I gave 4 AI tools the same graph—with & without storytelling with data principles

Author: storytelling with data · Watch on YouTube

Video authored by the channel shown, not by IANIA; all rights belong to the creator. Transcript and summary generated automatically; the original-language content is preserved.

Original language: English

Original video: I gave 4 AI tools the same graph—with & without storytelling with data principles · Channel: storytelling with data · 13:28 · Watch on YouTube

Topics: Data, databases and embeddings, AI developer tools, AI content generation

Summary

Cole compares how ChatGPT, Claude, Gemini, and PowerPoint Copilot redesign the same test-score graph, first with only a simple prompt and then with the Storytelling with Data plus AI primer. Without context, the tools suggest some useful visual changes, but one also makes a factual error, and none can identify the story behind the data. With the primer, several tools ask about the audience and purpose, help shape the message, and sometimes challenge whether the evidence supports the proposed conclusion. Cole concludes that AI can strengthen the design process and make it faster, but people still need to provide context, validate the output, and make the final decisions.

Key points

The graph shows monthly minimum, average, and maximum test scores across a school year, with scores falling early and improving in the spring.
In the first round, all four AI tools redesigned the graph from the same image and prompt, proposing changes such as line charts, less clutter, direct labels, and takeaway titles.
Copilot made an incorrect claim about the June score spread, illustrating why AI-generated output needs to be checked.
The free Storytelling with Data plus AI primer gives the tools guidance on context, visual choice, focus, clarity, audience, and asking questions.
After receiving the primer, ChatGPT, Claude, and Gemini asked for context before redesigning; Claude also challenged the causal interpretation and noted limitations in the evidence.
Gemini offered presentation talking points, while Copilot’s initial redesign missed the study-group context because it did not ask for it.
Claude’s PowerPoint add-in raised concerns about attributing improvement to the study groups and offered alternative ways to visualize the data.
AI can support stronger, faster communication, but it does not replace human knowledge of the audience, the data, or the desired outcome.

Chapters

Start	Chapter	Summary
00:00	The first AI redesigns	Cole gives four tools the same graph and simple prompt, reviews their visual suggestions, and flags an incorrect Copilot statistic.
02:36	Why context matters	Cole introduces the Storytelling with Data plus AI primer and explains how it guides AI to consider audience, purpose, and evidence.
04:09	Testing the primer with four tools	With added context about administrators and study groups, ChatGPT, Claude, and Gemini ask questions and create more audience-focused responses.
09:19	Copilot and the PowerPoint workflow	Copilot creates editable slides but misses the study-group context, while Claude’s PowerPoint add-in asks questions, challenges the claim, and explores alternative designs.
12:26	AI as a partner in the process	Cole reflects on how the primer improved the conversation and emphasizes that people must still supply context, evaluate evidence, and shape the final work.

Quotes

That’s a good reminder that we always need to check our AI generated output. — 02:04

Everything’s clearly labeled. Color is used more sparingly. — 06:13

It challenged the story I was telling. — 07:15

I wasn’t just getting a better graph. I was having a stronger conversation about the story I wanted to tell. — 12:57

The opportunity isn’t to hand the work over to AI. It’s to use AI to help us do the work better and faster. — 13:28

Concepts and entities

storytelling with data + AI primer, ChatGPT, Claude, Gemini, Copilot, Storytelling with Data, Contexto da comunicação, Visualização de dados, Validação de resultados

Full transcript (original language)

00:00 Can giving AI a better foundation help it create better graphs? I’m Cole and I put this question to the test using the same graph, the same prompt, and four different AI tools. Here’s the graph I used. It shows monthly test scores for a class over the course of a school year. The minimum, average, and maximum score each month. You can see scores decline early in the year and then improve considerably in the spring. For the first round, that’s all the AI gets, the graph and one very simple prompt. I used paid versions of chat GPT, Claude,

00:31 Gemini, and Copilot within PowerPoint. I started a fresh conversation with each tool and asked it to disregard anything it might know about me so prior context wouldn’t influence the results. Then I gave each the exact same prompt. Can you improve this graph? Please create a redesigned version. Chat GPT responded with this line graph. It kept the same colors as the original, but changed the chart type to make the trends easier to see. It also added a takeaway title. test scores rebounded after December and annotations calling attention to the December lows. Next, I turned to Claude,

01:03 which started by pointing out that the grouped bars were fighting the data. This is a trend over 10 months, it said, but the bars encourage comparison within each month rather than across time. It recommended a linebased view with shading between the minimum and maximum range. I like the shaded range. It gives us a different way to see the spread while keeping the average prominent. Claude also explained the changes it made, shifting to lines, simplifying the colors, removing borders, and adding a takeaway at the bottom. Gemini took a slightly different route. First, it gave

