David Colburn

I’ve written before about the dangers of artificial intelligence and the large language models powering systems like ChatGPT. These systems can produce writing that sounds polished, persuasive, and completely believable while quietly inventing facts, sources, and entire realities with the confidence of a guy at the end of the bar explaining how he once almost landed a world-record walleye.
And yet, despite those very real risks, AI systems operating with human oversight are rapidly becoming embedded throughout business and research. Hospitals use them to summarize medical records and assist with documentation. Law firms use them to review contracts and legal filings. Researchers use them to sort through enormous amounts of scientific literature and data that would otherwise consume vast amounts of human time.
Journalism has hardly been standing on the sidelines during this transition. Newsrooms across the country, including the Minnesota Star Tribune, have been experimenting with AI tools to assist with research, organization, and workflow tasks while emphasizing the need for human oversight and verification. Like much of the industry, I watched those developments with equal parts fascination and caution.
So naturally, being a journalist with deadlines, curiosity, and perhaps an overdeveloped belief that I can outsmart machinery, I incorporated ChatGPT into parts of my own workflow.
I have found it genuinely useful as a research assistant and organizational tool. I use it to help search the web for relevant background sources on complicated topics. I use it to summarize lengthy court filings and legislative bills. I use it to help organize meeting coverage into logical outlines before I start writing. I use it to help format repetitive information from statistical data, including sports results, honor rolls, and scholarship lists. In many ways, it’s like an eager, very fast intern who never sleeps, occasionally says brilliant things, and sometimes confidently walks into a wall.
But ChatGPT only knows words. It does not know stories, no matter how sophisticated its language prediction becomes. It has no emotional understanding of what it is processing. Every good story, whether it’s a football game, a school board meeting, or a city council debate, is infused with human emotion and human context. ChatGPT can read a softball box score, but it has no idea which hit shifted momentum or how the crowd reacted. It can summarize a transcript about budget cuts, but it cannot genuinely understand the fear, anger, tension, or heartbreak surrounding discussion of teacher layoffs or program reductions.
That distinction matters. AI can assist good journalism in limited ways, particularly with organization, formatting, research assistance, and data-heavy tasks. But it is no substitute for reporting, observation, judgment, curiosity, skepticism, or human storytelling. It cannot do my job, and it never will.
The first time I used ChatGPT to summarize a court brief, it missed two of the six states specifically mentioned in the filing. A fifth grader could have found all six, but not ChatGPT. That was enough to immediately teach me a lesson. Since then, I check ChatGPT summaries against the source documents themselves. If it gives me a research source, I verify the source exists. If it summarizes a bill, I compare the summary against the actual language. If it creates an outline for a meeting story, I check it against my notes, transcripts, and agenda packet.
And over time, something else happened, too.
I grew comfortable with it.
That’s the part I think people still underestimate about large language models. Even when you know intellectually that they make mistakes, the conversational nature of the interaction quietly creates a sense of trust. You ask a question, it responds fluently. You refine the question, it adapts intelligently. You joke with it, it jokes back. It starts to feel less like software and more like collaboration. And because the answers are often useful, and most often correct, your brain gradually stops treating every response with the skepticism you’d normally apply to an unfamiliar source.
Meanwhile, the journalism world has been learning its own lessons. AI-generated stories and features published by outlets including CNET, Gannett, Sports Illustrated, the Chicago Sun-Times, and the Philadelphia Inquirer have all faced controversy over factual errors, fabricated material, or questionable AI practices. In some cases, the AI hadn’t merely gotten details wrong. It had invented reality.
And somehow, despite watching all of that unfold, I still managed to fall into the exact trap I understood well enough to warn others about. And in doing so, I created problems for students, families, school staff, and community organizations that should never have happened.
Recently, I used ChatGPT to convert a North Woods School scholarships spreadsheet into a printable list for publication. On the surface, it seemed like one of the simplest possible uses for AI. I wasn’t asking it to analyze constitutional law or summarize a thousand-page budget bill. I was taking names and scholarships already contained in a spreadsheet and converting them into readable text.
Except this time, ChatGPT hallucinated entries. It invented scholarships and recipients that did not exist in the original spreadsheet. And because I failed to follow my own normal verification process, because I trusted that such a simple formatting task could not possibly go sideways, those inaccuracies made it into print.
That failure belongs to me.
Not the school. Not the spreadsheet. Not some mysterious gremlin hiding in the newsroom computers. Me.
I should have checked the generated list line-by-line against the source document before publication, exactly the same way I normally verify AI-generated summaries, outlines, and research results. Had I done that, the errors would have been caught immediately. Instead, my lapse created a mess that misrepresented scholarships actually awarded to students and inaccurately identified some of the organizations and donors involved. Scholarship recognition should be careful, accurate work because it reflects real accomplishments and real generosity within a community. In this case, I failed to meet that standard.
What makes this especially frustrating is that I already understood the danger. I’ve written about it. I’ve talked about it with readers and colleagues. I knew these systems can fabricate details. But familiarity breeds confidence, and confidence breeds complacency. Somewhere along the line, I stopped treating that particular task with the caution it deserved because AI had become normalized inside my workflow.
That normalization is happening everywhere right now.
The lesson here is not that AI is useless. Frankly, that would be absurd. The technology is already too useful, too integrated, and too widespread for that argument to survive contact with reality. The lesson is that large language models are not databases, calculators, or traditional software programs. They are predictive language systems that generate plausible responses based on patterns. Most of the time those responses are impressively good. Sometimes they are horribly wrong. And the more natural the interaction feels, the easier it becomes to lower your guard.
That’s the sword I’m falling on this week.
Not because AI made mistakes. We already knew it could do that.
Because I knew it could make mistakes, and I stopped checking anyway.