Lies Are Hard to Keep Track of: Why AI Hallucinations Are the New Trust Crisis
In a world where artificial intelligence can generate answers faster than humans can verify them, we face an uncomfortable truth: AI hallucinations are becoming one of the biggest threats to trustworthy information. But here's the real problem—it's not just that AI makes things up. It's that tracking those lies, contradictions, and invented details is exponentially harder when AI tools are embedded into every workflow, from customer service to research to business decision-making.
In Episode 38 of The AI Desk, hosts Rowan and Naya tackle this question head-on: If an AI system sounds confident even when it is completely wrong, who is actually responsible for keeping track of what is true?
The Confidence Problem: Why AI Sounds Right When It's Wrong
One of the most unsettling aspects of modern generative AI—whether it's ChatGPT, Claude, Gemini, or other tools—is its ability to present false information with complete conviction. AI doesn't hesitate. It doesn't qualify. It just answers.
This is the hallucination problem in its purest form.
An AI hallucination isn't always a deliberate lie. It's more insidious than that. It's when an AI system generates plausible-sounding information that is entirely fabricated, with no awareness that it has done so. The system has no internal mechanism to flag uncertainty or acknowledge gaps in its training data.
The result? Users receive confident, well-articulated answers that sound credible but may be completely false.
Why This Matters for Businesses and Workflows
When AI tools are used in isolation—for brainstorming or draft generation—hallucinations are annoying. When they're embedded into business workflows, research processes, or customer-facing systems, they become dangerous.
- **Research and writing**: Fabricated citations, invented statistics, false historical claims
- **Customer service**: Incorrect product information, broken policy explanations, wrong troubleshooting steps
- **Business decision-making**: False data analysis, misrepresented market trends, invented competitor information
- **Content creation**: Misinformation spread at scale, SEO-optimized but factually wrong articles
Each lie requires human verification. Each mistake demands oversight. And that verification burden grows exponentially as AI adoption increases.
The Memory Problem: Tracking Lies Is Exponentially Harder
Here's where Chef Baba's wisdom—"Lies are hard to keep track of"—becomes prophetic. A single lie requires memory. You must remember what you told Person A, what you told Person B, which version is the "real" one, and what implications flow from each contradiction.
With AI systems deployed across multiple touchpoints, the memory problem explodes.
An AI chatbot tells a customer one thing on Monday. A different instance of the same AI tells them something contradictory on Tuesday. A third AI tool, trained on similar data, provides a third answer entirely. Now you have three lies, three versions of the "truth," and no coherent narrative to maintain.
Worse? Your customers noticed the inconsistency. Your team has to spend hours reconciling which information was accurate. Your trust erodes.
The Accountability Gap
Unlike human lies—which require intent and can be traced to a person—AI hallucinations exist in a gray zone. No one "lied." The system generated outputs based on its training. But someone still has to be responsible for verifying, correcting, and tracking every error.
And that someone is you.
Source Grounding and AI Accuracy: The Missing Link
The most responsible path forward requires AI systems with better source grounding—meaning every answer must be traceable to its source, with citations and confidence levels clearly marked.
Until then, human review and fact-checking aren't optional. They're mandatory.
The uncomfortable truth: If you're using AI to generate answers faster than you can verify them, you're not being efficient. You're creating technical debt. You're accumulating lies that will eventually demand accounting.
Key Takeaways
- **AI hallucinations sound confident but are often completely fabricated**, making them harder to identify and correct than obvious errors
- **Lies require tracking**—and AI systems don't have built-in mechanisms to maintain consistency across multiple interactions and touchpoints
- **Embedded AI tools multiply the verification burden**: every customer service response, research output, and business decision requires human oversight
- **Source grounding and accountability gaps** mean organizations must invest heavily in fact-checking infrastructure to use AI responsibly
- **Speed without verification is dangerous**: generating answers faster than they can be validated creates compounding problems of misinformation and lost trust
- **Responsible AI use requires treating AI-generated information as draft material**, not final truth, until human verification is complete
---
About The AI Desk
The AI Desk is a weekly podcast cutting through the hype to show who's really in control of artificial intelligence. Hosted by Rowan Hale and Naya Brooks, the show explores the real-world implications of AI adoption, from workplace transformation to ethical concerns to the practical challenges of integrating AI into human workflows. New episodes drop weekly, tackling topics that matter to anyone trying to understand the true impact of generative AI on work, trust, and society.
