|
|
|
// YOUR AI EDGE
KPMG PUBLISHED A REPORT ON AI AND AI MADE HALF OF IT UP
Plus: a German court decides who pays when a chatbot lies, McDonald’s gives the drive-thru another robot, fake Reddit posts are quietly steering AI answers, and a self-driving CEO explains who she actually hires.
|
|
// TODAY'S SIGNAL
|
THE FIRM THAT AUDITS EVERYONE FORGOT TO AUDIT ITSELF
KPMG wrote a report about how companies use AI. AI wrote some of the companies into it that don’t exist. The report was called “Redefining excellence in the age of agentic AI.” It name-dropped real organizations as shining examples of AI adoption. Then those organizations started calling. UBS, the UK’s National Health Service, Swiss Federal Railways, and Transport for London all told the Financial Times the claims about them were either untrue or flat-out invented. KPMG quietly pulled the whole thing off its websites and opened an investigation into itself. A research group called GPTZero traced the errors back to AI hallucinations. So a firm whose entire business is checking other people’s work used a chatbot to write a report about chatbots, skipped the checking part, and published fiction with a logo on it. Last month EY yanked a report on loyalty programs loaded with fake footnotes. Two of the most trusted names in professional services, both caught shipping AI slop in the same quarter. Everyone knows AI lies. The new part is that the people who are supposed to be the adults in the room are pasting prompts in and hitting publish like the rest of us.
|
// CAREER MATH
The analyst who fact-checks the AI keeps their job longer than the one who trusts it. Make verification visible. When you hand work to your boss, say “I used AI for the first draft and checked every number against the source.” That one line puts you above two-thirds of your office, including, apparently, the partners at KPMG. The rare skill isn’t writing. It’s catching the confident lie before it leaves the building.
|
|
|
|
// SHORTCUT
Copy, paste, go:
|
"You just told me [PASTE THE CLAIM]. Act as a skeptical fact-checker. For each specific name, number, and date in that statement, tell me your confidence level (high, medium, low) and whether you might be inventing it. If you can’t verify something, say ‘I cannot confirm this’ instead of restating it."
|
|
|
THE WRONG ANSWER NOW HAS A PRICE TAG
For two years the answer to “what happens when AI makes something up about a real person” was basically nothing. A court in Munich just changed that. The Munich Regional Court issued a preliminary ruling that Google is liable for false statements its AI Overviews generate. The logic is clean: if you design it, train it, run it, and profit from it, you own what it says. The industry’s whole defense was that AI just remixes the web, so nobody’s really on the hook for the output. A judge looked at that and said no, the company is. This is one local court in Germany, not the Supreme Court. But it’s the first real crack, and every defamation lawyer on the planet just bookmarked it.
|
// STEP BACK
Every company rolling out an AI chatbot for customer service or marketing just inherited a new kind of risk. If your team is shipping AI-generated content with a brand name attached, someone needs to own accuracy, and that someone is increasingly a job. “AI quality control” is going from a nice-to-have to a line item. If you want to be useful in any content, comms, or legal-adjacent role, get fluent in where these tools break and how to catch it. That fluency is about to be very employable.
|
|
|
McDONALD’S IS GIVING THE ROBOT ANOTHER SHIFT
McDonald’s once had an AI drive-thru that added 260 chicken nuggets to a single order and tried to sell people bacon-topped ice cream. They killed it in 2024. The robot is back, and this time it brought Google. The new system is called ArchIQ, with a voice assistant named Archy, unveiled this month and built on Google’s tech. McDonald’s says early tests handled a million orders with humans stepping in only about 10% of the time. The first attempt with IBM was a meme. This one has the numbers to be a comeback. The difference isn’t that AI got perfect. It got good enough to take your order without ending up on YouTube. Watch this one, because the drive-thru is the test kitchen. Whatever works on a Big Mac order is coming to the call center, the help desk, and the front desk next.
|
// CONNECT THE DOTS
The same week KPMG got burned trusting AI too much, McDonald’s showed the version that works: narrow task, tight guardrails, a human ready to step in for the messy 10%. That’s the pattern worth copying at your own job. Don’t hand AI the whole project. Hand it the repetitive 90%, keep your hands on the part that needs judgment, and you get the speed without the bacon ice cream.
