AI in Science: Hype vs. Reality in Scientific Discovery
Leaders promise AI will cure all diseases.
But what is the actual progress?
This article explores AI’s real impact on science, from rapid data analysis to novel discoveries, while navigating the crucial human bottlenecks.
Where Is All the A.I.-Driven Scientific Progress? Separating Hype from Reality
The glow of the monitor was a familiar companion.
I remember it vividly – 2019, late nights blurring into early mornings, hunched over a colossal dataset.
My PhD, my future, felt tethered to the endless task of extracting meaning, drawing conclusions, and reading paper after paper.
The grind was real; the meager $40,000 annual stipend (New York Times, 2019) a constant reminder of the prize waiting beyond graduation.
It was a silent, often solitary battle, a testament to human perseverance in the face of scientific complexity.
Every scientist knows this feeling, this deep, almost spiritual connection to the data, painstakingly coaxing secrets from its raw form.
In short: AI is profoundly changing scientific discovery, accelerating data analysis and generating novel hypotheses at unprecedented speeds.
However, real-world bottlenecks like clinical trials mean AI is an accelerator, not a magic bullet, demanding a human-first approach to validation and ethical considerations.
This lived experience, this intimate dance with data, is precisely where the conversation around AI in scientific progress becomes so vital and urgent.
Today, the leaders of the biggest AI labs paint a picture of a world where artificial intelligence ushers in a new era of scientific discovery, promising to cure diseases and tackle the climate crisis with unprecedented speed (New York Times, 2025).
But what has AI actually done for science so far?
And how do we reconcile these grand claims with the messy, human-centric reality of research?
The AI Scientist: Supercharging Data Analysis
The core problem in modern science is not always a lack of data, but a bottleneck in processing and understanding it.
We are awash in information, yet often starved for the human hours it takes to make sense of it all.
Imagine the PhD student of yesterday, spending six months wrestling a single dataset to graduation.
Today, an AI agent named Kosmos is poised to dramatically alter that landscape, revolutionizing science automation.
Developed by Edison Scientific, a spin-off of the nonprofit FutureHouse, Kosmos is pitched as an AI scientist.
Its audacious claim?
It can accomplish the equivalent of six months of doctoral or postdoctoral-level research in a single 12-hour run (New York Times, 2025).
When Sam Rodriques, a scientist turned technologist behind Kosmos, first heard this metric, his reaction was blunt: There is no way that this is true (New York Times, 2025).
Yet, through rigorous testing with academic collaborators, Kosmos replicated findings that took human scientists months to uncover, overnight.
Its insights are right about 80 percent of the time (New York Times, 2025), a rate comparable to human performance.
A Mini Case: Unlocking Diabetes Secrets
Take, for instance, a recent breakthrough with Kosmos.
The human genome harbors millions of genetic variants linked to disease, but often, the why remains a mystery.
Scientists at Edison Scientific tasked Kosmos with raw data on genetic factors related to Type 2 diabetes.
Within its 12-hour run, Kosmos identified a previously unknown mechanism for a specific variant—a variant not even located within a gene.
It pinpointed a different protein binding, identified the gene being expressed (SSR1), and connected it to insulin secretion in the pancreas (New York Times, 2025).
This was a net new contribution to scientific literature, a significant step in novel scientific discovery, which a human scientist might not have made for a long time, if ever.
Beyond Analysis: Generative AI and Novel Biological Discoveries
AI’s role is not limited to sifting through existing data faster; it is also about creating the new.
The very essence of AI, as Rodriques explains, is building models (New York Times, 2025).
This principle extends to two fundamental categories in science: modeling the natural world, such as protein folding, and modeling the process of doing science itself.
Kosmos models the process of science, but it’s the modeling of the natural world that truly excites many in AI in science.
This year, 2025, has been marked by a significant surge in generative models.
These AI systems can produce examples of proteins, antibodies, or even organisms with desired characteristics from scratch (New York Times, 2025).
Companies like Chai Discovery and Nabla are pioneering de novo antibody design, offering the promise of clicking a button and having a targeted antibody for a specific disease.
The Arc Institute, for example, designed a bacteriophage – a virus that infects bacteria – entirely from scratch, a monumental achievement in synthetic biology for novel organisms (New York Times, 2025).
This capability represents a new frontier, dramatically cutting down the arduous, iterative design phases that previously characterized drug and material discovery.
