By Santiago Gisler, 15 September 2026
Clearly, I’ve become that grumpy old man waving a fist and yelling at the AI clouds. Whether my growing scepticism toward generative AI stems from a genuine, scientifically evaluated concern or just plain age remains debatable. I’d claim the former is true. But who am I to judge? We should let the evidence speak.
As generative AI tools increasingly creep into our everyday and professional lives, people naturally feel we should learn about large language models (LLMs) by implementing them; they claim we learn by doing. LLMs are a subset of generative AI that process and generate human-like language. Many scoff when I argue against using LLMs, especially when it comes to learning a new subject. These scoffers claim that students and professionals must learn how to use LLMs to develop skills and survive the future job market.
But isn’t learning to use LLMs to enhance our cognition putting the cart before the horse?
Regardless of whether LLM tools can realistically replace knowledge workers, many of us feel pressure to use these tools to meet tighter deadlines, diversify projects, or meet other job requirements. As a result, professionals and students have rapidly adopted LLMs and other generative AI tools in their work, especially amid the fear-mongering about AI replacing jobs.
The AI hype matters to professionals in fields that rely on critical thinking. Surely the rapid and ubiquitous adoption of LLMs in our daily and professional lives must affect our relationship with information; how we gather, analyse and disseminate it.
Our relationship with technology and LLMs
Humans have always criticized the impact that technology has on our cognitive abilities. Socrates criticized writing, Trithemius printing, and Nicholas Carr has written extensively about the effects of the internet on our thinking.
We may laugh at these objections today, having incorporated all of them into our daily lives. But even a vital process, such as writing things down, is a form of cognitive offloading, in which we rely on external tools to reduce cognitive load.
Cognitive offloading can help us focus on other tasks, sometimes more important or demanding. But it can also undermine our long-term thinking and memory. Let’s return to writing as an example: before we started writing things down, a person had to memorize an incredible amount of information before sharing it with others. In essence, being intellectually strong was an exercise of memory. We don’t need that anymore – at least not to the same extent.
Outsourcing critical thinking to LLMs
LLMs can be used as a cognitive offloading medium, like the pen, printer and internet, which can both support and undermine our cognitive engagement and skill development. How, then, do LLMs affect the higher-order cognitive skill known as critical thinking?
Much as the Swedish expression tells us that a loved child has many names (‘kärt barn har många namn’), critical thinking also has several definitions. For the sake of this article, let’s agree that critical thinking enables us to process information so that we can make reasonable, intellectually motivated decisions. It goes beyond acquiring, memorizing or using information.
Most of us have seen how users of these LLMs – sometimes ourselves – continuously offload not only their memory but also their analytical and creative thinking processes. This skewed delegation of thinking became especially obvious to me when I volunteered as a tutor for aspiring science communicators. As I corrected their summaries of a scientific article I’d assigned them, I discovered that more than one was beautifully, almost professionally written but contained closely related yet incorrect information.
This outcome should have been obvious from the start. If we don’t know the topic we ask LLMs to help us write about, we will struggle to identify hallucinations; after all, it reads like a professional piece of writing. The question is: how does ‘lazy’ LLM use affect our critical thinking in the long term?
One of the most cited studies evaluating the effect of LLM tools on critical thinking used a mixed-methods approach, yielding both quantitative survey data and qualitative findings from interviews with 666 participants. The study found a strong negative correlation between frequency of LLM use and critical thinking, with younger users most affected.
As with all research, we need to mind the nuances of the findings. We rarely see simple binary conclusions in any research, and that is also true about the cognitive impact of LLM use. Much like our other sources of cognitive offload, the effects depend on how we use them and the premises we adopt.
Some evidence indicates that users’ motivation and confidence can predict how cognitive offloading to LLMs correlates with critical thinking.
Self-reports from a university in China, for example, indicated that intrinsically motivated students, whose engagement was driven by interest and enjoyment, used LLMs as a thinking partner while maintaining cognitive agency. By contrast, extrinsically motivated students, motivated by external rewards or punishment, were more likely to use LLMs as a thinking replacement, without maintaining metacognitive oversight for genuine learning.
