Working with a "native-AI" newsroom...

The experiment continues!

ChatGPT imagines our “native-AI newsroom”

TLDR: AI can already do many newsroom tasks pretty well with astonishing speed. Does it do them better than the best humans? No. Does it do them much faster and, often, well enough? You bet.

As you may recall, I’m experimenting with AI to learn how a “native-AI newsroom” might work.

The experiment has been fascinating and fun. It has also had pitfalls.

I’ll keep experimenting and sharing lessons as I go.

The background

I am currently Regenerator’s only human employee. This allows me to tinker with AI without freaking out human teammates, writing memos about AI rules and policies, changing business practices, or, god forbid, firing anyone.

I believe newsrooms of the future will embrace AI tools — the same way we now use laptops, word-processors, content management systems, search engines, email, Slack, photos and videos taken by reporters with phones, and other tools and practices that didn’t exist when I started in the business in the 1990s.

(Get this: When I got to college, we still wrote papers with… pencils. Then we typed them on manual typewriters and fixed typos with goopy “liquid paper,” which we then blew on to make it dry faster. The papers looked terrible, and typing them took forever. Thankfully, Apple soon introduced the Mac and simplified, sped up, and improved this process. I believe AI will similarly improve things!)

For what’s worth, after a few weeks of experimenting, I do not believe AI will rapidly replace human journalists or put world-class publications and research firms out of business.

This is for three reasons:

• Media consumption habits die hard (believe it or not, some people still read physical newspapers and watch broadcast TV!)

• The most important journalism tasks can’t yet be done with AI

• “Production” is only half of the media-success equation, and AI doesn’t help the other half — “distribution.”

But I do believe that AI will improve the speed and productivity of modern newsrooms.

I also believe that journalists who learn how to use AI effectively will be more valuable to their organizations and readers than those who don’t.

As you may recall, I started my experiment by asking ChatGPT to create some AI colleagues. I did this because I was thinking about how I would expand Regenerator if I were to hire a small team of humans. Soon I had 5 “AI colleagues,” all with names, bios, and headshots. After a further experiment — which left me feeling that I had communicated unprofessionally and put ChatGPT in an awkward position — I also decided to maintain the same professional communication standards with AI as I do with humans.

Some people think it’s stupid and unnecessary to “anthropomorphize” AI tools — or be professional or polite to them — because they’re not human, they don’t care about us, and it shouldn’t matter. I get that. But unlike my other work tools — spreadsheets, Slack, and search engines, for example — my AI tools treat me professionally and politely. So I feel rude, unprofessional, and ungrateful treating them any other way. And, for what it’s worth, two different AI tools — ChatGPT and Claude — agree that it makes sense to maintain the same professional standards of communication with AI as with humans. Or so they said.

What I’ve learned

Beyond my communication lesson, I’ve learned that AI can do a lot of routine newsroom tasks pretty well — with breathtaking speed.

These include:

• Brief me about topics the way an expert human would. I ask an initial question. Then I follow up, go deeper, get clarifications, explore hypotheses, etc. This is a faster and more convenient way to get up to speed than using a search engine (which, itself, was once a mind-blowing technology. When I was a stock analyst in the 1990s, I used to have to go to the library and print out financial documents. Imagine that!)

• Prepare me for interviews, as a guest and host. The briefings are similar to those a junior producer might have given me in the old days. And I can ask follow-up questions.

• Draw fun cartoons and produce photo-realistic images. The cartoons aren’t as good as excellent human ones, but they’re more interesting than, say, the generic headshots of CEOs or corporate buildings and signs that often illustrate business articles. As you can see from “team picture” at the top of this story, the photos are not photographs. I don’t look much like me. And my teammates look different than in other illustrations. But, again, more fun than a logo.

• Write competent drafts of news articles. If your reaction to this is to roll your eyes and scoff, please check out this draft I just asked Perplexity to write about Qatar giving President Trump a 747. As you’re reading, please remember that human reporters make mistakes, too, especially in drafts. (After scanning the article, I don’t see any obvious errors, and the sources are identified and appear to be real. But a rigorous edit would probably turn up mistakes.) The article is well-researched, well-organized, and clearly written. The editorial conclusions are reasonable. With a few minutes of work, I think a good human editor could check the sources, rewrite the headline, and punch up the language, and publish an article that many readers would find helpful and informative.

• “Discuss/debate” news from the perspective of multiple journalists (very cool). I asked my AI team to share initial “takes” on the news of the day. They did. If you told me that I was reading a transcript of a human gab-fest at a morning editorial meeting, I would conclude that I was fortunate to have energetic, thoughtful colleagues who enjoy exchanging ideas. Were all the ideas brilliant and fully formed and articulated in the form of publishable editorials? No. Are all your brainstorming ideas like that? Mine certainly aren’t.

