AI Manifesto

Sergio Gutiérrez

May 11, 2026

A breakdown of how (and how little) I use AI at my studio, and the reasoning behind it.

The following page describes how I use AI (LLMs) at Antropia and the rationale behind some of these decisions.

This post is the result of my work and experiences using AI in different projects. It’s also a public declaration of what you can expect from my work regarding AI usage.

Inspired by

This post has been heavily influenced by David Bushell’s own AI policy.

Summary

I have worked on moderate-scale projects using AI in different tasks: code assistance, coding agents, code reviewing, task definition, and research.

From my experience AI can be as useful as a contextual search engine, and can be used to lower language barriers when used consciously. My current framing is using LLMs as a powerful search engine. Therefore, I take what it produces with the same pinch of salt I’d take a Google result.

Beyond that, I draw a clear line to prevent myself from depending on them. Surrendering my critical thinking or letting my skills rust is off the table. I also keep LLMs away from anything that involves another human on the other side. If you ask me something, you’ll get my take, not ChatGPT’s.

Having said that, I recognize the damage AI has caused to the environment, people’s security and privacy, and society as a whole.

For these reasons, I aim to minimize its use as much as humanly possible. And when using it, I prioritize the use of local models over small model providers, and over big-corps.

Argumentarium

The impact of AI in the world can be analyzed from different angles. In this section, I have collected the ones that resonate with me the most, and therefore have guided my decision on their usage.

1. Ethics

To me this was the most obvious since the whole LLM-ball started rolling. Companies have been scraping the entire internet without consent¹, and then raising their voices when others did the same with their models².

I’m not a hard advocate for copyright and intellectual property. To me, they are a subproduct of a system that rests on mutual distrust and commodification, and I believe we’d be better off with them gone entirely .

There were two ways the training of LLMs could have gone without raising so many eyebrows:

Following the rules: You ask for permission to get a large corpus from different platforms. You train the model and finally you sell it. This is OK to me because it follows the rules we have designed around ownership, intellectual property and copyright, and it does with consent at the center of the picture (otherwise it’d be considered stealing when analyzed from within the system itself). I don’t advocate for these particular rules, but I concede it’s a coherent take.
Aiming for the greater good: You take all the text you can, with or without permission, train your models and release them for free (or covering the training costs). It might sound surprising to some, but this would also have been OK to me. I’m aware if analyzed only from within the system, we’d also consider this theft, but when we see the bigger picture, we’d realize it would be for the benefit of society as a whole .

However, these companies followed a different path. That is, to follow the set of rules they cherry-picked to later lock the models behind a paywall. No consent, and no greater good comes out of it. It’s the intention and the inconsistency of their claims that makes it all wrong.

Outcome

I prefer the use of local / open-weights models and smaller competitors over the big ones that are purposefully disrupting society for profit reasons only.

2. Atrophy

I’ve used LLMs to assist on projects, large and small.

There is an invariant in all of them: I had to spend an entire week afterward to remember how to do the most basic coding without resorting to LLMs³. This is not new, it has been discussed ad nauseam⁴ ⁵ ⁶ by others, and I’m just here to confirm it.

The dependency AI creates is real, and it’s bad.

It’s bad because we have to pay for it twice, first with our wallets (highly subsidized), but also with our own skills that rust over time, deepening the dependency.

It’s even worse if I let it run uncontrolled, I lose critical thinking⁷, creative skills, and in the end, I can no longer feel satisfaction in my work, and I lose confidence in myself⁸.

I’m positive big AI companies know this, and it’s why they are force-feeding everyone with their grandiose narratives . Because if people try it, they will start depending on it, not because it’s particularly powerful, but because of FOMO and skill atrophy. We know how this all ends, it has already started, eventually they will worsen the experience⁹, raise prices¹⁰ and cash it all, it’s their playbook.

Outcome

I trace a clear line in what tasks and LLM can help me do. Surrendering my critical thinking and current skills is off the table. If I’m no expert in the matter in question, I ask the LLM to teach me how to do it by myself.

3. Quality

This is one of the most interesting topics to me, because it’s so nuanced, and still everyone is just throwing contradictory and self-serving claims like it’s the end of the world¹¹ ¹² ¹³. I don’t pretend to do the same, and so I will only talk about my personal experience, read it as a data point, not a trend.

As I said, I have worked in projects using LLMs in multiple steps of the workflow: code assistance, task generation, code reviewing, etc. One common pattern surfaced in all of them: things seemed great, until they didn’t.

