AI's Rapid Evolution: From Search to Task Automation

Silicon Valley's Plan to Automate Everything

The automated future just lurched a few steps closer. Over the past few weeks, nearly all of the major AI firms -- OpenAI, Anthropic, Google, xAI, Amazon, Microsoft, and Perplexity, among others -- have announced new products that are focused not on answering questions or making their human users somewhat more efficient, but on completing tasks themselves. They are being pitched for their ability to "reason" as people do and serve as "agents" that will eventually carry out complex work from start to finish. Humans will still nudge these models along, of course, but they are engineered to help fewer people do the work of many. Last month, Anthropic launched Claude Code, a coding program that can do much of a human software developer's job but far faster, "reducing development time and overhead." The program actively participates in the way that a colleague would, writing and deploying code, among other things. Google now has a widely available "workhorse model," and three separate AI companies have products named Deep Research, all of which quickly gather and synthesize huge amounts of information on a user's behalf. OpenAI touts its version's ability to "complete multi-step research tasks for you" and accomplish "in tens of minutes what would take a human many hours." AI companies have long been building and benefiting from the narrative that their products will eventually be able to automate major projects for their users, displacing jobs and perhaps even entire professions or sectors of society. As early as 2016, Sam Altman, who had recently co-founded OpenAI, wrote in a blog post that "as technology continues to eliminate traditional jobs," new economic models might be necessary, such as a universal basic income; he has warned repeatedly since then that AI will disrupt the labor market, telling my colleague Ross Andersen in 2023 that "jobs are definitely going to go away, full stop." Despite the foreboding nature of these comments, they have remained firmly in the realm of speculation. Two years ago, ChatGPT couldn't perform basic arithmetic, and critics have long harped on the technology's biases and mythomania. Chatbots and AI-powered image generators became known for helping kids cheat on homework and flooding the web with low-grade content. Meaningful applications quickly emerged in some professions -- coding, fielding customer-service queries, writing boilerplate copy -- but even the best AI models were clearly not capable enough to precipitate widespread job displacement. Since then, however, two transformations have taken place. First, AI search became standard. Chatbots exploded in popularity because they could lucidly -- though frequently inaccurately -- answer human questions. Billions of people were already accustomed to asking questions and finding information online, making this an obvious use case for AI models that might otherwise have seemed like research projects: Now 300 million people use ChatGPT every week, and more than 1 billion use Google's AI Overview, according to the companies. Further underscoring the products' relevance, media companies -- including The Atlantic -- signed lucrative deals with OpenAI and others to add their content to AI search, bringing both legitimacy and some additional scrutiny to the technology. Hundreds of millions were habituated to AI, and at least some portion have found the technology helpful. But although plain chatbots and AI search introduced a major cultural shift, their business prospects were always small potatoes for the tech giants. Compared with traditional search algorithms, AI algorithms are more expensive to run. And search is an old business model that generative AI could only enhance -- perhaps resulting in a few more clicks on paid advertisements or producing a bit more user data for targeting future advertisements. Refining and expanding generative AI to do more for the professional class -- not just students scrambling on term papers -- is where tech companies see the real financial opportunity. And they've been building toward seizing it. The second transformation that has led to this new phase of the AI era is simply that the technology, while still riddled with biases and inaccuracies, has legitimately improved. The slate of so-called reasoning models released in recent months, such as OpenAI's o3-mini and xAI's Grok 3, has impressed in particular. These AI products can be genuinely helpful, and their applications to advancing scientific research could prove lifesaving. Economists, doctors, coders, and other professionals are widely commenting on how these new models can expedite their work; a quarter of tech start-ups in this year's cohort at the prestigious incubator Y Combinator said that 95 percent of their code was generated with AI. Major firms -- McKinsey, Moderna, and Salesforce, to name just a handful -- are now using it in basically every aspect of their businesses. And the models continue getting cheaper, and faster, to deploy. Read: The GPT era is already ending Tech executives, in turn, have grown blunt about their hopes that AI will become good enough to do a human's work. In a Meta earnings call in late January, CEO Mark Zuckerberg said, "2025 will be the year when it becomes possible to build an AI engineering agent" that's as skilled as "a good, mid-level engineer." Dario Amodei, the CEO of Anthropic, recently said in a talk with the Council on Foreign Relations that AI will be "writing 90 percent of the code" just a few months from now -- although still with human specifications, he noted. But he continued, "We will eventually reach the point where the AIs can do everything that humans can," in every industry. (Amodei, it should be mentioned, is the ultimate techno-optimist; in October, he published a sprawling manifesto, titled "Machines of Loving Grace," that posited AI development could lead to "the