They Pay You to Think. And They Delegated the Thinking.

There's a chain of trust holding up almost everything you do in your professional life, one you almost never look at directly.

When a top consultancy publishes a report, you believe it. When your lawyer files a brief, you trust they actually read the cases they cite. When you search something on Google and a summary pops up at the top, you assume it's true. You don't verify every link. You can't. Life would be impossible if you had to check everything. That's why experts, seals of approval, and institutions exist: so you can delegate judgment to them and get on with your day.

Over the past few weeks, some strange things have happened on this front. The guardians of judgment, the ones who charge fortunes precisely for thinking well, were the first to delegate their own thinking to a machine. And nobody verified it — well, actually someone did, which is exactly how they got caught.

Case one: the consultancy that hallucinated about hallucinations

KPMG, one of the Big Four, one of the four largest and most prestigious consultancies on the planet, published a report titled "Redefining Excellence in the Age of Agentic AI." A document about how leading companies use artificial intelligence, complete with concrete case studies.

The problem: several of those cases were made up. AI hallucinations.

The report claimed Swiss bank UBS integrated AI agents into investment advice, risk management, and compliance. UBS responded that the claim was "factually incorrect." It claimed Swiss federal railways offered agents that optimized trips based on carbon impact. The railway confirmed it wasn't accurate. It claimed Transport for London used agents to predict congestion. The agency called it "misleading." It claimed Manchester's health system used AI for patient triage, when the original press release it pulled from was actually about a lung cancer tool, with no mention of any such agents.

Sit with that for a second:

One of the top consultancies in the world, one that charges its clients for advice on how to adopt AI responsibly, published a report on AI full of invented AI that nobody checked. The joke wrote itself.

And it's not an isolated case. Last month, EY, another Big Four firm, retracted a study over fake footnotes. Elite law firm Sullivan & Cromwell admitted in April that it filed a court brief with AI-generated inaccuracies. GPTZero, the company that caught KPMG's hallucinations, summed it up in a phrase worth keeping: these reports "poison the information well." And because they come from sources perceived as highly credible, they multiply the risk of "secondhand hallucinations." You cite KPMG in good faith, and unknowingly you're spreading something a machine made up.

Case two: the two lawyers who cited laws that don't exist

Mississippi, United States. A contract dispute between a lawyer and the city of Aberdeen. Nothing unusual, except for how both sides prepared their arguments.

They used AI. Both of them. And both cited rulings that don't exist. Fabricated case law, on both sides of the same case. The judge, Sherrion Aycock, described the situation as "unusual," which in judicial language is roughly equivalent to setting the building on fire and calling the weather mildly warm. Her line for the record was surgical: in an era of rampant, unverified AI use in the legal field, this case is an example of the risk of serving as a mere rubber stamp.

Rubber stamp. That's the image. Four lawyers who stopped doing the one thing that justifies their title, judgment, and limited themselves to stamping their signature on whatever the machine spat out. What was missing in that courtroom wasn't technology. It was judgment. The word that's supposed to distinguish a lawyer from an autocomplete. I'm sure Saul Goodman would never have done that.

Case three: when the one hallucinating is you, Google

And we arrive at the one that goes furthest, because it's no longer a company that messed up. It's a court putting a price tag on surrender.

The Munich Regional Court ruled against Google on May 28th. The AI summary feature that appears at the top of search results, AI Overviews, had falsely linked two publishers to scams and questionable business practices. The AI mixed information from genuinely shady companies with these two, which had nothing to do with it, and built an entire narrative: claims, a list of red flags, even tips on how not to get scammed. None of that was in the sources the summary itself linked to. The AI made it up.

What the court did is the important part. It classified those summaries as Google's own content, not a reorganization of third-party results. In the ruling's words, "new, independent statements generated by Google's own AI." Therefore, Google is directly liable for every false claim it generates. It barred Google from continuing to spread them and stuck it with 80% of the costs.

Google's defense is worth examining closely, because it's exactly the one all of us use when we delegate without thinking. It argued that users could check the linked sources to verify whether the summary was correct. The court's response was brutal: the possibility of disproving a claim by researching it yourself does not exempt whoever published it from liability. And it closed with a lethal data point: according to Pew, only 1% of users click through to sources in AI summaries.

In other words: Google built a product designed to give you the answer without clicking, then defended itself by saying you should have clicked to verify. The court didn't buy it.

The ruling isn't final, Google has already announced it will appeal. But the reasoning is already on the record, with precedent and arguments, ready for other courts to pick up.

The pattern connecting all three

These three cases aren't about technology failing. The AI did exactly what it does: generate plausible text. In all three, the failure was human. And it was the same failure.

Two weeks ago I wrote about cognitive surrender at the individual level: the moment you stop building your own answer and the AI's output becomes your output, without you forming your own opinion to compare it against. What these cases show is that the problem stopped being individual. It became institutional.

The institutions that for decades sold judgment as their core product, the consultancies, the law firms, the search engine you use for everything, surrendered first. And they did it for the same reason it happens to you on a Friday at six in the evening: forming your own view felt disproportionate to the task, and the machine handed you something that looked finished. The fast, happy path.

The difference is the scale of the damage. When you surrender on an email, you pay for it. When KPMG surrenders, the error gets cited across publications industry-wide. When the two lawyers surrender, the one who pays is the client who trusted them with the case, and those two lawyers too. When Google generates it, that false information reaches millions of people with the visual authority of an established fact, and only 1% will ever check.

Why this matters to you, today

Here's the practical part, because I'm not writing this so you distrust everything and freeze up.

First, as a consumer of other people's judgment: the seal of approval no longer guarantees the content. The fact that something comes from a top consultancy, an expensive professional, or Google's summary doesn't mean anyone verified it. The chain of trust has new holes in it. You don't need to become paranoid, but you do need to raise your skepticism at least one notch when something important depends on a single source, however prestigious.

Second, and more important, as a producer of your own judgment: if some of the top institutions in the world fell for this, so can you. The mechanism is identical. The only defense is the one I keep repeating: use AI to think, not instead of thinking. Before signing your name to something, read it as if someone else wrote it and you were the one accountable. Before deploying to production, review that code. You are the only one accountable for what you deliver under your name. The judge in Mississippi said it better than I could: the problem isn't using the tool, it's serving as a rubber stamp.

Judgment was the one thing those institutions had to sell that a machine can't replicate. The day they handed it over, they became replaceable by the very thing they were using. That's the trap, and it's the same one for them as it is for you.

They pay you to think. The day you delegate the thinking, there's nothing left to charge for.

PS: If this made you think, forward it to someone who's signing things without reading them.

Pablo / En Producción

---

References:

• Bratton, E. & Foley, S. "KPMG report contained AI hallucinations on benefits of AI." Financial Times, June 12, 2026.
• Regional Court of Munich (LG München I), Endurteil 28.05.2026, case 26 O 869/26. Verlagshaus24 v. Google. Coverage via The Decoder, ERP Today, BankInfoSecurity, June 2026.
• Sanctions order, US District Court, Northern District of Mississippi. Legal fees case, city of Aberdeen. Judge Sherrion Aycock.
• GPTZero. Edward Tian's statements on "secondhand hallucinations." Via Financial Times.
• Pew Research. Data on source-click rates in AI Overviews (1%).