I am finding myself neither on any of the sides, nor somewhere in the middle with my position. Rather, I am all over the place. In short: I wish it didn’t exist, but now that it’s here, I am searching for the “correct” opinion to have. I use it. It is incredibly useful. But it gives me a myriad of negative feelings all the same. I want to summarize some of the big reasons to dislike it in this article to structure my own thoughts a little bit more and come to a better consensus with myself on how to feel.
1. Resource Consumption
Let’s get the obvious out of the way. AI uses incredible amounts of resources: the raw materials used in the hardware, the electricity and water to run it, and the real estate required for all the data centers. While the world is already aching under the ever-increasing demand for the finite resources available, this adds an unnecessary burden on top.
However, my biggest gripe with this point is that it is disputed. While it should be easily measurable and there should be one right answer to the question of “how many resources are consumed by a datacenter?”, depending on where you look, you will hear about either side of the “argument”. For instance, do datacenters increase the electricity costs of the surrounding residential areas or do they even decrease said cost? For some reason I now have to care about this and I just don’t want to.
2. Legality and Ethics of Training Data Collection
Tech companies are using questionable methods to collect all the data they can get their hands on. How much of it is actually illegal is currently explored in court in many different lawsuits, e.g., The New York Times v. Microsoft and OpenAI. We can assume that the data collection didn’t consider licensing information and copyright law when scraping the web. This could be seen in early image generation models which sometimes created watermarks they had accidentally included in the training data.
Even when the legality is somewhat of a gray area, such as any text or images without explicit licenses attached to them, we reach the question of morality. Is it OK to take everything made by humans available on the internet and use it for AI training without the authors’ consent? No, it is not. There is no argument here. This is morally deplorable. There are two main arguments from AI bros against it:
(1) Humans do the same thing, just on a smaller scale. Everyone reads, looks at images, and watches videos. Then they are inspired to create something based on their experience. If you think this is the same thing, humanity is in trouble.
(2) It is “technically legal”. Exactly. That’s what you say when you know you have no moral leg to stand on.
3. Rights of AI-generated Content
The rights issue goes beyond the legality of the training data. The issue of the rights of the output generated by the AI is twofold: (1) the AI provider could potentially claim ownership of whatever goes in and out of their models, and (2) if a model reproduces something from the training data, what happens to the license of the original work and how can one even detect that this happened? Without clarity on this part, how can anyone proceed to use AI for production-level commercial applications?
On the input side, users readily hand over all their data to the model providers and sometimes even additionally to a middle-man. What could possibly go wrong?
On the output side, we will see a reckoning eventually when the laws catch up and people find mountains of copyright-infringing content all over the place.
4. The Rugpull
Model providers are not profitable. Adoption is widespread. This combination will inevitably lead to either a sharp increase in price, vastly exceeding the human resource costs saved by AI-generated layoffs, or the collapse of those companies, completely removing access to the now-essential models. This effect can be counteracted by hosting local LLMs with increasingly powerful open-weights models. But hosting the good models is also costly and prohibitively expensive for small companies. In any case, I fully expect a rugpull to occur, and the effects will be brutal. Source: trust me bro.
5. Security Nightmare
The security nightmare manifests in different ways. But not in the most obvious way: poor code quality. The reality of the situation is, AI code is above average. As such, it most likely produces a below-average amount of security vulnerabilities.
However, there is another big problem: usage by idiots, who have no clue about proper deployment. Many people in upper-management positions in non-IT fields are vibe coding all sorts of applications. The meme of the vibe coder telling people to “check out their app at http://localhost:3000” is a classic already. But some actually manage to make their app available to the internet at large somehow. Not even a single clue what an SSL certificate is. I’m somewhat afraid of this behavior and hope they will be the only ones suffering the consequences and no innocent user data will be involved. Unfortunately, there almost certainly will be collateral damage.
