It feels like just yesterday that ChatGPT burst onto the scene, and in the blink of an eye, it’s everywhere. It’s helping write emails, plan vacations, and even debug code. We’ve all gotten used to having a pretty smart assistant just a browser tab away.
But now, OpenAI is setting its sights on something much, much bigger than our daily to-do lists. They’re making a serious play for the world of science.
In October, the company quietly announced a brand-new team called "OpenAI for Science." Their goal? To figure out how these powerful language models can help scientists make the next big discovery. We’re already seeing stories pop up from mathematicians, physicists, and biologists who claim that AI, and specifically GPT-5, gave them a nudge in the right direction or helped them solve a problem that had them stumped.
So, what’s the deal? Is this the beginning of a new scientific revolution, or just another tech company trying to find a new market? Let's get into it.
Why the Sudden Pivot to Science?
First off, you might be thinking, "Aren't they a little late to the game?" And you'd be right. Google’s DeepMind has been focused on this for years with incredible projects like AlphaFold, which literally changed the game in biology. DeepMind's co-founder Demis Hassabis has said that using AI for science is the whole reason he started the company.
So, why is OpenAI jumping in with both feet now?
I think a big part of the answer lies with the person they put in charge: Kevin Weil. He’s a classic Silicon Valley product guy, having led product teams at both Twitter and Instagram. But here’s the twist: he started his career as a scientist. He was two-thirds of the way through a PhD in particle physics at Stanford before the tech world lured him away.
When I read his story, it clicked. This isn't just a corporate mandate for him; it's personal. "I thought I was going to be a physics professor for the rest of my life," he says. "I still read math books on vacation."
Weil connects this new focus directly to OpenAI’s core mission of building artificial general intelligence (AGI) for the good of humanity. He asks us to imagine the future impact: new medicines, new materials, a deeper understanding of reality itself. "Maybe the biggest, most positive impact we’re going to see from AGI will actually be from its ability to accelerate science," he argues.
And according to him, the technology is finally ready. "With GPT-5," he says, "we saw that becoming possible."
A Smarter Kind of AI
So what changed? Why is GPT-5 supposedly ready for the lab when older versions weren't?
It comes down to a new trick these models have learned: reasoning. Think of it like the difference between a student who memorizes answers and one who can actually show their work. The latest models can break down a complex problem into smaller steps and work through them logically. This has made them dramatically better at math and science.
Just a few years ago, we were all blown away that an AI could get a perfect score on the SATs. Now, they’re acing graduate-level physics problems and even performing at a gold-medal level in the International Math Olympiad, one of the toughest math competitions on the planet.
"These models are no longer just better than 90% of grad students," Weil says. "They’re really at the frontier of human abilities."
That’s a bold claim, but there are some numbers to back it up. There’s an industry benchmark called GPQA, which is basically a PhD-level exam with questions in biology, physics, and chemistry. Human experts typically score around 70%. GPT-4? It scored a dismal 39%. But the latest version, GPT-5.2, apparently scores a whopping 92%. That’s a massive leap.
A Little Too Much Hype?
Of course, with big claims comes big hype. And OpenAI definitely stumbled a bit right out of the gate.
Back in October, some senior folks at the company, including Weil, got on X (formerly Twitter) and boasted that GPT-5 had found solutions to several unsolved math problems. The math community was, let's say, skeptical. They quickly pointed out that what the AI had actually done was dig up existing solutions from old, obscure research papers—one of which was written in German.
It was still a cool feat, but it wasn't the groundbreaking discovery they made it out to be. The posts were quietly deleted.
Weil is much more careful with his words now. He points out that just finding answers that have been forgotten is a huge win. "We collectively stand on the shoulders of giants," he says, "and if LLMs can kind of accumulate that knowledge... that’s an acceleration all of its own."
He’s not claiming the AI is the next Einstein. The mission isn't to have the model make a world-changing discovery on its own. The real question, he says, is: "Does science actually happen faster because scientists plus models can do much more... than scientists alone? I think we’re already seeing that."
So, What Are Scientists Actually Doing With It?
This is where things get really interesting. OpenAI shared a bunch of stories from scientists who are already using GPT-5 in their day-to-day work. It seems to be really good at a few key things:
- Connecting the Dots: GPT-5 has read basically every scientific paper from the last 30 years. It can find connections between different fields that a human researcher might never see.
- Brainstorming Partner: It can help sketch out mathematical proofs or suggest new ways to test a hypothesis in the lab.
- The Ultimate Collaborator: As Weil puts it, you can’t easily find a thousand human experts in a thousand different fields to bounce ideas off of at 3 a.m. But with the model, you can. It doesn't sleep, and it doesn't get annoyed if you ask it ten things at once.
Robert Scherrer, a physics professor at Vanderbilt, is a great example. He started out just playing with ChatGPT for fun (he got it to rewrite the Gilligan’s Island theme song in the style of Beowulf), but then a colleague at OpenAI gave him access to the high-powered GPT-5 Pro.
"It managed to solve a problem that I and my graduate student could not solve despite working on it for several months," Scherrer says. He’s quick to add that it’s not perfect. "GPT-5 still makes dumb mistakes. Of course, I do too, but the mistakes GPT-5 makes are even dumber."
Derya Unutmaz, a biologist at the Jackson Laboratory, uses it to brainstorm ideas and analyze old data sets his team had already looked at. The model came up with fresh insights. He believes using these tools is no longer optional if you want to keep up.
But Can We Really Trust It?
This is the million-dollar question, isn't it? These models are notorious for "hallucinating"—making things up with incredible confidence.
In December, a quantum mechanics scientist named Jonathan Oppenheim pointed out a major error in a peer-reviewed paper where GPT-5 had supposedly generated the core idea. The problem? The AI's idea tested for the wrong thing.
Oppenheim’s analogy was perfect: "It’s like asking for a COVID test, and the LLM cheerfully hands you a test for chickenpox."
The scary part is that these mistakes can be so subtle that even experts can miss them. The AI is designed to sound confident and agreeable, which can lull you into a false sense of security. As Oppenheim put it, "A core issue is that LLMs are being trained to validate the user, while science needs tools that challenge us."
Weil and his team know this is a huge problem. But his perspective is a bit different. He shared something a former math professor on his team told him: "When I’m doing research, if I’m bouncing ideas off a colleague, I’m wrong 90% of the time and that’s kind of the point."
The goal isn't to get a perfect answer every time. It's about spitballing ideas until you stumble upon a grain of truth. So, OpenAI is working on making GPT-5 a bit more humble. Instead of saying, "Here's the answer," it might say, "Here's something to consider." They're also experimenting with using one AI model to act as a critic for another, essentially fact-checking its own work before showing it to the user.
The Race Is On
OpenAI isn't doing this in a vacuum. Google DeepMind is still a massive force, and other companies like Anthropic are building incredibly powerful models too. This is as much about planting a flag in a new, incredibly important territory as it is about pure scientific curiosity.
But Weil is betting that we're on the cusp of a major shift.
"I think 2026 will be for science what 2025 was for software engineering," he predicts. He explains that just a year ago, using AI to write code made you an early adopter. Today, if you're not using it, you're probably falling behind.
He believes the same thing is about to happen in science. "I think that in a year," he says, "if you’re a scientist and you’re not heavily using AI, you’ll be missing an opportunity to increase the quality and pace of your thinking."
It’s a bold vision, and it’s still very early days. This new partnership between scientists and AI will have its share of hype, mistakes, and maybe, just maybe, some truly incredible breakthroughs. It’s going to be fascinating to watch.
