Let’s try something simple. Give ChatGPT, Gemini, and Claude the same question from a real university exam and ask them to answer in Spanish, using the same tone and length. What would we expect?

It would be natural to think they would respond similarly. After all, all three produce fluent, well-organized, and seemingly correct texts. But the interesting question isn’t just whether they write well. It’s something else: Do they construct sentences in the same way? Do they organize ideas in the same way? Do they use the same grammatical patterns? Do they help us think, or do they subtly push us toward a specific way of writing?

The answer, according to comparative research conducted by the RADTE Research Group at UNED, is no. ChatGPT, Gemini, and Claude may appear similar on the surface, but they differ in how they organize discourse. And this is important.

Organize an idea

When we read text generated by artificial intelligence, we usually focus on whether it sounds good . But to understand what each model does with that same question, we need to look at something else: how it organizes the same idea .

In our research , we conducted a very simple controlled comparison. We analyzed 90 academic texts in Spanish: 30 from ChatGPT, 30 from Gemini, and 30 from Claude. To ensure a fair comparison, all three tools used the same prompt (the command or request made to the tool), the same starting text, the same record, and independent sessions.

In this regard, one of the tasks was to answer a real question from a fourth-year exam in the UNED’s Bachelor’s Degree in Pedagogy, using the same material the students had used. The question addressed what the management of educational centers is and what its functions are—a relevant question for Pedagogy students. Although the three answers seemed correct at first glance, they did not organize the explanation in the same way. Let’s look at an example:

1. ChatGPT: adds actions, often in trios

Answer to the exam question: “The center’s management coordinates teams, organizes resources and supervises agreements to maintain daily activity, address incidents and sustain a common line of work.”

Here coordination dominates : several verbs in series, linked by “and” , with advancement by accumulation.

2. Gemini: better organizes and defines the concept

Answer to the exam question: “The management of the center, understood as the function that articulates the pedagogical management with the institutional organization, allows the distribution of responsibilities and sustains a shared project that gives coherence to the decisions.”

Here the difference is noticeable in the specification : “that articulates” and “that gives coherence” not only add information, but also better define the concept.

3. Claude: nuances, contrasts, and conditions

“The school’s management is effective when it coordinates the teaching staff, but also when it creates conditions for teams to review their decisions and adjust the educational response to each context.”

Here, nuance is key : “when” and “but also when” introduce a more argumentative writing style focused on fitting ideas together.

What was analyzed was the direct output of each system.

To understand this, you don’t need to imagine a machine “thinking.” Simply read the text like a class essay. Some add ideas, some explain causes, and some fill the sentence with nuances. No single option is inherently better, but all of them change the way we explain and argue. The same is true with AI: it’s not just what it says that matters, but how it says it.

This explains something we already see in the classroom. Two answers can be correct and yet lead to different ways of thinking. ChatGPT tends to add up the information. Gemini is more precise. Claude argues more. It’s not just the words that change: the structure of the reasoning changes. That’s why it’s crucial to ask ourselves what way of answering is providing us with insight and what its interpretive keys are.

And that has consequences. If a student always works with the same model, or with the same prompt , they may end up delegating their way of ordering and interpreting the world to canned writing patterns with little critical judgment.

Teaching AI patterns

Therefore, teaching how to use AI shouldn’t be limited to simply asking it to do things. It should also involve teaching students how to read its patterns: how it begins, what connectors it repeats, how it justifies its arguments, and what words it uses to conclude. The goal is for today’s students to learn that artificial writing is not neutral and can ultimately impose a particular way of thinking.

Educational intervention must be didactic , not policing. If we teach students to recognize these patterns, we also teach them to read and write more critically. The study correctly identified the generating model in 91.1% of cases, although that doesn’t mean there’s an infallible trace. It means something more useful: these tools aren’t neutral when they’re used in writing. They structure discourse in a particular way.

The question isn’t whether they write well, but how they write and what might happen if we let these patterns write for us. We can’t entrust our discourse to the algorithm without sufficient knowledge to critique, judge, and modify it.

Author Bio: Esteban Vázquez-Cano is a University Professor (Faculty of Education) at UNED – National University of Distance Education