Is Human Laziness The Real Problem With LLMs?

We should know by now how to use these tools effectively.

The Laziness Problem

Last week, a report from AI detection company GPTZero revealed that over 100 AI-hallucinated citations appeared in more than 50 papers accepted at NeurIPS 2025, an academic AI research conference. The papers cited authors who don’t exist, papers that were never written, and journals that never published them. These weren’t rejected submissions—they beat out 15,000 other papers despite containing fabricated references.

Since the news of these false citations, many people have used this report as a cudgel to once again point out the defects in Generative AI. Generative AI hallucinates, the tools are unreliable, the technology isn’t ready. But that framing lets the actual culprits off the hook: the researchers who were too lazy to use GenAI properly.

The legal profession seems to be having similar problems with more cases accumulating of false citations in legal filings. A French researcher tracking such incidents has now documented over 800 cases where lawyers submitted AI-generated hallucinations to courts. In one recent California case, about nine of the 27 legal citations in a 10-page brief were incorrect in some way, including two completely non-existent cases.

We have known about the defects with Generative AI for almost 3 years, and every “Intro to GenAI” training course discusses its limitations and pitfalls. Heck, when you ask ChatGPT-5 about AI-generated fabrications it tells users that “verification and human oversight are non-negotiable.” Even the AI is telling us to check the output.

The Real Sin: Skipping the Work

What these incidents have in common isn’t the technology—it’s the workflow. In each case, someone used AI to generate output and then submitted it as their own without performing the fundamental task of verification. A researcher didn’t read the papers their paper cited (Remember the AI is not listed as an author, the researcher is). A lawyer didn’t confirm the cases their brief referenced. These aren’t AI or computer failures. They’re human failures.

I’ve written before about “workslop”—the term BetterUp Labs coined for AI-generated presentations, documents, and code snippets that look polished but lack substance. In their survey, 40 percent of respondents said they had received workslop from colleagues in the last month, and cleaning up each incident took nearly two hours on average.

Workslop is what happens when people try to use AI to turn a 40-hour task into 10 minutes of work. That’s not what AI is good for—at least not yet. Effective use of AI can take a 40-hour task and turn it into a 20- or 30-hour task. The AI handles the scaffolding and summarizing, but you still have to do the real work: reading, verifying, thinking, refining. If you skip that work, you get slop and you miss out on developing unique insights.

Writing is thinking. Research is thinking. The process of reading citations and understanding how they support your argument is how you develop understanding . When you outsource that to AI without ever engaging with it yourself, you’re not just risking factual errors—you’re cheating yourself out of benefits of the task and might miss out on some amazing new insight that only you could have thought of.

What Good AI Collaboration Looks Like For Me.

I developed a hands-on teaching activity called “The Blockchain Game” several years ago with then graduate student Riley Coy. Creating it required significant upfront investment: researching MIT’s Beer Game (the inspiration for this activity) and its pedagogical foundations, designing the simulation mechanics, building the PowerPoint slides and handouts, testing it with students and faculty, and refining based on feedback. That was hundreds of hours of work spread across years.

Recently, I wanted to write about the game for two different outlets—an academic paper and a practitioner article for Harvard Business Publishing’s Inspiring Minds. I used Claude to help draft both pieces. I uploaded my directions for the activity, powerpoint slides from the game, and pointed it to news articles and podcasts from when it was released. The AI searched for relevant citations about experiential learning and the Beer Game, drafted sections based on my descriptions, and helped me organize the narrative.

But here’s the critical part: I checked every reference the AI provided. I edited the text extensively, catching sequencing issues and missing details. When the AI suggested citations, I verified they existed and said what the AI claimed they said. I rewrote paragraphs, added nuance from my actual experience running the game, and incorporated recent research findings from a colleague’s study. The AI even helped me clean up my own edits when my revisions introduced grammatical errors.

The result? Two polished pieces I’m genuinely proud of and I would not have gotten done without GenAI given competing demands on my time. It turned a 20 hour task of writing into a 2 hour task. But the quality came from the collaboration, not from outsourcing. The AI accelerated my work; it didn’t replace it. And crucially, I could only collaborate effectively because I had already done the deep thinking and a lot of work. I knew the Beer Game literature. I understood blockchain concepts well enough to teach them. I had run the simulation dozens of times. The AI helped me articulate and organize knowledge I already possessed—it didn’t generate that knowledge from nothing.

