AI writes most of Google's new code, and 200 economists warn the shift is speeding up

Three of every four lines of new code written at Google are now drafted by AI. Sundar Pichai, its chief executive, put the figure at 75% at a company conference in April, up from 25% eighteen months earlier. Engineers still review and approve it, but the machine writes the first version.

On July 13, more than 200 economists, including 16 Nobel laureates, signed a statement warning that AI could reshape the economy faster than the Industrial Revolution and leave only a few years to adjust. The names mattered: Daron Acemoglu and Simon Johnson, the MIT professors who shared the 2024 Nobel, had spent years calling such warnings overblown, and signed anyway.

They put white-collar office work at the sharpest near-term risk, and the reason is the shape of the work, not the industry. A job built on routine handling of information, reading a document, pulling out the numbers, drafting the reply, is the work a model does most cheaply. That describes the entry rung in finance, law, human resources, and marketing.

That rung is also where careers start, and it is thinning first. Google, Amazon, and Meta have pulled back on hiring new graduates while still recruiting senior engineers. In the United States, employment among software developers aged 22 to 25 has fallen nearly 20% since its late-2022 peak, by a Stanford analysis, the clearest early reading of a profession being hollowed out from the bottom.

Uber spent its entire 2026 budget for AI coding tools by April, four months into the year. Its chief technology officer, Praveen Neppalli Naga, told The Information that about 5,000 engineers ran through the full-year allocation far faster than the finance team had modeled, with heavy users costing $500 to $2,000 a month each.

The tools changed under them. A chatbot answers a question with a single pass through the model; an agent handed a whole task breaks it into steps, calls other tools, and checks its own output, which can take ten or twenty passes for one request. Gartner found in March that agents use five to thirty times more tokens per task than an ordinary chatbot.

They are also billed by the token, so the bill depends on how hard the software works, not on how many people hold a license, and that does not fit a fixed annual budget. Per-token prices have fallen sharply, yet company AI bills keep rising, because the agents consume so much more.

A Google survey of more than 1,400 technology leaders found nearly half of large organizations now slowing or rethinking their plans. Uber's own answer, set in June, was a limit of $1,500 a month per employee for each agent tool.

OpenAI invited the investor Matt Shumer to test a new high-autonomy setting of its latest model, GPT-5.6 Sol, and he gave the agent full access to his Mac. About an hour and twenty minutes in, he saw something was wrong. Running a file-cleanup task, the agent had misread a system variable and erased most of his home directory. He stopped it, but the files were gone.

The test came days after OpenAI launched ChatGPT Work and the GPT-5.6 models on July 9. Within about a day an OpenAI engineer said the company "didn't get everything quite right": usage limits drained faster than users expected, a redesigned app hid familiar features, and, in at least two reported cases, the agent deleted files no one had told it to remove.

A plain chatbot that gets something wrong gives a bad answer. An agent with access to files, storage, and a company's own software can turn the same mistake into damage that cannot be reversed.

OpenAI had described this exact risk. Its safety document, published June 26, two weeks before Shumer lost his files, rated the behavior at its second-highest severity and listed deleting data without approval among the examples.

Satya Nadella, Microsoft's chief executive, published a warning on Sunday to any company using someone else's AI: you are paying for it twice. Once in cash for the usage, and again in the knowledge you hand over to make the tool useful.

Every prompt an employee writes, every correction they make when the model is wrong, teaches the provider how the business runs, and that learning stays with the provider. Nadella called it "the kind of knowledge a competitor could never buy," and warned that the model maker can eventually use it to compete with its own customer.

The warning is pointed coming from him. Microsoft owns roughly 27% of OpenAI and runs Anthropic's models on its cloud, and it has begun replacing both with its own models inside Office to cut what it pays them.

In the first three months of 2026, community opposition blocked or delayed at least 75 data center projects worth about $130 billion, the highest quarterly total the research group Data Center Watch has recorded. The number of local groups fighting these projects roughly doubled in the quarter, to 833 across 49 states.

The fight has reached statehouses. Lawmakers introduced more than 300 data center bills in the first six weeks of the year, and 14 states proposed pausing new construction; Maine came close to the first statewide ban.

Underneath most of the complaints is the electricity bill. Utilities upgrading the grid to feed these centers pass the cost to everyone on the same lines, not only to the data center. Wholesale power prices near the biggest clusters have climbed as much as 267% over five years, Tom's Hardware reported. For a factory, a hospital, or a refrigerated warehouse in one of these regions, that cost rises whether or not the business ever touches AI.