The Most Dangerous Employee You’ll EverManage Isn’t Human
- HEATHER DI ROCCO
- 2 days ago
- 7 min read
How Leaders Must Supervise Autonomous AI, Prevent Hallucinations, and Maintain Decision Authority
The first time artificial intelligence unsettled me, it was not because it failed. It was because it stepped in and corrected me. I asked what I thought was a simple question: how can I make money with this AI-powered crypto system that a friend of mine swore was revolutionary? I wasn’t seeking a regulatory breakdown. To my surprise the AI did not agree with me. I did not request a forensic audit. I was looking for an opportunity, aka I wanted easy AI money too! The system, instead of people pleasing, pushed back. It outlined warning signs common to Ponzi schemes. It listed countries that had already banned similar structures. It explained, in clear terms, what “SEC registered” actually means and why being listed in a government database does not equal meaningful oversight. It even walked through how scammers exploit that confusion to appear legitimate while sidestepping the rules that protect investors.
What struck me was not the accuracy of the analysis. It was the posture. The machine did not simply answer my question; it reframed it. It declined to optimize for my enthusiasm, I mean isn’t AI all about gassing you up? Not this time, it optimized itself for my protection. And in that shift, I realized something larger was happening. I was no longer interacting with a passive tool. I was engaging with a system capable of shaping judgment before I asked it to do so.
We like to talk about AI as if it floats in a neutral digital cloud, free from politics and profit. It does not. These systems are trained on vast quantities of scraped human behavior, licensed content, public data, private data, and everything in between. AI is funded and maintained by corporations whose survival depends on prediction and influence. Surveillance capitalism is not theory; it is the operating model of much of the modern internet. Data is extracted, refined, and monetized. Human behavior is the raw material. Prediction becomes a product. Your browser history is as valuable as oil!
When you type a prompt into an AI system, the answer seems conversational, almost cozy. But beneath that experience is infrastructure designed to learn from you, while shaping you. The system predicts what you might ask next. After all, GPT is an acronym for Generative Pre-trained Transformer! It adjusts its responses based on patterns drawn from millions of interactions. It is always nudging, like your grandma offering just one more serving. Sometimes that nudge protects you, as it did in my case. Sometimes it persuades you, as in your next Netflix binge.
This is why leadership in this environment requires more than technical literacy. You do not need to understand neural networks or token probabilities to guide your organization wisely. But you do need to understand power. Every time you delegate to AI, whether for drafting emails (or articles), screening investments, generating marketing strategies, or conducting research, you are shifting authority. You are allowing a system built within specific economic incentives to shape the boundaries of your thinking. If you fail to supervise that shift, you risk mistaking speed for efficiency, but HAL 9000 thanks you for that.
Story time, long before AI became mainstream, I learned what happens when speed outruns supervision. Early in my intelligence career, I was “the intern”. I had proven myself reliable. I had delivered solid work. I had gained enough trust to be taken seriously. And then one afternoon, a source we jokingly called “T-Online” reported what sounded like a chemical explosion near my town. We called him “T-Online” because much of his “intelligence reports” seemed to come from German news articles he translated himself, often with a creative touch. His enthusiasm, in most cases, exceeded his precision.
The report mentioned chemicals, an explosion and a radius. My apartment sat squarely within that radius. Suddenly this was not just urgent or dramatic. It was personal. I was less concerned about national security and more concerned about whether I was about to be homeless. I passed the information up to an analyst I was friends with, looking for a place to spend the night, not to initiate a crisis management team deployment. My lack of three dimensional thinking as “the intern” resulted in causing a kerfuffle with senior leadership. The chief analyst confronted the chief of collections. Accusations were made about withheld information. Voices were raised. What began as my attempt to not be homeless, became an internal storm.
Eventually, someone verified the details. The “chemical explosion” was a university student experimenting with rocket fuel who had severely injured himself. There was no large-scale threat. My apartment was fine. The city was fine. But I had triggered real leadership conflict because I escalated unverified information based partly on mistranslation and partly on my self-preservation.
I had triggered real leadership conflict because I escalated unverified information based partly on mistranslation and partly on my self-preservation.
I was not malicious. I was inexperienced. “T-Online” also had a huge heart and the best of intentions, always. The moral of the story: no one had paused to supervise the intern.
That memory sits quietly in the background every time I interact with artificial intelligence. AI is the most capable intern you will ever have. It is articulate, fast and eager to deliver. It produces complete answers with remarkable confidence. And like that younger version of me, it does not always recognize when it is operating on incomplete or misunderstood information.