Another study on knowledge workers found that high confidence in LLMs correlated with less critical thinking, while high self-confidence correlated with more critical thinking. The authors noted that although LLMs may improve work efficiency, they hamper users’ cognitive relationship with the information at hand and create an overreliance on the tools. Rather than gathering information ourselves, we settle for verifying what the tool has already gathered; we limit ourselves to integrating digested content rather than solving problems; and we start supervising the tools rather than performing the tasks.
And that’s a creeping problem to which we can all fall victim. I’ll stick my neck out and suggest that our increasing exposure to these tools and the grandiose projections of their potential, spewed by certain tech giants, may be incentivizing greater confidence in them – maybe even overreliance. We cannot rely on our intrinsic motivation guiding our rational use of these tools. We all have to perform under pressure sometimes, perhaps with a motivating carrot in front of us or a stick threatening our tender backs. An innocent cognitive offload here and there can’t hurt, right? Especially if we’re the captains running the ship… right?
The LLMs’ homogenizing effects can make us – or break us
Of course, it’s not all doom and gloom. We could argue that LLMs can somehow democratize information. In a world where English dominates academic and cultural propagation, LLMs can arguably help the dissemination of knowledge, especially for those with English as a second language.
We know that non-native English researchers spend significantly more time on every stage of communication, including reading, writing and sharing content. Non-native English-speaking women from low-income countries published up to 70% fewer English-language papers than male native English speakers from high-income countries.
Viewed this way, LLMs have the potential to democratize writing, levelling the playing field between more or less privileged parts of a population. They can also homogenize language, aligning voices until writers start to sound alike. As the Swedish saying goes: alike children play best together (‘lika barn leka bäst’), to shoehorn in my last Swedish expression for the piece.
However romantic that image may seem, the reality is that the written word, critical thinking, and, consequently, intellectual and societal progress also require diversity, disagreement and (brace yourselves) debate. That’s not to say we should resist the democratization of information access and the opportunity to disseminate it – on the contrary, it’s essential to democratize discourse. However, whether current LLMs are the proper vector for these initiatives is questionable.
LLMs can strengthen narratives and language at an individual level, perhaps making ideas more creative and satisfying. But the trend towards homogenization of language and ideas also has real-life consequences for intellectual and societal progress. LLM-generated texts are collectively less diverse in style, language and ideas. In other words, what makes our intellectual endeavours better individually also makes them worse collectively.
I’ve called this phenomenon a positive feedback loop – with a quantitative meaning to ‘positive’ – since ChatGPT’s launch. Our increasing cognitive offload to LLMs creates ever more machine-generated content, which fills the content landscape and, in turn, trains the next generation of LLMs. New, human outlier ideas, some insignificant, others paradigm-shifting, will eventually drown in an ever-expanding sea of LLM-trained LLMs or LLM-trained humans. Until…
We solve excessive LLM use with metacognition
One recurring theme seems to encapsulate a possible solution to the problem of reduced critical thinking stemming from cognitive offloading to LLMs (or generative AI, more generally): metacognitive training.
Some researchers argue that, rather than learning about the actual tools, we should learn to reflect on our thinking processes while using them. For example, at an individual level, knowledge professionals and students could explicitly question what cognitive work they offload, why they choose to offload it and how they will verify the tools’ outputs.
Remember how I questioned whether more LLM training wasn’t like putting the cart before the horse? That’s because study after study shows us that we can’t reduce LLMs’ cognitive effects by ‘just using them more’ – in fact, an overreliance on them worsens cognitive abilities.
From experience, I’ve seen that the most solid content comes from students and professionals who’ve taken the time to inform themselves about the specific topic or task. It allows them to think outside of the LLM-restricted box and colour their content with rough, yet vivid language and original, sometimes dissenting ideas. That’s what advances us as professionals and, before all, as social beings.
Sure, we could redesign models to include ‘cognitive forcing functions’, requiring LLM users to wait, interact or answer questions before seeing an output. But that depends on choices we don’t get to make – yet.
I believe in our cognitive integrity. Let’s stick with training that part for now. I’ll lower my fist now… though I’ll keep one eye on the clouds.
|
Blog post by: Santiago Gisler |
