A widely shared dog cancer story has become a test case for how AI-assisted medicine is discussed in public. Paul Conyngham, an AI consultant from Australia, used ChatGPT, AlphaFold, Grok, genome sequencing and researchers while searching for a possible treatment for his dog Rosie, who has incurable mast cell cancer.
The case is compelling because it shows a medical layperson using AI tools to navigate complex research territory. It is also risky as a public narrative because the central medical question remains unresolved: whether the AI-assisted vaccine actually helped.
What happened in Rosie’s case
Conyngham started the project back in November 2024. On ChatGPT’s recommendation, he had both Rosie’s healthy genome and tumor genome sequenced at the Ramaciotti Centre for Genomics at UNSW Sydney. The sequencing alone cost him $3,000.
He then used AI systems to identify a target protein and an already FDA-approved substance that could help. According to Conyngham, the final vaccine design was created using the Grok AI model.
After the drug was administered, Conyngham says the cancer has shrunk by about 75 percent. Rosie, however, has not been cured.
The story spread quickly on social media and was shared by OpenAI President Greg Brockman and Deepmind CEO Demis Hassabis, among others. Conyngham has since written up his process in detail and released his approach as an open-source method.
Why the claim is under scrutiny
The strongest criticism is not that AI played no role. It is that the case has been treated as if the vaccine’s effect is already established, when the source describes that as unproven.
Egan Peltan, co-founder of a biotech startup and a Stanford-trained PhD in chemical biology, pushed back on the hype. He says there is zero evidence that the AI-assisted work did anything for Rosie’s cancer.
A key complication is that Conyngham was simultaneously giving Rosie a PD-1 inhibitor, an approved immunotherapy drug that makes cancer cells visible to the immune system again. PD-1 inhibitors are described as one of the most effective cancer immunotherapies available.
That matters because if two interventions happen at the same time, an improvement cannot be cleanly assigned to one of them. The source says the most likely explanation for Rosie’s improvement is a response to the conventional drug, not the mRNA vaccine. Conyngham, however, says a chatbot also pointed him toward PD-1 in the first place.
How OpenAI leaders framed the story
The debate intensified after OpenAI CEO Sam Altman and VP of Science Kevin Weil used the story to promote their own narrative. Weil wrote that Paul had used ChatGPT and AlphaFold to create a personalized mRNA vaccine protocol for his dog’s cancer, calling it "a glimpse of the future, with AI accelerating personalized medicine."
Altman called it the "coolest meeting I had this week" and quoted Conyngham saying the chatbots gave him "the power of a research institute." The source notes that neither OpenAI executive mentioned that there is no evidence the vaccine actually worked or that the whole effort made any difference.
That omission is the core issue. AI may have helped Conyngham explore options, coordinate information, or move through research steps that would otherwise have been inaccessible. But a persuasive story about access is not the same thing as proof of medical efficacy.
What experts say still has to be proven
Patrick Heizer, a researcher in cell and gene therapy, warned against getting too excited. He says fighting tumors in the lab is the easy part. The real challenge is proving a therapy is both safe and effective in controlled human trials.
One major issue is precision. Proteins in the body often look very similar, so a therapy that targets a tumor protein could also hit similar proteins in healthy organs like the heart, Heizer explains.
There is also a translation problem. Results in animals do not move directly to humans, because mice and dogs have different proteins than people do. Heizer also points out that pharmaceutical companies and regulators have to ensure long-term safety over five or more years, something that is impossible to measure with short-lived lab animals.
Peltan also argues that AI’s role in the case is overblown and that all of this could have been done without ChatGPT. He estimates the real cost of treatment at $20,000 to $50,000. He also notes that personalized mRNA cancer vaccines have been in development for years with no clear success in large-scale trials.
His broader point is that the field needs Phase 3 results, not anecdotes, before anyone argues that regulators are standing in the way of life-saving treatments. Peltan calls the entire thing "storytelling for AGI true believers. Specifically, a story in search of venture money." He says Conyngham should provide evidence that the individual vaccine story actually worked before selling it to consumers.
The real takeaway from the AI medicine story
The useful lesson is narrower than the hype suggests. This case shows that AI tools can help a determined non-specialist engage with highly technical medical and genomic information. That is significant, especially when paired with researchers and real sequencing work.
But the unresolved questions are just as important:
- Did the personalized mRNA vaccine help Rosie, and if so, by how much?
- Was the improvement instead caused by the PD-1 inhibitor?
- What were the full costs and tradeoffs of the treatment path?
- How would safety and efficacy be proven beyond one case?
Until those questions are answered, the story is best understood as an example of AI-assisted exploration, not proof that AI created an effective cancer treatment. The difference matters because personalized medicine is not validated by a compelling anecdote. It is validated by evidence strong enough to separate hope from effect.