Экономист из Стэнфорда: ИИ сделает следующие 10 лет лучшими или худшими в истории

Summary

Stanford economist Eric Brynjolfsson argues that AI is already reshaping work, but its capabilities are advancing faster than businesses are translating them into measurable productivity and profits. The outcome will depend on how companies redesign work, retrain people, and distribute the value created.

Labor-market impact and task redesign

• AI exposure varies by occupation and task. Brynjolfsson describes jobs as bundles of tasks: AI may take over one part while leaving other parts—such as judgment, coordination, or client interaction—largely human-led. He notes that no occupation is currently made up entirely of tasks AI can perform.
• In research he discusses, employment among workers under 25 fell about 16% in the most AI-exposed occupations. The decline was also seen in occupations where AI is used to automate work, including programming and call centers. The interview does not establish that AI alone caused all of these losses.
• Less-exposed work, such as home-care assistance, was growing. Workers using AI to augment their work rather than fully automate it also appeared to fare better.
• Radiology is a counterexample to simple predictions that AI will replace an entire profession. Image reading is only one of a radiologist’s many tasks. More efficient and cheaper imaging can also increase demand for scans and for the other work radiologists do. Brynjolfsson says there are now more radiologists than before, with shortages in the field.
• Roles he identifies as particularly exposed include programming, call-center work, some sales and marketing tasks, junior development, and legal-assistant work. The risk is not uniform across every task in those jobs.

A practical framework for AI-enabled work

Brynjolfsson divides a project into three stages:

1. Formulate the task: Identify the real problem and define what a good result looks like.
2. Execute: Delegate suitable work to AI agents.
3. Evaluate: Check the output, catch errors or misunderstandings, and revise the request.

He expects many workers to manage a fleet of AI agents, spending more time on problem definition and quality control than on routine execution. Because agents can hallucinate or follow a request literally while missing its intent, he recommends treating AI as a collaborator rather than delegating most tasks completely.

Capabilities that complement AI include:

• Domain knowledge and technical understanding
• Sound judgment and the ability to evaluate results
• Initiative, problem-solving, and comfort with ambiguity
• Human interaction and practical skills that are harder to automate

What companies should do

• Connect AI projects to real business problems. Brynjolfsson describes a company hackathon whose winning project used AI to improve the cafeteria menu—an amusing idea, but not necessarily a strategic priority. He recommends focusing on business outcomes such as productivity, profit, sales, or valuable new products and services.
• Redesign processes, not just add tools. Turning AI capability into business value requires changes to workflows, employee training, and sometimes the creation of new products or services.
• Build training into junior roles. Infosys is presented as an example: it still hires junior employees but gives them more deliberate training in project management and broader work, rather than relying on routine tasks as informal on-the-job learning.
• Protect the talent pipeline. If companies eliminate junior roles, they may weaken the path to mid-level and senior roles. Brynjolfsson describes this as a coordination problem: an individual company may save money by not hiring juniors, while the industry as a whole needs people to gain experience and develop.
• Treat transition support as part of the change plan. Brynjolfsson warns that technological gains without retraining and support for displaced workers can provoke backlash. He argues that businesses and policymakers should plan for workforce transitions rather than assume workers will adapt unaided.

Entrepreneurship and individual action

Brynjolfsson encourages people to try using AI in their current work and learn how to direct and assess agents. In his teaching, he uses open-ended questions and asks students to define problems themselves rather than simply follow a step-by-step recipe.

He also recommends exploring unmet needs and potential products or services. A practical starting point is to set aside time to brainstorm problems, use a language model to develop ideas, and test several of them. He says early experimentation can now be inexpensive.

His broader advice is to become a value creator, not just an executor of instructions. He expects AI to lower the cost of experimentation and make entrepreneurial work accessible to more people.

Metrics, examples, and timelines

• 16%: Approximate employment decline among under-25 workers in the most AI-exposed occupations in the research discussed.
• 12%: Reported global decline in entry-level hiring on LinkedIn, which Ryan Roslansky attributed to expensive corporate financing—not AI.
• 1.3 million: New professions reportedly created by AI, according to the LinkedIn discussion.
• $18,000: Average annual US pay premium reported for jobs requiring AI skills.
• 75 million / 4 million: People listing “creator” on LinkedIn versus those listing it as their primary job, as cited in the interview.
• 3–5 years: Brynjolfsson’s estimate for AI capabilities to translate more substantially into business value. He expects the impact to become much more visible by 2030.
• Brynjolfsson says AI capability benchmarks are rising rapidly while current economy-wide effects remain modest. He expects businesses to close that gap gradually, as adoption requires time for process changes and retraining.
• He cites a historical comparison: it took roughly 30 years after electric motors first appeared in US factories for productivity gains to become noticeable.
• A Stanford measure called GDPB/GDPV in the subtitles aims to capture consumer benefit from products and services beyond what conventional GDP measures. The team assessed about 600 products and services. The reported value provided by language models and chatbots rose about 70% over nine months.
• Brynjolfsson says the value of free digital services can be substantial but is poorly reflected in GDP. He argues for tracking both conventional output and the benefits people receive.

Strategic outlook

Brynjolfsson is optimistic about potential productivity and prosperity gains, but concerned that wealth and power could become concentrated among a small number of companies or institutions. He argues that broad participation in value creation matters alongside growth. He also stresses that the future is not predetermined: businesses, workers, and policymakers can influence how AI is used and how its gains are shared.

Presenters and sources mentioned

• Presenter: Marina Mogilko
• Guest: Eric Brynjolfsson, Stanford economist and co-author of The Second Machine Age with Andrew McAfee
• Other sources or organizations referenced: Stanford AI Index; Brynjolfsson’s company Workhelix; Infosys; LinkedIn and former CEO Ryan Roslansky; researchers Geoffrey Hinton and Paul David; economist Bob Gordon; DeepMind