Feature
Artificial intelligence (AI) is being used by data science teams to work on both enterprise data strategy and their day-to-day tasks working with data.
As data scientists work with math, statistics and software coding, they turn to machine learning (ML) and AI to help analyze, predict and leverage data for the business.
Here, a group of data science pros share with us just how AI is helping them do their jobs.
Programming Assistance
By far, the most prevalent AI use case for data science they mentioned was writing code — within limits.
“I use AI mostly to help scaffold code and to outline documentation,” said Grantham Taylor, a principal associate in applied AI research at Capital One.
“I do not necessarily trust it to do any of these things accurately, but it is normally good enough to start. I would never build an entire project using AI — only tiny little pieces here and there.
“Similarly, I will also use AI to scan my projects and recommend documentation and occasionally recommend ways to simplify it otherwise. Once again, it is almost always wrong, but I am able to validate the changes afterwards, so it never really hurts.”
Marcus Elwin, a data scientist and machine learning engineer at Pocketlaw, said the company’s back-end and front-end is on Typescript.
“I'm mostly proficient in Python but have now been able to do more back-end and front-end work using tools such as Augument, GitHub Copilot and cursor,” Elwin said. “Apart from that, I use both ChatGPT and Claude for various queries, whether it be for small code snippets or explanation of legal terms.”
That said, AI can be useful for automating rote data science tasks, said Artem Shelamanov, a machine learning engineer at Neuroforge and the CTO of Yazero.
“Some steps are mostly consistent across different projects — , cleaning, processing and so on,” Shelamanov said. “If you're working on a computer vision problem, you would load pictures from a folder, re-size them, crop them, rotate them, process color channels, convert them to arrays and so on. These steps are almost always the same for these kinds of projects. I like to outsource these tasks to AI, because rewriting them every time is both boring and time-consuming.”
However, using AI to help generate code can lead to a data scientist not actually understanding what the code does, Shelamanov said.
“Newcomers to the field often use AI too much instead of researching and thinking for themselves,” Shelamanov said. “This can result in impressive projects on their resumes that they didn’t truly understand or code themselves. For recruiters, this can make such candidates appear more competent than others who actually learned and understood the concepts but didn’t use AI as much.”
Moreover, a data scientist who doesn’t understand what their code does could have more trouble debugging and adapting it, Shelamanov said.
“It’s important to verify the code properly, even if it seems to work at first glance,” Shelamanov said. “For example, you might copy a data processing step, check the output and assume it’s correct without thoroughly reviewing the code. This could lead to bugs that are difficult to trace later.”
Taylor said if ChatGPT is “barely able to take the square root of a number, would you really trust it to calculate the marginal probability of millions of customers defaulting on their loans, given some recent changes to their behavior?”
Related Article: 10 Top AI Coding Assistants
Automating Workflows
Xenia Mountrouidou, principal cyber data scientist at Expel, a managed detection and response provider, said she uses AI to automate security workflows for analysts.
“Using language models, we generate pipelines that write detection rules for detecting malicious behavior,” Mountrouidou said.
“The models that we use are foundational with some prompt engineering, such as chain of thought.
“I also use some packages, such as Pandas AI, to analyze and explore data as a black box.”
Related Article: AI's Transformation of Workflow Automation: Current Trends and Applications
Searching and Writing
Data scientists are also using AI to generate text and as a search engine.
“AI, or generative AI built from large language models, has been integrated into many of the tools I use, including search engines and text editors,” Miller said.
“So it is hard not to use AI to some extent. I use grammar and spelling checkers in my writing. I do not ask AI to write anything from scratch. I much prefer to put things in my own words.”
Miller also uses AI tools for more sophisticated search.
“I am looking for information, and I want to get that information efficiently,” Miller said. “AI-assisted search is often better organized than conventional search.”
Elwin said “when I need to do more structured search, I tend to often use Perplexity now instead of Google.”
Shelamanov agreed, saying Perplexity is “ideal for quick research.”
“It browses the web, finds various articles, websites and social media posts, and creates nice summaries,” Shelamanov said.
He also uses AI for writing text.
“Since English is not my native language, they help me catch grammar mistakes and improve phrasing to sound more natural,” Shelamanov said.
Related Article: Will AI-Powered Search Engines Ultimately End Traditional Search?
Disrupting Education
Data scientists working in the higher education field have a particular challenge, as students are quick to learn how to use a tool to their advantage.
“I have students solve problems on their own,” said Thomas W. Miller, faculty director for the master of science in data science in the School of Professional Studies at Northwestern University.
“After solving problems on their own, I ask that they set up a session with an AI assistant, such as ChatGPT, to see if the assistant can solve the problem. Students submit complete transcripts of their sessions with the AI assistant. It is surprising what AI assistants can do. Coding copilot tools can help a lot with programming tasks.”
Miller isn’t alone.
“Many data science faculty have had to modify discussion questions and assignments so that students are doing the work,” Miller said. “We have to test discussion questions and assignments to ensure that an AI assistant cannot do the work.”
Cautionary Views
Data scientists tend to be aware of the potential downsides of using AI as well, and a number worry that some users are getting too dependent on AI.
“Some organizations are far too trusting of AI,” Taylor said. “They might assume that it is extremely powerful, so if I give it all of my data, it will provide great results — without considering that, at the end of the day, it is largely just auto-complete.
“Sentiment analysis has been around for a long time, and no solution is perfect, but some organizations might choose to just engineer some prompt that requests an LLM to look at each body of text and output a number. And just like that, because it works, it is considered good enough, whereas there is no validation or review of performance, other than just eyeballing the results.”
Several data scientists also mentioned security issues.
“It’s tough to use AI because of privacy concerns, whether these are clients or personal data,” Mountrouidou said. “So I use AI with models that I deploy locally or in the cloud — I do not use the publicly deployed models with sensitive data.”
Ultimately, it’s important to remember that AI is a tool in data science.
“We need to ensure that creativity is not stifled by a heavy reliance on AI,” Miller said. “We need to protect what is personal, individual and human.”
Related Article: Why CIOs and CDOs Need to Work Together on Generative AI