Extend, Not Replace

This is what technology should do: extend human capabilities and creativity. A calculator doesn’t replace mathematical understanding—it lets someone who understands math solve problems faster. A word processor doesn’t replace writing skill—it lets a skilled writer revise more efficiently. AI should work the same way.

Unfortunately, LLMs have a flaw, in that it always provides an answer and often states its answer confidently. And an answer that matches with what is statistically the best sequence of words to be associated with each other but not necessarily what is factually correct. [BTW, that is why people are starting to look beyond LLMs and into other forms of AI.]

The problem arises when people try to use AI to do things they don’t know how to do themselves. If you don’t understand blockchain, you can’t effectively collaborate with an AI to write about it. If you haven’t read the relevant literature in your field, you can’t verify whether AI-generated citations are real or hallucinated. If you don’t know what good legal reasoning looks like, you can’t evaluate whether an AI’s brief makes sense.

These tools are most powerful in the hands of people who have already invested in developing knowledge and skill. The AI amplifies what you bring to the table. If you bring nothing, you get nothing worth having.

The AI Intern

Think of AI like a very capable intern. You wouldn’t hand an intern a research project, accept the results without reading them, and then put your name on the report—at least not if you wanted to keep your job. Yet that’s exactly what professionals are doing with AI every day.

A good intern can dramatically speed up your work by doing initial research, organizing material, and drafting outlines. But you still have to review the work, verify the facts, and ensure the final product meets professional standards. The intern’s efficiency doesn’t eliminate your responsibility; it frees you to focus on the higher-order thinking the task requires.

When professionals submit unverified AI output, they’re not being innovative. They’re being lazy and deserve to be called out.

Misaligned Incentives

Getting back to the case of the researchers submitting hallucinated citations in papers accepted at NeurIPS. The deeper problem is that the academic system reward quantity over quality, which encourages or forces lazy work. I am sure the researchers who were caught with hallucinated citations weren’t lazy people—they were just operating in an academic environment where there is pressure to be constantly publishing

The same dynamic plays out in business. Organizations push employees to adopt AI tools and demonstrate productivity gains, often without investing in training or establishing clear expectations for quality. What they get is slop—faster, but worthless.

A 2023 study I’ve cited before in this Substack, “Navigating the Jagged Technological Frontier: Field Experimental Evidence of the Effects of AI on Knowledge Worker Productivity and Quality,” found that using ChatGPT-4 enabled consultants to finish 12 percent more tasks, complete them 25 percent faster, and produce work that was 40 percent higher quality than colleagues without AI access. The key word is “quality.” Those gains came from using AI as a tool within a professional workflow—not as a replacement for professional judgment. And that was three years ago, ancient history in the world of AI.

AI is a remarkable tool for those willing to use it properly: as an amplifier of existing expertise, an accelerator of genuine effort, an extension of human capability. But for those hoping to skip the work entirely? It’s just a faster way to produce garbage.

What are you seeing in your industry? Let me know in the comments.

---

Book Me for Your Next Conference or Event

After nearly two decades teaching and building innovative programs at the University of Missouri, I’m now focusing more of my time on keynote speaking and workshops—helping organizations make sense of the emerging technologies reshaping work, learning, and leadership.

If your organization, conference, or leadership team is exploring how to navigate AI and emerging technologies responsibly—and productively—I’d love to talk.

For more details, topics, and booking information click here.

---

📥Recent Talks, News and Updates

• I gave several talks recently (Chamber of Commerce, Red River Estate Planning Council, Association of Government Accountants, etc.), and I have compiled a master list of all the studies and articles that I cited or used in these talks, organized by topic. Check it out here: www.profc.io/ai-links

• I presented a “Lunch and Learn” at the Daniel Boone Library in collaboration with the League of Women Voters on “AI and Democracy — Guardrails We Can Choose” on December 10th at Noon. You can view the video below.

👍 Products I Recommend

• Products a card game for workshop ideation and ice breakers (affiliate link). I use this in my workshops and classes regularly. Made by a former Mizzou student Aaron H.

📆 Upcoming Talks/Classes

• I am teaching a short course “Thinking Machines, Changing Minds: How AI Is Shaping Work and Wisdom” for Osher about AI during the Winter 2026 semester. Details and registration are available here.

• I will be presenting “AI Externalities” at Law, Technology, and Society: Charting the Next Frontier symposium on the MU campus on April 15th. Details coming soon.