This can be explained from multiple perspectives:

Lack of context: There always seems to be something missing in prompts. It can be something mundane like something you discussed with a colleague recently, or agreed with the team on how to do things from now on. Others are more substantial, like the number of users in your platform, or the last version of a framework. Skills and harnesses alleviate the problem for some time, but then the number of skills in your project is so large they don’t fit in a regular context window, or they are plainly ignored.
Project growth: LLMs can make it look like you are flying for small projects like scripts or internal tooling. Projects with a short lifespan benefit from them. But when projects grow, or have been AI-assisted for longer, LLM results are more often than not, wrong. And the more you use LLMs on them, the harder and more expensive the project becomes to maintain.
Appearance vs reality: Have you ever filed a ticket with an LLM? It sure looks amazing, it covers all the scenarios, even the ones you didn’t consider! Then you give that task to someone else and the entire thing crumbles. You thought you had it all covered because it was given to you, rather than thought. But when a human closes the loop, it’s missing stuff, it’s full of noise explaining obvious things, and others are just based on wrong assumptions. LLMs can produce things that look fantastic, in theory, but that when faced with the reality of things don’t make that much sense.
Lack of critical thinking: LLMs displace work from thinking things through to reviewing things. These are two separate tasks in our brain, and they use different paths. How many times did we have to remind ourselves that we are owners of the code we review? We do it because reviewing can be shallow, but building rarely is. It’s a way to force us to be thorough, because it’s in building we truly understand the nuances of doing things a certain way, or considering alternatives. With LLMs we lose that.

The pattern is pretty clear to me. Short-term impressions are usually wildly impressive. Then reality hits, and you are left with a hard dependency on something you never controlled. If that’s the new definition of quality , I don’t want it.

Outcome

I use LLM for coding if, and only if, the lifespan of what I’m building is really short. If the code is going to live longer than a week, I throw it away and do it myself.

4. Humanity

This is probably the more innocent take, but to me is the most important one, and the one I’ve heard most about from friends and fellow freelancers recently.

I heard of Claude conversations being thrown to each other to prove a point, where code reviews and tickets were entirely made by AIs. I had similar experiences, and it’s the most alienating feeling I have ever had.

What’s weird is the first question that popped in my head was “why would I talk to you if I could just ask ChatGPT their opinion?” to then suddenly realize it’s actually a non-question. It’s not that I can do things with machines, it’s that I want to do this with you, fellow human.

There is another level of it, and that’s when entire teams shift to AI-assisted workflows. It’s hard to see it at first, but it propagates like a virus.

You pick a PR to review, you read the description, and it’s just a list of things that have changed, no explanation on why they picked X library instead of Y, no evaluation of the costs of implementing that abstraction, “weird”, you think to yourself. You do the review nonetheless, you write some comments on some odd code that was probably a debugging fluke, you look up and all of a sudden you now have 3 more PRs to review.

You spent time and effort in generating or reviewing something for the next human in the line, out of respect and empathy. You realize the process was inherently human, and now it’s not. You are just a burden in the pipeline, adapt or die. Those using AI to do their work feel blazingly fast, because they are skipping the hard work of reaching consensus, exploring alternatives, asking themselves why. Now they are unloading that bag to the next human in the process, be it another fellow coder, a QA, or someone from support.

If you don’t use AI you are expelled, you can’t keep up, they moved the bottleneck up one step (or so they think), and that step is you.

This is a clear example of the prisoner’s dilemma, and everyone is losing it out of FOMO, or imposition from above.

But replacing the human was never the point of it, and if you think it is, maybe you should ask yourself why you work in the first place, apart from an act of survival .

To me, working that way feels miserable, and it misses the point so much it’s not worth it.

I wouldn’t have called the studio Antropia if it wasn’t to put humans at the core.

Outcome

I will always prioritize humans over LLM-generated content.

Conclusion

I left so many other valid points out, but this is not a thorough review.

After all this, you’d expect me to say I shouldn’t use LLMs at all. I do. Pretending otherwise would be dishonest.

I do use LLMs but from a closed-world assumption, forbidding its use by default, and only adding it to workflows when other alternatives have been considered. And I do it keeping some core principles in mind:

Humans first: I use LLMs as long as they don’t displace human connections. I will never send you what Claude told me. If you ask me something I will work under the assumption that you want my take, not Claude’s.
Known things: I use LLMs for things I already know how to do myself and understand completely. Even then, I test my skills from time to time, if only because it’s fun, and sometimes that’s the point.
Build over review work: When faced with a new problem, I work to come up with a solution first by myself, no AI assistance. This forces me to use the building brain path. I have discovered the review path is far shallower, and it doesn’t activate the reward of coming up with an original solution so strongly. Oh, and it makes it so much easier to reason about it and truly own what I do.
Lifting language barriers: English is not my main language, so I often ask LLMs to find me a specific expression, or a word I can’t remember. I have learned so many English expressions this way! As a Spanish myself, I also know there can be some cultural differences that I tend to avoid out of deference. That’s why I sometimes use LLMs to soften the tone or match theirs. I have also read of people using it to overcome their dyslexia¹⁴ which I find uplifting.
Safe-check: We are humans, and we make mistakes, and this is fine! But our entire sector premise is to reduce human errors as much as possible. This is why I sometimes run a single automated code, post, or document review through an LLM. I do this only as a last resort, and always after human review. It helps me lift my confidence (maybe falsely, I’m aware) before committing to something. I see it as a more thorough last-minute Google search.
Small models over big-corps: I have set a priority for LLM usage. That is to favor local models when possible, and then small model providers over mainstream ones.

That’s it, I’ll keep reviewing my usage from time to time, making sure it doesn’t remove the joy and pride I find doing my work.

Last reviewed: 11-05-2026

Summary

Argumentarium

1. Ethics

2. Atrophy

3. Quality

4. Humanity

Conclusion

Footnotes