defeat of most diseases, the growth in biological and cognitive freedom, the lifting of billions of people out of poverty to share in the new technologies, a renaissance of liberal democracy and human rights.") Altman has used similarly grand language recently, imagining countless virtual knowledge workers fanning out across industries. These bright visions have dimmed considerably when put into practice: Elon Musk and the Department of Government Efficiency's efforts to replace human civil servants with AI may be the clearest and most dramatic execution of this playbook yet, with massive job loss and little more than chaos to show for it so far. Meanwhile, all of generative-AI models' issues with bias, inaccuracy, and poor citations remain, even as the technology has advanced. OpenAI's image-generating technology still struggles at times to produce people with the right number of appendages. Salesforce is reportedly struggling to sell its AI agent, Agentforce, to customers because of issues with accuracy and concerns about the product's high cost, among other things. Nevertheless, the corporation has pressed on with its pitch, much as other AI companies have continued to iterate on and promote products with known issues. (In a recent earnings call, Salesforce CEO Marc Benioff said the firm has "3,000 paying Agentforce customers who are experiencing unprecedented levels of productivity.") In other words, flawed products won't stop tech companies' push to automate everything -- the AI-saturated future will be imperfect at best, but it is coming anyway. The industry's motivations are clear: Google's and Microsoft's cloud businesses, for instance, grew rapidly in 2024, driven substantially by their AI offerings. Meta's head of business AI, Clara Shih, recently told CNBC that the company expects "every business" to use AI agents, "the way that businesses today have websites and email addresses." OpenAI is reportedly considering charging $20,000 a month for access to what it describes as Ph.D.-level research agents. Google and Perplexity did not respond to a request for comment, and a Microsoft spokesperson declined to comment. An OpenAI spokesperson pointed me to an essay from September in which Altman wrote, "I have no fear that we'll run out of things to do." He could well be right; the Bureau of Labor Statistics projects AI to substantially increase the demand for computer and business occupations through 2033. A spokesperson for Anthropic referred me to the start-up's initiative to study and prepare for AI's effect on the labor market. The effort's first research paper analyzed millions of conversations with Anthropic's Claude model and found that it was used to "automate" human work in 43 percent of cases, such as identifying and fixing a software bug. Tech companies are revealing, more clearly than ever, their vision for a post-work future. ChatGPT started the generative-AI boom not with an incredible business success, but with a psychological one. The chatbot was and is still possibly losing the company money, but it exposed internet users around the world to the first popular computer program that could hold an intelligent conversation on any subject. The advent of AI search may have performed a similar role, presenting limited opportunity for immediate profits but habituating -- or perhaps inoculating -- millions of people to bots that can think, write, and live for you.

Column | 1 in 4 programming jobs have vanished. What happened?

A big jump in unemployment for programmers since 2022 may be the first sign that artificial intelligence is taking human jobs. More than a quarter of all computer programming jobs have vanished in the past two years, the worst downturn that industry has ever seen. Things are sufficiently abysmal that computer programming ranks among 10 hardest-hit occupations of 420-plus jobs for which we have data from the Bureau of Labor Statistics. Learning to code was supposed to save millions of would-have-been liberal arts majors. But today there are fewer programmers in the United States than at any point since 1980. That's a 45-year period in which America's total workforce has grown by about 75 percent! It's so long ago that millennials hadn't been invented, the oldest Gen Xers were barely in high school, and even many boomers were too young for their first real coding jobs. The numbers are equal parts apocalyptic and astonishing. Often when we discover this kind of jaw-dropping data, we find it's not as bad as it looks. As unabashed boosters of both computers and programming, we crossed our fingers that the pattern would repeat itself as we dug deeper. Skip to end of carousel Department of Data We here at the Department of Data are dedicated to exploring the weird and wondrous power of the data that defines our world. Read more. End of carousel We started by stepping back until our field of view widened. Most programmers work in an industry that specializes in writing software for corporate clients. And while overall employment has sputtered as that industry has retreated from its early-pandemic excesses, it hasn't seen the same devastating decline. More important, when we looked at who worked in that industry, we noticed that programmers were in the minority. They're dwarfed by, among other occupations, the software developers. The two terms are sometimes used interchangeably, but the jobs follow very different trajectories. Nationwide, software developers haven't struggled nearly as much as their programming brethren -- few other computer-related occupations have. So what makes programmers different? To answer that, we need to dive forehead first into everybody's favorite part of any analysis: arcane occupation definitions! Upon perusing the fine print, we saw that while programmers do in fact program, they "work from specifications drawn up by software and web developers or other individuals." That seems like a clue. In the real world, "developer" and "programmer" can