When AI Prose Is Worth Reading

What Makes a Publication Worth Subscribing To?

Think about what makes The Atlantic or The Economist worth reading. It is not primarily the bylines. You happily consume an article by someone you have never heard of, provided the piece is interesting, well-argued, and addresses a question you did not know you wanted answered. The publication's value lies in its editorial judgment — the ability to identify which questions matter, which angles are underexplored, and which expert should be commissioned to address them. You subscribe to the editor's taste, not to any particular writer's name.

And the editor's job does not end with the assignment. A good editor does not hand a writer a topic and then passively accept whatever comes back. The relationship is interactive and often adversarial in the best sense: the editor reads the draft critically, pushes back on weak arguments, demands better evidence, cuts the self-indulgent tangent, insists that paragraph seven actually contradicts paragraph three. The editor ensures that each piece meets the publication's standard, that the voice across multiple writers remains coherent, and that the final magazine is greater than the sum of its parts. Above all, the editor takes responsibility for every word published — not just the words she wrote herself, but the words she approved.

We bring this up because we believe the editor analogy is the right frame for understanding when AI-generated prose is worth reading — and when it is not.

The Yuck Factor Is About Slop, Not About AI

There is an instinctive recoil against AI-generated text, and surveys confirm it: the Reuters Institute found that audiences perceive news labeled as AI-generated to be less trustworthy, even when the content itself is rated as equally accurate and fair. A Graphite study found that by late 2024, over half of English-language web articles were primarily AI-generated, and by April 2025 the figure for newly created pages had climbed to 74%.

But look at what most of that content actually is: SEO slop, review spam, formulaic filler designed to game search algorithms. The yuck factor is a perfectly healthy reaction to this material. Here is the thing, though — you would have the same reaction if a human wrote it. We have all encountered the hastily written corporate blog post, the phoned-in analyst report, the thought-leadership piece that contains no actual thought. Bad writing has always been abundant. AI simply made it cheaper to produce.

The issue is not human versus AI. The issue is whether anyone with genuine knowledge and editorial judgment directed the work. A mediocre human writer working on autopilot produces output that is every bit as bland and statistically average as an unprompted LLM — and for the same reason: both are substituting formulaic patterns for actual thinking. Tom Friedman's columns famously open with some version of "I was talking to my taxi driver on the way to Davos" before pivoting to a grand thesis about globalization. Imitating Trump's rhetorical style has become a national pastime. ("Many people are saying this is the greatest column ever written. Believe me.") When a writer's method is this legible, an LLM can reproduce it almost perfectly — which tells you something not about AI but about how formulaic the original was.

The Editor-AI Relationship

Once you accept that the quality problem is about editorial direction rather than authorship, the productive question becomes: can a human serve as a genuine editor to an AI writer?

We think the answer is yes, provided you take the editorial role seriously. This means treating AI output with the same skepticism you would bring to a piece from a talented but sometimes unreliable freelancer. You read the draft critically. You challenge its assertions. You notice when it is hedging instead of committing to a position. You catch the moments when it drifts toward the statistically average take rather than the interesting one. And when it gets something genuinely right — an unexpected connection, a well-turned synthesis — you recognize that and build on it.

The "just make better prompts" crowd misses this entirely, because they treat the process as a single transaction: prompt in, article out. The reality is closer to a multi-round editorial negotiation. The prompt is not a magic spell; it is the beginning of a conversation. And the hard part — the part that requires taste, experience, and tacit knowledge — remains deciding what to investigate and how to frame the inquiry.

This is a specific instance of a principle we have been writing about for some time: AI shifts the locus of value from how to what. In every domain it touches — software, design, analysis, writing — AI commoditizes the execution while making the direction more important than ever. How you write a piece is no longer the bottleneck. Competent prose, clear structure, well-sourced evidence — these are now table stakes that a well-directed LLM can deliver reliably. What matters is what you choose to write about: which question you ask, which assumption you challenge, which two facts you juxtapose that nobody else thought to put in the same sentence. The editor's role has always been about the what. AI just made that role the only one that cannot be automated.

The Curiosity Bottleneck

If the what is now the scarce resource, then the real bottleneck is curiosity — specifically, the kind of disciplined, informed curiosity that generates questions worth answering.

Consider what we do at the Circular. Our most valuable contributions have never been in the elegance of our sentences. They are in connections that emerge from decades of geopolitical pattern recognition combined with a sharp instinct for how technology reshapes incentive structures. Which questions deserve 4,000 words? Which assumptions need challenging? Which data has nobody thought to juxtapose? These are acts of curiosity, not acts of composition, and they remain stubbornly human.

There is an analogy to software engineering that makes this concrete. Nobody opens an app and asks whether it was written by AI or by a human programmer. You care whether it works, whether it solves your problem, whether someone with good judgment designed it. The same standard should apply to analytical writing. The question is not who or what produced the prose. The question is whether someone who understood the problem directed the work and took responsibility for the result.

This is, if anything, a more optimistic conclusion than the doom-and-gloom narrative suggests. The people who were always coasting on formulaic analysis will be exposed; their output was already, in Ferguson's pointed phrase, "LLM-like" before LLMs existed. Meanwhile, the people whose real contribution was always in the framing — the question, the angle, the counterintuitive juxtaposition — will find that AI amplifies their reach considerably.

What This Means for Us — and for You

This essay is, in a sense, a declaration of intent.

For the reader, the implication is simple: stop asking whether a piece was "written by AI" and start asking whether someone with genuine expertise and curiosity directed its creation. The teacher who requires students to critique AI-generated essays rather than simply write their own has the right instinct: the skill that matters is judgment, not production.

For us, the implication is more specific. We have decided to lean into the editor model rather than resist it. Going forward, we intend to publish a series of articles that use AI as a capable writer commissioned by editors with strong opinions about what matters. Our job is not to produce every sentence ourselves. Our job is to identify the questions that help people make sense of a rapidly changing world, to direct the research, to read every draft with the critical eye we would bring to any commissioned piece — and to take full responsibility for the result. If something is wrong, it is our failure of editorial judgment, not the machine's. That is what editorship means.

Some of these articles will be deeply informed by decades of pattern recognition across geopolitical and financial markets. Others will reflect a close reading of where technology is quietly reshaping incentive structures. All of them will be editorially directed, researched with AI assistance, and held to the same standard of intellectual honesty that we have always tried to maintain. We will not pretend otherwise.

The value was always in knowing which questions to ask. Now we can finally focus on that.