|
|
|
| |
|
SOMEONE IS WRITING FAKE REDDIT POSTS TO STEER WHAT AI SAYS
Ask ChatGPT for a product recommendation and it often pulls from Reddit. Marketers figured that out. Now they’re seeding fake Reddit posts to land their stuff in the answer. It has a name: Answer Engine Optimization. Old-school SEO fought to rank on a results page. AEO skips that and aims straight at the sentence the AI hands you, using aged accounts and paid posters that look exactly like real people. The numbers explain the gold rush. Google’s AI Overviews cite Reddit in roughly one of every five answers, and just 100 subreddits drive most of the brand mentions AI repeats back. Seed a few high-traffic threads and you bend what millions of people get told. The scary version is health. Moderators of an 830,000-member biohacking community locked down posts after vendors flooded it pushing unregulated peptides, the kind of thing an AI might then confidently recommend a dose of. An AI answer sounds like a neutral synthesis of the whole internet. Sometimes it’s a synthesis of whoever paid to flood the comment section.
|
// THE MOVE
Treat AI product recommendations like a stranger’s hot tip, not a verdict. When a chatbot pushes a specific brand, supplement, or vendor, ask it for its sources and then go look at them yourself. For anything that touches your money or your body, that 30-second check is the whole game. The people who blindly buy what the AI suggests are the exact target this trick was built for.
|
|
|
THE SELF-DRIVING CEO WOULD RATHER HIRE THE ROOKIE THAN THE VETERAN
Raquel Urtasun trained under the man they call the Godfather of AI, then built Uber’s self-driving program. When she talks about who to hire right now, the answer should change how you think about your own résumé. Urtasun runs Waabi, a self-driving truck company, and she’d rather bet on someone early in their career than a 20-year industry veteran. Her reasoning, laid out in Fortune, is that the people without decades of “this is how we’ve always done it” adapt to AI faster. Experience can be a moat. It can also be a rut. A founder who needs the technology to work is telling you fluency with new tools now outranks a long track record. That’s a hiring signal worth taking to the bank, with one catch: it only works if you actually are the adaptable one. “Comfortable with AI” isn’t a vibe, it’s a portfolio of things you’ve genuinely built or fixed with these tools.
|
// THE REAL STORY
Being early in your career is a hiring advantage right now, but only if you can prove the adaptability. Pick one AI tool and get genuinely good at it this month, good enough to show a before-and-after at work. Then say so in interviews and reviews: “Here’s a process I rebuilt with AI and what it saved.” That’s what turns “less experienced” into “moves faster.” A blank slate plus a real example beats a long résumé that stopped at last year’s tools.
|
|
|
| |
|
// THE TOOLBOX
NOTEBOOKLM
NotebookLM is Google’s free research tool that only knows what you feed it. You upload your own documents, PDFs, notes, a messy pile of sources, and it answers questions using only those, with citations pointing to the exact line. It’s the opposite of a chatbot that scrapes the whole web and guesses. Think of it as a research assistant who’s read your stuff and refuses to make things up about it.
The caveat: it’s only as good as what you put in. Garbage sources in, confident garbage out. The win is that it shows its work, so you can actually check it, which is more than KPMG managed.
|
|
// WAIT... DOES THIS ACTUALLY WORK?
THE OBSCURE PROMPT OF THE DAY
prompts nobody asked for. results nobody expected. try it anyway.
|
"Take my to-do list below and reorganize it by how much I’ll dread each task. Put the thing I least want to do first. Then write one sentence of trash talk daring me to do it before lunch. [PASTE LIST]"
|
// OUR VERDICT
Sounds like a joke. Works like a personal trainer who read your diary. It correctly identified that “email the client back” was the thing I’d been avoiding for three days and called me a coward about it. I did it before lunch. Useful in the most annoying possible way.
SURPRISINGLY PRACTICAL: ★★★★½
|
|
// YOUR EDGE
|
01
Learn this: How to spot an AI hallucination before it ships. Check every name, number, and date an AI gives you against a real source. KPMG skipped that step and pulled a whole report. The skill that pays now isn’t generating text, it’s verifying it.
|
|
02
Watch this: McDonald’s ArchIQ numbers over the next few months. If a Google-built AI can take a drive-thru order with humans stepping in only 10% of the time, the same math is coming for call centers and help desks. The drive-thru is the public beta for white-collar automation.
|
|
03
Say this: "A German court just ruled the company that builds the AI is liable for what it says. So if we’re putting AI content out under our name, who owns checking it before it ships?"
|
|
|
// GOT THIS FROM A FRIEND?
Your edge on AI, twice a week. Free forever.
Subscribe →
|
|
|
|
