Thinking with AI: what to hand over and what to keep practising

Offloading changes where the work happens

A shopping list lets you stop rehearsing groceries while you plan dinner, and a calendar remembers appointments while you concentrate elsewhere. Risko and Gilbert (2016) call this cognitive offloading: using actions and things outside ourselves to reduce a task's mental demands. Their review covers what pushes people to offload and how people judge whether they need help. Those judgements can be mistaken, so a feeling of being unable to manage a task is an imperfect guide to when assistance is useful.

With an AI assistant, the choice includes handing over some of the reasoning itself. It can be reasonable to request a draft when the immediate purpose is finishing routine correspondence, yet writing that draft yourself provides practice in organising an argument. The difference depends on your purpose today. Choose a particular skill to preserve, such as checking a numerical claim or explaining a concept, and reserve some opportunities to perform it before receiving an answer.

An attempt creates something to learn from

Producing an answer can leave a different memory from reading one. In their experiments on the generation effect, Slamecka and Graf (1978) compared memory for words that participants generated from cues with memory for words presented for reading. Generating the words helped later memory under the conditions they tested. A cue such as a word and the initial letter of its synonym required participants to contribute something to the answer, making this a narrower task than writing an essay with an assistant.

Even an unsuccessful attempt can sometimes prepare later learning. Kornell, Hays and Bjork (2009) found that attempting answers before seeing them improved subsequent learning with fictional general knowledge questions and word associations. Their materials were chosen so that attempts would fail, and they excluded the rare correct guesses in the word association experiments. The findings support a pretesting effect in those tasks. They do not establish that an uninformed guess helps every kind of learning, or that correction can be omitted.

A practical use follows from this evidence, although applying it to AI is an inference. Before requesting an explanation of tides, write what you think causes them and identify the part you cannot explain. Once you consult a source, you have specific claims to inspect. Replace an error with the corrected account and explain why the correction changes your original answer.

Useful difficulty depends on the learner

A clear explanation can feel familiar while leaving you unable to reconstruct its reasoning later. Bjork and Bjork (2011) describe desirable difficulties, including spaced practice and retrieval, that can make practice less fluent while supporting later retention. They also distinguish learning from the performance visible during instruction. Following a solution while it remains on screen measures something different from solving a related problem after the screen has closed.

Difficulty becomes useful when you can engage with the task and learn from what follows. The same authors explain that lacking the necessary background knowledge can make a difficulty undesirable. If a mathematical explanation uses unfamiliar symbols throughout, begin with a worked explanation and a simpler example. After understanding that example, cover its final step and supply the missing reasoning yourself before checking it.

Automation can leave specific skills unpractised

Automation introduces another problem besides missed practice: people must decide how carefully to inspect its output. Parasuraman and Manzey's (2010) review of automation complacency and bias describes errors associated with imperfect decision aids. An operator might miss a problem the system failed to flag, or follow a recommendation despite contradictory information. Attention matters because checking competes with other demands, and having expertise does not remove every source of automation bias.

Evidence about skill fade also needs its boundaries. In a simulator study with a sample of 16 airline pilots, Casner, Geven, Recker and Schooler (2014) found instrument scanning and aircraft control mostly intact, while cognitive tasks involved in manual flight showed more difficulties, including tracking position and deciding the next navigational steps. Performance on these cognitive tasks was associated with how often pilots reported task-unrelated thought while the automation was in use. The study did not randomly assign years of automated flying, so it cannot establish that automation caused every observed difficulty.

These findings concern particular tasks and systems rather than a general decline caused by using AI. For ordinary work, the practical concern is whether you can still perform a needed step when assistance is unavailable. Check a calculation independently, explain the reasoning behind a recommendation or practise writing an opening paragraph before asking for revisions.

The Google effect has a disputed evidence base

Sparrow, Liu and Wegner (2011) reported experiments on expectations of future access to information. In some of their experiments, participants who expected information to remain available remembered less of its content, while another experiment examined how well people remembered where information was stored. The authors read this as evidence that anticipated access can change what people remember. Knowing where a record lives can be useful, although that knowledge will not supply the record's contents during a conversation without access.

Part of this evidence has faced replication concerns. Camerer and colleagues (2018) did not reproduce the original computer word priming result under their replication criteria. That experiment concerned whether difficult questions made computer related words more mentally accessible, rather than directly repeating the saved information memory tasks. Sparrow (2018) disputed aspects of the replication procedure. This leaves a contested finding within a wider set of claims, and neither paper measures the long term effects of present day AI assistants on an individual's thinking.

Try an attempt, a consultation and a comparison

Begin with a task small enough to attempt from what you know. For an argument, write a claim and the evidence you would need; for an estimate, record your assumptions and a plausible range. Preserve that attempt before consulting a tool. Otherwise, the answer you see can become so familiar that you lose track of what you could produce independently.

During consultation, ask for help with the gap you identified and verify factual claims against an appropriate source, then compare the answer with your attempt by naming a change you now understand. You might discover that your estimate omitted a category of costs, or that your argument relied on evidence about a different population. Close the explanation and reconstruct the corrected reasoning, returning to it later if you want to retain it.

If you want a separate place to practise answering before receiving feedback, Funga Wega offers short sessions of puzzles, comprehension and reasoning questions with review of earlier misses. Nothing in those sessions is generated or graded by AI. Choose a skill from your own work too, and try a fresh example after a delay to see what remains available without assistance.