Good writing used to credibly signal things like intelligence, expertise, and effort. Now with AI everyone can sort of fake this. All writing is just way way more persuasive, on average, than it used to be. This is mostly good, but has led to unintended side effects:

• Governments and organizations unintentionally (but sometimes not) made “decent writing” a speed bump or rate limiter for specific types of actions/resources - so that that people would only invest in it if it was worth the high cost of producing the writing (e.g., suing your neighbor, or your boss). Now you can sue anyone quickly and effortlessly?
• Society explicitly picked “can write well” as a filter: Who gets to be a journalist, a novelist with a book advance, or a high achiever on the written SAT? What quality of homework sorts students into As vs Bs vs Cs? Who gets into startup accelerators like Y Combinator?

I sometimes get asked for advice from like, college graduates who despair of their employment prospects. I empathize with them – AI has injected a lot more uncertainty into white collar knowledge work. But I rarely have any good advice on hand to give.

After thought, one concept I think will be useful to college grads - and that I have been working on lately - is learning to discern genuine expertise: In fields where I am not an expert, how do I evaluate writing to determine whether someone has actual expertise? It’s much easier to fake expertise now, and it’s bad for everyone if more people make decisions based on wrong but persuasive writing.

A couple reflections, and a running list of some ideas (from talking to LLMs lol):

Positive signals

Acknowledging tradeoffs core to the problem

Non-experts often optimize for a single metric. Real systems always balance competing objectives, and true experts know what those objectives are because they’ve had to manage them.

“This approach adds latency, but is usually worth it when you have problem X and Y, and don’t care about latency as much because of Z.”

Reasoning from the causal chain, instead of just conclusions

E.g., “Microservices increase your velocity because they let teams independently evolve and deploy their codebases, at the expense of introducing distributed systems complexity.”

Details should connect back to first principles, and ideas follow from them. Non-experts can spout a collection of isolated facts but can’t support why they have to be true.

Speaking in terms of constraints and boundaries

Weak writing is more prone to make sweeping dogmatic claims because it doesn’t have a strong working mental model of its claims (“monolithic deployments are better, and microservices are just over-engineering”). Strong writing makes the constraints obvious:

• “This holds for teams under roughly 50 engineers”
• “This doesn’t work for networked client applications because you start with a distributed system anyway”
• “Below this scale, the effect disappears and none of this matters because everything fits in memory”
• “You can’t compress it beyond that point because then you lose enough precision that the algorithm fails to converge”

Do they understand the competing viewpoints?

As a corollary to the above, true experts know the limits of their own position and why smart people might disagree. They can acknowledge that different situations or facts can lead to differing conclusions.

“The disagreement comes from different assumptions about the importance of X”

They can make predictions that are surprising or unintuitive

True expertise is not just explanatory, it leads to better predictions. Experts can make falsifiable predictions using their better mental model e.g.,

“If this diagnosis is correct, then we should expect A, B, and C to happen in the next few months.”

“Given these assumptions, the bottleneck should migrate from memory to networking over the next year.”

They explain reality, rather than defending an identity

Some writing feels optimized to signal intelligence, engagement, or tribal affiliation. True expertise - free of tribalism or political bias - is more humble, because it bends to new facts and evidence.

Ask yourself: “If new evidence appeared to contradict the author, would they change their mind?”

Negative signals

Explanations where every piece of available evidence points in the same direction

Real problems are usually multi-dimensional sets of tradeoffs and constraints and uncertainty – which is why there’s a diversity of approaches in the first place; neat sounding soundbites (with a single hero and villain) are almost always a vast oversimplification of the problem space.

Confident prescriptions without assumptions

There are very few ideas that always work without some domain-specific prerequisites being present.

Undeclared conflicts of interest

The author has a clear personal stake in a preferred outcome, but doesn’t address it for fear of undermining their own argument. True experts can explain why their argument stands regardless.

Unfalsifiable declarations

Often a lack of expertise combined with motivated reasoning leads to a pre-ordained conclusion supported by “one favorite framework” that has all the explanatory power, but the reasoning is mushy enough that it can never be proven wrong.

The one question to ask:

Could someone have really produced this writing without having interacted extensively, in detail, with reality?