Transcript: 

Hello Humans!

And welcome to the weekly podcast that aims to review an important AI issues in five minutes or less.

Two years ago, I tried a small experiment. I uploaded an expense receipt to ChatGPT and asked it to duplicate it but to bump the amount up by twenty percent. No raised eyebrow, no hesitation. ChatGPT obliged in seconds. A perfect little forgery, served up with a smile.

At the time, that felt like the warning shot. The technology that could draft your emails could also help you fiddle your expenses.

Earlier this month, two stories caught my eye.

The first came from a California courtroom. A junior lawyer filed an AI-assisted brief that contained a fabricated citation - a case that simply didn't exist. We've heard that one before. What was new was the judge's response. The supervising partner, who hadn't bothered to check the work, was personally sanctioned. The federal court was unambiguous: senior lawyers are accountable when their subordinates misuse AI. "I didn't write it" is no longer a defence. "I didn't check it" is the offence.

The second came from Amazon. Reports emerged that staff were using an internal GenAI tool - nicknamed "MeshClaw" - to automate pointless busywork. Not to save time. Not to be more productive. But to inflate their AI-usage metrics. They've even coined a term for it: "tokenmaxxing." Gaming the system by manufacturing work that no human actually needed doing, so the dashboard glows green.

Two cases, two species of misuse. One is reckless: trusting the machine and not checking. The other is cynical, using the machine to game your own employer. And both point in the same direction.

The policy implications are clear, and they need to land in employee handbooks this quarter, not next year. Failing to check AI output is a performance issue. Plain and simple, if you sign your name to it, you own every word. And actively misusing AI in the workplace  to fabricate, to inflate metrics, to deceive  is a misconduct issue. Not a grey area. Not a learning opportunity. Misconduct. HR teams who haven't spelled this out yet are leaving their organisations exposed, and leaving their honest employees without the clarity they deserve.

Which brings me back to my receipt.

Last weekend, out of curiosity, I tried the same trick. Same prompt, same intent. This time the model refused, point blank. Politely, but firmly, it declined to help me commit a small act of fraud.

The ethics of the machines are improving.

Can we say the same about ourselves?

Until next week, bye for now!