A few months back, after coming home from vacation, I gave artificial intelligence a genuine task. I stopped sorting my laundry, sat down with Claude, and requested that it examine the full extent of my genome. In less than half an hour, it completed what a group of around 30 individuals required nearly a year to accomplish in 2009.
This outcome shows just how advanced AI has become — along with a caution about where it’s headed. Standards must be set before people begin relying on these systems for their health.
Genome Analysis in 2009
Back in 2009, interpreting a whole genome was a massive undertaking. It involved 30 people, including me, working together to interpret the sequence of a colleague, friend, and fellow scientist named Stephen Quake. The team included experts in genetics, computer science, pharmacology, and clinical medicine.
A specialized engine was created for genome analysis, then used to examine millions of genetic variants while also reviewing the scientific literature and debating which findings would matter most for his individual health circumstances. The result was a study published in The Lancet in 2010, one of the first attempts at interpreting the complete genome and the first to turn that analysis into personalized medical recommendations.
A year after the fact, my genome was analyzed through a comparable framework, even if it was never made official and never published.
What Claude Did in 30 Minutes
This month, Claude tried the same thing — without the 30 people and nearly yearlong expenditure. I asked it to follow the same framework we had set forth back in 2009, in a prompt of 180 words. I told it to look for rare variants that could cause an inherited disorder, genetic abnormalities that could affect my response to different medications, and whether it had enough data to estimate my risk for common diseases.
The assignment was far from easy. The genome file dates back over a decade, was created with an older edition of the human reference genome, and omits data found in more recent sequencing standards. Such shortcomings might readily result in an analysis that is misleading or inaccurate.
Claude was able to recognize several key discoveries from our initial investigation. The single finding it marked as the most significant concerns my genome: I possess two copies of the APOE ε4 variant, which carries an elevated risk for Alzheimer’s disease. It did not detect any major gene-disrupting variant tied to a rare inherited condition, agreeing with the conclusion reached by our original group. Additionally, it noted variants in the DPYD and CYP2C19 genes, which can affect how we react to specific drugs.
It correctly warned me that my file lacked sufficient data before attempting to compute genetic risk scores for common diseases.
“A genome is not a definitive record.”
The whole exchange drew on only about 400,000 tokens and ran to roughly $5, with most of the half-hour passing while waiting for scientific databases to load. A job that once demanded nearly a year and a large, specialized team can now be attempted at home for roughly the price of a two-liter bottle of soda.
Why the Genome Is Not Definitive
There is a common belief that once a genome sequence has been read, it stands as a lasting record that never changes. The truth is more complicated: every genome analysis gets shaped by several things, including the reference genome it is compared against and the sequencing technology employed to read the DNA.
Over the last couple of decades, sequencing technologies have undergone major improvements. Researchers can now study parts of DNA that were previously beyond reach, and they can spot types of genetic variation that earlier techniques might have overlooked. Yet there are still some things these tools cannot do.
Most genomes in existence today come from a method called “short-read sequencing,” which involves breaking DNA down into millions of pieces and then relying on software to put those fragments back together in their proper places. While this process functions effectively throughout the majority of the genome, it tends to become less dependable in areas where sequences overlap, look alike, or carry extensive structural shifts.
These areas are not without significance. Harmful genetic changes can appear in places where standard sequencing has difficulty detecting them. A conclusion might state that no troubling variant was found, yet fail to explain that the pertinent gene was not fully examined.
The standard DNA sequence used for most comparisons also stands as a barrier. Built from a limited number of individuals, it fails to capture the full range and depth of human genomes. This inadequacy makes certain genetic variations harder to spot, particularly among groups that remain under-represented in genomic datasets.
The newer references show many different versions of the human genome through branching paths instead of forcing everyone’s DNA onto one template. That approach can improve variant detection, particularly in complex regions and across diverse populations. But laboratories, databases, and clinical systems have not adopted these newer references consistently.
