Skip to content
Artificial intelligence · Anthropic · Stanford UniversityEpisode 158 · 1 September 2026 · 19:20

Will We Soon Be Unable to Work Without ChatGPT? What AI Is Doing to Our Skills

Central question

What happens to a person's professional competence when AI takes over the practice that competence usually grows out of?

What you take away

The reader gets four measurable signals — a nineteen-percent drop in employment among young people in AI-exposed occupations, students scoring more than twenty percent lower without AI in a supervised exam, medical students halving their accuracy after a convincing but wrong model explanation, and sixty-nine percent of employers who now read a grammatical slip in a CV as a plus — plus a working formula of double competence instead of an argument for or against AI.

Main threads

What to watch for

1Take a task you used to do yourself and do it without a model: the difference in time and in confidence is your current skill drop.
2Separate two modes in your work and name them out loud: where a model amplifies a skill you already have, and where it replaces the practice that keeps that skill alive.
3Keep tasks you have to stay inside for a long time: depth grows out of time in the task, not out of the number of tasks closed.
4When checking a model's answer, start from “what if this is wrong” — a confident wording is not evidence of correctness; it is exactly what breaks your calibration.
5If you are hiring, ask not only whether the person can work with AI but what they can do without it — these are two different qualifications.
Signals to track afterwards
Whether employment among young specialists in AI-exposed occupations recovers or the nineteen-percent gap settles in.
Whether studies appear on working professionals rather than students: so far it is mostly learning that gets measured.
Whether a confidence calibration failure becomes a working concept beyond medicine — anywhere a model's error is equally hard to feel.
Whether other companies follow Anthropic in testing the two qualifications separately at the door.
Whether lines like “can work without AI” appear in CVs, and whether they come to mean anything to employers.
Most useful for
Specialists whose way of working changed markedly over the past two years and who want to know what changed in themselves.People entering a profession now, and those responsible for training them.Managers and hiring leads who need to know which skill they are buying.Doctors, lawyers and anyone whose mistake is expensive — on confidence calibration.Anyone who has already asked whether AI made them a stronger specialist or merely a faster one.

Key takeaways

02:37No Mass Layoffs, but the Door for Young People Narrowed

A Stanford University lab study across millions of American workers, with data through June 2026, shows no mass layoffs caused by AI. But employment among people aged 22–25 in occupations more exposed to AI is nineteen percent lower than among peers in occupations where AI is used less.

03:41An Expert Grows Out of Routine, and the Ladder Is Being Removed

Nobody is born an expert: depth accumulates from a large number of routine tasks, mistakes, repetitions and feedback. If there are fewer entry-level tasks for juniors, the training ladder disappears — and the market arrives at a state where seniors are still needed and there are ever fewer people who managed to become them.

04:52A Personal Example: Dictating Letters Instead of Writing Them

The author describes her own drop. Writing quickly and confidently in text was a well-developed skill of hers; now, under heavy load, she dictates letters and a model packages them in her logic, her words and her meanings. The speed is different — and so is the skill. Writing letters herself has become noticeably slower. The author immediately adds an honest caveat: perhaps it is not the skill that weakened but the willingness to spend time on a task that can be done much faster.

06:26What Is Lost Is Not the Ability but the Readiness to Stay in the Task

The hypothesis the author arrives at: what is lost is not the ability but the readiness to stay inside a hard task for a long time. It is exactly that time — writing something at length, composing at length, thinking one more time — that forms deep professional skill.

09:16With Students: Twenty Percent Lower Without AI in the Exam

Studies of learning among American and Chinese students and schoolchildren differ in design but agree on the main point. With AI a group solves faster and better. In a supervised exam without it, the same group scores twenty percent or more lower — the knowledge that stayed with the person turns out to be markedly less.

09:47The System Grows, the Human Capital Inside It Does Not

The episode's key formulation: the productive human-plus-AI system grows, and grows fast, but the human capital inside that system may be shrinking — or at any rate is not growing as fast as we expect.

10:59The Medical-Student Study: Accuracy Falls by Half

A 2026 study found on the medical portal PubMed: medical students were split into three equal groups and each given twenty-five hard diagnostic questions in the style of the American medical exam. The first group saw only the question, the second the question plus a correct model explanation verified by specialists, the third a deliberately convincing but incorrect explanation. In the third group the probability of a correct answer fell by more than half.

