You cannot feel it happening. Two people can use AI the same amount, in the same kind of session, and walk away with opposite effects on how they think and what they end up remembering. Research published in Frontiers in Psychology in 2026 explains why: cognitive offloading to AI splits into two distinct patterns, dependent offloading, where AI does your thinking, and autonomous offloading, where AI holds what you already thought. The first tracks with lower creativity, weaker independent judgment, and reduced intrinsic motivation. The second does not. The short version: keep the reasoning, offload the storage.
That distinction matters because most coverage of AI and memory has treated offloading as a single behavior, either a bad habit to break or a productivity trick to embrace. Neither framing survives contact with the newer research. A person who asks ChatGPT to draft an essay from scratch and pastes the result is doing something cognitively different from a person who stores a decision they already reasoned through so they can find it again next week. The tools look the same from the outside. The research says the effects are not.
The core distinction: dependent versus autonomous offloading
The Frontiers in Psychology study, led by Qiuhan Zhu and coauthors, surveyed how people use generative AI and modeled two separate pathways: dependent offloading, where the thinking happens inside the model, and autonomous offloading, where the thinking stays yours and only the storage moves to the AI. Dependent offloading means delegating core thinking to AI: accepting AI-generated answers with minimal evaluation, letting the AI structure the reasoning, and treating the output as a finished product rather than a draft. Autonomous offloading describes a different pattern: using AI output as a starting point, comparing it against your own reasoning, and folding it into thinking you still control.
The two pathways produce different downstream effects. Dependent offloading showed negative associations with four separate outcomes the study tracked: a person's later autonomous capability, creativity, depth of processing, and independent judgment. It worked through two mechanisms, what the researchers call cognitive agency transfer (handing decision-making authority to the AI) and a loss of intrinsic motivation. Autonomous offloading showed the opposite pattern: positive indirect effects on those same four outcomes, driven by preserved or increased intrinsic motivation, with no measurable agency transfer.
Here is the finding that matters more than the headline split: both modes produced comparable immediate benefits. A session leaning on dependent offloading can feel just as productive, in the moment, as one leaning on autonomous offloading. You cannot feel which kind you are doing. That is the part that should worry you.
The divergence only shows up downstream, in outcomes a person cannot observe from inside a single session. The researchers found that metacognitive monitoring, in plain terms, paying attention to whether you are still doing your own thinking, partially protected against agency transfer but did not protect against the motivational erosion. Noticing the pattern helps. It does not fully fix it.
What the critical thinking data adds
A separate 2026 study on Chinese university students, published through PubMed Central, surveyed 353 students and sorted them into four usage profiles based on how they engaged with generative AI, ranging from simple Q&A users, about a quarter of the sample, to critical co-thinkers, roughly one in six. The study's framing lines up with the dependent-versus-autonomous split: the central question it asks is not whether a student offloads routine cognitive work to AI, but whether they retain cognitive agency while doing it or gradually cede their thinking autonomy to the system.
This heterogeneity finding stands on its own. A single average correlation between AI use and critical thinking would flatten four genuinely different user profiles into one number, and the flattened number is what most casual coverage of AI and cognition reports. The actual usage pattern, not just the frequency of AI use, is what determines whether critical thinking gains or stalls.
The MIT Media Lab study everyone still cites, and what it actually measured
The most recirculated single study on this topic is MIT Media Lab's EEG research, informally known as Your Brain on ChatGPT, led by Nataliya Kosmyna and colleagues. It split 54 participants into three groups writing SAT-style essays: one using ChatGPT, one using a search engine, and one using no external tool at all, tracking brain connectivity with EEG across several sessions. The ChatGPT group recorded the weakest brain connectivity of the three groups and the lowest self-reported sense of ownership over their own essays. The recall gap was the starkest result: 83 percent of the ChatGPT group could not accurately quote a sentence from the essay they had just written, a far higher failure rate than the search-engine or no-tool groups. Reliance also compounded over the study: participants in the LLM group grew more likely to copy-paste output as sessions went on.
The study has real limits worth naming rather than skipping. The sample is small at 54 participants, the paper has circulated as a preprint rather than completing traditional peer review, and the task, essay writing under exam-like conditions, is a narrow slice of how people actually use AI day to day. Read next to the dependent-versus-autonomous research described above, though, the EEG data reads less like a verdict on AI itself and more like a detailed picture of what dependent offloading looks like inside the brain: an essay task where participants pasted an AI-generated answer with minimal evaluation is close to a textbook example of the dependent pathway the newer study defines, not a test of AI-assisted work in general.
Why adolescents are a separate case
Adolescents are still building executive function: the brain's capacity for planning, judgment, and self-control that AI offloading can substitute for. A 2026 paper in Frontiers in Developmental Psychology, by Jorge Pereira Campos and Tatiana Koff, makes the case directly: most cognitive-offloading research studies adults whose executive function had already finished developing, missing the population with the most at stake. Adolescents, roughly ages 10 to 20, need effortful practice at exactly the tasks generative AI is best at doing for them.
The authors name this the Cognitive Offloading Paradox: the tasks that benefit most from AI assistance in the short term, planning, drafting, structuring an argument, are the same tasks whose effortful repetition is thought to drive adolescent brain development in the first place. A 14-year-old who lets AI draft the outline for an essay skips the exact struggle, deciding what matters and in what order, that develops the planning circuitry an outline is supposed to exercise. Their review points to the same scaffold-versus-substitute line the adult research draws. AI that prompts a student to keep reasoning, asking questions or requesting revisions rather than handing over a finished answer, tracked with critical thinking gains. AI used to skip the reasoning step entirely was linked to what the researchers describe as metacognitive laziness and no measurable gain in what the student actually learned.
