AI Can’t Do Your Data Job

Human judgment remains crucial in data-related jobs despite AI advancements, says Dr. Jignesh Patel, co-founder of DataChat and Professor at Carnegie Mellon University.

5 Min Read
artificial intelligence digital illustration concept with human heads
Alamy

On Monday, April 1, The Daily Show host Jon Stewart questioned the AI industry’s promises to solve climate change and cure cancer. As Stewart wryly observed, AI leaders seem to be more interested in replacing people with AI for the sake of profit margins. “We have, as a society, been through technological advances before, and they all have promised a utopian life without drudgery,” said Stewart. “And the reality is they come for our jobs.”

In the U.S., many of the 1.3 million people working in data-related occupations share Stewart’s concerns. How can they not when every consultancy and bank predicts an AI job apocalypse? Analyst group Forrester, for instance, has estimated that 53% of computer and mathematical professions could be lost to AI by 2030.

Meanwhile, data jobs are among the fastest growing in the U.S. Indeed, the U.S. Bureau of Labor Statistics estimates that data science roles will grow 36% between 2021 and 2031.

Though AI will eliminate certain occupations, I think we overestimate its capabilities. The future lies somewhere between the miracles AI leaders promise and the labor upheaval Stewart decries. Automating a process or workflow with AI doesn’t eliminate the need for human judgment. 

A Task Is Not a Job

At Carnegie Mellon University, where I have the privilege to teach data science, you wouldn’t know that the data job apocalypse is nigh. Across majors, data science and statistics classes are among the most popular on campus. If data work is so threatened, why would students bother learning about it? 

Related:How CIOs Can Integrate AI Among Employees Without Impacting DEI

There is a difference between how data science teaches people to think and what data workers do to solve problems. Data science is discovering facts and identifying falsehoods in a digital society. From business to academia to politics, data analyses are used (and abused) to shape beliefs and decisions. Data science is therefore a tool for navigating life in complex civilizations.     

Sure, AIs increasingly perform the skills a data scientist, analyst, or engineer uses at work. For example, it can write in SQL and Python, coding languages long used in data science

Is coding then a useless skill? No. Data scientists need a background in those languages to reason about algorithms, code patterns, and models. Moreover, the data professional is still responsible for directing and vetting AI’s work so that it doesn’t lead a whole organization astray. As more and more organizations depend on AI, America’s 1.3 million data workers will be more and more essential. 

Related:Artificial General Intelligence: Are We There Yet?

LLMs Are Needy

The best-known LLMs, like GPT-4 and Gemini, are generalists that ingest broad information from the web. Yet human training was and still is integral to developing LLMs. They are needy tools that rely heavily on data-savvy professionals.

Leaders at OpenAI have said that reinforcement learning from human feedback (RLHF) was integral to training ChatGPT. Essentially, human trainers read GPT’s responses to prompts and then indicated which response was most preferable. They also rated responses on other criteria, like truthfulness (where it continues to struggle). Firms train LLMs this way because AI doesn’t “think” for itself. It merely synthesizes information in ways that sound intelligent, based on our feedback. 

Especially in technical, niche cases, LLMs cannot just learn from the internet. Think of all the inscrutable processes in medicine, finance, insurance, and HR that are largely undocumented and specific to a single organization. In those cases, human beings need to provide relevant prompts and accurate responses and correct AI as it learns.

Eventually, AI can come up with its own examples to learn from, but even then, it still needs a human in the loop. AI will only be as good as the data it ingests and the data professionals who coach it.

AI Isn’t Curious

AI is not an independent being that can fill interesting job roles. It’s a technology with an insatiable appetite for human-derived knowledge. When using AI for data analysis, people decide what data is worth crunching, whether it’s a dataset of app usage metrics or physical readings from sensors. In deciding what to measure, we assign meaning to data and form the arena in which AI plays. In a way, AI’s potential is a measure of our own ingenuity, creativity, and curiosity.

Progress and disruption in every field of life—AI included—requires crazy geniuses drawing upon emotion, experience, relationships, and serendipity to do and think the unexpected. They are not in the business of synthesis and parroting, like an AI. An AI “scientist” would be stuck ruminating about things we already know.

In the foreseeable future, we need people to question and challenge conventional knowledge, subject their ideas to scrutiny, and engage in science with courage. Those individuals are the lead actors. AI is their stage prop.

Don’t Be a Process

If you work in data and your sole duty is to execute rote, repetitive processes, then yes, AI might make your role obsolete. The more you exercise judgment at work, the less you should fear AI. 

Soon enough, everyone in your workplace will use AI to prep and clean data, run analyses, and create visualizations. You can’t compete with AI on the volume of analyses completed. Again, though, AI doesn’t know what those patterns mean, why they’re important, or how an organization ought to act upon them. AI’s usefulness in the real world depends on us.

AI performs tasks but does not destroy jobs. Only people can choose to do that.

About the Author

Jignesh Patel is a professor in the Computer Science Department at Carnegie Mellon University and the co-founder of DataChat, the no-code, generative AI platform for instant analytics. Patel’s research interests include analytics, AI, and scalable data platforms. He has supervised over 20 Ph.D. students, and his research papers have been selected as the best papers in several top database venues, including SIGMOD and VLDB. He is a fellow of the AAAS, ACM, and IEEE organizations. Additionally, he has received teaching awards at the University of Wisconsin and the University of Michigan, where he previously served as a professor. He is keenly interested in technology transfer from university research and has spun out four startups from his research group.

Sign up for the ITPro Today newsletter
Stay on top of the IT universe with commentary, news analysis, how-to's, and tips delivered to your inbox daily.

You May Also Like