Did ChatGPT Really Cure a Dog's Cancer?

ChatGPT cancer treatment stories don't get much more compelling than Rosie — a seven-year-old Australian Shar Pei given months to live who ended up receiving a personalized mRNA cancer vaccine, built with a patchwork of AI tools and university lab equipment, after her owner refused to accept the prognosis. The case went viral over the weekend after OpenAI co-founder Greg Brockman shared it with his hundreds of thousands of followers. And then the scrutiny started.

The Story Behind Rosie's Experimental Vaccine

Paul Conyngham, an Australian AI consultant, noticed strange lumps on Rosie's head back in 2022. Vets initially dismissed them as warts. They weren't. By the time a proper diagnosis came back, Rosie had late-stage cancer — and the vets told Conyngham there was nothing left to try. She had between one and six months to live.

What happened next is the part that went viral. Conyngham didn't accept that verdict. He built what he later described as a research pipeline out of consumer AI tools, starting with ChatGPT. The chatbot told him he needed genomic sequencing — one sample of healthy tissue, one from the tumor — and pointed him toward specific institutions and equipment. He followed the lead, eventually connecting with a director at UNSW who introduced him to Dr. Martin Smith, head of the Ramaciotti Centre for Genomics.

Dr. Smith agreed to sequence Rosie's genome for around $3,000. The sequencing ran 30-times depth on healthy tissue and 60-times depth on tumor tissue — the higher pass rate being necessary to isolate the specific mutations driving the cancer. The Centre returned 320 gigabytes of raw data: Rosie's complete biological fingerprint, expressed in strings of the letters A, T, C, and G across what UNSW described as the equivalent of 700,000 double-sided pages of text.

Conyngham then focused on c-KIT, a protein well-documented in the published literature on mast cell tumors in dogs. He used AlphaFold — Google DeepMind's protein structure prediction system — to model Rosie's version of the protein and compare it to a healthy baseline. The result looked wrong. Mutated in ways that matched what the literature predicted. From there, he searched for existing compounds that might target c-KIT or similar proteins, and found one: a drug already in human use for a different cancer.

We took her tumour, sequenced the DNA, we converted it from tissue to data, and we used that to find the problem in her DNA and then develop a cure based on that. ChatGPT assisted throughout that entire process.

What Did ChatGPT Actually Do?

Here's where the story gets messy — and where the headlines started to mislead. ChatGPT did not design the vaccine. The actual mRNA vaccine construct for Rosie was designed by Grok, Elon Musk's AI model. Conyngham acknowledged this directly in a separate post, while also noting that 'Gemini did a ton of the heavy lifting too.' ChatGPT's role was narrower: it helped him sift through scientific literature and identify researchers who might be able to assist.

That role — AI as research navigator — is genuinely useful. But it's a long way from 'ChatGPT cured a dog's cancer.' Prof. Palli Thordarson, Director of the UNSW RNA Institute and the scientist who actually assembled the mRNA vaccine, was careful about the framing. He confirmed the collaboration but was explicit about its limits.

The AlphaFold angle deserves scrutiny too. Conyngham used the protein modeling tool to render Rosie's c-KIT structure — but the confidence score on that output came back at 54.55, which UNSW structural biologist Dr. Kate Michie publicly described as low. She noted that AlphaFold 'can get stuff wrong' and that significant lab validation is required before acting on its outputs. Dr. Smith subsequently confirmed publicly that AlphaFold was not, in the end, used for the actual mRNA vaccine design at all.

This may not have cured Rosie. Bought time for sure, yes, but some of the tumours didn't respond.

What the Science Actually Shows

The treatment itself is not snake oil. The mRNA platform is supported by clinical research — the COVID vaccines were built on it, and personalized cancer vaccines based on the same platform are an active area of legitimate oncology research. The researchers involved — Dr. Smith, Prof. Thordarson — have recognized credentials and did real scientific work. Nobody was harmed. The underlying biology is established.

But the results were partial. Prof. Thordarson posted on March 15, 2026 that his team is still analyzing why some tumors failed to respond. His working hypothesis: those tumors may have mutated differently from the sample used to design the vaccine, which would explain why parts of the treatment worked and others did not. He also flagged that the treatment required co-administration of a checkpoint inhibitor — a standard immunotherapy drug — and that the overall costs, when you account for all the in-kind time and resources contributed by university labs, were 'quite high.' This was not a cheap or replicable garage experiment.

The case sits in a genuinely complicated middle ground. Rosie's condition appears to have improved — Conyngham has shared follow-up posts suggesting the tumors have responded. But improved is not cured. And the story as it spread across tech media credited ChatGPT with a role it did not play, while the researchers who did the actual scientific heavy lifting were footnoted.

Why AI Health Hype Has a Real Body Count

The Rosie story would be easier to dismiss if AI health hype didn't have a documented failure history. In 2017, internal IBM documents revealed that Watson for Oncology — a system IBM marketed as capable of recommending cancer treatments better than human oncologists — was generating recommendations its own engineers flagged as 'unsafe and incorrect.' MD Anderson Cancer Center walked away from the project after spending $62 million. IBM sold off Watson Health entirely in 2022.

Watson for Oncology failed not because the underlying idea was absurd but because the hype machine ran too far ahead of the science. Sound familiar? The Rosie case is different in important ways — no patients were harmed, the science is real, and Conyngham has been relatively transparent about the tools he used and their actual roles. But the viral framing — 'ChatGPT cured a dog's cancer' — repeats the same structural mistake: assigning AI credit for outcomes that depended on human scientific expertise.

When AI tools absorb credit for breakthroughs that were actually delivered by university researchers and specialized lab infrastructure, it warps public understanding of what the technology can do. People start believing that the limiting factor in cancer treatment is access to the right chatbot, not the decades of scientific training and millions in lab equipment sitting behind Dr. Smith and Prof. Thordarson. That misunderstanding has consequences — for research funding, for policy, and for the next person who decides to build a DIY cancer research pipeline and doesn't happen to have UNSW on speed dial.

Rosie, by most accounts, is doing better. That's the genuinely good news here. But the credit belongs to the scientists who showed up, not the chatbot that pointed the way.