A shopping list, a calendar, a calculator: using the world so your mind doesn’t have to. Psychologists call this cognitive offloading. It is ordinary, and often wise. The question with generative AI is not whether we offload — we always have — but whether we still do the thinking the skill depends on.
Risko & Gilbert, 2016 · Trends in Cognitive Sciences
Offloading is a strategy, not a defect
Cognitive offloading is “the use of physical action to alter the information processing requirements of a task so as to reduce cognitive demand.” We choose between doing the work internally and handing it to the environment based on a metacognitive bet: how good am I at this, versus how good is the tool? When that bet is wrong — overconfidence in the model, or underconfidence in ourselves — we offload work we still needed to do, or fail to offload work the tool would have helped with. The literature’s point, and Manta’s: offloading itself is neutral. The risk is an unchecked handoff.
Bastani, Bastani, Sungu, Ge, Kabakcı & Mariman, 2025 · Proceedings of the National Academy of Sciences
Unguarded genAI can help performance and hurt learning
A field experiment with nearly a thousand Turkish high-school mathematics students gave some of them GPT-4 during practice. Two interfaces: GPT Base, which behaved like ChatGPT, and GPT Tutor, which was prompted to give hints rather than answers. With the model in the room, scores jumped. On a later exam with no AI, students who had used GPT Base scored worse than classmates who never had it. They had used the model as a crutch — asking for solutions and copying them — and they did not notice the learning loss. The Tutor’s guardrails largely erased that harm.
+48%on practice with GPT Base — a ChatGPT-like tutor
−17%on the later unaided exam, versus students who never used AI
hints, not answersGPT Tutor largely mitigated the learning loss
Kosmyna et al., 2025 · MIT Media Lab · arXiv preprint
Writing with an LLM, the brain does less of the work
In an EEG study of essay writing, people worked unaided, with a search engine, or with an LLM. Neural coupling was strongest in the unaided group and weakest in the LLM group. People who had used the model felt less ownership of the essays and recalled less of what they had “written.” In a later session, those who switched from the LLM to writing alone still looked under-engaged compared with people who had started unaided. This is a preprint with a modest sample — 54 people, 18 in the follow-up — so treat it as a signal, not a verdict. It points the same way as the classroom evidence: the tool can take the load, and the skill may not form.
Gerlich, 2025 · Societies
Frequent AI use tracks with more offloading
A survey of 666 adults found that more frequent AI use was associated with lower critical-thinking scores, and that the link ran through self-reported cognitive offloading. Younger respondents reported more dependence. This is correlational — it does not prove that AI causes weaker thinking — but it matches the mechanism the experiments describe: the more the thinking is handed off, the less it is practised.
Manta’s bet sits on the classroom result: guardrails matter. You name which skills to keep. When a prompt looks like it is handing those over, Manta asks you to stay in the work — an attempt first, a question, a pause — the same kind of friction that kept Bastani’s Tutor from becoming a crutch. Offloaded tasks go through. You set the line.
