AI Dependency Is Deskilling Humans, Not Just Eroding Jobs

In 2021, a network of four endoscopy centers in Poland started using AI to help spot precancerous growths during colonoscopies. Researchers later compared the same highly experienced doctors’ performance on procedures done without AI assistance, before and after they’d gotten used to the tool. Unassisted detection of precancerous growths dropped from 28.4% to 22.4%. Real doctors, each with more than 2,000 procedures behind them, got measurably worse at a task they’d done for years, once a tool made it optional. The study, published in The Lancet Gastroenterology & Hepatology in 2025, is one of the first pieces of real-world clinical evidence for what researchers call deskilling, and it moved the conversation from theory to documented outcome (STAT News, EurekAlert). It’s an observational study, not a randomized trial, and the authors are upfront that other factors could be involved, but it’s the first real-world clinical data point in a pattern researchers are now seeing across fields.

That’s the part of the AI dependency conversation that gets buried under job-loss headlines. The real question isn’t only “will AI take this job.” It’s quieter and applies to nearly everyone reading this: what happens to a skill once a tool makes it optional, starting with the people who never got the chance to build it in the first place.

This Fear Is Old. The Evidence Is New.

Every tool that removes cognitive effort has triggered this same argument. Socrates warned in Plato’s Phaedrus that writing would destroy memory, because people would rely on marks on a page instead of recall. Calculators were supposed to end mathematical reasoning. Wikipedia was supposed to end research skill. None of those fears fully landed. Calculators didn’t kill mathematical thinking; math education shifted from arithmetic drilling toward conceptual understanding.

But some versions of this fear were correct. A 2020 study published in Scientific Reports followed regular drivers and found that heavier lifetime GPS use predicted measurably worse spatial memory on independent navigation tasks, replicated again in 2024 (Nature Scientific Reports). Self-guided navigation actively engages the hippocampus, building a mental map and weighing alternatives. Following turn-by-turn directions along the identical route doesn’t.

So “offloading a task to a tool erodes the skill behind it” isn’t universally true. It’s true in some domains and false in others, and the honest answer for AI is that the evidence so far leans toward the GPS outcome, not the calculator one. A widely cited 2025 study out of SBS Swiss Business School, surveying 666 people across age groups, found a negative correlation between frequent AI tool use and critical thinking performance, driven largely by cognitive offloading. Younger users (17 to 25) showed both the heaviest AI dependence and the lowest critical thinking scores, while higher education levels acted as a partial buffer (phys.org).

Four Different Problems Wearing One Name

“AI dependency” gets used as one phenomenon. It’s actually at least four, and they don’t share a cause or a fix.

Skills never built. This is the developer who ships working code without being able to explain or debug it, because AI wrote it before they ever wrote it themselves. A 2026 survey of senior engineers found only 16% believe junior developers fully understand the AI-generated code they submit, 57% said juniors understand it “to some extent,” and 23% said rarely (BairesDev). Separately, Stanford’s Digital Economy Lab found that, as of mid-2025, employment for software developers aged 22 to 25 was down nearly 20% from its late-2022 peak, a signal that the entry-level jobs that used to be where this skill got built are shrinking at the same time the incentive to build it is disappearing.

Skills left to rot. This is the colonoscopy example. It’s also the writer who stops drafting from a blank page. A preliminary MIT Media Lab study, still circulating as a preprint since mid-2025, used EEG to compare people writing essays unaided, with a search engine, or with ChatGPT. Brain-only participants showed the strongest, most distributed neural connectivity while writing. ChatGPT users showed the weakest. Eighty-three percent of the ChatGPT group couldn’t quote a single sentence from the essay they’d just submitted, and self-reported ownership of the work was lowest in that group (arXiv preprint). Because it’s a preprint, not yet published in a peer-reviewed journal, treat the specific numbers as preliminary, but the direction lines up with everything else here.

Original thought getting priced out. This one isn’t really a skill problem; it’s an incentive problem AI has accelerated rather than invented. Content mills existed before generative AI; SEO farms optimized for volume over originality for a decade before ChatGPT. What’s changed is the cost of producing filler dropped to nearly zero, so the economic pressure against slowing down for an original idea got sharper. The fix here isn’t “use less AI”; it’s realigning what gets rewarded, which is a publishing and platform problem more than a personal-skill one.