12:14Confidence Calibration Failure

The term the author came across in connection with this study. In plain words: a person stops sensing a wrong answer as less reliable and starts relying entirely on what the system produces. Such things, the author notes, are better tested in a learning setting than where the question is already one of life and death.

13:14The Most Dangerous Specialist Is Not the One Who Does Not Know

The formulation the whole thread about doubt converges on: the most dangerous specialist is not the one who does not know, but the one whose confidence has stopped telling knowledge from error.

13:39Signal One: Companies Are Looking for AI-first Specialists

On one side, employers are looking for AI-first and AI-enabled people — those who apply models competently in their work, get results fast, pass that knowledge on and build new processes around the tool.

14:01Signal Two: a Grammatical Error as a Sign of a Person

On the other side, employers find it ever harder to tell which skill they are buying. In a survey of two hundred and fifty British employers, sixty-nine percent said they would rather see even a grammatical error in a CV or application: it lets them think a live person wrote the text rather than a model writing all of it.

15:50Anthropic Tests Two Qualifications at the Door

Anthropic is the company that built Claude, and Claude is used there every day; it does look for candidates who work competently with models. At the same time, a candidate prepares the CV and the answers themselves, takes the test task without models unless otherwise agreed, and gets no AI help during the live interview. One of the most AI-native companies effectively tests two different qualifications.

18:06The Main Asset Is Double Competence

The author's conclusion: the main professional asset is not AI literacy in itself but double competence — the ability to amplify yourself with a model while keeping your own depth and practice, so that there is something to answer for the result with.

What this episode is about

A solo episode by Anna Volchek, continuing the conversation about saved time. Last time the question was who gets the time AI saves; here it is turned around: did we borrow that time from our own competence? The prompt for the topic came from the author's son after the previous podcast — could it be that with AI none of us are getting smarter?

The first block is about entering a profession. A Stanford University lab study across millions of American workers, with data through June 2026, shows no mass layoffs caused by AI. But employment among people aged 22–25 in occupations more exposed to AI came out nineteen percent lower than among their peers in occupations where AI is used less. The author connects this to what the channel has said about programmers: there are fewer entry-level tasks for juniors, and an expert grows precisely out of routine, mistakes and repetition. Remove that training ladder and you get a market where seniors are needed and there is nowhere for them to come from.

The second block is personal and therefore precise. The author describes how she stopped writing letters by hand: she dictates, and a model packages the text in her logic and her words. The speed is different — so is the skill. An honest caveat follows: perhaps it is not the skill that weakened but the willingness to spend time on a task that can be done faster. She found no research on exactly that, but did find that dictation and typing distribute planning, pauses and formulation differently. Hence her hypothesis: what is lost is not the ability but the readiness to stay inside a hard task for a long time — and that time is what makes a skill deep.

The third block is the numbers. Studies of learning among American and Chinese students and schoolchildren differ in design but agree on one thing: with AI a group solves faster and better, and in a supervised exam without AI the same group scores twenty percent or more lower. The human-plus-AI system grows; the human capital inside it does not. A 2026 study on PubMed goes further: medical students were split into three groups and given twenty-five diagnostic questions; among those who received a convincing but incorrect model explanation, the probability of a correct answer fell by more than half. The term the author came across is a confidence calibration failure: a person stops sensing a wrong answer as less reliable.

The episode closes on two labour-market signals pointing in opposite directions. Companies are looking for AI-first and AI-enabled specialists — and at the same time understand less and less which skill they are buying: of two hundred and fifty British employers surveyed, sixty-nine percent would rather see a grammatical error in a CV as a sign of a live person. Anthropic, one of the most AI-native companies, where Claude is used every day, nonetheless requires candidates to prepare their CV and take the test task and interview without models unless explicitly permitted otherwise. That is, it tests both qualifications at the door.

The episode's conclusion is neither that AI harms nor that it saves. It is that the productivity of the human-plus-AI system and the competence of the human inside it are two different quantities, and they grow at different speeds. Hence the main asset the author names: not AI literacy in itself but double competence — the ability to amplify yourself with a model while keeping your own depth and practice, so that there is something to answer for the result with. This is not a call to return to a world without AI but a demand to manage the process while there is still a person inside it doing the managing.

Episode transcript

The episode is in Russian; below is an English reading guide to the transcript (the full EN transcript is a machine translation). Voice matching applied to 30 segments: 30 identified, 0 mixed, 0 probable, and 0 unresolved.

Loading…