A tool built specifically to catch the difference
PAUSE, a tool described in a 2026 paper by Mahbub Ul Alam of Uppsala University, is a free, browser-based self-reflection instrument designed for AI-associated cognitive offloading. It walks a person through their own AI habits across four domains, reasoning and critical thinking, creativity and originality, research and learning, and social and communicative capacity, and returns a descriptive, non-diagnostic reflection on the pattern it finds. The tool runs entirely client-side: no account, no server, no data leaves the browser. Its design explicitly does not assume AI use is inherently harmful. It is structured to surface the substitution pattern the 2026 research flags, not to discourage AI use in general.
That PAUSE exists at all is itself a data point. A tool that helps someone check whether their own AI habits lean dependent or autonomous only makes sense once the research has drawn that line clearly enough to check against. Before this dependent-versus-autonomous research existed, the honest self-check question was simply, 'Am I using AI too much?' After it, the more useful question is narrower: when you offload to AI, are you handing over the thinking, or just the storage?
Dependent offloading versus autonomous offloading, side by side
| What you're comparing | Dependent offloading | Autonomous offloading |
|---|---|---|
| The one-line verdict | AI does your thinking | AI holds what you already thought |
| What gets outsourced | The reasoning itself: the AI drafts the argument, the decision, or the conclusion | Storage and retrieval of facts, decisions, or ideas you already reasoned through |
| How the output gets used | Accepted with minimal evaluation, often as a finished product | Treated as a starting point, compared against your own thinking |
| Effect on critical thinking (per the Frontiers in Psychology study) | Negative, via agency transfer and lower intrinsic motivation | Positive indirect effect, via preserved intrinsic motivation |
| Immediate feel in the moment | Feels just as productive as autonomous use; the divergence isn't visible in-session | Feels productive, and downstream outcomes back that feeling up |
| Practical example | Pasting an AI-written essay or decision memo without reworking it | Storing a conclusion you already reached so you can retrieve it later without re-deriving it |
What this means if you already rely on AI daily
The practical test the research suggests is not how often you use AI. It is what you are asking it to hold. If you are asking a model to store something you decided, a conclusion from a meeting, a preference you settled on, a fact you looked up and verified, so you can retrieve it later instead of re-deriving it, that is closer to the autonomous pattern the research above found no harm in. If you are asking a model to decide something for you and then treating the answer as settled without checking it against your own reasoning, that is the dependent pattern the same research ties to lower creativity and weaker independent judgment over time.
The line is not always obvious from inside a single session, which is exactly what the Frontiers in Psychology study found: both patterns feel equally productive in the moment. The check that works is retrospective. Ask, after the fact, whether the AI did your thinking or held onto something you had already thought through. Ask whether you could reconstruct the reasoning yourself if the AI output disappeared tomorrow. If the answer is no, that is a sign the offloading has drifted toward the dependent end, not the autonomous one.
Where MemX fits, and where it doesn't
MemX (memx.app) is built around the autonomous side of this distinction, not the dependent side. It stores what you choose to save: a decision you already made, a fact you already verified, a preference you already settled, so you can pull it back up in ChatGPT, Claude, and Gemini instead of re-explaining it every session. It does not draft your reasoning or hand you a conclusion to accept unchecked. The retrieval step still requires you to do something with what comes back, which is the same requirement that research found separates offloading that helps from offloading that erodes. Storage that you control and reasoning that you still do are not the same thing as letting a model think for you, and MemX is designed to do only the first one. What it stores stays private by architecture: per-user isolation, encryption at rest, and on-device processing, so the record of what you chose to remember is not a record anyone else can casually pull up.
None of this makes MemX, or any memory tool, a fix for dependent offloading. A person who outsources their thinking to a chatbot and then stores the unexamined output in a memory layer has simply made that output easier to retrieve. That does not make it trustworthy. The research points to a habit to build, not a product to install: keep the reasoning, offload the storage. Tools can make the second half easier. Only the person doing the thinking can hold onto the first half.
01Does using AI actually make your memory worse?
Not uniformly. 2026 research in Frontiers in Psychology found dependent offloading, letting AI do your reasoning, correlates with worse outcomes on creativity and judgment, while autonomous offloading, storing facts and decisions you already reasoned through, showed no such harm and even a small positive effect through preserved motivation.
02Does using AI always hurt your thinking?
No, not always. The research splits AI reliance into two patterns: dependent offloading, where you accept AI-generated answers with minimal evaluation and let AI structure your reasoning, and autonomous offloading, where you use AI as a scaffold while you still do the comparing, evaluating, and deciding. Only the first is linked to worse outcomes.
03Is the MIT 'Your Brain on ChatGPT' study still relevant?
Yes, as a detailed look at one specific case, essay writing where participants leaned on ChatGPT output with little revision. It is a small, not-yet-peer-reviewed study, and newer offloading research suggests its findings describe dependent-style use specifically, not AI assistance in general.
04Is AI worse for teenagers' brains than for adults?
Research in Frontiers in Developmental Psychology argues yes, because executive function, the brain capacity most benefited from effortful practice, is still developing through adolescence. AI that substitutes for that practice may interfere with development in a way it does not for adults whose executive function is already formed.
05How do I know if I'm too dependent on AI?
Ask whether you could reconstruct the reasoning behind an AI-assisted decision without the AI's help. If not, that leans dependent. Free, privacy-preserving tools like PAUSE offer a structured self-check across reasoning, creativity, research, and communication habits.