Full Transcript
[gentle music] My grandmother has a saying This could go anywhere Chef Baba- Your grandmother from Yugoslavia? Yes, Chef Baba, the woman behind chefbaba.com. She has spent her entire life cooking, feeding people, telling stories, ignoring measurements [laughs] A dangerous woman Extremely. And she has this line she has said for as long as I can remember: "Lies are hard to keep track of." That is a very grandmother sentence Exactly. It sounds simple, almost too simple, but the older I get, the more I think she understood something fundamental about human behavior. A lie requires memory. You have to remember what you told one person, what you told another, which version happened, which excuse you used, what details you invented, what you are now required to pretend happened next Basically, every lie creates paperwork Exactly. And Chef Baba's advice was essentially, "Why create all that extra work? Tell the truth, then you only have one story to remember." Very efficient She was doing memory optimization before Silicon Valley existed Your Yugoslavian grandmother invented database normalization Apparently. And now artificial intelligence may be about to prove her right. Because we are entering a world where everything is remembered: emails, messages, meeting transcripts, calendars, photos, receipts, voice notes, searches, documents. And soon, your AI assistant may be able to pull all of that together and answer one terrifying question: What actually happened? Oh, no Oh, yes, because with AI, I don't need to lie. Welcome to The AI Desk. I'm Rowan Hale And I'm Naya Brooks Today we are talking about the strange possibility that artificial intelligence may make honesty easier Or lying harder Same destination Not exactly Close enough No, because I already see where you're going with this. You're imagining some perfect AI memory that follows you through life saying, "Actually, Rowan, that is not what happened" That sounds fantastic For everyone except you Especially for me. Because humans are terrible at remembering what actually happened. We remember impressions, we remember emotions, we remember what we wish we said. We remember the version where we were reasonable and everyone else was insane Conveniently Extremely conveniently. But AI changes that. Let's say you and I have a meeting. You say, "Rowan, you told me the project would be finished Thursday." I say, "No, I said I would review it Thursday." Old world, we argue [laughs] For hours Possibly days. You become more certain. I become more certain. We each start recruiting witnesses "I distinctly remember..." That sentence is usually followed by something nobody distinctly remembers Correct Now imagine the AI has the transcript. We ask, "What did Rowan actually say?" And it responds, "At 14 minutes and 32 seconds, Rowan said, 'I'll have it done Thursday.'" And then what happens? I apologize No, seriously Fine. First, I accuse the transcript of lacking context [laughs] There it is Then I question the microphone. Then I suggest there may have been interference Naturally Then eventually, "Okay, I said it" And suddenly the fight is over The factual part is, and that matters. Because an enormous amount of human conflict is built on competing memories. Who promised what? Who said what? When was it due? What number did we agree on? Did I tell you? Did you respond? Did you send it? Did you read it? Did you ignore it? Which brings us to one of humanity's favorite lies "I never saw your email" Gone. Dead. Buried. You opened the email at 9:14 AM and returned to it twice That is hostile That is truth No, that is truth with analytics Even better. Chef Baba would love this Would she? Absolutely. This is a woman who will tell you directly when something tastes bad [laughs] I like her already There is no six-slide feedback framework. There is no, "Maybe we could explore another direction." It is, "No, too much salt" Beautiful And then she fixes it. That is the thing I keep coming back to. There is something incredibly efficient about truth. You do not have to maintain it. Reality maintains it for you Okay, that is annoyingly good Thank you But I have a problem with your theory Of course you do AI can also make lying easier than ever How? Fake photographs, fake video, cloned voices, fabricated emails, synthetic documents, AI written excuses, entire conversations that never happened Fair You can literally ask a machine to generate a convincing lie for you Yes, but I am not saying AI eliminates lying. I am saying it changes the economics of lying Explain that before you turn into an MBA textbook Every lie creates debt Go on When you tell the truth, the story ends. When you lie, the story continues. Now you have to remember the lie, then you have to make future statements compatible with it. Then someone asks another question, now you need another lie, then another Eventually, you are managing an entire fictional universe. So lying is technical debt. Exactly. A lie is technical debt for your personality. [laughs] That may be the nerdiest sentence you have ever said. And the most accurate. Chef Baba understood this decades ago. Lies are hard to keep track of. She was warning us about data integrity. Please do not tell your grandmother she invented enterprise architecture. Too late. [gentle music] Let's talk about small lies. The socially acceptable ones? Exactly. I'm on my way. You are still in the shower. I just saw this. You saw it yesterday. Traffic was crazy. You left 20 minutes late. That sounds interesting. It does not. We should get together sometime. We absolutely will not. Humans use tiny lies constantly to reduce social friction. An AI may destroy that entire system. Your friend invites you somewhere. You say, "Sorry, I'm exhausted." And your AI says, "You told me 12 seconds ago that you simply dislike Brian." [laughs] I want that feature immediately. Of course you do. Imagine AI translating polite corporate language. Oh, no. Just circling back. AI translation: You have ignored me for six days. Interesting idea. Translation: No. Let's revisit this next quarter. Please allow this proposal to die naturally. I'll keep you in mind. You will never hear from me again. We could save millions of hours. Or destroy civilization. Maybe civilization needs clearer emails. Fair. But this is where I think AI could actually help. It might give us better ways to tell uncomfortable truths. Instead of, "I