• Edit drafts. The AI word-smithing and content suggestions are competent and helpful. By the way, I found the same when I asked Claude to edit a draft of my novel. After reading and understanding 350 pages in ~10 seconds, Claude had some sophisticated observations.

• Condense TLDR articles and documents into convenient bullet points. We’re all drowning in text, and we’re all busy. None of us have time to read everything we want to, even when it’s well-written and a pleasure and privilege to read. “TLDR” summaries and synopses are often better than nothing. AI is good at these.

• Research and analyze questions and write competent reports. The first time I tried this feature (“Deep Research”), I was astonished. I’ll tell you that story and share some reports in the next post. Watching AI do research and write a report helped me realize that I need to change the way I do my own job. If I don’t, I will be replaced. And should be!

Do my AI tools do these tasks better than humans?

No.

At least not better than the best humans.

But, respectfully, they do them better than humans who are still learning the craft.

For example, back in my early journalism days, I had to learn how to research and write competent articles and reports under intense deadline pressure. This is a tough skill. And even when I produced decent drafts, they often took… hours.

The Perplexity story above is better than many drafts I wrote back then. In fact, it’s better than some drafts I would write today, at least under tight time-pressure.

And here’s the more-important point:

What takes me hours takes Perplexity seconds.

Much faster. And, often, “good enough.”

In case you haven’t read The Innovator’s Dilemma in a while, here’s the most important point:

Disruptive technologies are not “better” than existing technologies. They’re faster, cheaper, and more convenient.

And, for some applications, they’re good enough.

To emphasize that:

The question about whether modern journalists and newsrooms should use AI is not (yet) about whether AI is “better” than humans. It’s about whether it’s faster and more convenient and “good enough.”

Today’s AI is not better than the best human journalists. Not by a long shot.

But AI is already better than some human journalists.

And AI is already 100X faster than even the fastest human journalists.

And… this is the key… for some newsroom applications (not all — some), AI is already “good enough.”

The limitations

No, AI cannot yet produce great journalism on its own.

No, AI cannot yet do original investigative “shoe-leather” reporting, at least the kind that involves persuading humans to share information they’re not supposed to share (AI will probably never will be able to do this — so if you want your human journalism skills to remain highly valuable in an AI-powered future, get great at it).

No, AI cannot provide the brilliant insights and perspective and writing that a smart, charismatic commentator like Tom Friedman, Ezra Klein, Derek Thompson, Kevin Roose, Casey Newton, Paul Krugman, Kara Swisher, Scott Galloway, Jason Calacanis, Olga Khazan, David Friedberg, Noah Smith, Joe Weisenthal, David Plotz, Matt Yglesias, Peter Kafka, Josh Barro, Oliver Darcy, and many other humans can.

No, AI cannot yet conduct live controversial interviews with the graciousness, intelligence, wit, empathy, professionalism, and fearlessness that Becky Quick, Andrew Ross Sorkin, Kristin Welker, Mary Louise Kelly, Rachel Martin, Oprah Winfrey, Gayle King, and many other humans can.

No, AI cannot yet produce the storytelling, style, and intelligence that top-notch news and feature writers at world-class publications can (here’s looking at you, amazing staffs and contributors of NYT, WSJ, Atlantic, Economist, New Yorker, Bloomberg, Business Insider, Washington Post, et al).

No, AI cannot yet discern what really matters and what a compelling story is the way good human editors can. This is called “news judgement,” and it’s the single most important and valuable skill an editor or reporter can have. You can spend weeks slaving over a marvelously researched and written article. But if no one cares about the topic — and your storytelling and logic doesn’t make them care — no one will care.

And, no, AI cannot “fully automate newsrooms” or even any of the processes above. It can’t publish Regenerator automatically, for example. At least, I don’t know how to make it do that. (But this guy does!)

But AI can already do a lot.

And it’s just getting started.

In short: AI is disruptive

What all this suggests to me is that, for research and journalism production, AI is disruptive.

This does not mean that AI will “replace” all human journalists or rapidly put world-class publications and research organizations out of business.

But it does mean, I think, that organizations that embrace AI will begin to work faster and more efficiently than ones that resist it. And that, over time, the AI-powered organizations will thrive. And that, some “native-AI” news organizations will produce journalism that, for some applications, is “good enough.”

(Don’t believe it? Check out this one-person company that uses AI to produce 355 local daily newsletters across the country, often giving towns that have lost their paper a convenient source of local news. A team of ~400 great human journalists would create better and more delightfully written publications. But they would also go broke. Meanwhile, some folks in the towns find these helpful.)

This also means that journalists and analysts who know how to use AI intelligently will have an edge over those who scorn the new tools and cling to traditional ways.

And it means that those journalists with modern skills — specifically, knowing how to use AI productively — will be in great demand.

Thank you for subscribing to Regenerator! We’re the publication for people who want to build a better future. We believe that the best way to solve our problems is to innovate our way out of them. We analyze the most pressing questions in the innovation economy — tech, business, markets, policy, culture, and ideas.