School without the struggle. Between May and December 2025 alone, the share of middle school, high school, and college students using AI for homework rose from 48% to 62%, according to nationally representative RAND survey data. Growth was steepest among the youngest users; middle schoolers went from 30% to 46% in the same seven months. By 2026, one survey put overall student AI use for homework at 84%, with only three in ten schools having any policy governing it (RAND, Fortune). Notably, 60% of students using it still say they’re worried about what it’s doing to their thinking. This is a pedagogy design question before it’s an AI question: is the tool being used as a tutor that forces retrieval, or as an answer machine that skips it?

The Real Mechanism Isn’t AI. It’s Missing Struggle.

Learning science has a name for what all four of these share: desirable difficulty. Skills get encoded through retrieval, error correction, and effort, not through exposure to a correct answer. Any tool that supplies the correct answer before the struggle happens prevents the skill from forming, whether that tool is a copied homework answer, a calculator used before arithmetic is understood, or a language model.

What’s different about AI is scope. A calculator replaced one mechanical subtask. AI can replace the judgment and synthesis layer itself, in writing, coding, diagnosis, translation, and analysis simultaneously, which is why the deskilling signal is showing up across so many unrelated fields at once. In a global survey of 70 C-suite and senior executives, half already report observing deskilling inside their own organizations, and more than 60% expect it to become a material threat within three to five years (BCG). This is a scale and sequencing problem, not a light switch. The professional who spent a decade building judgment before AI arrived and now uses it to move faster is in a fundamentally different position than the one who never built that judgment at all. Same tool, opposite outcome, depending on what happened first.

Translation Already Ran This Experiment

This is where it stops being abstract for SEAtongue. Machine translation post-editing (MTPE) has become the dominant mode of production in professional translation, and among professionals using MT or LLMs in their workflow, 90 to 98% still perform some level of post-editing rather than shipping raw output, per RWS Group’s 2025 industry survey of nearly 2,000 language professionals (Slator). Human review hasn’t disappeared. But the quality of that review depends entirely on whether the human doing it still has the judgment to catch what the machine missed, and that judgment is exactly what’s at risk.

Research on post-editing performance shows it genuinely engages a post-editor’s cognition; it isn’t passive rubber-stamping; eye-tracking studies confirm real mental effort goes into it. The actual risk isn’t that post-editing is mindless. It’s that a translator who only ever post-edits, and never builds from-scratch translation competency, loses the reference point needed to judge whether the machine’s output is actually good. Translation training programs haven’t caught up either. Whether post-editing gets taught as its own discipline, separate from translation itself, is still largely left to individual instructors rather than built into curricula (ScienceDirect). Some in the field describe the working reality bluntly as “post-edit or perish.”

The economic version of the same effect shows up outside the industry too. A CEPR analysis found that regions with heavier Google Translate adoption saw job postings requiring foreign-language fluency grow measurably slower than comparable regions, roughly 1.4 percentage points less for Spanish-language postings (CEPR). And in a separate Slator survey of 260 freelance linguists, more than half said they’d thought about changing careers, citing falling income alongside a sense of lost competence, not just lost work (Slator).

None of this argues for avoiding MT. It argues for treating post-editing as a skill that has to be deliberately built and maintained, not assumed to come free with the tool. That’s the same distinction SEAtongue applies operationally: combining MT with dedicated post-editing expertise rather than treating raw output as a finished product, and designing localization workflows where human judgment sits at the checkpoints that actually matter.

What Actually Works

None of this is an argument to avoid AI. It’s an argument against sequencing it before the fundamentals exist. A few things the evidence above actually supports:

Build the skill before you automate it. The junior developer, translator, or writer who spends time producing unassisted work before leaning on AI has something to check the output against. The one who skips straight to AI output has no way to know when it’s wrong.

Treat AI output as a draft to interrogate, not a finished answer. The MIT data suggests the ownership and retention gap opens up specifically when people accept output rather than actively working with it.

Build review into the process deliberately, the way translation already had to. Post-editing works when it’s treated as its own discipline with its own training, not as an afterthought bolted onto a faster workflow.

The people and organizations who’ll matter in five years won’t be the ones who avoided AI. They’ll be the ones who kept building the judgment AI can’t fake, and used the tool to move faster once that judgment already existed.