Misleading graphics, deception and the ethics of truthful encoding

A chart can carry a false value, a true value that suggests a false magnitude, or a faithful picture that still leaves a belief the data does not support. The field named the second case as the deceptive visualisation, and the evidence has since redistributed the emphasis. The distortions that carry the name mislead by very different amounts (Pandey et al., 2015), and the largest failure measured so far is the reversal of a message, not its exaggeration. Some of the most reliable ways a chart misleads sit in the title above it (Kong et al., 2018), and the common violations of honest design are habit rather than the work of bad actors (Szafir, 2018). This seventh of ten units, after the sixth told the story, argues that the remedy is design and audit, and an obligation to answer for what a chart does once published.

The craft inheritance, and what it cannot test

Tufte (1983) supplied the detection rule. The lie factor compares the size of the effect a design shows with the size of the effect in the data, and the data-ink ratio treats ink that carries no data as waste. It argues by example, with no experiment on whether a reader's inference changes. Cairo (2019) took the reader's side, arguing that citizens must be able to decode charts that deceive, intentionally or unintentionally.

Tufte's later book supplies the case behind his rule. In Visual Explanations (1997) he examines the thirteen charts that Morton Thiokol sent to NASA before the Challenger launch. Those charts grouped the damage data by flight and motor rather than plotting it against launch temperature, so the link between cold and damage stayed invisible. His own display of the same record puts every launch below 66 degrees Fahrenheit on the damaged side. The case shows the evidence was in the record and the arrangement hid it, but it does not measure a reader, and engineers who worked on the programme have since contested the reading.

The first controlled test, and the spread of severity

Pandey and colleagues (2015) ran the test the craft tradition implied. They defined deception at the level of the message. The definition splits it into exaggeration, where the extent of a true fact is distorted, and reversal, where the reader takes the opposite message from the same data. They wrote the definition without motive, so a distortion could be studied whether or not anyone meant it. They built control and distorted versions of the same charts and ran crowdsourced studies. Every technique moved the reader, by very different amounts:

• Inverted axis (line-area chart, reversal): 78.95 per cent of readers took the wrong message, against 2.5 per cent in the control, p < 0.0001.
• Aspect ratio (line chart, exaggeration): mean response 3.19 against 1.39 on a five-point scale, r = 0.66.
• Truncated axis (bar chart, exaggeration): mean response 2.77 against 1.45, r = 0.37.
• Area as quantity (bubble chart, exaggeration): mean response 2.71 against 1.71, r = 0.34.

All three exaggeration techniques were significant at p < 0.001, so the label was not empty. Education, chart familiarity, visual ability and need for cognition settled nothing.

Faithful to the data, and still misleading

Newman and Scholl (2012) showed that the failure is not confined to distorted scales. Readers saw a bar depicting a mean and judged the likelihood that a test point belonged to the distribution behind it, and points inside the bar were judged likelier than equally likely points the same distance above it. The bias survived error bars, both axis directions, memory and online perception, and several populations.

Correll and Heer (2017) generalised the point towards intent. They treated the designer as an intermediary who can attack a reader, and named four categories of manoeuvre: breaks of convention, data manipulation, obfuscation, and nudging. A chart can be a straightforward mapping and still leave a false impression. The binary phrase deceptive visualisation conflates poor design with malicious design.

McNutt and colleagues (2020) took the analyst with no intent at all, naming visualisation mirages: apparent findings that arise from the data, the specification or the analyst's assumptions rather than from the phenomenon. Their remedy is borrowed from software testing. Metamorphic testing perturbs the data and the specification and asks whether the conclusion survives a change that should not matter, and they report that it reliably surfaces mirages with little prior knowledge of the domain.

The frame sits in the title, and the reader cannot report it

Borkin and colleagues (2016) tracked the eye movements of 33 participants over 393 visualisations and found titles drew the most fixation and were recalled most often.

Kong and colleagues (2018) varied the frame and held the chart constant. Their first study collected 888 participant-written titles and found five news frames plus an open-ended frame and a statistics frame, the last singled out as the subtle and most dangerous one. In the second, slant changed the perceived main message, and readers derived opposing messages from the same chart (chi-square = 27.06, df = 2, p < 0.001). Of the readers, 83 per cent reported that the visualisation was unbiased, while 50 per cent of the coded recalled messages were slanted and 77 per cent of those matched the title. Self-report did not detect the effect. The title also travels where the chart does not, since a screen reader reads it first, so the frame reaches a reader who never sees the display.

