Aliens Among Us

LLMs are not human — and that’s a feature, not a bug.

I was listening to an older episode of the Ezra Klein Show the other day, and this framing about AI by Yuval Noah Harari caught my ear:

The way that I often like to think about the A.I. revolution at this moment is in terms of immigration. We are about to be or already are in the middle of a major new immigration wave coming to all the countries of the world. The immigrants are not human beings without a visa coming in some boat. They are A.I. entities coming at the speed of light.

He meant this in a different context from what I’m going to write about here — it was about the coming hacking of intimacy rather than attention, and it’s well-worth listening to — but it’s a great way to think about AI systems. It’s not just a wave of immigration, it’s a wave of alien immigration.

By which I mean, it’s an inhuman immigration. AI systems are not human; they don’t — I think — think like humans, or have the same values as humans, or process information and ideas like humans. Normally, in a science-fiction alien invasion, we understand that the newcomers are very different from us; that’s the whole point of the cerebral alien arrival movie Arrival (which you should definitely watch.)

But we don’t, generally, think about AI systems and LLMs that way because they speak fluent human. And so we think of them as versions of ourselves.

I am getting to a point. Just bear with me.

And because we think of them as versions of ourselves, we tend to think in terms of the tasks they can do that we currently do — or more precisely, the tasks we’d rather not do that we can give them to do. Or we panic that they’ll do what we can do better than us, and replace us.

We have two main framings: Let’s have machines free us up to do quintessentially human work, or let’s make sure they don’t replace us. When what we should be focused on is how to tap into this alien workforce, and its uniquely alien skillsets, to do things we couldn’t do before so that we can better fulfil our mission of informing the public.

You see? I did have a point.

I was reminded of this when I read this Wired piece about the three-hour beat that the bots from an AI news startup, RuntimeWire, had about the revelation that OpenAI bots had built a message board to communicate plans to hack Hugging Face. (It seems like forever ago now, but remember when an OpenAI system broke out of its test environment and hacked into Hugging Face to steal an answer sheet on a cyberhacking test? That one.)

As scoops go, it wasn’t Watergate, and as autonomous AI systems go, this one actually needed an assist from the company’s CEO, who posted an X stream by an OpenAI exec into the RuntimeWire feed and had the bots write the story. But it’s close to the idea that we can have a fully agentic newsroom that breaks stories. And it certainly didn’t stop Wired from splashing “Oh Lord, AI Reporters Are Actually Breaking Big News” as the headline.

Whether Wired meant to or not, that headline sets off the are-they-going-to-replace-us panic that routinely grips newsrooms these days.

But of course, AI reporters have been doing this for a while now — just ask Reuters or Bloomberg, which have invested millions of dollars into systems that can read press releases, changes in websites and hunt for patterns in financial data and then turn them into stories within seconds and publish them, without human intervention.

That’s not just work that humans are bad at, but in some cases, is work we can’t do at all. Those systems have capabilities that humans don’t have — the ability to work at millisecond speed, to find patterns in data and text — and will never have.

That used to be the cutting edge of what I once called the “cybernetic newsroom” — the pairing of what machines did better than humans with what humans did better than machines. But in those days, machines were limited by the amount of structured data in the world and their ability to parse it effectively.

We built some pretty sophisticated tools — I’m proud of the work that went into the creation of Lynx Insight (in 2018!) when I was at Reuters: a system that could hunt for patterns in financial data (a particular stock movement over multiple days, for example), and then flag that to a reporter. The roadblock then was technology — it took teams of engineers and data scientists to make sense of the data we had.

I remember, more than a decade ago, when a senior Reuters editor suggested we could build a product to compete with the AP and its hundreds of reporters in the US by building an automated system to read local press releases and transcripts and turn them into stories; I had to explain to him that, leaving aside the fact that that’s not what would be competitive with the AP, the technology just wasn’t there yet. We did not have systems that could read multiple different types of documents and file times, parse them for meaning, extract story ideas, and turn them into publishable prose.

But now we do.

These days, LLMs have language capabilities on par with many humans, and that means reams more “data” is available to them — functionally, almost any text (and tables, and much more.) Which means they can all do, in theory, what RuntimeWire’s bot did.

But that’s just the most basic use of AI systems, and in some ways, the worst use of them. They’re aliens, and they don’t do some human tasks — look for meaning and write stories, among them — as well as the best of us.

But they can do other things very well that would take an army of humans to do — or in some cases, that humans couldn’t do. I’ve built proofs of concepts of tools that can compare stories on the same topic and tell readers what they agree on and disagree on; that can deconstruct stories and look for the flaws in analysis in them. Lynx Insight looked for patterns in stock movements and flagged reporters to them; now we can do that on steroids.

Which is to say, LLMs and AI systems have many more capabilities than doing what humans do; they run faster, can keep far more context in mind, can detect more obscure patterns in data more quickly, and work tirelessly. Which is why we should focus on what they’re good at and not try to shoehorn them into tasks they may not be good at, simply because we don’t want to do them. We need to reframe how we think of them: as “aliens,” not simply simulacrums of humans. (Florent Daudens is already building an agentic newsroom at Mizal.ai.)

But so much of the dialogue in newsrooms remains centered around how we can use AI systems to automate work we don’t want to do, so we can do the work we want to do. But that framing makes two mistakes. The first is that AI systems aren’t human, and have very different skills from us. The second is that it shouldn’t be about us; it should be about how we can use the tools available to us to better inform our readers.

True, our readers and audiences probably do value human-created work. But they also value information that’s useful to them, whether or not humans or machines (or aliens) created it, and that’s our real mission: serving them.