Let’s talk about the second opinion you asked for — and why it came back in a different voice with exactly the same words.
Picture a woman in a specialist’s waiting room. Scan done, first opinion delivered, and something in her gut says get another read. So she does it properly. Pays cash, takes a morning off, waits three weeks for the appointment.
She opens the second letter in the parking lot.
The recommendation is identical. Same conclusion, same order of points, same caveat about further imaging not being indicated at this stage. And for about ten seconds she’s relieved. Then she reads it again and notices the sentence rhythm. The way the caveat is worded. It’s not how her first doctor talks. It’s not how this doctor talks either.
It’s how the software talks.
Nobody did anything wrong. That’s the part worth sitting with. The second specialist dictated into an assistant, the assistant pulled the referral and the first letter into context, and drafted a response. The doctor edited a line, added a paragraph, signed. Perfectly reasonable workflow. Efficient. The kind of thing you’d be foolish to ban.
And the second opinion still wasn’t a second opinion.
Because here’s what a second opinion actually was, for the whole history of professional life. Not a smarter person. A separate one. Different training, different failures, different memory of the case that went wrong in 2019, different colleagues they’d have to face if they got it wrong. Independence was never a personality trait. It was structural. Two minds that arrived at a conclusion from different directions, so that when they agreed, the agreement carried information.
The entire architecture of professional trust runs on that. Audit exists because the auditor didn’t write the books. Peer review exists because the reviewer didn’t write the paper. Appeals exist because the appeal judge wasn’t the trial judge. You get three quotes from builders not because you want three prices, but because you want three people to have looked at your wall and seen different problems.
All of it assumes one thing: that error is uncorrelated. Two readers miss different things. That’s the mechanism. Not intelligence. Distance.
Now think about what happens when the distance closes.
Every professional’s first pass now starts in roughly the same place. A handful of models, trained on the same internet, tuned toward the same helpful, balanced, three-considerations-and-a-recommendation shape. The doctor’s draft, the lawyer’s memo, the auditor’s finding, the consultant’s deck, the kid’s essay, the minister’s speech. Same substrate. Different letterheads.
I’m not arguing the outputs are identical. They’re not, and anyone who’s used these tools knows it. The argument is narrower and worse: the deviations are shrinking. The strange read, the “this doesn’t fit the pattern”, the thing an experienced person says that nobody else said — that’s the first thing a drafting assistant smooths out, because it’s the thing that looks least like the rest of the corpus.
And when the deviations shrink, errors start to correlate.
That’s the quiet danger, and it doesn’t announce itself. Correlated error doesn’t feel like error. It feels like consensus. Three professionals agreeing on the same conclusion feels like proof. It can also be one model agreeing with itself three times, in three different fonts.
You can already see the tell in how advice sounds. Smoother. More confident. Fewer loose ends. Fewer people saying “I don’t know, but something about this is off.” And the dissent that does survive costs more to produce than accepting the draft, so it gets rarer for the most ordinary reason in the world: it’s expensive to argue with a document that’s already written.
Second-order effect, and this is the one that should worry anyone running a company. The person with the unusual read becomes friction. They’re slower. They ask for the file again. They say the numbers look too clean. On a dashboard, that person is a cost centre. So they get managed out, or they get so buried in review queues that they stop bothering. And the firm keeps the consensus and loses the only thing that was ever going to catch the consensus being wrong.
So what do you actually do about it.
If you’re the one asking, ask a different question. Not “what do you think” — “what did you see before you saw my file?” If the answer is “your referral and the previous letter”, you didn’t buy a second opinion. You bought a second draft.
If you’re the one giving advice, form your own read first. One paragraph, in your own words, before you open anything. It takes four minutes and it’s the only thing that keeps you a separate observer instead of a reviewer of somebody else’s output. Then compare. The gap between your paragraph and the machine’s is the actual value you brought to the room, and it’s invisible if you never wrote it down.
If you run a review process, design for independence or stop pretending. The second reviewer should not see the first output before forming a view. Otherwise you’ve built a machine for producing agreement with extra steps and a signature on top.
And keep one person whose job explicitly includes thinking without the assistant. Their time will look like waste on every report you generate. Defend it anyway, because it’s the only line item that produces the thing you can’t buy back later.
You’ve got this. The lever is smaller than it sounds — one paragraph written before the draft arrives, one question asked about what the adviser saw first.
The machine can give you a thousand opinions before lunch. What it cannot give you is a second one. A second opinion is a person who could be wrong in a way that costs them something, who carries a different history into the room, and who might, on an ordinary Tuesday, tell you the thing no model was going to say because it wasn’t in the corpus yet.
That’s what she drove across town for. Three weeks and a few thousand rand for a genuinely separate mind.
She didn’t get it. Neither will most people, unless somebody starts checking where the judgment came from before it arrived in a nice letter.
Ask for the second opinion. Then check that you actually got one.