Learning from AI summaries leads to shallower knowledge than web search

Results of a set of experiments found that individuals learning about a topic from large language model summaries develop shallower knowledge compared to when they learn through standard web search. Individuals who learned from large language models felt less invested in forming their advice, and created advice that was sparser and less original compared to advice based on learning through web search. The research was published in PNAS Nexus.

Large language models (LLMs) are artificial intelligence systems designed to interpret and generate human language by learning statistical patterns from vast collections of text. They are typically based on deep learning architectures, which allow them to process context and relationships between words over long passages. The most popular large language models today include those developed by OpenAI (GPT series used in ChatGPT), Google (Gemini), Anthropic (Claude), and Meta (LLaMA).

The development of large language models has progressed rapidly over the last decade due to advances in computing power, the availability of large datasets, and improvements in training algorithms. Early models focused mainly on simple text prediction, while modern models can perform complex reasoning, summarization, translation, and dialogue. Training usually involves two main stages: large-scale pretraining on general text and fine-tuning on more specific tasks or with human feedback.

These models are widely used in applications such as chatbots, virtual assistants, search engines, and automated customer support. In education and research, they assist with writing, coding, literature reviews, and data exploration. In business and industry, they are used for document analysis, marketing content generation, and decision support. Despite their usefulness, large language models sometimes produce errors, biases, or misleading information because they do not truly understand the world but rely on patterns learned from the materials used for their training.

Study authors Shiri Melumad and Jin Ho Yun note that many people use summaries of various materials generated by LLMs as learning tools. However, when learning from LLM summaries, users no longer need to exert the effort of gathering and distilling different informational sources on their own. The study authors hypothesized that this lower effort in assembling knowledge from LLM summaries might suppress the depth of knowledge users gain compared to learning through traditional web search, resulting in shallower knowledge. In turn, this shallower knowledge would result in less investment in giving advice based on that knowledge, and in sparser and less unique advice content. Such advice would then be seen as less informative and persuasive.

Free daily newsletter

The study authors conducted a series of experiments to verify elements of their model. The first experiment involved 1,104 participants recruited via Prolific. They were told to imagine that a friend was seeking advice on how to plant a vegetable garden. One group of participants had to learn about this through Google search, while the other learned from ChatGPT. They would then give advice.

The second experiment involved 1,979 participants recruited via Prolific. It was the same as the first experiment, but the participants were limited to typing just one query. The query did not result in a typical search or response generation. Instead, participants were all given the same results formulated either as a series of linked websites or a summary of ChatGPT-style suggestions.

The third experiment was similar to experiment one, but the two groups of participants either used Google search or Google’s “AI Overview” (and not ChatGPT). They were to give advice about leading a healthier lifestyle. In this way, the platform was held constant. Participants in the fourth experiment rated various characteristics of the advice produced in the third study.

Results of these experiments showed that participants who used LLM summaries spent less time learning and reported learning fewer new things. They invested less thought and spent less time writing their advice. As a result, they felt lower ownership of the advice they produced. Overall, this supported the idea that learning from LLM summaries results in shallower learning and lower investment in acquiring knowledge and using it.

Participants learning from web searches and websites produced richer advice with more original content. Their advice texts were longer, more dissimilar to each other, and more semantically unique.

“A theory is proposed that because LLM summaries lessen the need to discover and synthesize information from original sources—steps essential for deep learning—users may develop shallower knowledge compared with learning from web links. When subsequently forming advice on the topic, this manifests in advice that is sparser, less original—and less likely to be adopted by recipients. Results from seven experiments support these predictions, showing that these differences arise even when LLM summaries are augmented by real-time web links, for example. Hence, learning from LLM syntheses (vs. web links) can, at times, limit the development of deeper, more original knowledge,” the study authors concluded.

The study contributes to the scientific understanding of how people learn using LLMs. However, it should be noted that the initial experiments involved hypothetical scenarios (advising a friend), though later experiments confirmed the results held even when the topics were of high personal relevance to the participants.

Additionally, the experiments involved paid participants—individuals who were likely primarily motivated by the award for participation, which was not dependent on the quality of the advice they produced. Results of studies looking into real-world learning situations where participants feel responsible for the outcomes of learning and have a personal stake in the quality of advice they produce may differ.

The paper, “Experimental evidence of the effects of large language models versus web search on depth of learning,” was authored by Shiri Melumad and Jin Ho Yun.