待翻譯：The Economist Problem and AI

AI 服務暫時不可用，以下為來源正文，待恢復後補全翻譯。

The Economist Problem and AI | Pith & Pip Pith & Pip Our Work Our Team Our Blog Our Writing Get in Touch SaaS Writing Consultation & Production Pith & Pip SaaS Writing Consultation & Production The Economist Problem and AI Matthew Guay Wednesday, August 5, 2026 They're good at covering the basics. They appear worse, the more expertise you have on a subject. The Economist keeps readers generally informed of what’s happening in the world around them, in a way that increasingly reminds me of the bell-curve of AI’s outputs. The Economist is not the paper with headlines that bleed and lead, nor is it the best place for detailed analysis that analyzes what caused a thing to happen and the second-order effects. The Economist is, instead, the paper to read for a well-rounded perspective. It’s more detailed than an average newspaper, less like an academic journal, written professionally yet approachable. It’s slow news, produced once a week, giving events a bit of time to settle before they’re reported on. I like that. It fits my general goals of trying to avoid clickbait and take a calmer, less emotional approach to news. Plus, it keeps me informed generally of things outside my general knowledge and expertise. I'm unlikely otherwise to think about the economy of Kazakhstan or the minutiae of oil trades. Yet The Economist has a problem. If you’re an expert in a topic covered in its pages, you might find that its coverage is merely cursory. Generally, the broad strokes are correct, and more detailed and on the money than you might expect from another newspaper. What it isn’t is an expert opinion on every plausible topic, and so you read The Economist for the things that are outside of your locale and industry. For the things closer to your heart, you likely have better sources: industry journals, trade magazines, Substacks, and the like. It’s an effect not limited to The Economist; “Everything you read in the newspapers is absolutely true except for the rare story of which you happen to have firsthand knowledge,” journalist Erwin Knoll is said to have quipped, and the effect is only amplified by The Economist’s broad range of coverage. The Economist lets you know generally what’s going on in the world and converse with a mid-level knowledge of that topic without having to dig all the way in. And that’s enough for most topics (or, at least, as the Gell-Mann amnesia effect theory posits, it feels like it is—enough that I still turn to it for general world knowledge). Which is strikingly similar to AI. AI feels brilliant at the tasks other people do. It feels considerably worse at doing the tasks in which you’re an expert. And while it’s easy to discount its skills at our expert tasks, it’s equally easy for those experiences to not temper usage of AI for other tasks in which we’re less capable. Take it from me, who whipped up an internal dashboard with Cursor in an hour, one that would have taken me significant time to hand-code. Who also coded a quick bit of data-parsing Javascript for a Zapier workflow with AI, found the first output wanting, and poked the AI until it generated what I wanted. AI’s good enough at such tasks for me to hand them off without a second thought. But I also feel like AI’s terrible, relatively speaking, at consistently finding accurate quotes, doing deep research, testing software, and writing prose—the tasks to which I’ve dedicated the majority of my working hours for well over a decade. I can tell at a glance when it’s making up a Steve Jobs quote, or at least know well enough to double-check the source each time. It can both find rare needles-in-the-haystack in research sometimes and fail to find things a cursory Google search would uncover at other times. And so I use AI for those tasks but take its responses with a grain of salt. But then, perhaps scarily, I’ll ask AI for one-off medical advice, and forget to similarly discount its answers based on its prior quote-hunting hallucinations. Yet here’s the thing: AI is genuinely good at some tasks. Scoped tasks, especially. AI is great as fuzzy search, good at finding “this and things similar to this” in a text that you’d never find with standard exact-word search. It’s good as a first-pass proofreader, as long as you ask it for suggestions and don’t let it rewrite your text for you. It just feels worse—or rather, its generalizations and mistakes are more apparent—the more expertise you have in a field. Thus how impressive AI seems at proofreading, programming, or prose is likely inversely correlated with how proficient you are in those skills. The Economist doesn’t have to publish inaccuracies to run up against Knoll’s observation. It’s that everything contains multitudes, and that there is more nuance to almost everything than any cursory coverage could detail. Thus the expert reader recognizes the causes and potential effects, and feels that the piece doesn’t cover enough detail to make the full case. But when you don’t know enough about the topic, you’re dealing with unknown unknowns, and if you give the publication the benefit of the doubt you’ll assume its coverage is good enough. The same goes for AI results, where it both is generally useful at a broad range of tasks (and frighteningly good at an increasingly large number of specialist tasks, like solving the previously-unsolved Erdős planar unit distance problem), it fails to accurately perform other tasks either consistently or in one-off glitches that are hard to square with how well it can perform other times (like counting the number of r’s in strawberry). Which sets up the slightly counterintuitive notion that, as Sean Goedecke put it, LLMs reward expertise. Asking an LLM a question wouldn’t seem to be a skills problem, any more than getting more detail out reading The Economist. But asking the right question and reading between the lines are increasingly the more important skills. Ironically, those least impressed with AI today may stand to gain the most from it. They’re the experts who can recognize its flaws, spot hallucinations at a glance, and refine their questions until they steer the AI to do what they’d intended. The writer who knows what needs to be edited in a piece, or who can notice hallucinated quotations, is best positioned to prompt an AI in a way that will have a higher success rate. The developer who’s shipped full-stack applications before has an incalculable advantage at solving the issues that crop up in developing with AI, versus a non-developer asking an AI to make an app in general. That lends a quick rubric: Use AI only for the tasks and topics you’re well versed in, the opposite of what the models encourage by pushing you to ask questions about topics in which you have little experience. Use AI as a springboard for curiosity, as a starting point to level up, not as the end-all at first answer. The question remains how curious we’ll remain, and if we’ll outsource our worldview to the machines or use them to hone our own curiosity and intuition and level-up faster than would have otherwise been possible. In much the same way that The Economist rewards a curious reader with both a wider understanding of the world and a starting point to dive in and research more and, perhaps, in the process come to discount The Economist’s initial generalized reading, AI gives humanity an infinite curiosity machine, a calculator to figure out almost anything. But when solving the tasks on which we spend our waking hours, expertise and its resultant taste means you’ll not settle for AI’s averaged-common-denominator output. That’s your edge. Let's write your software's story, together. Chat with Pith & Pip ©2025 Pith & Pip, LLC LinkedIn Twitter Email us Let's write your software's story, together. Chat with Pith & Pip ©2025 Pith & Pip, LLC LinkedIn Twitter Email us