My Take On Llms And Other Glorified Autocorrects

· 27 October 2026

As with every new cycle of “new shining technology”, I approach the surrounding hype with skepticism.

Most often technologies are just a compound of other pieces, ideas that somebody else had, assembled together and, in some cases, productized. In that sense, technologies are often born after long debates, adaptations. Trevor Pinch and Wiebe Bijker explained this very well, analyzing the history of the bicycle, arguing that technological evolution is not linear or purely logical, but rather a reflection of social negotiation, conflict, and cultural values among diverse social groups. These values ultimately steer the specific tasks the technology is built to attain, however repressive or violent those tasks may be. The vision and goals given by its creator and adaptor are always embedded within the system.

For example, let’s think for a second about the difference between a knife and a sword: both can be carved out of steel or iron and have cutting capabilities. However, the former offers a flexibility that fits the daily life of the common person in the kitchen, in crafting something, and much more. The same blade can be extended alongside a stronger hilt, and here you have a war instrument. The need to specialize existing technologies for a specific purpose is a political goal, embodied in the artifact of a sword.

In our current globalized economic and political system, the technological solutionism paradigm works as a way to expand those technologies in every possible corner of the globe to fix every possible problem under the idea of democratization.

Some pseudo-transhumanists overcharge these technologies with meaning and start having (sometimes) wild ideas specifically formed to respond to human needs, interests, power, habits, politics. Most often, everyone else around it is caught in the crossfire (and often manipulated).

I feel skeptical about the voices that reach the mainstream. One can see many experts and a general market push to have EVERYONE moving in that frame, a frame decided mostly on market needs. Everybody wants to move in that realm of power and interests, even if {insert new shiny thing here} is not really their thing. We are seeing this with AI, we might soon see this with quantum cryptography, just like we saw that with blockchain. Here, psychological, reputational, and egotical dynamics predominate.

More critically, at the market level, I suspect AI is probably a bubble that’s about to burst (and of which we will keep some watered-down benefits), so I’ll try to not get too used to it. Also, I don’t want to get deep in the financial rabbit hole, but look at the expectations surrounding these tools. The SpaceX IPO was the largest in history with $75 billion raised and Nasdaq rewrote its index rules to fast-track it (and the coming OpenAI and Anthropic IPOs) into the Nasdaq 100 within 15 trading days instead of three months. Meanwhile, S&P refused to do the same.

But I only see the hype doing the work here, not financial fundamentals. While SpaceX generates revenue, the AI labs themselves still have to demonstrate they can convert this scale of investment into durable returns (perhaps betting on orbiting data centers isn’t such a great idea). Their business models remain unproven against the trillions in valuation being demanded.

In the future, I expect large companies to have B2B agreements with providers for specific automated AI/ML tasks. LLMs in particular, I’m not sure. You and me as mere mortals won’t get the same (read, cheap) access to LLMs in the future, that’s sure. So don’t get too used to it, especially for important, unstructured tasks.

Let’s get to the point. I use LLMs, but I also recognize the level of exploitation they can trigger and justify. I think they are a very useful tool to rearrange textual elements. When it comes to problem solving and creative work, LLMs are able to systematically apply existing techniques they have been trained with to new problems, and quickly evaluate potential and underexplored options. That’s basically a glorified autocorrect. For copyediting, checking typos, and such, those are great tools.

But when you have to work on any assignment, your human contribution should still be there, not to merely correct, but to enrich in meaning. Humans are meaning-driven. If we replace the inner nature to seek that meaning, that’s where we lose ourselves. The more people work on an assignment, the more it reflects the minds, perspectives and values of those that worked on it. The ability to tell a story, interpret other actors’ meaning and reality, optimize a representative model of objects in a system, express your emotions and feelings. All those are contexts where AI would maybe simulate a decent output, but in reality will never be capable of expressing what’s really on your mind. Enrichment in meaning is uniquely human, while a machine merely assembles statistically plausible outputs.

Ultimately, we risk losing the ability to forge our own connections between reality and abstraction, which is how we truly learn and find satisfaction. Authorship is about originating, owning and defending meaning.

That’s why AI can not be a valid author. As FSFE recently explained, meaning and intellectual value potentially copyrightable comes from selection, guidance, organization, and all added human work. This does not reduce itself to a “write a creative prompt for my prompt (make no mistakes!1!)”, as that only covers generation, which is the true realm of LLMs, not the human. In a age where it seems like you can delegate everything, the role of humans for choosing desired collective paths will be even more political and catered to making sure everyone can reap the benefits of these tools.

Since authorship requires human agency, the question of who gets to exercise that agency becomes critical. Inclusion among those producing AI is the next mandatory step: inclusion of marginalized groups, all workers, women, teenagers, the elderly. True digital inclusion does not mean “adoption of a product”, not even “better training on the data of those groups harvested with unclear consent”. It means holding companies, developers, and governments accountable for the impact these technologies have on those very groups. If these groups will not be able to access LLM tools in a way that protects and preserves their agency, then they will only become a tool of exploitation.

Do we really want AI in every single corner of our schools? Do we want to massively deploy vibe-coded airplane control systems, or healthcare management systems? Do we want AI agents running entire portions of key internet, energy or defense infrastructure? Or can we instead deploy less-powerful, less resource-intensive local models to assist with specific, detailed tasks as defined by collective needs? Such an approach would return control to communities, reduce environmental costs, and ensure that the technology serves us rather than consuming us.

Fixing AI is a very difficult task, I understand.

In the meanwhile, what I can do is to always disclose my use of AI. Each piece will specify the model used and the specific role it played (e.g., brainstorming, drafting, or editing).

Keep reading, writing, thinking, debating. You don’t want to become isolated and have an LLM as your personal buddy. Thinking is still our job. Imagining and building a future for the 99% is still our job.

# Transparency Disclosure
This piece was written by me, with core ideas, critical stance, and final conclusions entirely my own. I used an AI assistant (Lumo 2.0) as an editor to:

* Rewrite certain portions of text for clarity and flow, while preserving my original voice and word choice.
* Correct typos and grammar.
* Verify specific financial claims (SpaceX IPO, Nasdaq rules) via web search.
* Translate some Italian sections into English.