Contested question: a failure of design, an act, or a failure of reading?

The first position holds that most misleading charts are accidents of unexamined defaults, so the remedy is a design discipline. Szafir (2018) states it for practitioners, reviewing five ways visualisations mislead and attaching a fix to each. Pandey and colleagues (2015) wrote their definition without motive, so that a distortion could be studied whether or not it was meant.

The second position takes intent and consequence as the unit of analysis. Correll and Heer (2017) argue that persuasion is ubiquitous, that the space of techniques is far larger than the famous handful, and that a designer is well placed to attack a reader's inference, while Correll (2019) sets out moral obligations covering provenance, framing, dual use and the treatment of participants. Here a taxonomy is the wrong instrument: the same chart is a defect in one pipeline and an attack in another.

The third position treats the remedy as literacy, and it is the oldest. Cairo (2019) and Bergstrom and West (2021) argue that a reader can be taught to interrogate a chart and a quantitative claim. The evidence against relying on this alone is direct. Kong and colleagues (2018) found readers shifted by a title while they still reported the chart as unbiased, so the skill would have to work below self-report, and Pandey and colleagues (2015) found no definite conclusions from education or need for cognition.

What would settle it is a study the literature does not contain. If defaults are the cause, a checklist applied before publication should reduce measured distortion and reader error across a corpus of real charts. If intent binds, what reduces harm is accountability for outcomes. If literacy is the remedy, trained readers should resist framing and untrained readers should not. A decisive design would rate published charts by expert judgement and by measured reader effect, then measure a delayed outcome by decision rather than by self-report.

The two positions rest on different evidence. The design position rests on a practitioner review and one controlled study of four distortion techniques (Pandey et al., 2015). The intent position rests on a taxonomy and an argument (Correll and Heer, 2017; Correll, 2019) rather than on a study. The literacy position rests on two books and is contradicted by the framing experiments (Kong et al., 2018). While the dispute is open, treat distortion as a failure to avoid rather than a motive to assume.

The strongest counter-argument

The strongest objection is that the thesis, locating the problem in defaults, understates the adversarial case. The largest effect in the evidence is the inverted axis, the technique least plausibly produced by an unexamined default, and reversal rather than exaggeration is what does the damage; if the intent to persuade is ubiquitous (Correll and Heer, 2017), a discipline of habits and checklists addresses the smaller half of the problem. The reply keeps the thesis. Severity does not track intent, because a mis-set direction flag and a deliberately inverted axis produce the identical reader error, which is why Pandey and colleagues (2015) defined deception without motive.

Where the area is heading

McNutt and colleagues (2020) make the audit automatic by perturbing the data and the specification and asking whether the finding survives, turning an ethical question into a runnable test. Correll (2019) supplies the obligations a research group should be able to answer for. Szafir (2018) attaches a fix to each of five failure modes. The reading side is least settled. Bergstrom and West (2021) treat critical reading as teachable, while Kong and colleagues (2018) show both why it is needed and why self-report will not tell us whether it worked.

Mahbub and colleagues (2025) tested how misleading designs affect vision-language models, and found most of the ten models they examined deceived by truncated or inverted axes and similar tactics, across more than sixteen thousand responses. The deception question no longer concerns human readers alone.

The sources this unit rests on

Each note states what the cited document did and found, from the resolved record and abstract. Where a plan note and the paper disagree, the paper governs.

