About eight years ago, I gave a talk where I predicted that a lot of the day-to-day work of a data scientist would be replaced by AI. I reasoned that this would leave more time for the high-value tasks of interpreting the results and making business decisions.

Imagine my surprise when instead of coming only for the data analysis, AI started coming for its interpretation too.

I believe that's a good thing—it means that anyone has access to a tool to help them interpret data. And data is essential to even creative endeavors, whether it's reading dashboards or interpreting charts. So you may as well let ChatGPT give you the highlights from it.

But, like everything GPT, while it will give you a starting point, make sure you confirm any important points with an expert. ChatGPT can make mistakes or lead you astray.

Here’s how to use ChatGPT’s new powers of data analysis.

How to upload charts and diagrams to ChatGPT

The process for inputting a chart or diagram for ChatGPT to analyze is the same as for image analysis: Just navigate to the chat box (on desktop or mobile) and click the paperclip icon.

You can also use ChatGPT’s power of data analytics on many other data formats—try uploading an Excel spreadsheet!

Read more: ChatGPT can see, hear, and speak now — here's how to make the best of it

Using ChatGPT to interpret charts

Got a chart you're having trouble making sense of?  Give it to ChatGPT and it can tell you the main findings from it. 

When I tested it on a demo of Buzzsprout's downloads analytics chart, it did a pretty good job.

I first asked for a simple interpretation, which summarized the high-level results (3331 downloads in the last 30 days) from the text title and correctly interpreted what the chart was about, that it was showing the number of downloads each day. 

It noted that there were noticeable peaks and troughs "which suggest significant variability in daily downloads."  It identified the two spikes in downloads and suggested that it was possible that new episodes were released on those days or that promotional activities increased traffic on those days—both pretty good, if basic, interpretations of the chart. 

It noted that there was a consistent base level of downloads each day too, and suggested that this might be a result of a consistent listener base for the podcast—a stat I wouldn't have thought to look at immediately, but an interesting observation.

ChatGPT also suggested a few other avenues for further analysis, including looking at the days that were low points to see why they were so low, looking for day-of-week trends in the data, and looking at the downloads in different time scales (last 7 days, last 90 days, all time) in order to take the current data in context.

Overall, I was impressed with its interpretation of the information. But when I tested it on a busy marketing dashboard, cracks began to show. It started making mistakes and misinterpreted the numbers. My hypothesis was that I was just giving it too much information at once.

The lesson here is to simplify the input as much as possible.

Asking ChatGPT questions about the diagram

Given a chart, you can ask ChatGPT specific questions about it too. To test it with the Buzzsprout graph, I told ChatGPT that I did a marketing push on Tuesday the 7th and asked if it worked. ChatGPT gave me a positive response, noting that there was indeed a significant spike in downloads on that day.

When I asked how many downloads the podcast got, it did a pretty good job estimating the peak at around 400 downloads, but it struggled reading the corresponding date. ChatGPT's image analysis, at least at the moment, really struggles to read text that isn't perfectly horizontal.  

Tip: If you notice it having trouble reading the text, you can ask it to rotate the graph and then read it. More on that later.

When I asked if the podcast download stats were good, ChatGPT avoided the question and told me it depends as "good can vary widely depending on several factors" including podcast age and reach, frequency of episodes, whether the podcast is niche or mainstream, promotion and marketing. It also told me that engagement also matters, not just reach, and that my own goals play a role in whether or not it's "good." Okay, fair point.

Interpreting other types of diagrams

ChatGPT can also interpret and answer questions about common kinds of diagrams, like organizational charts. 

I tried this with the Des Moines Public Schools Org Chart. It was able to accurately name the departments of the organization and determined whether a specific department (Internal Audit) fell under another specific department (Financial Services).

But I found that ChatGPT sometimes stops explicitly referring to the image you provide it when answering its questions, and as a result it can come up with an incorrect answer. So if you're \asking lots of different questions, make sure you ask it to specifically reference the image. Otherwise, it can rely on its own internal assumptions and "knowledge" instead of actually analyzing the image and giving an incorrect answer. When I explicitly asked it to look at the image again, it was able to figure it out.

ChatGPT can also interpret what the graphic means. If you give it a flowchart, it has a rudimentary understanding on which parts connect and how to interpret this. This gives you the ability to explain complex relationships to ChatGPT in a much easier way than trying to type it all out.

You can use this as a way to input a workflow or process into ChatGPT instead of trying to explain it in words.

Tip: ChatGPT can also write Python code to do more sophisticated image processing.

Getting ChatGPT to help with data visualization & design

ChatGPT has a pretty good understanding of best practices when it comes to data visualization, so if you're making charts and graphs for a presentation, ChatGPT can help in a couple of ways. 

When I gave it three chart options to represent four quarters of gross margin data and asked which one was the best, it gave a nuanced recommendation to present it as a column chart, followed by a line chart (which, it noted, might be misleading because it implies a continuous change in the gross margin quarter to quarter), and lastly a bar chart, noting that this is not really best practice and mostly useful if comparison is the main goal of your analysis. Overall, I thought it was a pretty good recommendation.

Given data, it is also able to create a graph for you using Python code. To test it, I asked it to pull data from a Tableau marketing dashboard. It successfully found the data and graphed it. It did make a minor mistake on the graph labels, saying it was in thousands of dollars when it wasn't.

When asked to make it more visually appealing, it made a few good aesthetic changes, like putting the marketing spend in dollars, centering the labels, and adding gridlines.

From there you can ask it to make specific changes you want, such as alternating different colors. 

But here's what I’ve come to love about the visualization tool: Although it skews positive, it will absolutely tell you if your design sucks, and even give you tips on how you might improve it.  As always, you can feel free to agree or disagree with its feedback, and decide when it's a better idea to get real feedback from real experts.

When I asked for feedback on my rudimentary flow chart above, it gave a lot of constructive criticism, including that it "lacks visual appeal" and it could be "streamlined to reduce visual clutter."  

Helping ChatGPT read your chart: Rotation, rotation, rotation

If you're working with the image interpretation tools in ChatGPT, it's extremely important to get the text to be as close to horizontal as possible since this iteration really struggles to read any text that isn't perfect. I found that reading the dates from a chart tripped it up quite often.

But there’s an easy fix: You can specifically ask it to try rotating the image to better read it. 

Conclusion

Uploading diagrams and charts is a neat way to input information into ChatGPT, and asking it to analyze those images is a handy use of the chatbot. Overall, it's a pretty good data intern and ChatGPT data analysis can give you a head start on any kind of interpretation or analysis task. That ChatGPT can write code to manipulate the image is pretty cool too. I expect this to become even more powerful as we go along.