01:33 me the data in a table, outlined the changes it wanted to make, and rendered this rough visual. When it offered to create a higher resolution, I said I’d like it to do that. Like Claude, it moved from bars to lines and used a shaded band for the min to max range with the average more prominent. It also simplified the percentage formatting and directly labeled the endpoints. Finally, I gave the same graph and prompt to copilot within PowerPoint. One practical advantage here is that the result is editable in PowerPoint rather than an

02:04 image. It added a takeaway title, average score rebounded to 88% by June and pulled several numbers into callouts on the right. One of those is actually incorrect though. The June spread is 35 points, not 23. That’s a good reminder that we always need to check our AI generated output. AI redesigned the graph and there are good ideas in every one of these. A different chart choice, less clutter, more focus, takeaway titles, direct labeling. But there is still a big problem. None of these tools

02:36 know why this graph exists. Who is it for? What happened? What does the data mean? And what do I want the audience to understand or do? But what if we prime AI to not just give us a graph and instead help us think more critically about the situation and everything that impacts how we design our graphs and slides? That’s exactly why we created the storytelling with data plus AI primer. The storytelling with data plus AI primer is a free resource we created

03:07 to give you and your AI partner a shared foundation in the core principles we teach at Storytelling with Data. It starts with an excerpt from our latest book, Storytelling with Data Before and After, covering principles like understanding the context, choosing an appropriate visual, eliminating clutter, focusing attention, and using words and story effectively. Then we translate those ideas specifically for AI. There’s a page of guidance directed at your AI partner that encourages it to ask questions, clarify the message, consider

03:38 the audience, challenge your thinking, and stay grounded in what the data actually supports. Finally, there are conversation starters you can use to put those principles into practice. Whether you are exploring data, creating or improving a graph, building a slide, or preparing to present, the idea isn’t to give AI a rigid set of rules. It’s to give it a better foundation for working with you. Let’s see what happens when I give each of our four AI tools this foundation before asking them to improve that graph. In turn, I uploaded the

04:09 primer to each tool along with this prompt which is copied directly from the first page of the primer. Review the attached guide from storytelling with data and use these principles when we explore ideas or discuss, review, and create graphs, slides, and presentations. Prioritize simplicity, clarity, and audience focus in all future communications. Then I gave it the exact same graph and asked the exact same question as before. Can you improve this graph? Please create a redesigned version. Immediately things were different. This time I didn’t get a graph. I got questions.

04:41 ChatgPT starts by pointing out that the three statistics have equal visual weight. But before redesigning it asks a question. What do you want your audience to take away from these scores? It suggests a few possible stories it sees in the data. Then outlies some changes it might make. It ends by asking for the context it needs. Who is the audience? What happened during this period? And what do you want them to know or do as a result? These questions come directly from the context setting foundation in the primer. I responded, “The audience

05:12 is school administrators. I’m the teacher of this class presenting an update on student performance over the school year. Scores fell considerably early in the year, reaching a low in December. I tried a number of different teaching and student support strategies. And in March, I introduced mixability study groups, after which scores improved sharply and continued to improve through June. I want the administration to understand the impact the study groups appear to have had, particularly for lower performing students. I plan to introduce them

05:43 earlier next school year, and I’d like to share this strategy with other teachers so they can consider using it with their students. Now that it has context, chat GPT frames the core message as mixed ability study groups helped turn performance around, especially for struggling students. It recommends a line graph with three lines, then outlines the design changes it plans to make. And it points out that my recommendation belongs on the slide, too. Taking all of that into account, it designs this. There’s a lot that I like

06:13 here. We have a takeaway title and subtitle that tell us what to look for. Everything’s clearly labeled. Color is used more sparingly. The timing of the study groups is marked directly on the graph. And at the bottom, it brings us all the way back to what the audience should do next. Claude also came back to me with words rather than a graph. It refers to the SWD approach. I also love this line. I don’t just want to make it prettier without knowing what it needs to do. It identifies a few improvements that it

06:43 can make regardless of context. removing clutter, creating a clearer takeaway, and reconsidering the grouped bars. But the best fix depends on what I’m trying to say. It asks me for the main point I want people to take away, and I tell it, “The study groups introduced in March appear to have had significant positive impact, particularly for lower performing students.” With that, Claude outlines its approach before building anything. switch to a line graph, emphasize the minimum and mute the other series, mark the introduction of study groups, use a takeaway title, and

07:15 declutter. Then it builds the graph. The result is clean and focused. The takeaway title and sparing use of color direct our attention to the improvement among lower performing students, and the March marker connects that change to the timing of the study groups. But Claude did something else I found particularly interesting. It challenged the story I was telling. It pointed out that scores had already begun recovering before March. So, I should be prepared for my audience to ask whether the study groups drove the acceleration or whether they coincided with a recovery already

07:45 underway. I love this. It’s thinking beyond the graph to the communication and anticipating where my audience might question the story that I’m telling them. And in this case, I have that additional context. I came back after winter break re-energized, tried a lot of different things to boost learning, understanding, and ultimately scores. So, a lot of things were helping a little bit, but it wasn’t until I introduced those study groups in March that scores sharply increased. Gemini follows a similar path. It first identifies issues with the existing

08:16 graph and recommends changes, switching to a line graph, decluttering, directly labeling the data, creating visual focus, and using a takeaway title. But before redesigning, it also asks about the audience and what I want them to know and do. After I answered that, Gemini shared the key adjustments it would make for this audience. But then it goes beyond the graph and gives me talking points for the presentation itself. The plot scores declined early in the year. Tension. Lower performing students were really struggling. Resolution. After the study groups were introduced, their scores rose

08:47 marketkedly. And finally, the proposal. introduce the strategy earlier next year and share the approach with other teachers. This is pretty good. It’s not simply redesigning the graph. It’s helping me to think about the story I might communicate around it. Here’s the resulting graph. Feels a little busy to me, but I actually like many of the components. There’s a takeaway title paired with a descriptive graph title. Everything is labeled directly. The intervention point is clearly marked. The recommended action is right on the graph. I think some refinement of the typography and hierarchy would make this stronger, but there’s a lot that I like

09:19 here. Last to test was Copilot. It was actually the only one out of the bunch that didn’t ask me any questions before redesigning. I was using Copilot integrated directly into PowerPoint, which meant the output was editable slides instead of static images, which is what I got from the other tools. Copilot designed a three slide progression. First, a blank graph. If I’m presenting live, this would give me the opportunity to explain to my audience what we’ll be looking at before they can jump to the detail. How did test scores change from September to June? That’s what we’ll be looking at today for my classes monthly minimum,

09:51 average, and maximum test scores, which is what we’ll see plotted on the vertical y-axis with time on the x-axis, ranging from the beginning of the school year, September at the left, to the end of the year, June, at the right. First half of the year was not looking great. Test scores took a nose dive into December, but scores increased steadily from there. Copilot didn’t include any detail on the study groups because remember it didn’t ask me for any context. So I didn’t know that important information. The resulting story misses the key thing that I want to

10:22 communicate. On the plus side, remember these are editable slides. So I can add that context and either make this three slide progression work for that or use the final one as a starting point to design a single standalone slide. The integration with PowerPoint could be so useful that I wanted to test it one more time and see if I’d have different results using the Claude addin. Claude did have a conversation with me before proceeding. It outlined what’s working, what’s getting in the way, and the bigger question, what is the takeaway?

10:53 Again, I love Claude’s framing here. Two things would let me build the right one rather than a prettier version of the wrong one. It asks two questions in the chat window that are embedded in PowerPoint. You can see those and my responses here. the audience is leadership and it’s a recovery story. I added the additional context that my audience is school administrators. So I want to understand that the mixed ability study groups introduced in March appear to have had a market positive impact particularly for lower scoring students and that I’d like their support sharing this strategy with other

11:23 teachers. Claude pushed back on a couple of things. We saw this first potential issue earlier that the increase predates the introduction of the study groups. That’s fine. I have an explanation for that. Claude raises another good point that this is a single classroom observation without any comparison beyond the before and after. So we should be careful about promoting it as a proven effect. It even raised the idea of another action could be to test it more thoroughly. In its redesign, Claude first landed on a line graph explaining that it eliminated the maximum line because the story is more about

11:54 improvement for the lowcoring students. I actually don’t love that. It feels like important comparative context has been lost by not showing the upper end of the range. But Claude also offered another idea. Focus on the range between low and high. After I confirmed I’d be interested in seeing it, it created a second slide so I could compare. This is pretty awesome. Not as a stopping point, but as a starting point for me to jump in and iterate. And because this is native in PowerPoint, I have full control over the details and can continue to modify from here. As a

12:26 bonus, it’s working from the same data as my original PowerPoint graph, making the data easier to validate. After about 20 minutes of fine-tuning, here’s where I landed. With AI only, we might get a better graph. But when I gave AI, the storytelling with data, philosophy and principles, it did more than improve the graph. It asked me questions. Who’s my audience? What do I want them to know and do? It helped me clarify the story, challenged my thinking, anticipated questions my audience might have, and suggested different ways to visualize

12:57 the data based on what I was trying to communicate. So, I wasn’t just getting a better graph. I was having a stronger conversation about the story I wanted to tell. And that helped me get to my final version in a fraction of the time. AI didn’t give me the finished product. And I don’t think we should expect it to. There’s too much that we bring to the process. our knowledge of the audience, the context, the nuances in the data, and what we’re ultimately trying to accomplish. The opportunity isn’t to hand the work over to AI. It’s to use AI

13:28 to help us do the work better and faster. If you want to try this yourself, download our free storytelling with data plus AI primer at storytellingwithdata.com/ai. Gives you the principles and conversation starters I used throughout this process.