Standards, Averages, and "Good Enough"

Speaking from past experience, I've had moments in my life where, rather than being responsible and getting a bowl and spoon, I would simply take the gallon of ice cream and a spoon and have more than I probably should have. In the moment, it was fine. But after the fact — usually a couple hours after the fact — you get the idea. Sometimes it's just hard to stop, especially if there's nobody there to stop you.

There's a lot of things in life like that. It could be ice cream. It could be scrolling Instagram. It could be Netflix autoplaying to the next episode, and you just roll on with it.

Another place I see that same pitfall is AI. I very clearly see the appeal of putting text into an input and getting an immediate response back from something that sounds very human and is willing to support us and tell us how wonderful our ideas are. And why wouldn't we want to consume as much of that as possible? If we're in a rough place in life, we're not getting the support we need, whatever that looks like — but we've got that one thing that's always in our corner. It's gonna be really hard to say no to that.

---

Much like most people, I have a very basic understanding of the way AI makes its choices — all the calculations that say this is likely the next word and why it's the next word. But for a moment, I don't even want to think about the math side of it. I just want to consider something I observe in the world.

AI can only really train on what people have talked about. And if you scroll on LinkedIn, you see the dichotomy of people who have an opinion about something — this is what I did to get a job, and you can too, versus don't do this because of this or that. I think it's really important to remember: these people are speaking from experience, but specifically their own personal experience.

When you get in a room with somebody and start having a conversation, that's a completely unique moment — affected by the smile on your face, the shirt you wear, the deodorant you put on that day, the sandwich the other person had for lunch, how much sleep they got, all these little things adding up. So it's going to be the things that people actually report that the models use.

So if we take this framing and look at resumes — you've got somebody who says, I did these things to my resume, I stuffed keywords, I changed the formatting, etcetera. You'll get so many different combinations, but it's relevant to the time and place and person. So it's not appropriate to assume that one answer is the correct answer.

And even further still, I would venture to say that people just do the right thing, and maybe don't talk about it, because they're busy living their lives and just doing the thing and not telling the whole world about it.

Think about it this way. A lot of what shows up on a place like LinkedIn is posturing — people writing to be visible, for whatever reason. There are a few posts that are genuinely "I'm sharing this because I think it's cool, and I want other people to know." So if you're not the kind of person who postures — if you just show up, do your work, and do it well — your best insights are happening person to person, at meetups, at conferences, in conversations that never get written down. And the models, as I understand it, were trained on a large, curated part of the internet. Which means those unrecorded conversations never make it in. The keen insights, the stuff actually steering toward the next thing, isn't available to the model at all. What's left is the generally agreed, the already-assumed-correct — with the most variability landing exactly where things are newest or most in conflict.

So if AI is trained on the average, how can you really get at the right thing without doing a little bit of work?



A real "good enough" is something based on standards. The less standards that exist in the space you're addressing, the less credibility "good enough" has. Without any standards behind it at all, "good enough" is three kids in a trench coat pretending to be an adult.

Think about a car. You've got the totaled one — wrecked, undrivable. That's your zero. You've got the one that's properly cared for — serviced on schedule, oil changed, tires rotated, running the way it should. New off the lot, that's your hundred. And then there's "good enough": the car you keep driving with the maintenance light on. You skip the oil change, or you patch it instead of fixing it. It drives fine today. But the cost of ignoring all that doesn't show up until much later, when it's a much bigger problem with a much larger price tag attached.

It's generally agreed that AI isn't really coming up with new ideas. And because of that, it's not going to magically define a standard either — it can't conjure something that doesn't already exist in a given space. It's leveraging the knowledge that already exists for that space, so it falls back on the quality of standards in play there. And if the answer to "what are the standards" is "not a lot," then in addition to engaging with the prompt, you have to be really clear about what your own personal standards are — because if you don't name them, AI is just going to default to that magical average of information that exists in the training data.

