Cognitive offloading: what AI and reminders mean for learning

What do you need to be able to do afterwards?

Consider a train journey. Saving the departure time is sensible if your purpose is to catch the train. If you are learning to plan journeys independently, you may also want to practise comparing routes, allowing time for changes and checking the timetable.

The information you store and the skill you practise serve different purposes. There is no special virtue in memorising the departure time if a reliable record does the job. But having a saved itinerary does not show that you can plan the next journey.

Before opening a tool, finish this sentence: “When this is over, I need to be able to…” You might need to explain a decision to a colleague, remember a concept for an exam, or simply arrive somewhere on time. That gives you a more useful basis for choosing assistance than a rule about how many times a day you should use it.

Also think about the circumstances in which you will need the knowledge. A presentation with notes and a conversation without them make different demands. A skill needed during a power cut or an emergency deserves a different backup from a restaurant recommendation.

How can you use AI without skipping the learning?

Suppose you are learning to interpret a chart. One approach is to ask for the complete interpretation and copy it into your notes. Another is to write your own attempt first: what the axes mean, the main pattern and one uncertainty. You can then ask for feedback on those points.

Afterwards, try a different chart without the explanation in front of you. Check your answer against a trustworthy teaching source or someone qualified to help. If you struggle, identify the step: reading the scale, comparing groups, or deciding whether the data justify a conclusion. “I don't understand graphs” is much harder to act on than “I confused the number of people with the percentage.”

You do not need to struggle indefinitely before seeking help. A worked example can be useful: follow the steps, identify why each is needed, then see whether you can apply the idea to another example yourself. If the explanation remains unclear, ask about the particular step rather than requesting the whole answer again.

If you keep reopening the assistant before finishing an attempt, decide in advance when you will consult it. You might finish a rough outline or try the first two steps before asking for help. If unrelated notifications repeatedly interrupt that attempt, our phone-habit guide covers ways to reduce those interruptions.

There are also tasks for which learning is not your aim. You might ask a tool to alphabetise a list or suggest headings and then check its output. Deciding to delegate such work does not oblige you to create an exercise in independent recall. Save that effort for knowledge and skills you actually need.

Can writing things down free attention for other work?

In two preregistered experiments published in 2023, researchers investigated whether saving information could help people remember something else afterwards.

Participants studied two word lists. Saving the first list helped them remember the second, but only when the second list was tested first. When they returned to the saved first list before recalling the second, the benefit disappeared. The authors suggest that seeing the first list again may have brought its words back into competition with the second list. That is a possible explanation for the effect, rather than a proven account of what happened in each person's memory.

In the second experiment, people chose which list to save. When they knew the second list would be tested first, they tended to save the first list for later. Their choices therefore favoured the arrangement that had helped memory. Saving information can help, but the order in which you need it matters.

For an everyday illustration, imagine keeping a short list beside your work. Write down the groceries or errands that occur to you while working, then return to the work. You have a place to recover those details later without making them the main task now.

What happens when the aid disappears?

A 2019 study by Kelly and Risko tested a cost of relying on external storage. Across two experiments, people completed memory-recall trials using an aid. The final trial removed it. One group had expected to keep using the aid; another had known it would be unavailable.

The group expecting the aid recalled less of the offloaded information when they suddenly had to rely on memory.

This gives you a practical question to ask about an aid: what happens if you cannot reach it? A saved address can be an adequate backup for a journey. For knowledge you need to explain or use without a device, include some practice without the saved answer. Those are different requirements, and you can plan for each.

Did AI help students perform, learn or both?

A 2025 randomised study at one Turkish high school gives a clearer example of the distinction between performance and learning. Nearly 1,000 maths students had access to a general GPT-4 assistant, a tutor designed with teacher-provided hints and safeguards, or no AI.

During assisted practice, the general-assistant and tutor groups scored about 48% and 127% higher respectively than the control group. On a subsequent assessment without assistance, the general-assistant group scored about 17% lower. These are relative differences in grades, not percentage-point changes. The tutor safeguards largely removed the negative effect, without showing a clear improvement over the control group on the unaided assessment.

