Lesson 4 of 7 · 7 min read

Evidence under every claim

Put real data under each action title, mark gaps as [TBD], compare on a shared baseline, and cut anything that does not prove the claim.

Every claim needs real data

Lesson 3 gave every slide an action title. This lesson puts proof under each one. The rule has three parts: use only real data, mark every gap as [TBD], and put a source line under each chart.

AI tools fill gaps with plausible numbers. One former bank deck builder gave the same sales-pitch brief to three tools. The brief included no prices, and each tool invented its own. A Hacker News user was asked to review a boss’s AI-made deck that was mostly right but had misleading points. Checking it would take hours.

An unchecked number moves the work to your audience. So give the assistant your data, tell it to add nothing, and write [TBD] wherever data is missing. A [TBD] in a draft is honest. An invented number in a final deck is not.

Label every estimate as an estimate. Say whose estimate it is and what it rests on. A forecast drawn in the same style as measured data looks like a fact.

A source line names who produced the data, what it measures and the period it covers. “Source: order system, weekly on-time shipments, November to January” is enough. A reader who doubts a number should know where to check it.

Ask for charts the reader can edit, built from your table, not pictures of charts. When a number changes, you change the table and the chart follows. Lesson 6 covers how each assistant builds them.

This part of the rule is expert consensus. Practitioner guides, including one from Plus AI, call for real data, one chart per claim and a source under each chart. No controlled study has tested it.

Compare with position and length

Choose the chart by what the reader must compare. In 1984, William Cleveland and Robert McGill ranked how accurately people read quantities from different chart elements. Position on a shared scale came first, then length. Angle, area, volume and color came last.

Jeffrey Heer and Michael Bostock repeated the tests in 2010 with online participants. Accuracy was a little better than in 1984, but the order held. This rule is experimental evidence, and it has been replicated.

In practice, show the comparison that matters as bars or points on a shared baseline. A pie chart asks the reader to compare angles. A 3D bar adds depth and perspective, so the reader compares shapes instead of lengths. Start bars at zero, so each length matches its number.

Use a line for change over time. Each point on a line sits on a shared scale, so the reader sees the trend and the size of each change.

Label bars and points directly instead of using a color legend. Color sits near the bottom of the ranking, so the reader should not need it to read a value.

Cut what does not support the claim

Interesting material that does not support the point lowers learning. Researchers call it a seductive detail. Sundararajan and Adesope pooled 58 studies, with more than 7,500 students, in a 2020 meta-analysis. Students who learned with seductive details scored lower than students who learned without them.

The harm was larger when the detail stayed on screen, such as a static joke or image. It was also larger when the detail sat next to a relevant diagram. This rule is experimental evidence, drawn from many studies.

Imagery is not the problem. Unrelated imagery is. In a 2010 study, Bateman and colleagues showed 20 people charts embellished by the designer Nigel Holmes, and plain versions of the same charts. People read both kinds equally accurately. About two weeks later, they remembered more about the embellished charts.

This is experimental evidence, but small: one study with 20 people. It supports imagery tied to the subject of the data. It does not support decoration for its own sake. A stock photo of a handshake fails that test. A photo of your own backlog may pass it.

Key words on the slide, script in the notes

Lesson 1 moved your script to the speaker notes, based on Mayer, Heiser and Lonn. Their students learned less when on-screen text repeated the narration word for word.

That finding does not mean slides should carry no words. Adesope and Nesbit published a meta-analysis of this question in 2012, as described in a later review. Short on-screen key words led to better learning than full on-screen text. Key words beside a chart help. Full sentences that repeat your talk do not.

Both findings are experimental evidence. Keep the action title, the labels on the chart and a few key words that tell the audience what to look at. Put everything else in the notes or the appendix.

Check each slide against its title

Deckary’s consulting guide pairs the titles-only test with a second check. Each slide’s evidence must prove its own title. Read one slide at a time, and ask whether the chart alone shows what the title claims.

If it does not, change the chart or narrow the title. Never change the data to fit the title. This check is expert consensus, and Lesson 7 builds it into the final critique.

Do this check yourself before you ask an assistant to critique the deck. You know what the data shows. The assistant knows only what you gave it.