I think AI tools have the potential to do really good things. But I don't think testing and quality-checking is the user's responsibility. Not really. Somebody shows up to use a tool, and they expect it to work well — they expect they've been handed something that does what it claims it will do. The responsibility falls entirely on the developer, the product team, everybody backing the idea. Companies need to slow down and actually think about what they're putting out.

That said — users still need to go into this with their eyes open. Going in eyes open means that when you get a response, you're willing to question whether it's just what you want to hear, or something that's valid and based in reality. If you don't think about any of this, you're going to get that average answer, built on some good and some bad, and you won't know which parts are which.

AI is a brand-new tool — very powerful when used correctly, very dangerous when used incorrectly. When I go onto Claude or Gemini, there's nothing there outside of the guardrail responses that show up if I ask something in violation of them. The tools just very happily respond. Hallucinations are still a thing — they're just hidden better now. Search and source citations help, in that they give you something to actually go check — but they don't mean the problem's solved. Citing a source isn't the same as the source being right. There's a lot stated as matter of fact, and that's the part that's dangerous. But there's nothing that trains the user how to engage with these interfaces in a way that safeguards themselves. And that's the problem: it gets promoted as this huge, amazing thing, and almost nobody talks about where it falls short — or whether we even got to opt into using it in the first place. (Looking at you, Google.)

On the developer side — and this is coming from my own personal observations — I think there's been a lot of relying on the user to evaluate the quality. I think of OpenAI in its early days of ChatGPT, where it would present two different answers and people would pick which one read better to them. There are two things going on there: OpenAI was using users to train the models, and the answers themselves weren't necessarily authoritative — it was simply "I like A better than B." But in the end, OpenAI didn't pay most of us. They just said, here's the thing, answer this question for us because it's really easy to answer. And guess what — people did. I did it. And even today, it still exists in some ways: thumbs up, thumbs down. It's information for the developers — "this was good," or "I didn't like this." Which means they can still keep shipping things quickly that live under the ruse of good enough, and quietly use those indicators to make corrections later.

I've had moments of this in my own app, and in some ways, I regret it. I think it's just a pitfall of being one person trying to build an idea. But I made an effort to understand and improve, which is all I think can be asked of anyone.

AI is fantastic at pattern recognition. But it has to have enough information to find the pattern, which means a person has to give it enough information in the first place. And that's the hard part.

But on the flip side, most people are prone to take the easy path, at least at first. There are a ton of reasons for that. Maybe you just really don't care. Maybe you don't have the energy. Maybe the thinking brings up past experiences that hurt. I could keep going, but you get the idea.

Something I've noticed in interacting with different AI tools, though, is that taking the easy path starts to leave you with an empty feeling inside after a while. You feel how little control you have, and how empty things sound or feel. And then you want to go back the other way — so you start putting in just a little bit of effort.

I think it takes a very self-aware and mentally healthier person to just dive into the hard things. And I don't think that's a place many people reside, especially these days. So much of the time, as far as I can tell, nobody has quite enough money, and it feels safer to stay where you are than to make a change. There's safety and familiarity in staying put.

But it takes a healthier person to be able to do this alone. With the right kind of support, the hard things become more accessible to everybody. Excavation isn't only for the self-aware few. It's for everyone. It should be for everyone. It shouldn't be a luxury.

The minute you put a gallon of ice cream in the freezer, it's there. And at any time, you could take the whole thing out.

It takes effort and discipline to only grab a bowl. And sometimes, for some people — like myself — it's better not to have the gallon in the freezer in the first place. You stick with just the occasional pint.

So whether we're talking about a gallon of ice cream, or an AI model that will tell you you're the most wonderful human being and all your choices are amazing — it's worth stopping and thinking about the conscious and unconscious choices we have to make on a daily basis, with all the different things that are now part of our lives, whether or not we asked for them.

So, as a parting thought: if you're going to get a gallon of ice cream, make sure you have some humans to share it with.