Are LLMs Influencing the Way We Think?

How should we relate to the LLMs that increasingly power our lives? One approach has been to dub them “tools” and treat them like any other instrument that extends human capacity. That tool metaphor has proved popular: In 2023, designers Caspar Lam and Yujune Park called AI image-generation platforms “just another tool” for creative expression. Designer Mia Blume went a step further, comparing generative AI models to “super pencils.”

Suggestions From a Hammer?

In one sense, these comparisons are fair. While we might hear the word “tool” and immediately think of building tools like hammers, drills, and screwdrivers, the word’s technical definition has always been broader: basically, a tool is “any object that extends a person’s capabilities and lets them carry out particular functions.” But is that all an LLM does? When you use a hammer to hang a picture, it doesn’t whisper, “What about a mirror instead?” Likewise, your cereal bowl doesn’t suggest that oatmeal might be a healthier choice, and your electric razor doesn’t ask if you’d like to see samples of popular beard styles for 2026. That’s because our relationship to tools is user-directed and unidirectional – we act on them, but they don’t generally act on us.

The way we interact with LLMs isn’t so straightforward. They perform the tasks we ask them to, yes, but in doing so, they also shape how those tasks unfold by deciding what information we encounter and how we think about it. For this reason, an LLM is best understood as a mediator, or something that enters between us and an activity and changes our relationship to that activity. Consider how we navigate. A paper map is a tool that provides information about a specific geographic area and features like roads, rivers, lakes, and parks that we can use to choose a route. A GPS app recommends a route, directs our attention turn by turn, recalculates when conditions change, and offers alternatives based on what it thinks we care about (time, distance, tolls, etc.). When we use Apple Maps to get to a destination, we’re still driving and choosing where to go, but the GPS shapes how we get there. LLMs play a similar role in our thinking.

When you use a hammer to hang a picture, it doesn’t whisper, ‘What about a mirror instead?’… An LLM is best understood as a mediator—something that enters between us and an activity and changes our relationship to that activity.

Reactive Writing

The implications of this – that an LLM mediates our thinking – depend on one’s context. Because I’m a professor of communications, I think about what this means for our writing. We might assume that when we use an LLM to draft a document, for instance, it’s just helping us get our thoughts on paper and overcome what novelist Colum McCann calls “the terror of the white page.” But we’d be wrong. For one thing, research confirms something we all probably feel instinctively: LLMs don’t write like people. A recent paper by researchers at Carnegie Mellon University found that humans tend to write in a more interactive, conversational style, even when they’re writing in more formal genres – think shorter sentences, more questions and directives, and lots of pronouns. LLM-generated writing, by contrast, is informationally dense and impersonal, characterized by lots of nouns and adjectives, and an absence of personal pronouns.

There’s evidence that also shows how LLM use influences our thinking in addition to style. In a recent article, researchers studying human-AI interaction asked participants to write about social media while receiving suggestions from an AI programmed to emphasize either the benefits or harms of social media use. Those suggestions influenced which ideas participants ultimately wrote about, even when they were writing their own text and not simply accepting AI-generated options: people exposed to positive suggestions focused more on social media’s benefits, while those exposed to critical suggestions focused more on its harms. Participants who were not exposed to AI suggestions, by contrast, consistently explored a wider range of ideas about social media. According to the research team, AI use during drafting shifted participants’ focus away from asking, “What do I think about this?” and toward asking, “Do I agree with this suggestion?” – a phenomenon called “reactive writing.”

If LLMs mediate drafting in ways we might not notice, what about revising? Surely if we’ve done the writing ourselves, asking an LLM to help us improve it is a straightforward use of the technology as a tool; the LLM just helps us express them more clearly. But even here, the tool metaphor obscures what is actually happening. A study by a team of university and Google DeepMind researchers compared how people and LLMs revised the same set of essays in response to the same expert feedback. The differences were considerable: Human writers tended to make targeted changes that preserved most of their original language, while LLMs replaced far more of it, in one instance changing an essay’s vocabulary at nearly three times the rate of human revisers. These changes often added more nouns and adjectives and removed pronouns, making the writing less conversational and more impersonal.

More significantly, those changes in language sometimes also changed meaning. Across three different LLMs, the researchers found that AI revisions produced larger and more consistent “semantic shifts,” or changes in a sentence’s meaning, than human revisions did. This happened even when the models were explicitly instructed to make minimal edits or to correct grammar without changing the content. In some cases, the LLM fundamentally changed the conclusion the writer had originally reached.

What do writers make of the fact that LLMs subtly influence their writing? The same study found that people who relied heavily on an LLM were less likely to feel that their essays were creative or reflected their own voice, yet they were no less satisfied with the results. In fact, heavy LLM users reported slightly higher satisfaction with their final essays than writers who received limited or no AI assistance, even as they recognized that something of their own voice and creativity had been lost in the process.

Is “Good Enough” Good Enough?

What we’re seeing here is probably a form of what Carnegie Mellon legend Herb Simon calls “satisficing,” our human tendency to settle for “good enough” rather than the best possible output. But there may be other forces at work, too. French philosopher Jacques Ellul argued that modern technological societies increasingly make efficiency the standard by which we judge how things should be done. Once a new technology proves faster or more efficient, choosing a slower alternative becomes increasingly difficult to justify. LLMs enter a culture already primed to value exactly what they offer (more polished writing produced with less time and effort), so it’s perhaps not surprising if we find ourselves accepting writing that is good enough even as we recognize that it’s a bit flat and characterless.

But “good enough” writing comes with a cost. If LLMs influence which ideas we consider when drafting and reshape those ideas when revising, more of what we communicate could reflect model-driven behavior, not patterns of human cognition. Over time, that could narrow the range of ideas we write about as well as how we express those ideas. In other words, writing with an LLM will probably make our writing more homogenous, but it might do the same to our thinking. After all, writing isn’t simply a way of recording thoughts we’ve already had; it’s one of the ways we develop those thoughts in the first place. Writing researcher Janet Emig made this case nearly fifty years ago, arguing that writing is itself a “mode of learning” through which we analyze, synthesize, and make sense of what we know. If millions of us increasingly turn to the same handful of models to help us think, some of the differences in how we understand and make sense of the world may begin to disappear from our writing.