rewrite this title Did ChatGPT Really Cure a Dog’s Cancer? It’s Complicated – Decrypt

OpenAI co-founder Greg Brockman amplified a widely shared story over the weekend about a dog treated with a personalized mRNA cancer vaccine developed with help from ChatGPT, drawing attention across tech and AI communities.

The case centres on Rosie, a seven-year-old Shar Pei owned by Australian AI consultant Paul Conyngham.

According to posts circulating online, Rosie had been given only months to live before receiving the experimental treatment, which Conyngham said was developed with assistance from the AI chatbot.

“Back in 2022, I noticed strange lumps on her head,” Conyngham wrote in a November 2024 thread documenting the journey from the beginning. “What the vet deemed as ‘just warts’ ended up being late-stage cancer.” Vets estimated Rosie had between one and six months left and told Conyngham there was nothing more they could do.



The account spread rapidly after Brockman shared it with his hundreds of thousands of followers, prompting coverage across several technology outlets.

While the treatment itself appears genuine, the role ChatGPT is credited with in developing the vaccine has been debated, with some researchers questioning how much of the process could realistically be handled by a large language model.

Pushing ahead

Conyngham said he didn’t give up on Rosie. Instead, he decided to build a research pipeline out of consumer AI tools. He started with ChatGPT, using it to design a plan of attack.

The model told him he needed genomic sequencing, one sample of healthy tissue and one from the tumor, and pointed him toward specific institutions and equipment.

“The most ironic thing is that in a previous chat session with ChatGPT, it said that I should attempt to reach out to Elita or Dr. Martin and that I should use an Illumina machine,” he wrote at the time.

So he followed that lead.

A director at UNSW connected him to Dr. Martin Smith, head of the Ramaciotti Centre for Genomics, who agreed to sequence Rosie’s genome for around $3,000.

Ten days. Thirty-times depth in healthy tissue, 60-times in tumor: the higher pass rate needed to isolate the mutations driving the cancer. The Centre returned 320 gigabytes of raw data.

Genomic information is expressed in strings of the letters A, T, C, and G, so experts essentially ended up with a stack of 700,000 double-sided pages full of only those four letters, the University of New South Wales reported in June of last year. That was Rosie’s genome, her biological fingerprint.

He then focused on c-KIT, a protein well-documented in the published literature on mast cell tumors in dogs.

Using Google’s AlphaFold, he modeled Rosie’s version of the protein and compared it against the healthy baseline. It looked wrong, mutated in ways that matched what the literature predicted. He then searched for existing compounds that might attack c-KIT or proteins similar to it, and found one: a drug already in use in the U.S. to treat a different cancer in humans.

“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,” Conyngham told the Australian Today Show on Saturday. “ChatGPT assisted throughout that entire process.”

AI’s true role

Even so, there’s a big gap between ChatGPT finding a cure for cancer and ChatGPT assisting in research.

Conyngham eventually connected with Prof. Palli Thordarson, Director of the UNSW RNA Institute. “Prof. @martinalexsmith performed the DNA/RNA sequencing to convert Rosie’s tissue into raw data,” Conyngham posted. “Prof @PalliThordarson assembled the mRNA vaccine,” he added in another tweet.

Thordarson confirmed this in his own thread: “Proud with @UNSWRNA to have been involved & making the mRNA-LNP for Rosie,” he wrote on X on Sunday. “The intersection of RNA technology, genomic & AI poses an opportunity to change the way we do medicine and make access more equitable.”

But Dr. Smith wasn’t a man behind a ChatGPT screen. He was a professor running a university RNA institute, doing what his lab was built to do.

And when Conyngham identified the final vaccine construct—the specific molecular blueprint that would be encoded into the mRNA—he revealed which tool designed it. Not AlphaFold. Not ChatGPT. “The final vaccine construct for Rose was designed by Grok.”

That said, he recognized in a separate post that “Gemini did a ton of the heavy lifting too.”

ChatGPT was used to sift through scientific papers and identify researchers who might be able to help. The chatbot pointed to the Ramaciotti Centre and suggested sequencing equipment suited to the task, functioning largely as a tool for navigating the research literature. That role can be useful, but it differs from designing a vaccine or performing scientific analysis.

AlphaFold, a deep-learning system from Google DeepMind, predicts three-dimensional protein structures from amino acid sequences. It’s not the first model trained on biological data: other open-source initiatives like Ankh or AlphaGenome work on similar premises.

Conyngham used Alphafold to model Rosie’s c-KIT protein. The rendering carried a confidence score of 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 work is needed to validate any output. Dr. Smith, the UNSW genomics director, confirmed publicly in the same thread that AlphaFold was not, in fact, used for the mRNA vaccine design at all.

Dr. Thordarson was careful about the framing, too.

“This may not have cured Rosie,” he wrote on X. “Bought time for sure, yes, but some of the tumours didn’t respond.”

His team is now checking whether those tumors mutated differently, which would explain why parts of the treatment worked and others did not. The vaccine also did not work in isolation.

“The treatment required co-administration of a checkpoint inhibitor,” Thordarson noted, “likely to be with all personalized cancer vaccines.”

The use of AI for cancer treatment has not always been a history of success.

In 2017, internal IBM documents revealed that Watson for Oncology, marketed as a system that could recommend cancer treatments better than human oncologists, was generating what its own engineers flagged as “unsafe and incorrect” recommendations.

MD Anderson Cancer Center abandoned the project after spending $62 million on it. IBM sold off Watson Health in its entirety in 2022.

The Rosie case doesn’t fall into the category of AI failures. No one was harmed, the underlying science is established, and the researchers involved have recognized credentials.

The mRNA platform itself is supported by clinical research. The unease lies more in how the story has been framed. When AI tools receive credit for work carried out by scientists and research institutions, it can blur public understanding of what the technology actually does.

The researchers who performed the sequencing, produced the vaccine, and managed the safety protocols risk fading into the background.

The episode offers a reminder that AI can assist with tasks such as navigating scientific literature, but it remains far from replacing the expertise and infrastructure required to design and produce medical treatments.