Emma Pierson Is Wrong About AI and Cancer
The Atlantic’s case for slowing AI treats cancer progress as all or nothing. Clinical evidence supports targeted safeguards and faster research.
Elena Marsh covers science policy for Stanford Tech Review, with a focus on federal research funding, biosecurity oversight, and how Washington governs the labs it pays for.
Emma Pierson's essay in The Atlantic, "I'd Rather Risk Cancer Than See AI Move This Fast", makes an unusually personal case for slowing artificial intelligence. Pierson carries a mutation that puts her at high risk of several cancers. She has already undergone surgery that ended her natural fertility. She writes that she would accept more time under that risk if a slower pace of AI development gave society time to handle unemployment, inequality, surveillance, autonomous weapons, and the loss of human meaning.
Her candor deserves respect. Her conclusion deserves forceful rejection.
A person may choose to bear a medical risk. That choice cannot become a social mandate for millions of other patients, families, and researchers. The argument also sets the wrong evidentiary bar for medical progress. AI does not need to "cure cancer" in one cinematic breakthrough to justify faster research. Earlier detection, better trial design, faster target identification, and more productive clinicians can save lives long before anyone claims a universal cure.
The responsible policy is targeted control: test models before release, restrict genuinely dangerous capabilities, monitor deployment, and punish harmful uses. A broad slowdown taxes every beneficial application, including applications whose risks are already governed by clinical trials, medical-device review, privacy law, and professional oversight.
A private sacrifice cannot settle a public tradeoff
Pierson's essay gains moral force from her willingness to accept cancer risk herself. It loses moral force when that sacrifice is used to support a general slowdown that other people did not choose.
The International Agency for Research on Cancer estimated 20 million new cases and 9.7 million cancer deaths in 2022. That annual toll works out to roughly 26,600 cancer deaths every day. AI would not have prevented all of them. The number establishes the scale of the constituency missing from a first-person wager: patients with aggressive disease, families facing hereditary risk, clinicians with too few specialists, and health systems that cannot afford endless manual review.
The ethical question therefore extends beyond whether one researcher would wait. It asks who gets to impose waiting on everyone else, which advances would be delayed, and what evidence would ever be sufficient to restart. "Slow down" sounds prudent until it must become an operational rule. Would it cap computing power? Delay open scientific models? Restrict hospital decision-support systems? Freeze protein-design tools? Block code assistants that help biostatisticians analyze trials? Each choice carries its own risks and benefits. Treating them as one undifferentiated object called "AI progress" hides the decision that policy actually has to make.
Cancer progress arrives as a series of gains
Pierson correctly observes that biology cannot run at silicon speed. Patients are not simulations, clinical trials take time, and ethical constraints are essential. Her inference goes too far. A bottlenecked process can still accelerate when several stages around the bottleneck become faster and more accurate.
Drug development includes target selection, molecule generation, toxicity prediction, protocol design, patient matching, image interpretation, statistical analysis, manufacturing, and regulatory review. AI can improve several of those stages while laboratory experiments and trials remain deliberately human-governed. Faster computation does not abolish the clinical process. It gives that process better candidates and better information.
There is already prospective evidence. In the randomized Mammography Screening with Artificial Intelligence trial, AI-supported screening detected 338 cancers among 53,043 participants, compared with 262 among 52,872 participants receiving standard screening. That equals 6.4 versus 5.0 cancers per 1,000 people, a 29% higher detection rate. The AI-supported group also required 44.2% fewer screen readings, with no significant increase in false positives.
Those are measured clinical gains rather than a lab chief's prophecy. They matter precisely because cancer outcomes depend on time. More small, lymph-node-negative tumors were found in the AI-supported arm. Earlier detection can mean less aggressive treatment and a better chance of survival.
The drug pipeline offers another concrete signal. A 2025 Nature Medicine paper reported a randomized phase 2a trial of rentosertib, a molecule created against an AI-discovered target for idiopathic pulmonary fibrosis. The disease was pulmonary fibrosis rather than cancer, and the study was early and small. Its importance lies in crossing the boundary the slowdown argument treats as remote: an AI-generated therapeutic hypothesis moved through chemistry, preclinical work, and into a randomized human trial.
Progress looks like this before it looks revolutionary. A better screen finds more treatable disease. A stronger target-ranking system sends fewer dead ends into the lab. An automated analysis lets a research team test five hypotheses in the time previously required for one. The cumulative effect can shorten development even when no machine is allowed to experiment freely on a patient.