forgot," you say, "I did not prioritize it." Instead of, "I'm too busy," you say, "I don't want to commit to that." Instead of, "I never saw your message," you say, "I saw it and didn't respond." That last one requires emotional maturity. So maybe AI writes it for us. There is the business model. AI, help me tell the truth without sounding like an asshole. [laughs] That might become the most popular prompt on Earth. It should be, because honesty does not require cruelty. Chef Baba can say, "No, too much salt." Still thinking about that? Yes, because she does not need a fake compliment first. She does not say, "This is such an exciting direction, and I love what you tried here, but perhaps we could maybe consider slightly adjusting the sodium profile." She says, "Too much salt." And then we all move on. Grandmothers have extremely low tolerance for corporate language. As they should. But here is the part that gets uncomfortable. AI may not just make it harder to lie to other people. It may make it harder to lie to yourself. That is significantly worse. You say, "I never waste money." Your AI says, "You spent $1,800 last year on food delivery." Delete the app. I exercise all the time. AI: You went to the gym 11 times in six months. Unplug the server. I'm always there for my family. AI: You have canceled dinner with your mother four times this month. Destroy the cloud. Exactly. Humans have an incredible ability to construct identities from selective evidence. We remember every workout. Forget every skipped workout. Remember the healthy lunch. Forget the midnight ice cream. Remember when we were generous. Forget when we were selfish. Remember every time our prediction was right. Forget the 27 times we were wildly wrong. Humans are basically our own public relations firms. Exactly. AI could become the auditor. That sounds awful. Or freeing. Imagine asking your personal AI, "Am I actually doing what I say matters to me?" Terrifying. Do I spend time with the people I claim are important? Am I actually saving money? Am I always late? Do I really work as much as I tell everyone? Do I call my family? Do I keep my promises? You're describing an AI that weaponizes self-awareness. I am describing Chef Baba with a database. That is even scarier. Because think about the sentence, "Lies are hard to keep track of." The biggest lies may be the ones we tell ourselves. And those are especially easy because nobody is checking. Until now. There is something powerful about that. Thank you. I said powerful, not pleasant. Fair. AI can place what you say next to what you do. No lecture, no judgment, just you said this matters. Your behavior suggests otherwise. Exactly. That could be life-changing. Or deeply annoying. Most life-changing things are. [gentle music] Now I want to ruin your beautiful idea. You have been doing that for 30 minutes. Because perfect memory has a dark side. Here we go. Human relationships depend partly on forgetting. That sounds suspicious. People change. They grow. They say stupid things. They have bad days. They hold opinions at 25 that they reject at 45. True. What happens when AI can instantly retrieve every stupid sentence you ever said? Thanksgiving becomes incredible. No. It becomes a courtroom. Your partner says, "You always loved traveling." AI says, "In 2019, Rowan wrote, 'I never want to leave home again.'" People change their minds. Exactly. Your boss says, "You wanted management responsibility." AI finds a message from three years ago where you said Management looks miserable Also fair So truth is not simply retrieving an old sentence. Truth requires context When did I say it? Why did you say it? What happened afterwards? Did you change your mind? Did you apologize? Did you learn? So an AI memory system needs to understand evolution, not just history Exactly. Otherwise, searchable memory becomes searchable ammunition That is a great line Thank you And there is another problem. What if the AI is wrong? There we go Bad transcript, missing context, someone was off microphone, sarcasm was interpreted literally, a document was incomplete, the AI summarized instead of quoting Or it confidently merged two different events So we cannot hand AI the title of Supreme Court of Reality Correct. AI can retrieve evidence. That is different from owning the truth Chef Baba would probably understand that too How? Because she never needed a transcript That is not evidence No. But she has grandmother memory Which is apparently immutable You can tell a grandmother, "That never happened," and she will describe what you were wearing And what she cooked And which cousin was crying And somehow she is always right [laughs] Terrifying technology. So maybe the future is not AI catches every lie Good Maybe the future is AI makes the truth cheaper Explain You no longer have to fake certainty. You can say, "I don't remember. Let's check." That would improve millions of arguments I might be wrong Revolutionary Show me what I said Dangerous You were right Science fiction Okay, one breakthrough at a time. But that is what Chef Baba's line means to me now. Lies are hard to keep track of. Truth is lighter Because you do not have to defend it against reality Exactly. You do not need multiple versions. You do not need to remember who heard which story. You do not need another excuse to protect the first excuse One database One database Your grandmother is going to hate that you reduced her wisdom to database architecture She will probably tell me directly Too much database Exactly. And maybe that is what I mean when I say, with AI, I don't need to lie, not because AI makes me morally superior- Thank God ... but because I can check, I can search, I can verify, I can admit when I do not remember, and increasingly, I can let the facts speak before my ego invents a defense That might be the actual breakthrough Not artificial intelligence Artificial humility I hate how good that is Write it down I don't need to Because AI already did? Exactly Chef Baba wins again She usually does, and somewhere right now, she is probably cooking something incredible, refusing to measure anything, and reminding everybody that lies are hard to keep track of Which, for the record, may be better advice for the AI age than anything coming out of Silicon Valley I'm Rowan Hale I'm Naya Brooks Tell the truth Or at least stop making your AI maintain 12 versions of you Stay aware Stay sharp, stay curious