• Tufte, E. R. (1983). The Visual Display of Quantitative Information. URL: https://www.edwardtufte.com/book/the-visual-display-of-quantitative-information/ Set out the data-ink ratio, small multiples and the lie factor, proposing that deception be detected by comparing design variation with data variation. It argues by example and reports no experiment on reader accuracy.
• Tufte, E. R. (1997). Visual Explanations: Images and Quantities, Evidence and Narrative. https://www.edwardtufte.com/book/visual-explanations-images-and-quantities-evidence-and-narrative/ Examined how a display carries or conceals a causal claim, through the pre-launch Challenger charts and Snow's cholera map. He argued that the O-ring evidence was in the record and the arrangement hid it, an argument from documents rather than a test of readers.
• Cairo, A. (2019). How Charts Lie: Getting Smarter about Visual Information. URL: https://wwnorton.com/books/9780393358421 Catalogued how charts deceive, intentionally or unintentionally, across news and public data, and argued that citizens must be able to decode the visual information presented to them. The clearest statement of the literacy remedy, which the experiments complicate.
• Pandey, A. V., Rall, K., Satterthwaite, M. L., Nov, O. and Bertini, E. (2015). How Deceptive are Deceptive Visualizations? An Empirical Analysis of Common Distortion Techniques. doi:10.1145/2702123.2702608 Defined message-level deception as exaggeration and reversal, built control and distorted charts for four techniques, and ran crowdsourced between-subjects studies. The exaggeration techniques all had significant effects but differed in size, the inverted axis reversed 78.95 per cent of readers, and individual differences gave no definite conclusions.
• Newman, G. E. and Scholl, B. J. (2012). Bar graphs depicting averages are perceptually misinterpreted: The within-the-bar bias. doi:10.3758/s13423-012-0247-5 Ran experiments in which readers judged the likelihood that a test point belonged to the distribution behind a bar showing a mean. Points inside the bar were judged likelier than equally distant points above it across error bars, both axis directions, memory and online perception, and several populations.
• Correll, M. and Heer, J. (2017). Black hat visualization. URL: https://idl.cs.washington.edu/files/2017-BlackHatVis-DECISIVe.pdf Cast visualisation design as an attack surface and named four categories of manoeuvre: breaks of convention, data manipulation, obfuscation and nudging. Their key claim is that fidelity to the data is not the same property as accurate perception, so a chart can be correct and still mislead.
• Correll, M. (2019). Ethical Dimensions of Visualization Research. doi:10.1145/3290605.3300418 Surveyed the moral components of designing, building and researching visualisations and proposed additional obligations, covering data provenance, the framing of results, dual use and the treatment of participants. The unit uses it to widen accurate encoding into responsibility for consequences.
• McNutt, A., Kindlmann, G. and Correll, M. (2020). Surfacing Visualization Mirages. doi:10.1145/3313831.3376420 Named visualisation mirages as apparent findings produced by the data, the specification or the analyst's assumptions rather than the phenomenon, at every stage of the analytics pipeline. They adapted metamorphic testing and report that it reliably surfaces mirages with little prior knowledge of the domain.
• Borkin, M. A., Bylinskii, Z., Kim, N. W., Bainbridge, C. M., Yeh, C. S., Borkin, D., Pfister, H. and Oliva, A. (2016). Beyond Memorability: Visualization Recognition and Recall. doi:10.1109/tvcg.2015.2467732 Labelled 393 visualisations, tracked the eye movements of 33 participants and collected thousands of descriptions across recognition and recall. Titles drew the most attention and were recalled most often, and memorable visualisations also conveyed their message.
• Kong, H.-K., Liu, Z. and Karahalios, K. (2018). Frames and Slants in Titles of Visualizations on Controversial Topics. doi:10.1145/3173574.3174012 Collected 888 participant-written titles, finding news frames including a statistics frame, then varied title slant for the same chart. Slant changed the perceived main message (chi-square = 27.06, df = 2, p < 0.001) while attitude change was not significant, and readers reported no bias although half of the coded recalled messages were slanted.
• Mahbub, R., Islam, M. S., Laskar, M. T. R., Rahman, M., Nayeem, M. T. and Hoque, E. (2025). The Perils of Chart Deception: How Misleading Visualizations Affect Vision-Language Models. doi:10.1109/vis60296.2025.00006 Evaluated ten vision-language models on eight types of misleading chart across more than sixteen thousand responses and found that most were deceived by them. It moves the deception question onto machine readers, which now interpret charts for non-expert users.
• Szafir, D. A. (2018). The good, the bad, and the biased: five ways visualizations can mislead (and how to fix them). doi:10.1145/3231772 Reviewed five common ways visualisations mislead readers, including misleading scales, inappropriate encodings and framing, and attached a fix to each. The clearest practitioner statement of the defaults argument, and a synthesis rather than new evidence.
• Bergstrom, C. T. and West, J. D. (2021). Calling Bullshit: The Art of Scepticism in a Data-Driven World. URL: https://www.penguinrandomhouse.com/books/563882/calling-bullshit-by-carl-t-bergstrom-and-jevin-d-west/ Unpack selection bias and muddled data visualisation among a set of tools for interrogating quantitative claims, and argue that the sceptical skills need no great technical expertise. The wider data-literacy frame, against which the framing findings read as a caution.