Encyclopaedia, Google, ChatGPT: do we learn differently?

In spring 2023, in Germany, 91 students were randomly assigned to find information with Google or with ChatGPT on the nanoparticles in sunscreens, then to write, without notes, a reasoned recommendation. The students started with comparable knowledge. Those who had used ChatGPT reported a lower cognitive load, across its 3 components. Their recommendations contained fewer relevant arguments: 1.20 on average, against 1.87 with Google, out of 7 possible.

In 2025, 2 American researchers repeated the comparison on a larger scale: 7 experiments, more than 10,000 participants, asked to find information on a practical subject, then to write advice for a friend. In the 1st, those who had searched with ChatGPT agreed, at 3.43 out of 5, with "I learned new things about the topic", against 3.86 with Google, and their advice was 85 words long, against 95.

The result holds when the information is the same. In a 2nd experiment, participants saw exactly the same facts, presented either as a list of links or as a written answer: with the written answer, the advice remained shorter, contained fewer facts and was more alike from one participant to another.

These experiments measure a search session, not a school year. The learning is partly self-reported by participants; the quality of the texts, for its part, is rated by others. The German study covers 91 students and a version of ChatGPT from 2023. In the 3rd experiment by Melumad and Yun, carried out in a laboratory at a university on the US East Coast with Google's AI summaries, the advice is shorter, and the gap in self-reported learning is no longer significant, although participants still judge their knowledge less complete. Readers who did not know where the advice came from judged the advice derived from the AI summary less useful.

They also do not say that a search engine teaches better than a textbook. They compare 2 ways of searching during the same session.

For the search engine, the most cited studies concern memory and the idea people have of their own knowledge. The 2011 study published in Science, on the memory of what one thinks can be found online, and the replication of its experiments are read in Is AI making us stupid?.

In 2015, 9 experiments carried out at Yale University showed that after searching for explanations online, people consider themselves more knowledgeable: they confuse access to information with their own understanding. In 2017, 3 experiments showed that those who used Google to answer difficult questions then turn to it more often, even for easy questions, than those who had first answered from memory.

None of the studies found measures these 2 effects with an AI assistant, or over several months.

In autumn 2025, 57% of American teenagers also said they used an assistant to look for information. The last 2 measurements do not cover exactly the same question: ChatGPT alone in 2023 and 2024, all assistants together in 2025. For France, we have not found a public survey with the same method.

The search engine is itself moving closer to the assistant. In March 2025, among 900 American adults whose browsing was tracked, a click on a result followed 8% of visits to a Google page displaying an AI summary, against 15% without a summary.