For centuries, learning something new in depth required access to books, experts, or educational institutions. The democratisation of access to information that the internet brought was an enormous shift — but it also introduced the opposite problem: too much information, poorly organised, of variable quality, and with no mechanism for distinguishing the essential from the peripheral.
Language models have introduced a new possibility: an interlocutor who can explain complex concepts, connect ideas from different fields, answer follow-up questions, and adapt the level of depth to what is needed. For someone who wants to learn about a new topic, this is extraordinarily useful. But only if used with judgement.
How learning something new has changed
The traditional approach to an unfamiliar topic usually follows a pattern of search and selection: find sources, evaluate their quality, read the most promising materials, take notes, and try to build your own understanding. It is a slow but effective process, because the friction involved forces you to actively process information.
Internet search accelerated the initial location phase but did not improve the comprehension phase. A Wikipedia article, a blog post, and an academic publication on the same topic may require very different reading skills and prior context — and navigating between them requires already knowing something about the territory.
What language models offer is something qualitatively different: the ability to ask questions in natural language, receive responses calibrated to the asker’s level, and continue an exploratory conversation in which each answer generates new questions. It is closer to talking to someone who knows a great deal about the subject than to searching in a library.
This changes the learning process. But it does not simplify it as much as it might seem.
What AI does well when you research
Language models are particularly useful at several stages of the learning process.
For getting an initial overview, they are almost unmatched. If you ask a model to explain the fundamental concepts of behavioural economics for someone with no economics background, it will do so with a clarity and adaptability that is hard to find in any single written text. It can adjust the level of detail, use analogies, respond to “I didn’t understand that part,” and reformulate without impatience.
For identifying what you don’t know, AI is also very useful. Asking for a map of the key concepts in a field, the main schools of thought, or the current debates in a discipline helps you build a mental schema of the terrain before going deeper. Knowing which terms to search for, which authors matter, and where the controversies lie is an enormous advantage when starting from zero.
For translating technical language, its value is also significant. A statistical concept, a mathematical derivation, a legal term, or a philosophical definition can each be explained in multiple ways depending on the context of the person asking. The ability to reformulate in accessible terms without losing the essence is one of these models’ strongest suits.
What AI does poorly (and how to compensate)
The central problem is the excessive confidence that the format of responses can induce. Language models generate text with a fluency and coherence that closely resembles that of someone who genuinely knows what they are talking about. But that surface coherence does not guarantee factual accuracy.
“Hallucinations” — plausible but incorrect responses — are a known and persistent phenomenon in these systems. A model can cite invented statistics with the same confidence with which it cites real ones. It can attribute ideas to incorrect authors. It can simplify to the point of distortion. It does not do this out of bad faith, but because its functional objective is to generate coherent text — not to verify every claim before producing it.
The second problem is currency. Most models have a knowledge cut-off date and cannot access information published after that point. For topics where recent developments are relevant — scientific research, economic policy, technology — this is a significant limitation.
The compensation for both problems is active verification. Treating the model’s responses as a starting point, not as a definitive source. Confirming important data in primary sources. Finding the original paper when the model cites a study. Cross-checking central claims with independent sources.
A concrete workflow for researching with AI
One way to integrate AI into the research process without sacrificing rigour is to use it in the orientation and synthesis phases, reserving verification for primary sources.
The first step is to use AI to get a map of the territory. Ask for a general introduction to the topic, the fundamental concepts, the most influential authors or works, and the main open questions. Not to learn definitively, but to know where to look.
The second step is to read quality primary or secondary sources on the aspects that matter most. Here AI acts as support: if a paragraph is unclear, ask the model to explain it. If a concept seems to connect to another, explore that connection in conversation.
The third step is to use AI to synthesise and organise what you have learned. Tell it what you have read and ask it to identify gaps, inconsistencies, or connections you may have missed. This function as a critical interlocutor is one of the most valuable, as long as the raw material of the conversation is your own knowledge — not just what the model itself generated.
AI does not replace your own judgement
The most sophisticated trap in using AI to learn is delegating not only the search but also the evaluation. If the model says something is important, we assume it is. If the model simplifies, we accept the simplification. If the model has an implicit perspective on a debate, we absorb it without noticing.
Critical thinking — the ability to evaluate the quality of an argument, identify its assumptions, compare it with alternative perspectives — cannot be outsourced. It is precisely what language models cannot do for us, because they have no interests, no embodied experience, and nothing at stake when generating a response.
What AI can do is accelerate the information-gathering phase and reduce the initial friction of learning something new. What it cannot do is replace the work of thinking about that information, contextualising it in one’s own experience, and arriving at one’s own conclusions. That part remains irreducibly human.
