Colin Versteeg

Resume

Posts

Good Enough to Exist or the Paradox of Taste

10 Nov 2026

Recently, some indie game devs got in a bit of a hubbub after admitting that they had used an AI to help debug a difficult Unreal Engine bug, which upset some fans who subject every game they play to a purity test. To those fans, if games can’t be made without AI, they aren’t good enough to exist. On the other hand, to most developers at this point, any AI-generated bug fix that compiles, passes the tests and fixes the bug is obviously better than not having that bug fix. I think this makes a lot of backlash to AI (and specifically who is the source of the backlash) clear.

This story inspired me to tweet about a particular fear of mine that luckily didn’t come to pass:

Basically, Republicans (especially the Trump administration) are all in on AI, especially generated images for memes, and would presumably be willing to use AI for basically anything. On the other hand, a lot of Democrats are opposed to the technology, the companies promoting it, and especially its impact in the labor market.

I was (and honestly still am) concerned that this could lead to a serious disadvantage for the Democrats, if the promises to revolutionize white-collar work materialize, as a lot of political campaigning (running data analysis, coming up with marketing strategy, determining what issues are salient with voters) is pretty standard white-collar marketing work. But this goes a little further for me — why are Republicans willing to use the models and Democrats aren’t?

Good Enough

It’s important to remember that what we mostly have right now is generative AI — you can generate what you want. You can have documents, tests, tasks, reports, diagrams or sonnets that otherwise would take more effort than it was worth to produce. So basically, I think that the main limiting factor on uptake of AI is whether what you generate meets your standards of “good enough to exist”. Even a year ago, I was dismissive of “vibecoding” for things that I’m a specialist in — but now I write basically no code by hand.

But basically anytime I receive text “written” by AI from someone at work, I find it unbearable; why is that? My opinion is that for this content, it has to reach past your point being good enough to be better than there being nothing. For a lot of text I see, I find it so viscerally poorly written that I think less of the person who created it. But most code — the standards don’t really matter anymore.

A team of 22-year-old Republican social media operatives were never going to get a budget to make “GOD EMPEROR TRUMP” memes, but they certainly think that those should exist and be shared, so the standard is pretty low.
But when it comes to actual hard political work like polling, maybe they aren’t actually good enough to exist at all? Or maybe the limitation isn’t generating more ideas or memes or reports synthesizing expected voter behavior.

AI Girlfriends and Book Reviews

Recently the company Character.AI folded, after having been a zombie for a few years following the departure of much of its tech talent to DeepMind and then Thinking Machines.

A few years ago I interviewed there, and as part of trying to get familiar with the product, I thought about ways that it could be fun or enjoyable for someone like me. I’m a big podcast listener, and especially a fan of sports podcasts — so I tried to create a “Character” of Bill Simmons, by throwing a ton of his ESPN columns into the service. For me it didn’t work, and didn’t create anything that sounded like him, it didn’t even approach my dream of getting to enjoy a free-flowing conversation occasionally touching on mature subjects with him and House defending the glory of the Washington R Words.

But for many (especially young) people, Character and other similar fictional-friend or AI-girlfriend services are things that they love and in many cases are addicted to. I’d rather be alone with my own thoughts than talk to a bad facsimile of someone I respect, but for a lot of people it’s better than nothing. I hope as these kids grow up they figure out that physical and interconnected social connections will make these emotional relationships bad substitutes for human connection or we’ll have a serious issue.

And for me, I think the line of “good enough” content comes from something nearly as powerful as large language models: online kudos. The pseudonymous author Scott Alexander has a distinct writing style that has influenced a lot of online writers. He also frequently runs book review contests, where he publishes the favorite on his blog. I think a lot of them tend to be aggressively stylistically similar to him in both prose and structure; they summarize the content of the book, expand it to make some point about society and then re-compact it to come to a conclusion about the overall quality of the book and its arguments. So while I don’t think these guest authors reflect the views of Scott or Astral Codex Ten, for me a “Reader Book Review” about a book or topic I’m interested in exists as just the same thing as something Scott would write himself. He’s created in his community a collection of independent agents that generate Scott-like content on topics that are their own interests, something I think he’d find amusing.

Future

I still like to write for pleasure occasionally (see this blog), and I really like to read for pleasure. I imagine this is why I get so viscerally turned off by AI writing; I find the prose offensive and boring, and don’t think the person who created it respected me or the ideas they had enough to write it themselves.

But interestingly, when I write for pleasure, I think I tend to write in more of a seed fashion, starting with a thesis and following the ideas from my head as they come out — I’m finding my ideas as they leave my pen. I don’t really expect anyone to ever read anything I write (hell, I don’t think I’ve read most of my blog posts after publishing), I just need to write it. An hour or so with a keyboard and I’m suddenly Noel Gallagher with a bigger ego and no coke problem.

On the other hand, at work, I try to be more of a structured communicator with clear sections and subsections, breaking the ideas and therefore the documents down like a case study ready to be filled in. So maybe we’ll reach the point where this time next year I’m writing about how I don’t write at work anymore, I just create an outline and hand it off to Claude or Muse Spark. We’ll see.

