"AI Doctors Are Here": A Father's Account of Fighting His Son's Burkitt Leukemia With GPT-5 Pro and Gemini 3

Nathan Labenz opens by calling this the most personal episode he has ever recorded. His 6-year-old son, Ernie, had been a healthy boy who spent a weekend jumping on trampolines and in bounce houses. He was diagnosed with Burkitt lymphoma, later reclassified as Burkitt leukemia after cancer was found in the bone marrow.

The disease is among the fastest-growing cancers, with a doubling time as short as 24 hours, but also one of the most responsive to treatment. Doctors at Children's Hospital of Michigan in Detroit started chemotherapy a day before the final pathology report. Ten days in, Labenz says, the central mass had shrunk by roughly 80%, and bone marrow and spinal fluid tests had come back clear, which he calls a best-case response.

Labenz says he waited to share the story until he reached that milestone. His stance throughout is that in a serious situation you should use both human and AI doctors aggressively. He frames the episode as raw recollection mixed with practical lessons for AI enthusiasts and for anyone facing a medical crisis.

He explains cancer to his son as cells growing out of control, and sees the same exponential logic that people discuss in AI. Early symptoms were ambiguous: a week of tummy aches in early September, then knee pain after the trampoline weekend. His wife worried sooner, and he says the first thought that it might be leukemia came partly from her AI use.

The family left on vacation on October 22, with no strong reason for alarm. Over the next days Ernie had waves of pain, restless nights, and on the drive home from a food festival in Louisiana screamed about pain in his tooth and head. An urgent care visit with an X-ray of the knee showed nothing, partly because he seemed calm by the time he was examined.

From this Labenz draws practical tips. Film a child's worst moments and show the doctor, and keep a detailed symptom log so a provider can read the whole history in black and white. On the last full day, with a low-grade fever, their pediatrician advised the ER, where blood work was mostly normal except one value.

That value was lactate dehydrogenase (LDH), a marker of cell turnover, at about 1500, roughly three times the reference range. The lab noted hemolysis of the sample and repeated the test with the same result and the same note. The ER doctors said they had ruled out anything life-threatening, and the family flew home.

That night Labenz compiled the records from the hospital portal, copying results by hand because he did not trust an agent for a one-off task. He wrote up the medical history and ran one big prompt through GPT-5 Pro deep research, Claude Opus 4.1, Gemini 2.5 and Grok 4. He calls the result a pivot moment.

The models differed in tone but agreed that the LDH was a red flag and that hemolysis would not explain such a high number. They told him to seek further care within 24 to 48 hours, and leukemia was on their lists. Claude was the most alarmist and told him it looked like leukemia, which triggered what he describes as a mild panic attack. He preferred GPT-5 Pro's longer, more clinical response, though he concedes Claude's urgency may have been appropriate.

Labenz judges that the models were more accurate than every doctor they had seen so far. His wife used regular GPT-5 with screenshots from the portal, and that version repeated the doctors' mistake because the screenshot carried the hemolysis note. His lessons are to use the Pro tier for anything that matters and to present the same information in several ways. He also says humans and AI are both needed, because you can get many AI opinions but only a few human ones.

Their pediatrician said the likeliest outcome was that it would pass, but he arranged an oncology visit at Children's Hospital the next morning for safety. The oncologist's first guess was something autoimmune. When the doctor pressed on a spot in Ernie's lower left abdomen he recoiled in pain, and the parents realized the same thing had happened at the pediatrician's office the day before.

Looking back, Labenz wishes he had asked an AI to interview him or guide a physical exam of his son. On Halloween, his wife pushed forcefully for an immediate ultrasound, and the hospital agreed to do it right away. The preliminary read led to an MRI request, and the full report that appeared in the portal over the weekend, once run through AI, made clear it could be cancer.

He also suggests recording appointments and having AI transcribe them later, since the family only got brief preliminary verbal readouts. The report described a mass of roughly 5 by 6 by 4 centimeters in the belly plus possible involvement of the liver, spleen and kidneys.

