This one has been living rent-free in my head for a while.
Back in The Convenience Trap I said it flat out: AI doesn’t hallucinate. Not really. I want to push that further today, because I think the word isn’t just wrong. I think it’s actively making people worse at using AI.
Here’s my take: when an AI agent gives you garbage, it’s usually not broken. It’s unmanaged.
Let’s Get the Word Right First
Go look up the actual definition. Merriam-Webster’s medical dictionary defines a hallucination as a sensory perception, like a sound or an image, that happens with no actual external stimulus, usually caused by a neurological disturbance or by drugs.
So: senses. Brain chemistry. Your own brain tricking you into perceiving something that isn’t there.
An LLM has none of that. No eyes, no ears, no neurotransmitters, no bad batch of mushrooms. It isn’t perceiving anything, wrongly or otherwise.
I’m not the only one annoyed by this. A 2023 paper in PLOS Digital Health argued that calling these errors hallucinations implies the model is perceiving, and that this misrepresents what’s actually going on. Neurologists Gerald Wiest and Oliver Turnbull made the same case, arguing confabulation is a more precise medical term. A working paper out of Suffolk University Law School went further, pointing out that LLMs process tokens and don’t experience anything, so the word simply doesn’t fit.
But here’s where I part ways a little, even with the “confabulation” crowd. Confabulation still frames it as a disorder. I don’t think it’s a disorder at all.
AI says things based on what it knows, what it’s been told, and what it learned. That’s it. And just like a human can learn the wrong thing and say the wrong thing with total confidence, so can AI.
We don’t say your new co-worker “hallucinated” when they got the process wrong on day three. We say nobody showed them how it’s done here.
What’s Actually Going On: The Straight-A Student
This isn’t just my opinion. In September 2025, OpenAI and Georgia Tech researchers published a paper called Why Language Models Hallucinate. Their opening analogy is basically my whole argument: like students facing hard exam questions, models guess when they’re unsure instead of admitting it.
Why? Because that’s how we trained and graded them. Most benchmarks score accuracy only. A blank answer gets zero. A guess might get points. Their example: ask a model for someone’s birthday it doesn’t know, and guessing “September 10” gives it a 1-in-365 shot, while “I don’t know” guarantees a zero. So over thousands of test questions, the guesser looks better on the scoreboard.
Read that again and tell me it doesn’t describe a brilliant university student who has aced every exam and never held a job.
That’s the mental model I want you to walk away with: AI is a really smart graduate with zero experience. Wildly capable. Has read everything. Has never once worked at your company, on your codebase, with your customers, under your constraints.
Meet Your New Junior Hire
So think about how you’d treat that grad on day one.
You wouldn’t hand them a laptop, say “build the thing,” walk out and come back Friday expecting production-ready work. (Except, that did happen to me, and I learned a ton by doing that… maybe we have something here, then again maybe not) You’d:
- Explain the goal and why it matters
- Hand them the blueprints, the standards, the “how we do it here”
- Point them at the right information, not all the information
- Check in early, before they’ve gone 40 hours in the wrong direction
- Review their work, give feedback, and let them try again
Now look at how a lot of people use AI agents. One-line prompt. No context. No examples. No definition of “done.” Fire it into the wind, YOLO, then get pissed off when it comes back with something confidently wrong and call it a hallucination. There is an entire swath of people running around talking about REDUCING prompts – I just can’t get behind that.
That’s not an AI failure. That’s a management failure.
Interestingly, the people building these models are saying the same thing in engineering terms. Anthropic’s guide to context engineering for AI agents describes the job as finding the smallest set of high-signal information that gets you the outcome you want. They even warn about “context rot”: stuff too much in, and recall gets worse. Their advice is to write instructions at the “right altitude”, not so rigid it’s a script and not so vague it’s useless. Isn’t that how we talk to our resources?
That’s… just a good brief. It’s what every decent manager does with a new hire. Developer Manveer Chawla put it nicely in his junior-developer workflow: you wouldn’t drop a 500-page manual on an intern’s desk and expect them to remember page 412.
I know this one from experience, because I was that guy.
