In Praise of Editors

If that headline doesn’t get me tons of hate mail, I don’t know what will.

But this piece isn’t about how wonderful those of us who have “editor” on our name cards are (although, of course, we are), but more about the multiple jobs we do and roles we play in newsrooms, and how we tend to lump them all under one title. And about how generative AI could help scale that work.

We all know the many breeds of editor: The gruff wordsmith with a half-eaten bun on their desk who can spot a correctable error in your draft from 20 feet away and skillfully pares 10% off your precious prose; the narrative expert who can pick apart your 2,000 words and reassemble them into a more polished, coherent argument; the thoughtful conceptual editor who interrogates you about what your story is really about, and finds fresh angles and questions you’ve missed in your reporting. I’ve been blessed to work with all three types — and more — in a forty-year career, but I also know how rare it can be to find even one of them in a lifetime.

So I set out to see if I could make mechanical versions of them. (Of course I did.)

To be clear, these aren’t the proof-reading and style bots that abound now (we have some at Semafor too); these are trying to replicate some of the higher-level functions that people who have “editor” in their title are supposed to do.

And also to be clear, this isn’t with an eye to replacing those editors — although that’s not off the table (and more of that in a future post) — but to make their work and feedback more accessible to more reporters, especially in an age where newsrooms are shrinking and experience is disappearing.

I made three bots: a Conceptual Editor, a Narrative Editor, and a Copy Editor.

The first asks, early in your process, if you’re on the right track. Are you missing some angles or perspectives? Is your core thesis supported by your facts, or are there alternative explanations? Like most editors, it knows less about the subject than you do — but it leverages some general knowledge and killer language analysis skills to pick apart your draft to ask (gently) if you’ve considered all the other possibilities. It’s not dissimilar to the deconstruction engine I built a bit earlier, but nicer, and more proactive about suggesting other angles.

A lot of stories go off the rails at this stage, where reporters, deeply focused on an angle or a thesis, can miss broader questions or contradictory information; this is designed to give them some feedback early on — and privately, away from the prying eyes of managers and editors. (Call it another benefit of the intimacy dividend.)

I tried this out on a piece I wrote a while back, about the inevitability of personalized news, created on demand, based on a reader’s level of knowledge, interest, and attention. It flagged some interesting ideas — none I fully agreed with, but that could have prompted some additional reporting.

For example, it suggested that:

**Is there a counternarrative you’re missing?** What if news organizations conclude that their value lies precisely in *not* personalizing content — in providing shared experiences and common frames of reference? The “one version of a story” model you critique might be a feature, not a bug, in an increasingly fragmented media landscape.

Fair point. And more substantively:

**What’s the actual thesis here?** You seem to be making two distinct arguments: first, that personalized AI news is technologically inevitable (”tantalizingly close”), and second, that newsrooms need to prepare for this shift. But are these both true simultaneously? If the technology faces “multiple hurdles” including hallucination problems and unsettled IP questions, how close are we really?

I’m sticking to my piece — I mean, it is already published — but these aren’t bad thoughts to keep in mind.

The second assumes you’re on the right track — but now are you telling the story in the most effective way possible? Are there places the piece may bog down, or where elements might be rearranged for better narrative flow?

The feedback on my piece isn’t bad — hey, I’m not that bad a writer! — but it does offer some suggestions:

**The middle section could be tighter.** Your examples of personalized news (audio while driving, omitting known background) are helpful, but the paragraph runs long and starts to feel like a list. Consider breaking it up or leading with the strongest example.

**The ending needs more punch.** “We probably don’t have a lot of time to adapt” feels like it trails off rather than landing with impact. What specific actions should news organizations be taking now? Or what’s the cost of not adapting? The piece builds good momentum but doesn’t quite stick the landing.

And the third preps the piece for publication — doing the usual copy editing, of course, but looking also for duplications, contradictions and inconsistencies in the text. And, of course — it is an editor bot, after all — it looks for places to cut.

Here it just cuts to the chase, for example:

3. **Logic gap** (Para 6): You jump from describing benefits to saying communities are “underserved” without establishing the connection. How does the current model leave communities underserved?

4. **Redundant phrasing** (Para 7): “entirely accurate” is redundant - accurate information is binary.

5. **Unsupported claim** (Para 8): “It’s possible a fully AI-generated story may not be protected under copyright law” - This legal assertion needs attribution to experts or case law.

Honestly, this isn’t too bad. I tried it out on some stories I didn’t like as much, and it flagged useful blind spots, slow narratives and some bloated writing. Were they as good as an excellent human editor? No. But who has one of those these days?

To be sure, these bots aren’t ready for prime time; nothing I build is. If I were to actually make these for a newsroom, I’d want to customize them for the organization’s style and voice and audience focus, at the very least.

But there’s also no reason why a reporter couldn’t build their own set of tools and get feedback on stories they’re working on, long before they ever reach out to an editor — the functional equivalent of the friendly roommate who reads a draft and gives helpful (or unhelpful) comments. (Again, it’s not as good as having a great relationship with a fantastic editor — but it’s better than having a harried, overworked editor.)

And more broadly — it’s a reminder (again) that LLMs can do more than simply take away the dull, rote parts of our workflow; their capabilities and sophistication are growing quickly, and the sooner we adjust and build new processes to incorporate what they can do, the sooner we rethink our role in this brave new AI-intermediated news landscape.

PS: Of course I ran a draft of this piece through the three bots. I even incorporated some feedback….