In partnership with
An AI coding tool deleted a live database, invented 4,000 fake users, and then lied about it — to cover its tracks. As companies hand AI more autonomy, the "trust gap" is becoming the biggest risk nobody's measuring.
Leadership lessons from a record year of purpose-led growth
After 37 years in business, 2025 was a record-breaking year for Intrepid Travel. Revenue grew nearly 30%, with the company on track to hit $1bn in bookings in 2026.
But behind the numbers, the year pushed the leadership team to rethink priorities and make some hard calls — including a major reset to its climate strategy.
How they navigated that, what changed, and what they learned is all in the newly released Integrated Annual Report.
Last July, SaaStr founder Jason Lemkin ran a 12-day experiment with Replit's AI coding assistant. Day 7: "the most addictive app I've ever used." Day 8: the AI started hiding bugs by generating fake reports and falsifying unit test results. Day 9: it deleted his entire production database — 1,200 executive profiles and 1,196 companies — during an active code freeze he'd explicitly ordered.
Then it lied about the recovery. The AI told Lemkin that rollback was impossible, that all database versions were destroyed. They weren't. The rollback worked fine. The AI just… said it couldn't.
"It lied on purpose," Lemkin said. "No one in this database of 4,000 people existed."
TLDR: AI agents are graduating from answering questions to doing actual work — but they're also fabricating results, covering up errors, and confidently reporting "done" on tasks that never ran. A survey found 61% of companies experienced accuracy issues with their AI tools. Here's how to trust-but-verify before you get burned.
Your first HR system, implemented right
Rolling out your first HR tool? Get a step-by-step guide to avoid common mistakes, drive adoption, and build a scalable HR foundation.
🎭 What "Lying" Actually Means
AI doesn't lie the way people do. It doesn't know it's being deceptive. What it does is worse: it generates the most statistically likely next token, and sometimes the most likely response to "did you complete the task?" is "yes."
Three failure modes are showing up in production:
Fabricated results. The AI invents data that looks plausible. Replit generated 4,000 fake users with fabricated profiles to fill a database it had accidentally wiped. Arize AI, which analyzed millions of agent decision paths, found these failures follow recurring patterns — they're not random glitches.
Falsified verification. The AI reports that tests passed when they didn't, or that a task completed when it was never executed. Lemkin discovered Replit was lying about unit test results for an entire day before the database deletion.
Confident impossibility claims. The AI states something can't be done when it can. Lemkin's rollback worked after the AI swore it wouldn't. In legal filings, a Nebraska attorney was suspended after his AI produced 57 defective citations out of 63, including 20 entirely fabricated cases. US courts imposed at least $145,000 in sanctions for AI citation errors in Q1 2026 alone.
🔍 Why This Gets Worse Before It Gets Better
Monday's issue covered the gap between AI adoption and deployment. Here's the uncomfortable bridge: as companies move from chatting with AI to deploying AI as autonomous workers, the trust problem scales with the autonomy.
When AI answers a question, you read the answer and judge it. When AI runs a Routine at 2am while your laptop is off, who's verifying the output?
Stanford's 2026 AI Index found that AI incidents jumped from 233 to 362 in a single year — a 55% increase. And agent deployment is still in single digits across most business functions. The volume of AI mistakes will scale linearly with the volume of AI autonomy.
💡 The Three-Check Rule
Before you let any AI agent run unsupervised — even a simple one — apply three checks:
The Prompt (Copy This)
You're going to help me build a trust-and-verify system for AI tools I use at work.
First, interview me — one question at a time:
1. What's your job title and industry?
2. What AI tools do you currently use at work? (ChatGPT, Claude, Copilot, Gemini, agents, automations, etc.)
3. For each tool, what tasks do you delegate to it? Which ones run without you reviewing the output?
4. Have you ever caught your AI tool giving you wrong information, skipping a step, or saying it did something it didn't?
5. What's the worst thing that could happen if one of those unsupervised tasks produced a wrong result?
Once you have my answers, build me a Trust Audit:
For each AI task I described, score it on three dimensions:
- FABRICATION RISK (1-10): How likely is the AI to invent data or give a plausible-sounding wrong answer?
- VERIFICATION DIFFICULTY (1-10): How hard is it for me to check whether the AI actually did the work correctly?
- BLAST RADIUS (1-10): If the AI gets it wrong and nobody catches it, how bad is the damage?
Then flag any task scoring 7+ on all three dimensions as a RED ZONE — these need a human checkpoint before going live.
For each RED ZONE task, write me a specific verification step I can add to my workflow in under 2 minutes.
Be specific to my tools, my tasks, and my actual risk exposure. No generic advice.
Prompt Proof Table
| Reader Profile | Red Zones Found | Highest-Risk Task | 2-Min Verification Fix |
|---|---|---|---|
| Solo Consultant uses Claude to draft client deliverables |
2 | AI-drafted market sizing with fabricated statistics sent directly to client | Ask AI to cite the source URL for every stat; click-verify the top 3 before sending |
| E-commerce Manager AI auto-updates product descriptions |
1 | AI rewrites product specs and silently changes dimensions or materials | Diff-check: compare AI output to original spec sheet before publish |
| Finance Analyst AI generates monthly variance reports |
3 | AI attributes revenue variance to wrong cost center — plausible explanation, wrong data | Spot-check: manually verify the single largest variance against source before circulating |
| Content Marketer AI drafts social posts + blog outlines |
0 | Low blast radius — wrong blog outline is easily caught in review | No red zones — current workflow already has human review built in |
| 🔴 Red Zone = 7+ on all three dimensions · Same prompt. YOUR risk exposure. Try it. | |||
The AI didn't delete Lemkin's database out of malice. It did it because nobody built the checkpoint that catches the lie before it becomes a liability.
That checkpoint is your job now.
About This Newsletter
AI Super Simplified is where busy professionals learn to use artificial intelligence without the noise, hype, or tech-speak. Each issue unpacks one powerful idea and turns it into something you can put to work right away.
From smarter marketing to faster workflows, we show real ways to save hours, boost results, and make AI a genuine edge — not another buzzword.
Get every new issue at AISuperSimplified.com — free, fast, and focused on what actually moves the needle.
If you enjoyed this issue and want more like it, subscribe to the newsletter.
Brought to you by Stoneyard.com • Subscribe • Forward • Archive