When I first started building with AI, I called myself an AI Slop Builder, and honestly, the name was earned. I was pasting code out of chat windows into my editor and losing track of what changed after dozens of prompts. The AI forgot decisions we’d made the session before. My files were scattered across browser tabs and folders named things like “project-final-FINAL-v3.” And when it gave me something broken? I blamed the AI.
Looking back, the AI wasn’t the problem. I was a terrible manager. No brief, no memory, no plan, no review. I was handing a brilliant new hire a sticky note and walking away.
So I learned. I started writing a short GitHub issue before asking AI to build anything, so every task came with a real assignment and a permanent record of decisions. I stopped treating the chat window like a code vending machine and told it to act as a manager that plans the work and delegates to focused subagents. Then I wrote the rules down in a CLAUDE.md file so every agent knew “how we do it here” from day one, and kept them simple on purpose.
Same AI. Different leadership. Not long after, I shipped something real: LilySPOT, open-source firmware for a LilyGO mobile hotspot, built by a Manager, Planner, Coder and Tester agent team working from that CLAUDE.md.
The models didn’t suddenly stop “hallucinating.” I started managing. That’s what turned my slop into a meal.
Garbage In, Garbage Out Never Left
We’ve been saying GIGO in IT since before most of us were born. AI didn’t repeal it. It just made the garbage look really polished.
| What a junior employee needs | What that looks like for an AI agent |
|---|---|
| Onboarding and company context | System prompt, project files, a CLAUDE.md or README |
| A clear assignment with a definition of done | A specific task with acceptance criteria |
| Blueprints and examples of good work | Specs, templates, and worked examples |
| Access to the right resources | The right tools, docs and data, not everything |
| Check-ins before going too far | Plan first, approve, then build |
| Code review and feedback | Tests, reviews, and iteration |
| Permission to say “I’m not sure” | Explicit instructions to ask or flag uncertainty |
That last row matters more than people think. Remember the OpenAI finding: models guess because we rewarded guessing. So tell yours it’s okay not to know. Tell it to ask clarifying questions. A good manager makes it safe for a junior to say “I don’t know yet”. Do the same for your agent.
So Are Good People Managers Better at AI?
This is the question I really wanted an answer to. Turns out researchers have actually tested it.
In an NBER working paper called Measuring Human Leadership Skills with AI Agents, Ben Weidmann, Yixian Xu and David Deming (Harvard) had leaders run collaborative problem-solving tasks with teams of humans and with teams of AI agents. The result:
- Leaders who did well with AI teams also did well with human teams, with a correlation of 0.81.
- Even after controlling for “hard” skills like task ability and fluid intelligence, it was still 0.69. In other words, this is about soft leadership skills, not just being smart.
- The behaviours that predicted success were the same in both cases: leaders asked more questions, had more back-and-forth conversation, and used more “we” and “us”.
Let that sink in. The skills that make you good at leading people are measurably the same skills that make you good at leading AI.
Wharton professor Ethan Mollick landed in the same place from a different direction. In Management as AI Superpower he describes a four-day class where executive MBA students, mostly non-coders, built working startup prototypes with AI agents. His conclusion is that working with agents comes down to “management 101 skills”: knowing what good looks like, setting clear goals, and giving useful feedback. He ends with a line I love: the skills we write off as “soft” turned out to be the hard ones.
UCSB professor Paul Leonardi, speaking to ServiceNow’s Workflow, framed it in a way that should be on a poster: AI agents don’t eliminate the need for management, they relocate it. Human work leans on shared experience. AI needs explicit instruction.
So is it true that experienced leaders get better outcomes from AI? The evidence says probably, yes. Not because managers are magic, but because they’ve spent years practicing the exact skill AI demands: getting what’s in your head into someone else’s actions.
The Catch: Manage It Like a Junior. Don’t Blame It Like One.
I’d be doing you a disservice if I didn’t include the counterpoint, because it’s a good one.
In May 2026, BCG Henderson Institute researchers published Why You Shouldn’t Treat AI Agents Like Employees in Harvard Business Review. In a study of more than 1,200 managers, when AI was framed as an “employee” instead of a tool, managers caught 18% fewer errors and individual accountability dropped by 9 percentage points. People started quietly treating the AI as the one responsible.