The MRI was scheduled for Monday under general anesthesia, and the family debated going to the ER over the weekend. Doctors and the AI both leaned toward neuroblastoma, a slower-growing cancer, and the attending oncologist judged that a few days would not matter. Labenz now thinks the doctors anchored too much on the most likely scenario and underweighted less likely ones that demanded urgency.

After the Monday MRI came a Tuesday biopsy, bone marrow aspiration and a CT scan, and Ernie was admitted. Labenz describes the following days as the scariest of his life, including a long cry when his son first went under anesthesia. By Wednesday Ernie was swollen, weak and unsteady, and the parents pressed the team to start treatment because the weekend meant a diagnosis might not come until Monday.

When Ernie said he could not move his legs, the oncologist's concern visibly jumped. He decided to start chemotherapy before the official diagnosis, starting with a steroid that Wednesday and heavier drugs on Thursday. The surgeon had said it was definitely not neuroblastoma, so lymphoma treatments were similar enough. Labenz notes that organs were still working well, and he believes organ failure could have come within days.

On November 7 the team said it looked like Burkitt lymphoma: very aggressive but very treatable, with cure the most likely outcome and a poor prognosis if it returned. That moved the family from panic to a managed situation with a schedule. The first chemotherapy phase is gentler to avoid tumor lysis syndrome.

Labenz says there were three things parents could do, and AI helped with all of them. The first was double-checking the care non-stop. Labs came every six hours, the drugs were new, and the picture changed fast. At one point a diuretic removed close to a gallon of urine from a boy who weighs about 50 pounds.

He kept a long thread with GPT-5 Pro and pasted PDFs printed from the patient portal about 20 times in ten days. He says it was on the level of the human doctors in its analysis and confirmed that the treatment was highly protocolized and state of the art. It also helped answer whether they should move centers, with the answer that for a known cancer with an established protocol, the care available elsewhere would be no better.

On the margins it differed from the doctors. GPT-5 Pro wanted calcium and magnesium replenished sooner, would have waited and re-measured before treating a slightly elevated blood pressure, and wanted a few more tests. Its explanations let Labenz have an informed conversation, and he agreed with the doctor's call to treat. Overall, he says, it gave him a great deal of confidence that the team was on top of things.

Prompted by a post from Anthropic's Amanda Askell about regularly using 100-page prompts, Labenz says today's models handle long context well. His new standard prompt includes the parents' history of the initial symptoms plus MRI, CT, PET, spinal fluid, bone marrow and pathology reports. Around ten PDFs can go into one prompt, which gives a holistic view of the case.

He also asks the long-running thread to turn its history into a handoff document for a new attending physician. The result was a couple of thousand words of dense, useful material that works well as input for another prompt. Without AI, he imagines constant Googling and anxiety, whereas with it he felt empowered to advocate effectively.

The second lever was minimizing infection risk while Ernie's immune system is wiped out. Labenz and his wife both work from home, so they can run a COVID-style quarantine, and he will skip travel and speaking events for months. The treatment roadmap is six months, and doctors worry more about bacterial and fungal infection than viral.

On the day they returned from New Orleans he found about two inches of water in the basement, which raised the question of mold. He used GPT-5 Pro deep research to find local remediation companies and to interpret the air-sample test results. The results showed mildly elevated mold in the basement and the rest of the house fine.

The practical problem was combining mold remediation with oncology and the specifics of his son's case, something no human meeting would easily deliver. He ended up with a plan to run eight to 10 HEPA filters around the house, keep Ernie out of the basement during work, and possibly stay with his parents for a couple of days. The air would then have fewer spores than the autumn outdoors, and he says the analysis gave him confidence he would not have had otherwise.

The third lever was preparing for a possible relapse, since relapsed Burkitt carries a bleak prognosis. Labenz describes the treatment path: the prephase, two intensive rounds, two moderate mop-up rounds and two mild consolidation rounds. He says he might shave his head in solidarity if his son loses his hair. Genetic testing still pending could show a related subtype with a different prognosis.

For this planning he used GPT-5 Pro alongside what he believed was a Gemini 3 checkpoint in Google's AI Studio, which he declines to benchmark publicly out of respect for the testing request. He found it about as sharp as GPT-5 Pro but five to ten times faster, with focused answers that were easier to digest. By his account it quickly became the model he was most excited to read, with GPT-5 Pro serving as the checker.