Post Script

For entertainment’s sake (and because I’m not a Luddite, just a snob), I asked Gemini to summarize the above into an outline, and then claude to expand it out to article length. Honestly I think it just managed to compress each sentence I wrote into a point. Here’s the result:

# Good Enough to Exist: Why AI Adoption Splits Along Political Lines

A little while ago, a group of indie game developers got into hot water. Their offense was using AI tools to help fix bugs in Unreal Engine. Some fans reacted as if the developers had betrayed the craft, demanding a kind of purity in how games get made. The developers, understandably, cared more about whether the game worked.

That small skirmish captures a much bigger divide, and it's one I think is worth taking seriously: who adopts AI, who resists it, and why.

## The Political Divide over AI

Look at how AI is being used in American politics and a pattern shows up quickly. Republicans, including the Trump administration, have leaned heavily into AI. Meme generation and rapid-fire content creation are the obvious examples, and the technology fits neatly into a style of political communication that prizes speed, volume, and provocation.

Democrats, by contrast, tend to show more resistance. The reasons are understandable. Concerns about labor market disruption and about the concentration of power in a handful of large technology companies are central to the left's worldview, so skepticism of AI isn't an accident. It follows from values.

But there's a strategic risk here. Much of modern political work is white-collar knowledge work: marketing, data analysis, strategy, polling, and campaign operations. If AI transforms how that work gets done, as it already seems to be doing in other industries, a party that abstains on principle could find itself at a real disadvantage. Principled resistance and competitive readiness don't always point in the same direction.

## The "Good Enough" Standard

So what determines whether someone adopts AI? I think the answer comes down to a simple threshold: is the output good enough to exist?

That phrasing matters. The comparison isn't between AI output and the best a skilled human could produce. It's between AI output and having nothing at all. If the choice is a passable draft, a working function, or a quick graphic versus a blank page, the bar for "good enough" drops considerably. And that bar looks very different depending on the domain.

**Coding** is the clearest case. The criteria are functional: does it compile, and does it fix the bug? If yes, the standard is met. I've experienced this personally. I've shifted toward using AI heavily for code precisely because the evaluation is so objective. Nobody reading the final product cares whether a human typed every line.

**Writing** is a different story. Poorly written AI text tends to be off-putting, and readers can sense it. It often reads as a lack of respect for them, as if the writer couldn't be bothered to put in the effort. Here the threshold for "good enough" is much higher, because the reader's trust is part of what's at stake.

**Political operations** fall somewhere in between. The bar for churning out memes is low, which is why young social media teams can use AI so effectively. The output is disposable, the feedback loop is immediate, and a miss costs almost nothing. Higher-stakes work like polling and campaign strategy may be subject to different constraints, since errors there are costlier and harder to spot.

## Value Thresholds Beyond Productivity

The "good enough" framework extends past work tasks, and the results are interesting.

### AI Companionship

I tried to build an AI version of a favorite podcast host, hoping to recreate the feeling of listening to someone whose voice and thinking I enjoy. It fell short of what I'd hoped for. The result didn't clear my threshold.

Yet many younger users clearly feel differently. For them, AI companions are engaging and valuable despite their limitations. That doesn't make the worry about replacing real-world social connection go away. If anything, it sharpens it: a low threshold for "good enough" in relationships could have consequences that a low threshold for meme-making never will.

### Communities as Generative Models

There's an analogy I find useful here. Scott Alexander's *Astral Codex Ten* book review contests work a bit like a generative model. Participants write in a particular, recognizable style tuned to what the community appreciates, and the result is content shaped to meet that community's quality bar. Human writers, in effect, generate tailored output for an audience with known expectations.

It's a reminder that "good enough" is often defined socially, and that humans already do something like generation when they write to fit a community's standards.

## Writing for Pleasure vs. Writing at Work

All of this has made me think harder about my own writing, which splits into two quite different modes.

**Writing for pleasure** is unfiltered and exploratory. I often don't know exactly what I think until I've worked through it on the page, and the pieces tend to be driven by a thesis I'm discovering as I go. This is where I resist AI most firmly. The point isn't just the output. Human care and authentic expression are the whole value, and handing that off would defeat the purpose.

**Writing at work** is another animal. It's structured, section-based, and analytical. The goal is to communicate clearly and efficiently, not to explore. And here I'm open to change. I can imagine a future where I hand a structured outline to a model like Claude or Muse Spark and get back a polished workplace communication. Since the structure and the thinking are already mine, the AI is handling execution rather than substituting for expression.

## Conclusion

The fight over AI often gets framed as a moral binary, for or against. But what I keep noticing is that people are running a quieter calculation: does this output clear my bar for existing at all? In coding, the bar is functional and easy to clear. In personal writing, it's deeply tied to human authenticity and hard to clear. In politics, it varies enormously between a throwaway meme and a polling strategy.

Understanding where those thresholds sit, for individuals, professions, and parties, may tell us more about who adopts AI and who gets left behind than any debate over purity ever will.

Post Post Script

  1. Reading the claudeslop I’m realizing I might have failed to write a conclusion writing this blog.
  2. Maybe reading the AI writing has already killed my brain and so I’m unwilling to take a position.
  3. Obviously, people’s writing is just derivative of what it’s been trained to do through reading and writing.
  4. Maybe the standard American Academic Essay is the real mind-killer.
  5. I wish I was as good at writing as Robert Caro.