AI Scores Up to 67% on Banker and Lawyer Tasks: Is a "Dead Economy" Coming?

People have long argued about the "dead internet": the idea that most content online is written by bots. The theory now has numbers behind it. The agency Graphite analyzed 43,000 English-language pages from Common Crawl and found that in November 2024 machine-written text outnumbered text written by humans. Now the same question is being asked about work: what if office tasks are handed over to AI en masse? Lawyer Owen McGrann called this scenario the "dead economy." The scenario is still hypothetical, but it has a measurable basis.

What models can already do in place of white-collar workers

Two benchmarks show where the boundary lies. APEX from Mercor tests the work of an investment banker, a consultant, a lawyer at a major firm and a primary care doctor, with 100 tasks per profession. A typical task: analyze a financial model in a spreadsheet, run the calculations and present conclusions. GPT 5, Gemini 3 Pro, Opus 4.5 and newer models meet up to 67% of the control criteria.

OpenAI's GDPVal is broader: 1,320 tasks for dozens of professions, compiled by practitioners with an average of 14 years of experience. The prompts come with attached documents, spreadsheets and drawings. For example, you need to design tooling for a cable machine test stand so that the winding and unwinding cycle is handled by one operator instead of two. Experts compared neural network and human solutions blind. The best models were rated as good as a human's in slightly under half of cases. Just a few months ago, according to the head of frontier evaluations at OpenAI, no model could compete with industry experts.

Still, a benchmark remains a benchmark: these are isolated tasks with ready-made inputs, not real work with unclear requirements and responsibility for the result.

Why layoffs benefit one company and harm the market

For an individual firm, the logic is simple. Block cut 4,000 people in February, about 40% of its staff, and its stock rose by more than 20%. The market rewards such decisions.

Researchers from the University of Pennsylvania and Boston University modeled what happens if everyone starts doing this. If more people are laid off than new jobs appear, solvent demand disappears. Laid-off workers spend less, their former employers' customers see revenue fall, those companies cut staff too, and the circle closes. No single company can stop this alone. Economist Zoë Hitzig adds a psychological factor: many CEOs cut staff simply because their neighbors do.

Back in May 2025, Dario Amodei predicted that within one to five years AI would displace up to half of entry-level office workers, and US unemployment would rise to 10–20%. He himself is betting against the market adapting successfully.

There is also a fiscal risk. OpenAI's April industrial policy document says that if there are fewer wages, there will be fewer taxes, and those taxes support pensions and healthcare.

What is proposed and where the skepticism lies

OpenAI calls its ideas material for discussion, not a finished program. They include: shifting taxes from individual income to capital and corporate profit, creating a public fund that invests in long-term assets and shares income among residents, and introducing benefits and training vouchers that kick in when unemployment exceeds a threshold.

McGrann goes further: state stakes in AI infrastructure, strict antitrust control, a tax regime for automated labor. The researchers from the University of Pennsylvania and Boston University believe that in their model neither wage adjustments, nor a capital tax, nor employee equity stakes, nor universal basic income help. Only a Pigouvian tax tied to the volume of automated tasks works. It forces a company to compensate third parties for the harm done, like an emissions tax.

Skeptics object that scare stories are convenient for management, and that managers should have been the first to be replaced by AI. Daron Acemoglu notes that in 1987–2016 technology also displaced people, and new tasks did not appear fast enough. The economy did not die.

What practitioners should do about it

There are three takeaways. First, treat benchmarks like APEX and GDPVal as a map of automatable tasks. Take tasks, not whole professions: 67% on criteria does not equal replacing a specialist. Second, if you are building a product for corporate clients, do not base your payback calculation solely on headcount reduction. A scenario where clients cut people en masse also hits their ability to pay. Third, watch regulation: taxes on automation are still only a topic of discussion, but they could change the cost of deploying AI.

The whole story has one limitation: these are scenarios and models, not actual labor market statistics. But the fact that the decision to cut staff has stopped being a purely internal company matter is already visible in the numbers.