Theresa Wilkinson

Articles & Publications

A Polished AI Analysis Can Still Be Wrong

I analyzed my survey results before I gave the data to ChatGPT. My initial analysis focused on reviewing the responses, calculating the basic results, and coding the open-ended answers. I did not conduct statistical tests during that first review. I asked ChatGPT to help with those tests and provide another analysis of the data.

Some of it did. I used ChatGPT to help identify and work through statistical tests, then independently verified the calculations and interpretations. It also helped organize the results and draft text for the presentation.

But then I started checking the work.

I found incorrect percentages, means, rankings, and totals. Some open-ended responses were coded incorrectly or undercounted. Categories changed from one explanation to the next. In one case, a finding from one survey question was copied into the slide for another question.

The results looked professional. They were presented in organized tables, charts, and confidently written findings. Nothing about the presentation suggested that the information might be wrong.

That was the problem.

I Had Already Seen This in My Survey

My survey asked 34 UX professionals how much they trusted AI to help with different UX research activities. The strongest finding was not that they refused to use AI. Most were already using it.

However:

The participants were willing to use AI, but they expected to check its work. That is exactly what happened to me.

The Mistakes Were Easy to Miss

An obvious error is usually easy to catch. These errors were harder because they appeared inside polished work. A table contained exact numbers. A ranking placed every activity in order. A paragraph explained what the results meant. The information looked finished.

But some of it was not correct.

One incorrect number can also create several additional problems. It can affect a percentage, change a ranking, alter a chart, and lead to an inaccurate key finding. If that finding is copied into the executive summary or an article, the original error becomes much harder to locate.

I caught the problems because I had already analyzed the responses and continued comparing the AI-generated results with the source data.

If I had accepted the polished output without checking it, some of those mistakes would have remained in the final report.

AI Still Helped

This experience does not mean AI was useless. It helped me:

But I could not assume that the calculations, coding, or interpretations were correct.

Sometimes AI saved time. Sometimes checking and correcting its work took more time than doing the work myself.

What I Check Now

I have learned to verify:

I also compare the written finding with the data displayed on the same slide. If the chart and the finding do not tell the same story, I go back to the source information.

The Researcher Is Still Responsible

AI can calculate, organize, and draft. It can also make mistakes while sounding completely confident. The person conducting the research is still responsible for deciding whether the evidence supports the conclusion.

That is what the UX professionals in my survey said, and it is what I experienced while analyzing their answers.

View LinkedIn Version

Back to Portfolio

Contact

Email: theresaw@columbus.rr.com

LinkedIn: theresa-wilkinson