The tools helped produce answers during practice, but their designs left different amounts of learning behind. This was a specific school, subject and set of tools. It supports asking what someone can do after the assistant closes, rather than judging learning from the assisted work alone.

You can see why that matters when helping a child with homework or learning a skill yourself. A correct solution is useful feedback. If it arrives before you have tried the problem, however, you still need some way to find out which steps you understand. Re-reading the finished answer cannot answer that question on its own.

Does making your own outline first make a difference?

A September 2026 study assigned 900 Chinese-speaking students and early-career workers to three writing workflows: make an outline before using AI, adopt AI output directly, or write without AI. The main analysis included 659 people, averaging 24.6 years old, and used Qwen3.5 for two policy-essay topics.

On a new essay without assistance 15–20 minutes later, direct adopters scored 0.38 points lower than the no-AI group on a seven-point reasoning scale. The outline-first group scored 0.21 points higher than the no-AI group. That small advantage did not remain statistically convincing after accounting for the several outcomes tested. The authors interpreted it as maintaining performance, rather than demonstrating an improvement.

The study was exploratory: it was not formally preregistered, and some analytical choices were made after seeing the data. The groups also differed in effort, revision and time spent working. Exclusions included incomplete tasks, failed attention checks and suspected AI use in the no-AI group; broader analyses gave similar results for the direct-adoption disadvantage. These were immediate writing outcomes in existing AI users. In this short test, making an outline first avoided the drop seen with direct adoption of AI text; it was not shown to make people better writers than working without AI.

Do the brain studies show damage from AI use?

The 2025 Your Brain on ChatGPT preprint involved 54 people across three essay-writing sessions, with 18 completing a fourth. Participants used an AI assistant, a search engine or no tool. The AI group reported less ownership of their essays and struggled more to quote from what they had written. The researchers also compared patterns in recorded brain activity: their connectivity measure described relationships between the recorded signals, rather than damage to brain tissue.

EEG records the brain's electrical activity through sensors on the scalp. The MIT team's own explanation stresses that this preliminary work, still a preprint, involved a small, geographically restricted sample. It cannot establish brain damage or dementia risk. The useful question it raises is how people engage with a particular writing task when different tools are available.

Another much-discussed 2025 study by Michael Gerlich used surveys and interviews with 666 participants. More frequent AI use was associated with lower critical-thinking scores. Younger participants also showed greater reliance on AI and lower scores than older participants.

Because this compared people's existing habits, it leaves open why the groups differed. People who use AI more may differ in their education, circumstances or approach to tasks. The association alone cannot show that using a chatbot caused someone's thinking to deteriorate.

Check whether the answer is right as well as understandable

You can understand an explanation that is wrong. You can also hold a correct explanation without understanding it. These are separate failure points.

If an AI answer cites research, open the paper. Check who was studied, what they were compared with and what the researchers measured. A study of immediate essay scores cannot answer a question about decades of memory change. If the paper is inaccessible, treat the claim as unverified rather than assuming that a convincing citation settles it.

For numerical work, check the inputs and units as well as the result. Ask whether a percentage describes a relative change or a change in percentage points. A large-looking number may be mathematically accurate but still describe a narrow task.

For health questions, keep the original clinical advice available. An assistant may help organise questions for an appointment, but generated text is not a sound basis for changing a prescribed dose. If you cannot assess a recommendation safely, take the question to the clinician or pharmacist responsible for your care.

Keep reminders that help you manage daily life

Cambridge University Hospitals' memory guidance includes calendars, lists, timers, alarms and voice notes. A debate about AI is not a reason to abandon a memory aid that makes daily life manageable, especially one agreed with a professional.

For an important task, keep the information somewhere you can use without reopening the chatbot. Save the original instructions or a checked document in an accessible format, and know where it is. If an assistant summarises a clinician's instructions, keep the original alongside the summary so you can check a detail or resolve a discrepancy.

Our everyday memory guide explains how to take in new information, practise recalling names and instructions, and choose what to write down.

If memory problems are affecting your day-to-day life, the NHS advises seeing a GP. Do not assume a change is caused by AI use or try to diagnose it by taking away your usual supports. Bring examples of what has become difficult; someone who knows you may be able to help describe the changes.