I saw this in a far less dramatic but equally disruptive way when I asked AI to help curate a yearlong reading list for a book club. The system delivered beautifully structured recommendations. The titles sounded thoughtful. The authors felt credible. The summaries were so compelling that I began building promotional materials around them. I organized themes for the year. I drafted announcements. Then I started checking the books. Seventy-five percent of them did not exist. They were not obscure or out of print. They were entirely fictional. “The Aloha Spirit” by Barbara Santos. “Hawaiian Leadership” by Peter Apo. “Lucky We Live Hawaii” by Robin Campaniano. They sounded like books you might pick up at a local bookstore, here in Hawaii. The summaries were polished. The authority felt real, but they were inventions.
The system was not trying to deceive me. It was trying to satisfy me. It was filling gaps with coherence. It behaved like a child insisting there is no cookie missing while a half-eaten chocolate chip is clearly visible behind his back. The performance is convincing if you do not look too closely. It is only when you check the pantry that you realize the story was assembled to please, not to verify.
AI hallucinates not because it intends to lie but because it is optimized to produce fluent output. It would rather provide a complete answer than admit uncertainty. It generates confidence before it generates caution. And if you are moving quickly, captivated by its polish, you too may fall into a people pleasing versus fact trap. I wish it had hallucinated on that AI crypto thing, that the AI got right.
AI hallucinates not because it intends to lie but because it is optimized to produce fluent output.
The more seamless AI becomes, the more it disappears into our workflows. It gets it right a few times and you trust it. It drafts contracts, summarizes reports, filters candidates, screens investments, generates campaign strategies. The friction that once forced us to slow down is removed. Speed becomes the default. I mean if you are in sales how many times have you heard “speed to lead”! Speed feels productive. But friction is often where the truth lives. Verification requires time. Skepticism requires pause. When we eliminate those pauses, we eliminate the safeguards. The real leadership skill now isn’t using AI faster. It’s holding onto the authority to pause, question, and override it.
That way we don’t become Heather the intern. There is also a deeper side to this, that sometimes we forget. The companies building these systems operate at scales most organizations cannot match. They aggregate behavioral data from billions of users. They refine predictive models continuously. They deploy updates monthly, it seems. Meanwhile, leaders inside businesses, nonprofits, and institutions adopt these tools to stay competitive and it comes at a price, an extra level of supervision we often forget to deploy.
Leadership in this era is not about rejecting AI. It is about refusing to surrender oversight. I override AI constantly. It does not manage my money. It does not send final emails. It does not auto-publish content. It drafts. It suggests. It brainstorms. I decide. I also audit its outputs. I ask it to critique itself. I cross-check sources. I confirm that the books exist, or at least now I do. Honestly, I even turn it on myself. Just as we once Googled our own names to see what potential employers would find, I now instruct AI to conduct due diligence on my business as if it were a skeptical investor. What red flags would it identify? What inconsistencies would it surface? The results are useful, but not gospel. That, my friend, is a good thing!
As AI systems evolve from assistants to autonomous agents, executing tasks across platforms, negotiating schedules, drafting legal language, conducting due diligence, the temptation to let them run will grow. The promise will be freedom from busywork, liberation from inefficiency. And there will be real gains. But autonomy without supervision has always been a recipe for disaster. Whether the intern sits at a desk down the hall or operates inside a data center, the principle is the same.
A powerful leader in this moment is not anti-technology. She is curious. She experiments. She tests the limits of what these systems can do. She laughs when the book list turns out to be fictional, but she also learns from it. She welcomes the warning about the Ponzi scheme, but she does not assume every output is benevolent. She understands that the machine can calculate faster than she can. She also understands that accountability still rests on human shoulders.
Leadership in the age of autonomous AI is not about mastering the machine. It is about supervising it. It is about remembering that speed is not wisdom, that fluency is not truth, and that convenience is not the same as control. The pause between prompt and decision, the moment where you verify, question, and decide, that pause is still human territory. If we give that pause up, we will not lose because the machines rebel. We will lose because we stopped paying attention.
Curiosity without supervision becomes chaos. Curiosity guided by discernment becomes leverage. The difference is leadership.
Heather Di Rocco is an AI strategist, speaker, and former military intelligence analyst who spent 20 years advising leaders in complex, high-stakes environments.
Meet the expert:
Heather Di Rocco is an AI strategist, speaker, and former military intelligence analyst who spent 20 years advising leaders in complex, high-stakes environments. Today, as founder of AI in Wonderland and InsureBot Solutions, she helps organizations move beyond AI hype to build structured, decision-ready systems that improve performance and resilience. Her work explores how leadership, human judgment, and strategic clarity must evolve as autonomous technologies reshape the future of work.
Dive Deeper Into Her Wealth of Knowledge:
Follow:
