Science has a repetition problem.
Before any research finding gets treated as trustworthy, someone else is supposed to be able to redo the work and get the same result. That process is called replication, and it is one of the least glamorous, most important jobs in all of science. It is also slow, tedious, and expensive — which makes it exactly the kind of task AI labs are now racing to automate.
Enter Inherent, a British AI lab founded by DeepMind alumni. The company announced that its AI “teammate,” an agent called Faraday, outperformed offerings from Anthropic and OpenAI at replicating research in 2026. That is a bold claim, and it deserves both some excitement and some careful unpacking. So let’s do that — in plain English, as always.
What does it mean for an AI to “replicate research”?
Imagine a scientist publishes a paper. Replicating it means reading the paper, understanding exactly what the researchers did, rebuilding their experiment or analysis from scratch, and checking whether you get the same outcome. For a human, that can take weeks or months. You need to interpret ambiguous descriptions, fill in missing details, write code, run tests, and debug when things inevitably go sideways.
That combination of reading, reasoning, coding, and troubleshooting is why replication has become a popular test for AI agents. It is not a single skill. It is a long chain of skills that all have to work together, more like a project than a puzzle. If an agent can pull it off, that says something meaningful about its ability to handle real, messy, multi-step work.
Why the “teammate” label matters
Notice that Inherent does not call Faraday a chatbot or an assistant. It calls it a teammate. That word choice is doing a lot of work, and I think it signals where the whole field is heading.
An assistant waits for instructions and answers questions. A teammate takes on a chunk of the project, works through it over an extended stretch of time, and comes back with results. The difference is autonomy and duration. Replicating a research paper is not something you do in one chat message — it is sustained, goal-directed effort. Positioning Faraday as a teammate tells us Inherent is building for that longer horizon.
David versus two Goliaths
The most eye-catching part of the announcement is the comparison. Anthropic and OpenAI are two of the biggest names in AI, with enormous resources behind them. A smaller lab claiming to beat both at a demanding task is a genuinely interesting story.
It also fits a broader pattern. DeepMind alumni have founded dozens of European startups in recent months, and these founders bring serious research pedigree with them. The idea that frontier-level capability belongs exclusively to a handful of giant American labs keeps getting challenged, and Inherent’s claim is another data point in that direction.
The caveats you should keep in your back pocket
Now for the friendly-explainer fine print, because I would be doing you a disservice without it.
- This is the company’s own claim. Announcements about outperforming rivals come from the party with the most to gain. That does not make the claim false, but independent verification is what turns a press release into a fact.
- “Outperformed” needs definition. Better at which papers? Measured how? Details like these determine whether a result is a narrow win or a broad one, and they matter enormously.
- Benchmarks are snapshots. The big labs ship updates constantly. A lead in this field can be real and still be temporary.
Why non-technical readers should care anyway
Even with those caveats, I think this story matters for anyone watching AI agents mature. Replication is a proxy for something bigger: can an AI take a complex, loosely specified goal and carry it through to completion without a human holding its hand at every step?
If the answer is increasingly yes, the effects reach far beyond science. The same capabilities that let an agent rebuild an experiment — careful reading, planning, coding, self-correction — are the capabilities that would let agents handle substantial work in ordinary businesses too.
And for science itself, the upside is real. Replication is chronically under-resourced because it is unrewarding work for human researchers. An AI teammate that genuinely handles it well would not steal anyone’s job — it would do the job almost nobody wants, and make published research more trustworthy in the process.
So keep an eye on Inherent and Faraday. Whether or not this particular claim holds up under scrutiny, the trend
🕒 Published: