoulos, A. N. and Bates, S. (2023). Conformal Prediction: A Gentle Introduction. Foundations and Trends in Machine Learning. doi:10.1561/9781638281597 It introduces conformal prediction, which wraps any pre-trained model to produce sets guaranteed to contain the truth with a user-specified probability, such as 90 per cent. Label coverage is guaranteed in advance; what a reader infers from the set is not.
Cao, S., Liu, A. and Huang, C.-M. (2024). Designing for Appropriate Reliance: The Roles of AI Uncertainty Presentation, Initial User Decision, and User Demographics in AI-Assisted Decision-Making. Proceedings of the ACM on Human-Computer Interaction 8(CSCW1), 1-32. doi:10.1145/3637318 Calibrated uncertainty presented alone was inadequate in a skin cancer screening task, while a frequency format let users adjust their reliance and reduced confirmation bias. Reliance also varied with initial decision, age and statistical familiarity.
Duke, M. C. (2023). Probability and confidence: How to improve communication of uncertainty about uncertainty in intelligence analysis. Journal of Behavioural Decision Making 37(1), e2364. doi:10.1002/bdm.2364 Two experiments, with 275 and 796 participants, and a third with 510, compared ways of conveying second-order probability. Ranges and lexicons improved interpretations of probability phrases but not of confidence, while stating the source improved confidence-decision correspondence.
Fernandes, M., Walls, L., Munson, S., et al. (2018). Uncertainty Displays Using Quantile Dotplots or CDFs Improve Transit Decision-Making. Proceedings of the 2018 CHI Conference on Human Factors in Computing Systems, 1-12. doi:10.1145/3173574.3173718 An incentivised bus-catching experiment scored decision quality rather than probability extraction. Dotplots with 50 outcomes gave payoffs 97 per cent of optimal, five percentage points above the control, with a within-subject standard deviation of three percentage points, four points below the control.
Gigerenzer, G. and Hoffrage, U. (1995). How to improve Bayesian reasoning without instruction: Frequency formats. Psychological Review 102(4), 684-704. doi:10.1037/0033-295x.102.4.684 The authors argue that claims about Bayesian reasoning cannot be evaluated without specifying the information format, and show the computation is simpler in natural frequencies. Statistically naive participants derived up to 50 per cent of their inferences by Bayesian algorithms in frequency form.
Greis, M., Joshi, A., Singer, K., et al. (2018). Uncertainty Visualization Influences how Humans Aggregate Discrepant Information. Proceedings of the 2018 CHI Conference on Human Factors in Computing Systems. doi:10.1145/3173574.3174079 Participants combined two disagreeing sensor readings into one estimate under four uncertainty displays of increasing detail. Uncertainty information improved the estimates, most where the sensors differed in reliability, and richer displays helped more. The sensors are synthetic, so the study measures aggregation rather than a product decision.
Guo, C., Pleiss, G., Sun, Y., et al. (2017). On Calibration of Modern Neural Networks. arXiv preprint arXiv:1706.04599. doi:10.48550/arXiv.1706.04599 The authors measured whether the probabilities from modern neural networks match the true correctness likelihood. The networks were poorly calibrated, architecture and regularisation affected calibration, and temperature scaling was surprisingly effective, so a displayed confidence is a raw score until recalibrated.
Hullman, J., Resnick, P. and Adar, E. (2015). Hypothetical Outcome Plots Outperform Error Bars and Violin Plots for Inferences about Reliability of Variable Ordering. PLOS ONE 10(11), e0142444. doi:10.1371/journal.pone.0142444 Hypothetical outcome plots animate individual draws rather than summarising a distribution. Participants judged orderings of two and three quantities more accurately than with error bars or violin plots, but similarly for a single quantity.
Kay, M., Kola, T., Hullman, J. R., et al. (2016). When (ish) is My Bus? Proceedings of the 2016 CHI Conference on Human Factors in Computing Systems, 5092-5103. doi:10.1145/2858036.2858558 Requirements were derived from a literature review, a survey of transit-application users and an iterative design process. Dotplots reduced the variance of probabilistic estimates by about 1.15 times compared with density plots, a gain in precision rather than preference.
Li, J., Yang, Y., Liao, Q. V., et al. (2025). As Confidence Aligns: Understanding the Effect of AI Confidence on Human Self-confidence in Human-AI Decision Making. Proceedings of the 2025 CHI Conference on Human Factors in Computing Systems, 1-16. doi:10.1145/3706598.3713336 In a randomised behavioural experiment with 270 participants, self-confidence aligned with the AI's confidence and the alignment persisted after the AI was withdrawn, while real-time correctness feedback reduced it. For most participants the alignment impaired their calibration.
Ovadia, Y., Fertig, E., Ren, J., et al. (2019). Can You Trust Your Model's Uncertainty? Evaluating Predictive Uncertainty Under Dataset Shift. arXiv preprint arXiv:1906.02530. doi:10.48550/arXiv.1906.02530 A benchmark under increasing dataset shift found temperature scaling achieving low expected calibration error at low shift while the error increased significantly as shift increased, so in-distribution calibration does not guarantee calibration under shift. Methods marginalising over models held up better.
Spiegelhalter, D., Pearson, M. and Short, I. (2011). Visualizing Uncertainty About the Future. Science 333(6048), 1393-1400. doi:10.1126/science.1191181 A review of practice in visualising future uncertainty across sport, weather, climate, health, economics and politics. The authors report limited experimental evidence on how formats are processed, effectiveness depending on numeracy for some graphics, and deep uncertainties remaining hard to convey.
van der Bles, A. M., van der Linden, S., Freeman, A. L. J., et al. (2019). Communicating uncertainty about facts, numbers and science. Royal Society Open Science 6(5), 181870. doi:10.1098/rsos.181870 An interdisciplinary review proposing three objects of uncertainty, facts, numbers and science, and two levels, direct and indirect, with nine expressions of direct uncertainty. Many studies conflate object and level, and effects vary between individuals and formats.
van der Bles, A. M., van der Linden, S., Freeman, A. L. J., et al. (2020). The effects of communicating uncertainty on public trust in facts and numbers. Proceedings of the National Academy of Sciences 117(14), 7672-7683. doi:10.1073/pnas.1913678117 Five experiments, including a preregistered replication with a national sample and a BBC News website field experiment, with 5,633 participants, tested whether communicating uncertainty reduces trust. Perceived uncertainty rose, but trust fell only slightly, and mostly for verbal communication.