The more widely talked-about security nightmare is the incredible ability of models like Mythos to find vulnerabilities and write proof-of-concept exploits, even through multiple layers of abstraction. It was so good, in fact, that the public still does not have access to the full capabilities of the model. Instead, they handed out a crippled version of it in the form of Fable. On top of all that, apparently the companies claiming that “coding is solved” can’t even set up a proper sandbox to test their agents and they end up breaking out. At least OpenAI’s story is rather believable here. But as soon as Anthropic heard about it, they wanted a piece of the pie and claimed all sorts of breaks of containment. Sure, Jan. No matter what the companies say, the agents certainly have the ability to carry out cyberattacks similar to what is described, so this is very unsettling in a world where users are already vulnerable to bad actors as soon as they go online.
6. Dead Internet Theory
The Dead Internet theory in its original conspiracy form is obviously bogus, but it evolved into a description of the real consequences of GenAI getting better: many things you see on the web are AI-generated and mostly have no value. I look at it, discard it, and lament the lost seconds. The seconds add up. Too many things are no longer real. Further, because so many things are fake, you can no longer believe anything you see. Brains are trained to be skeptical of everything, even if there are no obvious signs of AI in what they are looking at. Consuming content is becoming insufferable and exhausting.
7. Sense of Accomplishment
When I vibe code something I can not be proud of the result anymore. I can get stuff done and that might be helpful, for instance to make a quick prototype to assess if a project is worth making for real, but it is slightly demoralizing to start personal projects in this environment. I keep myself motivated by having the AI guide me through it and removing impediments with my learning-by-doing agent skill. Yet, it has become increasingly difficult to keep up the will to make stuff by hand when seeing myriads of results from other people around, de-valueing anything I make, at least in my head.
8. Disruption to the Labor Market
Finding a job in the software industry is very difficult now because CEOs are too hyped for the technology without understanding any of it. In some places this was already felt and caused chaos with partial reversals of firings and such, but the environment is irreversibly broken. The situation is obviously much worse for artists. In that sense, this is very different from the industrial revolution, which people constantly liken to the current transformation. We are suddenly eliminating creative jobs and things humans actually enjoy doing instead of hard manual labor. At some point we took a wrong turn there.
The biggest issue in the labor market right now is for entry-level jobs. Junior personnel is replaced by AI and only seniors still have value. Where exactly do people expect the future seniors to come from? On top of all that, job postings are flooded by AI applications and then evaluated by AI because how else are they going to get through this mountain of paperwork?
9. Are You Really Saving Time?
To properly use agentic assistance, it costs an immense amount of work to establish a workflow and connect all the pieces: AIs that write the code, AIs that check the PRs, AIs that audit the security, etc. Making sure this all works together and monitoring the operation will take time. Then the entire frontier AI landscape changes about once a week. Are you actually getting faster or are you just spending your time on something else? The promised 10x development output is not reflected in a 10x product or feature output. It seems to be closer to 10%.
10. The Velocity Is Too High
If you have an established agentic workflow and keep it running as fast as possible by feeding in tasks and only surface-level testing the result manually, shipping features can get very fast. Given you manage to materialize a 10x in productivity, there will be a reckoning where something will fatally break because changes were implemented so quickly that (1) they could not be reviewed properly and (2) the humans in the loop did not understand the implications enough and could not fix the error when shit hit the fan because they are too far removed from the system they didn’t build themselves. There is no one who understands the code enough to pick up the pieces. Someone will have to. That takes time. And there goes the 10x…
Additionally, someone who ships 10x will burn out faster. Most senior IT people I know already had a burnout at some point in their careers. Speeding things up might increase the likelihood and speed up the timeline for this scenario. Not exactly a bright future to look forward to.
11. You Are Forced to Participate
It is impossible to keep up with the current desk-job reality without using some form of AI for assistance, even if nobody is explicitly telling you to do so. Management will implicitly increase expectations from everyone and you will not be able to keep up without AI assistance.
Conclusion
For enterprise-level work (not just code), I would not rely on the uncertainty of an LLM output.
AI is good for small personal projects, prototypes, and educational purposes if used responsibly.
Catch-22: Ideally, it would be used for increased productivity in a business, but that’s the last place where I would personally use it.
We live in the best and worst timeline simultaneously.