The Economist

‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌ ‌

April 16th 2024

SUBSCRIBER ONLY

Off the Charts

The best of our data journalism

Alex Selby-Boothroyd
Head of data journalism

The date for Britain’s next general election has yet to be decided, but the outcome for the ruling Conservative Party is looking like a done deal. Our poll tracker currently puts them 20 percentage points behind Labour and it is almost inconceivable that they would win a majority of seats when the country does go to the polls. Pundits disagree on whether they will suffer just a heavy defeat or a total wipeout. To give us more insight we built a prediction model trained on historic voting behaviour between 1959 and 2019. Even without taking into account tactical voting, which could end up hurting the Conservatives even more, it makes for sober reading for prime minister Rishi Sunak and his 346 Conservative MPs: our median prediction is that they would win only 198 seats if an election were held tomorrow.

More of our data reporting this week:

Michelle Hennessy looks at the recent phenomenon of crowdsourced UFO sightings. Around 15,000 objects have been spotted in the skies over the past 18 months, mostly in America.
Ahead of the Indian elections which start on Friday, Leo Mirani takes us on a cartographic tour of 2,500 years of Indian history.
Sub-Saharan Africa is experiencing a baby boom. Ann Hanna finds that by the end of this century, half the world’s children will be born in the region.
Below, members of the data team offer advice on using AI—or more specifically, large language models like ChatGPT—in your data work.

For any advice, questions or feedback for our data team, send an email to [email protected]. Finally, next week we will launch Middle East Dispatch, our latest newsletter, bringing together our coverage of a complex and consequential region. You can sign up here.

**Advice from The Economist data team™️—Using AI for data analysis**

Alex Selby-Boothroyd, Head of data journalism

AI is great at tackling the small, tedious tasks that can free up time for more valuable things. Despite working with spreadsheets for longer than I care to remember, I still rarely get a complicated formula right on the first—or even fifth—attempt. But ChatGPT has no problem keeping track of how many brackets to use and where they go. I recently wanted to extract the date from a bunch of URLs like this:

economist.com/graphic-detail/2024/02/09/five-reasons-why-indonesias-election-matters

I gave Bing’s Copilot this prompt:

I have a URL in cell A1 that contains a date in the format YYYY/MM/DD that always starts after the second forward slash. Please give me a formula that extracts the date and converts it to the format DDD DD MMM YYYY

And five seconds later it returned this formula, which worked perfectly first time:

=TEXT(DATEVALUE(MID(A1, FIND("/", A1, FIND("/", A1) + 1) + 1, 10)), "ddd dd mmm yyyy")

Ainslie Johnstone, Data journalist

Whenever I ask AI to help with coding problems I always ask it to give its suggestions in my preferred language and coding style—I’ll say “solve this problem, in R, using tidyverse”.

Even if the AI’s first attempt runs without any errors (which it often does not), it is really important that I am able to check that it is doing what I want. And, if the AI’s first solution doesn’t work, seeing the suggestion in a format I understand helps me troubleshoot and sometimes even sparks an idea for how I might do it better.

Andreas Moor, Interactive data journalist

Generative AI can be used to explore the syntax of programming languages or data API resources—this is especially useful for obscure tools or ones that you rarely use. Instead of having to read through reams of technical documentation, it allows you to take a “bottom-up” approach: ask for a concrete example that you can then adapt to your needs and learn how the tool works along the way; or familiarise yourself with the tool by having the AI explain a code snippet to you (see below). What’s more, the chat interface also allows you to ask follow-up questions.

Evan Hensleigh, Special projects editor

Though my colleagues have given some good examples above, I remain sceptical about using generative AI for data analysis. If you’re going to use large language models (LLMs) for data work, you need to be aware of the limitations and likely pitfalls. LLMs are good at regurgitating, remixing or repackaging information they’ve seen, but they can’t process data or make insights. Ask one for the three most interesting takeaways from a table of data and its answer may well be poor, made up or both. And remember that LLMs are fairly bad at maths: these models won’t be able to calculate the trend from a set of points, but they will be able to give you the Python or R code to do so correctly.

LLMs can also be useful for structuring unstructured data. An LLM can process a large body of text and organise it into topics. It will get some things wrong, but probably not enough to throw off your analysis. And modern LLMs, like GPT4 Vision, are reasonably good at reading data from PDFs or images and converting them into a machine-readable format. But remember to always check their work carefully.