As an educator, I find myself in a constant battle with my students to not rely too much on AI. It’s not just that I want them to continue thinking for themselves. It’s also the fact that, you know, AI is wrong a lot. Yet, what it spits out looks so well-formatted that it gives all who use it a false sense of security. Don’t believe me? Let me give you an example from my own recent work.

I am currently working on an issue brief for The Center for Retirement Research. The question is simple: how has employment for older workers — defined in my brief as those ages 55 plus — recovered since the pandemic? One group I am looking at is those with a bachelor’s degree or more. Little research has focused on how this group’s employment has evolved since COVID, so I figured I might start by asking Chat-GPT what it thought.

CHAT-GPT’s AI Response

Here is my prompt to Chat-GPT and then the first part of its lengthy response:

OK, pretty clear answer. Well formatted. It’s also completetely wrong.

The Actual Data

To answer my question, I went to The Current Population Survey and downloaded data between 2016 and 2025 on the employment of workers ages 55 plus. I did two different analyses. First, I just compared the percent change in employment among these workers relative to Q4 2019, the last quarter prior to the pandemic. Then, I also did a “regression-controlled” comparison. This regression was designed to account for the fact that even within the 55 plus age group, the average age is increasing. Since older people work less, without controls, it could look like employment hasn’t recovered, when apples-to-apples it actually has. The figure below shows the results for college-educated workers. Either way, they are working less, not more.

Figure. Employment Relative to Q4 2019 for College Educated Workers Ages 55 Plus

Note: Both raw and adjusted deviation calculated using a regression with quarterly indicators and Q4 2019 as the base case. The “Regression-adjusted” line additionally controls for age indicators, race (Hispanic, Black, Asian, Other), gender, and any physical or cognitive difficulty.
Source: Author’s calculations from the IPUMS-CPS, 2016 – 2025.

What’s more, another part of Chat-GPT’s answer is wrong too. It said that the recovery among college-educated workers has been greater than for less educated ones. In fact, when I look at the data, those with a high school degree or less are actually working more than they were before the pandemic once their aging is taken into account. To be sure, those individuals still work less than those with a college degree — they always have — but their employment has recovered more since the pandemic.

The Message: Be Careful When Using AI…It’s Developers Won’t Be

I’ve written before about a simple fact in the development of AI. It’s developers have no incentive at all to be careful. They will stop at nothing to be the winner in the race to develop this technology.

This fact leaves it up to the users of AI to be careful. Every semester I have my students “fact-check” an AI chat bot. What they have found is that these bots are good at answering basic, well-established facts. But, they struggle with things that are less well known, like the issue I am writing about. Yet, it sounds so confident.

The pretty formatting of AI chat bots gives them an aura of authority. Be careful — when it comes to emerging data, AI doesn’t earn the I.