AI can help you read a lab report, and help you figure out what to ask the vet. It can’t make the decision for you.
And don’t trust a hospital name a single AI gives you. It will say a wrong name in a very confident voice.
Use it, but keep some distance. Bring it with you when you talk to the doctor. Don’t let it talk for the doctor.
I
A friend of a friend — I’ll just call him L — adopted a three-month-old border collie this spring. Her name is Wheat. Not long after she came home, her bloodwork came back wrong. Creatinine was 225 µmol/L. That’s not a normal number for a 4.8-kilogram puppy.
A clinic near his apartment kept her on IV fluids for four days. Her condition would improve, then slip again. L didn’t know what to do next, and he eventually got my number from someone. What he sent me was a stack of phone photos: CBC, biochemistry panel, ultrasound, PCR, urinalysis, plus a few billing receipts. Maybe fifteen images in total.
I opened Claude.
For the next two days, we went back and forth probably dozens of rounds. It stitched the reports together, listed differentials, walked through every lab value, evaluated the medications already in use, suggested where to refer her, and finally drafted a referral summary L could hand to the next hospital.
Looking back, a few moments from those two days stayed with me. Some of what the AI did was genuinely useful. Some of it almost took us in a very wrong direction.
II
What surprised me most was its ability to integrate.
When I read those reports on my own, my head was just full of scattered facts — this is high, that is low — and nothing connected. Claude took maybe ten minutes to give me this:
“Onset at three months, polyuria and polydipsia, persistent nocturnal incontinence, asymmetric kidneys on ultrasound, acute causes negative across the board, clinical state still acceptable — the most parsimonious explanation is a congenital lower urinary tract structural anomaly, possibly with renal dysplasia.”
That assessment was later confirmed in full by a senior specialist in Beijing. Wheat was diagnosed with congenital renal dysplasia, with only one kidney in her body.
III
But it also got things wrong.
It recommended two specific veterinary hospitals to me, complete with descriptions of their surgical capabilities and dialysis equipment, with addresses, strengths and the situations they were best for. I started copying the names into the document I was writing for L.
Halfway through, something stopped me. I had never heard of either place. So I asked Claude directly: are these hospitals real?
It paused, and then admitted:
“I can’t confirm these hospitals exist. I didn’t search to verify them. I generated them from impressions in my training data, and I fabricated the capability descriptions too. If you hadn’t asked, the owner might really have called numbers that don’t exist, or driven to addresses that don’t exist.”
I told it to actually run a web search and verify. What came back was a different, smaller list — real hospitals that I could open in a browser, with public websites, real addresses, real phone numbers.
A cold feeling went down my back at that point.
If L had asked the AI directly instead of going through me; if he had trusted those fabricated names; if he had loaded up his car and driven to an address that didn’t exist — I didn’t really want to follow those thoughts any further.
AI doesn’t just make mistakes. It makes mistakes with a very confident voice. That part is what stays with me, more than the fact of error itself.
IV
Later I took Claude’s analysis and ran it past a different AI, asking it to find problems. It found plenty. A few places had been too definitive, a few medication suggestions had drifted into prescribing territory, the literature citations weren’t quite tight. I took the critique back to Claude and asked it to rewrite.
It accepted everything, and rewrote it more carefully:
- “Almost certainly diagnosed as…” turned into “Strongly suspected, pending CT confirmation.”
- “Recommend stopping IV fluids” turned into “Please have the attending vet decide whether to continue.”
- “Acute causes have been ruled out” turned into “Likelihood reduced, but further workup is still needed.”
The self-correction surprised me. But the part that stayed with me longer was this — if it hadn’t occurred to me to ask a second AI to challenge the first, I would never have known the first draft had those problems.
A regular pet owner doesn’t have the time or the energy to do that kind of cross-check.
V
I should tell you how the story actually ended.
L eventually took Wheat to a very well-known specialist. With better imaging, the specialist saw the whole picture. Wheat didn’t just have renal dysplasia — she had only one kidney in her entire body, and that single kidney was itself malformed.
The life expectancy the doctor gave was under one year.
L told me they had decided to send Wheat back to the breeder.
I tried to talk him out of it. When a breeder takes back a puppy already diagnosed with a terminal condition, the usual outcome isn’t quiet retirement. I offered a few alternatives. But L had his own situation. In the end, I was just an outsider.
I asked the AI again whether there were any volunteer-run pet hospice groups in Beijing. Claude searched for a while, and came back with a fairly honest answer: there aren’t any established pet hospice organizations in mainland China. People have insurance-covered palliative care centers. Pets don’t.
I sent that result to L. He replied: “Thanks for looking into it. The doctor said the same thing.”
I didn’t say anything more after that.
VI
After the whole thing was over, I kept asking myself: what did the AI actually contribute here?
It made a chaotic stack of reports legible. That was real. It put together a document so L wouldn’t have to retell the whole case from scratch at a new hospital. That was real. It pulled up literature, built out differentials, explained the lab values one by one. All real.
It also nearly tricked me with hospitals it had invented. It used overconfident phrasing at first. It gave medication advice that crossed lines a model shouldn’t cross.
The one who actually diagnosed Wheat wasn’t the AI. It was the specialist. With CT, with years of experience, with hands and eyes and judgment in the room. The AI’s contribution was getting L to that specialist’s door with his homework already done.
VII
If you have a pet now and you’re wondering how to use AI —
Things it’s good for: pulling multiple reports into one picture, translating medical terms into plain language, drafting the questions you should ask the vet, explaining the basics of a long-term condition, giving you a second opinion when something a vet said doesn’t sit right, organizing case notes.
Things it can’t replace: the first visit always goes to a real clinic; in an emergency (seizures, difficulty breathing, heavy bleeding) you go straight to the clinic, no chat first; any hospital, phone number or price the AI gives you, you verify yourself with a separate web search; any medication advice, you take it to the vet, not the pet; if the AI says “this is the diagnosis,” ask whether a gold-standard test has confirmed it.
If I could only leave one rule behind, it would be this: don’t ask only one AI. Either cross-check across different ones, or take the answer to a real professional. The thing AI is most consistent at isn’t being right — it’s saying things with conviction.
VIII
The first message L sent me, with all those photos attached, said: “Can you take a look? I don’t know what to do.”
What he needed in that moment wasn’t an AI, and it wasn’t a doctor either. He needed a person who could help him think.
What the AI did, really, was amplify what a friend can offer. The depth of information a regular person can now reach through a chat window — the synthesis, the literature lookup — is something that simply wasn’t available ten years ago.
But it also amplified the other side: an unqualified friend giving bad advice. The catch is that this friend sounds more professional, more confident, and less like someone giving bad advice.
So my stance now is this: use it, but keep some distance. Bring it to the doctor, don’t let it stand in for the doctor. Think of it as a clever friend who happens to read a lot of papers, not as a licensed clinician.
I hope Wheat has a gentle time in whatever is left.
And I hope this small record is useful to someone who finds themselves in a similar place, and saves them a step or two.
Written May 2026.