Worked example · Larkfield Supply

Morgan puts evidence under each claim

Morgan’s assistant drew charts for the draft deck from Morgan’s notes. Morgan checks each chart against its action title from Lesson 3.

The overtime slide has a pie chart of overtime hours, split by month. Its title says “Overtime ran 410 hours but did not hold the on-time rate.”
What would you do next?Does the pie chart prove the title? What would you use instead?Think it through, then show the answer
No. It shows which month had the most overtime, which is not the claim. The claim is about the on-time rate, so Morgan uses a line of the weekly on-time rate across the 13 weeks. The key words “410 hours of overtime, about $24,000” sit beside the line.
Morgan pulls the weekly rates from the order system. The line falls from 97% to 88% over the peak. The source line reads: “Source: Larkfield order system, weekly on-time shipments, November to January last year.”
The estimate slide has a 3D bar chart that compares last peak’s 88% with 96%, on an axis that starts at 85%. Morgan replaces it with a pair of flat bars on a shared baseline that starts at zero.
Morgan labels the second bar “About 96%: Morgan’s estimate” and draws it in outline, not solid. Its source line says the figure is Morgan’s own estimate, based on last peak’s weekly order volumes.
What would you do next?The draft also has a stat tile saying the second shift would “improve customer satisfaction by 15%.” What should Morgan do?Think it through, then show the answer
Cut it. Larkfield has no customer satisfaction data, so the assistant invented the figure. The real evidence is the two written complaints, which go on the customer slide with their dates.
Morgan finds a gap. Nobody has confirmed that four people can be hired before November. The cost slide now carries “Hiring lead time: [TBD], HR to confirm” until HR answers.That answers the question you guessed at: Nobody has confirmed that Morgan can hire four people before November. Guess: what goes on the cost slide where the hiring lead time would be?
Morgan removes a stock photo of a city skyline from the first slide. Morgan keeps a photo of last December’s loading dock full of unshipped orders on the customer slide, because it shows the problem the slide claims.

Every slide in Morgan’s deck now has one claim, one piece of evidence and a source line. The only estimate is labeled as Morgan’s, and the one gap is marked [TBD]. Lesson 5 checks the deck for AI tells.

Starter prompts

Chart from my data only

Here is my action title:

[paste title]

Here is the data that supports it, with its source:

[paste table or numbers, and where they came from]

Recommend one chart that proves the title. Prefer bars or points on a shared baseline, or a line for change over time. Avoid pie charts and 3D effects. Use only these numbers and add none. Write a source line to go under the chart. If the data does not prove the title, say so.

Find unsupported claims

Here is the text of my deck, slide by slide:

[paste deck text]

List every number, claim and comparison. For each one, say whether the slide shows a source. Mark anything without a source as [TBD: source]. Flag any figure that looks like an estimate but is not labeled as one. Do not fix, round or replace any number.

Cut what the claim does not need

Here is one slide. Title: [paste title]. Body: [paste content].

List anything on it that does not help prove the title: decoration, side facts, or text that repeats what I will say. Suggest at most three key words to keep beside the chart. Say whether each remaining item should move to the speaker notes or the appendix. Do not add new content.

When not to use AI

Producing the numbers. An assistant can chart, check and label your data, but every figure must come from your own records.
Making an estimate look more certain. If you ask an assistant to strengthen a forecast, it may drop the label that marks it as an estimate. Keep the label yourself.
Choosing which data to leave out. An assistant can suggest cuts, but you decide whether a dropped number would change the audience’s view.

Recall from memoryName the three parts of the rule for evidence under every claim.Answer it without looking back, then show the answer

Use only real data, mark every gap as [TBD], and put a source line under each chart. A [TBD] in a draft is honest. An invented number in a final deck is not.

Exercise · your project

Put real evidence under every claim in your own deck.

For each action title, find the data that proves it, and note where it came from.
Choose a chart that compares by position or length on a shared baseline, or a line for change over time.
Add a source line under each chart, and label every estimate with whose it is.
Mark every gap as [TBD], and note who can fill it.
Remove any image or fact that does not support the slide’s claim.

Mark this lesson as done