AlphaFold is evidence of leverage, not disappointment
Pierson points to AlphaFold as a formidable system that has yet to produce a revolution in drug development. The observation is fair. The implied standard is unfair.
AlphaFold solved a foundational prediction problem and made the result broadly usable. The AlphaFold Protein Structure Database now contains more than 241 million predicted structures. It has become infrastructure for researchers studying proteins across medicine and biology. Infrastructure rarely cures a disease by itself. Sequencers, microscopes, and public genomic databases do not either. They expand what scientists can ask and reduce the cost of answering it.
The lag between a research tool and an approved medicine is also built into the safety system Pierson wants to preserve. Targets must be validated. Compounds must be synthesized and tested. Trials must establish safety and efficacy. Regulatory review takes evidence. Calling that lag proof of limited AI value creates a trap: rapid deployment would be reckless, while careful translation is offered as evidence that the tool has failed to transform medicine.
A sound assessment measures intermediate outcomes as well as final approvals. How many structures became available? How many experimental paths were narrowed? How many researchers gained access? How many candidate molecules entered testing? Final patient outcomes remain the decisive endpoint, and the road to that endpoint contains measurable progress.
A failed rollout supports stronger gates
Pierson uses Anthropic's troubled Fable 5 release as evidence that institutions cannot handle fast-moving frontier models. A chaotic release does support stricter release discipline. It supports capability testing, staged access, incident reporting, use restrictions, and clear authority to suspend dangerous systems.
Those controls already have a policy vocabulary. The US National Institute of Standards and Technology's AI Risk Management Framework organizes risk work around governing, mapping, measuring, and managing. Its generative-AI guidance highlights pre-deployment testing, content provenance, governance, and incident disclosure. Regulators also apply domain-specific evidence standards. The US Food and Drug Administration maintains a public list of AI-enabled medical devices that have met applicable premarket requirements for their intended uses.
This risk-based approach is more demanding than a slogan. A model that materially enables bioweapon design should face a different access regime from a mammography assistant. An autonomous weapons system should face a different legal regime from a protein-structure database. A hospital tool should be validated on the patients and workflows where it will operate. High-impact deployments require monitoring, audit trails, human override, and incident response.
The broad slowdown proposal collapses these categories. It also creates a governance problem of its own. Progress would continue across countries, companies, universities, and open communities at different speeds. A vague pause is easiest to obey for visible, accountable institutions. Less transparent actors gain relative freedom. Effective governance must specify the capability, actor, use, and enforcement mechanism.
Human meaning is not a scarcity AI consumes
The essay's deepest fear concerns meaning. Pierson imagines research, writing, and intellectual discovery becoming hollow once machines can outperform people. That fear is real, and it deserves cultural responses alongside economic ones. It still provides a weak basis for delaying medical and scientific tools.
Human value has never depended on being the fastest available calculator, translator, chess player, or pattern recognizer. Scientists already use instruments that perceive what unaided humans cannot. Writers use dictionaries, editors, search engines, and software. Physicians use imaging, genomic assays, and clinical decision systems. Meaning comes from choosing questions, caring about outcomes, accepting responsibility, and forming relationships around the work.
AI may change the texture of mastery and employment. Society should respond with labor policy, education, competition enforcement, income support, privacy protection, and limits on surveillance and weapons. Patients should not become collateral in an effort to preserve one generation's preferred relationship with intellectual labor.
Accelerate the benefit, govern the danger
Pierson is right about uncertainty, institutional weakness, and the possibility of severe harm. The essay then converts those warnings into a single speed control for a technology with radically different uses. That move replaces risk analysis with risk aversion.
The stronger principle is simple: regulate the dangerous capability at the point where danger arises, and let beneficial science advance under evidence-based controls. Require pre-release evaluations for frontier systems. License or restrict access to capabilities that materially amplify biological or cyber harm. Preserve human authority in weapons. Audit high-impact employment and surveillance systems. Demand clinical validation from medical AI. Fund independent testing and give regulators the power to act.
Cancer research does not need a promise of a 95% mortality reduction to deserve urgency. A 29% improvement in detection within a large randomized study is already meaningful. A candidate drug reaching phase 2a is already meaningful. Hundreds of millions of accessible protein predictions are already meaningful. Each gain remains incomplete, and each can compound with the next.
Pierson has every right to choose what risks she would accept for herself. Public policy owes equal moral attention to the people who would choose faster progress, especially those whose disease will not wait for society to resolve every question about work, identity, and power. Their time counts too.