He also shares a practical caution about document transcription. Faxed or photographed printouts are error-prone, and GPT-5 Pro was not great at them, so he used Gemini 2.5 Pro and then checked carefully. The plan the models helped shape includes getting the genetic subtype, establishing relationships with two leading oncology centers through remote second opinion services that may cost under $1,000, and learning which trials exist.

On advanced therapies such as CAR-T and immune-recruiting drugs, the Gemini-style model explained why they should wait for relapse. The current protocol works extremely well, and chemotherapy suppresses the immune system so immunotherapies cannot work yet. Labenz says it also explained why several of his customization ideas were not good ideas now but might be later.

The most striking thing the AI surfaced, according to Labenz, was ctDNA, or circulating tumor DNA. The tissue sample can be sent for a genetic fingerprint, and later blood tests can detect cancer at a very sensitive level, which he gives as about five cells in a million. Their oncologist knew of it but said it is not standard, and an oncologist friend who works with solid tumors was skeptical.

He argues that for a lymphoma with a 24-hour doubling time, the standard monitoring of monthly blood draws and quarterly scans is inadequate. Most relapses are found because the patient is already sick. Burkitt usually returns within six months and almost always within two years, so ctDNA testing would only be needed for a limited period. It might even reduce scan radiation and avoid biopsies of leftover scar tissue.

The model's framing was that if ctDNA turns positive on a Tuesday, you want a next-generation drug infusing by Thursday. Treating a small amount of disease can mean milder side effects such as cytokine storm and fewer chances for cancer cells to evolve escape routes. Labenz concedes doctors will say there is no data proving this, but it makes sense to him.

He notes that at eight days from a confident diagnosis, while spending around 18 hours a day at the hospital and with his wife living there around the clock, he could already map the subtypes, drug classes, trials and second-opinion options. He thanks family members who stayed with the couple's other two children. Starting from zero and being conversant within days is, he says, remarkable.

Labenz says the experience makes the "AI is a bubble" debate feel beside the point: AI has been indispensable to him. He pays $200 a month for GPT-5 Pro, while a deep research report estimated that insurers might pay $500,000 to $1 million for six months of care. That puts the AI cost somewhere between a tenth of a percent and 2% of the total. He says he would pay $10,000 a month and still feel fine.

He argues AI could even reduce costs by minimizing complications such as infections and by sparing the family a costly move to another center. The models can't do procedures, physical exams or prescribing, but on analysis, reasoning and explanation he says "you really do have AI doctors now." In his view, the strongest revealed preference for AI is what people do when trying to care for their child.

He says there are no doomers in the pediatric oncology unit. He stands by his decision to sign the call for a ban on superintelligence, and still thinks an uncontrolled intelligence explosion is a bigger threat to his son than the cancer. Even so, he now feels more deeply that delaying an AI oncology researcher who could take an N-of-one case would cost lives.

Labenz turns to policy. He asks whether the US and China can race to cure cancer, imagining a gold medal tracker for cancer cures, and says Western leaders show a failure of imagination when they leave talk of longer healthy lifespans to authoritarians. He thinks the right response is to articulate a positive vision of healthier lives and deliver on it.

He praises the medical system that treated his family: a protocolized treatment, quick scheduling and willingness to start chemo early. Even so, he predicts that standardized medicine will become handcuffs as AI lets patients reason over far more information than any single doctor can. In his view patients will soon arrive with plans that plausibly beat the standard of care, even though he has not found one in his own case because the protocol is so strong.

He wants "right to try" reform, a liability standard where going beyond standard of care is not automatically negligence if done carefully, and much better data access. He thinks that getting records from the hospital has been a battle and that richer data would feed better analysis. He believes doctors will come around, and says his oncologists now trust that his questions reflect real homework.

He closes by saying he is grateful for the collective effort behind an effective treatment. He imagines the pain of facing this when it might be solvable by AI within a few years, and says there will likely be more AI-and-cancer episodes. The prognosis remains good but the road ahead, with harsh chemotherapy, is rough.