The Cost of Public Access
What was once unthinkable is now possible through accessibility, opening doors for the general public. As these doors swing wide, the scientific community carries a duty to establish clear standards for what constitutes a medically reliable genome.
The comparison is stark:
| Task | 2009 Approach | Claude’s Approach |
|---|---|---|
| Time | Nearly a year | Half an hour |
| Cost | Large, specialized team | ~$5 |
| People | 30+ experts | One AI session |
| Output | Published in The Lancet | Personalized results |
The contrast shows how much things have shifted. A project once carried out by professionals has become something nearly anyone can try at home, costing next to nothing.
Standards Are Needed Now
The public now has easy access to AI tools that can interpret genomes, and those interpretations could shape the direction of someone’s life and their health choices going forward. Because of that, there has never been a more pressing moment to establish standards for what counts as a high-quality medical genomic interpretation.
Simple genetic alterations are caught well by current quality checks. A changed DNA letter, or a small piece of DNA taken away or put in, shows up without much trouble. But medically important changes can be harder to find. That likeness applies to checking a book: spotting a misspelled word is plain enough, while finding a sentence or paragraph that seems misplaced requires more care.
The reliability of a medical-grade genome test depends on its ability to identify every genetic change that could influence a person’s health. The reference genome employed, the sequencing technology, and the demographic makeup of the data all need to be taken into account when judging the test.
The technology is ready. The standards are not.
The Warning From the Experiment
This story has left the writer amazed at what it describes. A decade-old genome was analyzed by an AI in 30 minutes for $5, and the result matched decades of expert work. The writer holds that public access now calls for clear standards of reliability.
It comes down to whether the scientific community can keep pace with the technology. That is already far ahead, and the public is already making use of it. Standards, where they exist, remain scattered across laboratories, databases, and clinical systems that have not adopted them consistently.
There’s not a second to spare. The doors stand wide, and the AI waits inside. Yet before any action follows, the scientific world must settle upon a shared standard for what qualifies as a trustworthy reading of a genome.
Recognizing that the genome does not serve as a final authority is essential. It functions instead as a work in progress, continually molded by the instruments employed to study it, and those instruments remain incomplete.
Before anything else can move forward, there must be a common set of rules for what counts as a trustworthy reading of a person’s genome. Getting everyone to agree on these rules will take bringing together sequencing labs, reference genome projects, and clinical systems into a single, coordinated effort — a difficult task, but one that can be done.
Transparency sits at the core of the third step. Each genomic interpretation must reveal the reference genome it depends on, the sequencing technology deployed, and the demographic makeup of the data behind it. Building trust is what transparency achieves, and trust is the foundation upon which any medical decision rests.
Continuous improvement is the fourth step. As sequencing technologies keep getting better and new references keep expanding coverage, the standards must evolve along with them.
In the fifth stage, we reach humility. The AI duplicated our initial group’s results within half an hour, yet it did not take our place. It handled the laborious tasks, but a person was still required to read over the output and caution me about the bounds of my document.
The Bottom Line
It took the AI just 30 minutes to analyze my genome for $5, matching our original team’s results. The report turned up the APOE ε4 variant, marked the DPYD and CYP2C19 variants, and noted the limits of my file. All of it was produced using a fraction of the resources we put into 2009.
The accomplishment lies in what the machine can produce on its own: it can match the output of 30 people in half an hour. Yet the machine’s own results remain incomplete without human judgment, which must step in to explain where the data falls short and what the limits of the information really are.
The people are arriving. The artificial intelligence has been prepared. The rules, however, have not yet caught up.
Before any medical guidelines can be set for what counts as a medically reliable genome, the scientific community must settle on clear standards for defining it. There is no moment left to wait.
Source material: “Opinion: Claude analyzed my genome in 30 minutes. Now we need standards for the results,” STAT.
Get the Notebook.
The day's best stories and every fresh verdict, in plain English, in your inbox by seven. One email a day, no more.