At first glance that sounds like it contradicts everything above. I don’t think it does. It sharpens it.
There’s a difference between managing AI like a junior and anthropomorphizing it like a co-worker. Give it a name, an email address and a spot on the org chart, and people stop reviewing its work. Give it a brief, a blueprint and a review process, and the work gets better.
A Microsoft study of product managers put it bluntly: accountability must not be delegated to non-human actors. A junior can’t be fired for your bad brief, and neither can an agent. You own the output. That’s actually the most “manager” thing about this whole idea.
Your AI Onboarding Checklist
If you take one thing from this post, steal this:
- Give it the why. Purpose changes decisions, for people and for agents.
- Define done. What does good look like? Show an example.
- Hand over the blueprint. Specs, standards, conventions, constraints. Write them down once and reuse them.
- Curate, don’t dump. Relevant context beats all the context.
- Make it safe to say “I don’t know.” Tell it to ask questions and flag uncertainty.
- Plan before build. Make it show its approach before it writes a line.
- Review everything. Trust is earned, same as with a new hire.
- Keep the accountability. Your name is on the result, not the model’s.
The Bottom Line
AI doesn’t hallucinate. It does exactly what a brilliant, inexperienced new grad does when nobody manages them: it fills the gaps with its best guess and says it with a straight face.
The fix isn’t a better model. It’s better leadership. AI needs just as much leadership as people do, and maybe more, because it will never walk over to your desk and admit it’s lost unless you’ve told it to.
Give it that leadership, and the output stops being slop. It becomes substantial, high-quality work, at a pace no junior team could ever match.
Stop YOLOing your agents. Start managing them.
Your turn: Have you noticed that the people who are great at leading teams are also the ones getting the best results from AI? Or have you seen the opposite? Tell me in the comments.
Catch up on the series:
- I’m an AI Slop Builder. Here Are Two Things That Actually Made Me Better At It.
- The Convenience Trap: How AI Consumer Tools Could Be the Death of Critical Thinking
- I Finally Shipped Something Real, And AI Helped Me Build It Right
- AI is Great – Simplicity Reigns Supreme.
Sources
- Merriam-Webster Medical Dictionary, “hallucination” – https://merriam-webster.com/medical/hallucination
- Hallucination or Confabulation? Neuroanatomy as metaphor in Large Language Models, PLOS Digital Health (2023) – https://pmc.ncbi.nlm.nih.gov/articles/PMC10619792/
- Wiest & Turnbull, Faulty Artificial Intelligence, or the Sleep of Reason – https://ctisus.com/learning/journalclub/345639
- Suffolk University Law School working paper on “confabulation” – https://dc.suffolk.edu/suls-faculty/396
- Kalai, Nachum, Vempala & Zhang, Why Language Models Hallucinate (OpenAI / Georgia Tech, 2025) – https://openai.com/index/why-language-models-hallucinate/
- Anthropic, Effective context engineering for AI agents – https://www.anthropic.com/engineering/effective-context-engineering-for-ai-agents
- Manveer Chawla, How to manage AI coding agents like a junior developer – https://manveerc.substack.com/p/manage-ai-junior-developer
- Weidmann, Xu & Deming, Measuring Human Leadership Skills with AI Agents, NBER WP 33662 – https://www.nber.org/digest/202507/leadership-skills-managing-ai-agents-vs-humans
- Ethan Mollick, Management as AI Superpower (One Useful Thing, Jan 2026) – https://www.oneusefulthing.org/p/management-as-ai-superpower
- ServiceNow Workflow, Do autonomous AI agents need human managers? (Paul Leonardi, UCSB) – https://www.servicenow.com/workflow/ethics/do-autonomous-ai-agents-need-human-managers.html
- Kropp, Bedard, Hsu & Krayer, Why You Shouldn’t Treat AI Agents Like Employees, HBR (May 2026) – https://www.bcg.com/news/6may2026-why-you-shouldnt-treat-ai-agents-employees
- Product Manager Practices for Delegating Work to Generative AI (Microsoft) – https://arxiv.org/pdf/2510.02504