The Roadblocks Ahead: Why AI Won’t Cure All Diseases in a Decade
Despite these incredible leaps in AI-driven scientific discovery, some of the more audacious claims — like curing all diseases within a decade or two — meet with a healthy dose of scientific skepticism.
Decade is crazy, states Sam Rodriques (New York Times, 2025), a stance he’s willing to take because, as he puts it, if I’m wrong, everyone wins.
This perspective highlights the realistic AI bottlenecks faced in medicine.
The hard truth is that the primary bottlenecks in medical progress are not solely about initial discovery; they are about the intricate, time-consuming, and staggeringly expensive journey to bring a treatment to patients.
Clinical trials, for instance, can cost hundreds of millions of dollars (New York Times, 2025) and inherently take years.
You need to manufacture the molecule at scale, ensure it is human-grade, recruit enough patients (which can be difficult for rare diseases), and then wait for observable results.
Casey Newton rightly notes, There’s no A.I. shortcut for almost any of that, at least not right now (New York Times, 2025).
AI’s strength lies in optimizing the planning of these experiments, ensuring they are the best experiments we could possibly be running given all available knowledge, but it cannot bypass the physical and biological realities of the human body.
A Glimpse into an AI-Accelerated Scientific Future
So, where does this leave us?
Not in a world of instant cures, but one of unprecedented acceleration.
The year 2025 has been the year of agents (New York Times, 2025), with AI tools beginning to infiltrate labs.
By 2027, Rodriques projects that the majority of high-quality scientific hypotheses could be generated by AI agents (New York Times, 2025).
This will significantly enhance scientific discovery by providing more focused starting points for human researchers.
This future, however, demands a human-first approach, one where AI elevates, rather than replaces, the human scientist.
We must remain vigilant about risks like AI hallucinations, and actively work to preserve serendipity – those beautiful, accidental discoveries like penicillin, born from noise or mistakes.
Kevin Roose insightfully pondered if we almost want your like A.I. scientist model to hallucinate a little bit (New York Times, 2025), recognizing that randomness can fuel evolution.
The goal is to amplify brilliant minds, freeing them from mundane tasks to pursue the truly profound, not replace them.
My own memories of grappling with that dataset back in 2019 are a stark contrast to what the next generation of scientists can expect.
Imagine my younger self, simply handing over that gigantic dataset to a Kosmos-like agent, receiving a trove of validated insights in mere hours, freeing me to design the next crucial experiment rather than just analyzing the last.
The future of scientific discovery is dynamically evolving, propelled by AI, but steered by human curiosity and oversight.
It’s a future where the partnership between human and machine unlocks truths we haven’t even dared to imagine, balancing innovation with careful validation in the pursuit of science.
Frequently Asked Questions
- What is an AI scientist like Kosmos?
Kosmos is an AI agent developed by Edison Scientific that can take a research objective and a dataset, then perform extensive analysis, such as writing 42,000 lines of code and reading 1,500 papers.
It generates deep scientific insights, potentially equivalent to six months of a human scientist’s work in 12 hours (New York Times, 2025).
- Can AI make genuinely new scientific discoveries?
Yes, tools like Kosmos have already demonstrated the ability to make net new contributions to the scientific literature, such as identifying previously unknown mechanisms for genetic variants related to Type 2 diabetes.
Generative AI is also creating novel proteins, antibodies, and even organisms from scratch (New York Times, 2025).
- Why are AI-driven cures for all diseases not imminent despite AI’s power?
The primary bottlenecks are not discovery but the lengthy and costly processes of clinical trials, regulatory approval (e.g., FDA), manufacturing drugs at scale, and patient recruitment.
These experimental phases inherently take significant time, even with optimal planning aided by AI (New York Times, 2025).
- How much does an advanced AI science tool like Kosmos cost?
Kosmos costs $200 per prompt or run, which is considered a promotional price and is expected to increase.
While seemingly high, this cost is often seen as negligible by scientists compared to the thousands of dollars spent on gathering experimental data, typically $5,000 to $10,000 (New York Times, 2025).
References
References for this article include the New York Times Hard Fork podcast (2019, https://www.nytimes.com/podcasts) and the New York Times podcast, Listen to Hard Fork: Where Is All the A.I.-Driven Scientific Progress? (2025, https://www.nytimes.com/podcasts).