Check your understanding & apply it · 4 questions, 1 essay

1A clinical team displays a probability of 0.83 from a model that has never been recalibrated, on a population that has shifted since training. Diagnose what that display is a claim about.

2A screening tool states a chance of about seventy per cent in prose. A designer proposes a quantile dotplot with 50 outcomes beside a frequency-framed statement. Predict what the change does to the reader.

3A team reports that its conformal prediction sets guarantee 90 per cent coverage and concludes that the interface work is done. How does the unit judge that conclusion?

4A product shows the model's confidence beside every output because users say the figure helps them think. Diagnose that design against the unit's evidence.

5Essay A radiology department will deploy a model that flags chest radiographs as urgent, and the proposed interface shows one number per study, for example '87% confident'. The clinical lead asks you to review the plan and to say what should be displayed instead, how that display should reach a reader who cannot see a chart, and how the deployment should be judged. Write the review. Ground each part of it in named sources from this unit.

The single number is the first thing to question, and Guo and colleagues (2017) explain why the figure is not yet what it appears to be: the networks they examined were poorly calibrated and overstated their correctness, so a score becomes a usable probability only after recalibration. The recalibration does not travel either. Ovadia and colleagues (2019) found that temperature scaling achieved low expected calibration error at low dataset shift, but that the error increased significantly as shift increased, so a number calibrated on an in-distribution validation set is not guaranteed valid when the scanner, the referral population or the coding practice changes. The department should therefore state which distribution the number is calibrated on and monitor shift, and Angelopoulos and Bates (2023) show what the stronger alternative looks like, since a conformal prediction set carries a coverage guarantee in advance rather than a point estimate. The representation matters as much as the quantity. Gigerenzer and Hoffrage (1995) found that statistically naive participants derived up to 50 per cent of their inferences by Bayesian algorithms in frequency formats, and Cao, Liu and Huang (2024) found that calibrated uncertainty presented alone was inadequate in a screening task while a frequency format allowed users to adjust their reliance. I would ask for the decision the interface must support to be named first, since Fernandes and colleagues (2018) showed that decision quality moved with the representation, dotplots reaching 97 per cent of optimal payoffs five percentage points above control, while the same uncertainty stated in words performed worst. Ancker and colleagues (2006) add that reasoning accuracy, behaviour and preference are separate outcomes, so the department cannot infer one from another. The confidence figure also acts on the reader, not only on the case: Li and colleagues (2025) found that self-confidence aligned with the AI's confidence, persisted after the AI was withdrawn, and impaired calibration for most participants, so showing one number per study may erode the radiologist's own sense of certainty. For judgement I would fix the quantity, vary the representation, and score clinical decision quality and the override rate, reporting effects by experience and by statistical familiarity, because the audience position in this unit, from Spiegelhalter, Pearson and Short (2011), Cao, Liu and Huang (2024) and Ancker and colleagues (2006), is that no single format serves every reader equally. Finally, the department should expect that disclosing uncertainty is affordable: van der Bles and colleagues (2020) found across five experiments with 5,633 participants that trust fell only slightly, and mostly where uncertainty was communicated verbally. The display also carries an access condition. A chart is read only by a person who can see it, so the graphic should be published beside a textual equivalent of the same quantity, and Duke (2023) supplies the evidence that a numerical range in the text improves how a reader interprets a probability. A ward on a slow connection is the same case in another form, because the richest display is the one that takes longest to load.