seem almost interchangeable. But in the world of government statistics, where we have legal permanent residency, there's a clear distinction. In the government's schema, programmers do the grunt work while the much more numerous -- and much faster-growing -- software developers enjoy a broader remit. They figure out what clients need, design solutions and work with folks such as programmers and hardware engineers to implement them. Their pay reflects this gap in responsibilities. The median programmer earned $99,700 in 2023, compared with $132,270 for the median developer. And while 27.5 percent of programming jobs vanished, jobs for developers have only fallen 0.3 percent, similar to the broader industry. So it's not just industry-wide headwinds holding programming back. What could account for the difference between the coder collapse and everyone else? Upon reflection, the solution seemed so obvious that, in the grand tradition of M. Hercule Poirot, we felt like imbeciles for not spotting it sooner. At the end of 2022, just before programmers plummeted, OpenAI released ChatGPT, the artificial intelligence chatbot that -- much like many data columnists -- has always seemed more adept at coding than conversation. Could this be the first concrete evidence of generative AI replacing workers in the real world? After all, the BLS kinda defines programmers as human coding machines. Other folks feed guidance in one end, and code comes out the other. That sounds an awful lot like what chatbots do. It's not the first time programmers have born the brunt of automation. For as long as the BLS has differentiated the two professions, programmers have been the black sheep, always struggling while developer jobs multiplied and multiplied again. With every innovation that made coding easier or less necessary -- services to handle common tasks, offshoring, free open-source tools, servers and computing on the cloud -- developers took on more of the work once left to pure programmers. We didn't want to jump the gun. We've seen endless warnings that AI would take our jobs, but until now there's been precious little evidence that the taking had begun. Research from folks such as Northwestern University economist Dimitris Papanikolaou and his collaborators has found the job market effects of earlier generations of AI and machine learning to be quite muted. The tools make workers more efficient and perhaps even redundant, Papanikolau told us, but that same efficiency boost also causes the firm to grow, and the growing firm hires more workers. (Not unlike how the laborsaving cotton gin counterintuitively increased demand for the work of enslaved Africans.) But, he said, it's entirely possible that some occupations within those firms, such as programmers, might be losing ground as jobs shift to occupations for which AI is more of a complement than a substitute. For further insight, we rang up our old friend Mark Muro at the Brookings Institution. Muro's been sussing out strategies to measure the impact of automation and AI on the labor market for nearly a decade now. He began with the eight sweetest words in the English language: "What you're seeing makes a lot of sense." As AI replaces rote coding tasks and people rely more on snippets generated by models, "the first inroads are going to be for the more routine programming," Muro told us. "Without getting hysterical," he added, "the unemployment jump for programming really does look at least partly like an early, visible labor market effect of AI." He said his estimates, based on data from OpenAI, showed that the broad occupation category including programmers was among the most vulnerable to AI. Similarly, Indeed Hiring Lab economist Allison Shrivastava found that job postings in the category containing programmers and developers mentioned AI-related terms more often than any other sector on Indeed. But to tease out the difference between programmers and developers using the latest and greatest evidence, Muro pointed us to a recent report from the California AI outfit Anthropic. Anthropic researchers analyzed about a million (anonymized) conversations with Claude, the firm's answer to OpenAI's ChatGPT, from late 2024 and early 2025. They matched those chatbot queries to specific activities from achingly detailed task lists the Labor Department has built for each occupation -- phlebotomists' tasks include "Dispose of blood or other biohazard fluids," while professional divers may "Descend into water" or "Drill holes in rock and rig explosives," and programmers "Write, update, and maintain computer programs," among other things. When they calculated the share of all queries used for tasks related to each of more than 700 occupations, they found people used AI to perform the tasks usually assigned to computer programmers more than those of any other job. Software developers came next, ahead of almost every other profession. To be sure, Anthropic's Alex Tamkin emphasized that while he sees the potential for AI to have large effects on the labor market, his team's current analysis wasn't designed to determine which jobs would be replaced with AI. It just looks at which jobs involve tasks that people are using AI for. In fact, in a majority of cases (57 percent) people are using AI to augment their work rather than to automate it entirely. "Usage tilts more towards augmentation -- which is things like having the AI check your work, asking questions to teach you things, iterating on a piece of work -- rather than automation," Tamkin told us. "And that suggests that right now AI is, on balance, used more as a tool to help you with the work you're doing rather than automating small chunks of it." The overlap between programming job losses, the rise of AI and programmers' very specific job description certainly seems suggestive, but all our sources reminded us that we can't lay the entire coder-dammerung at the feet of the chatbots. Programmers have been hammered hardest, but firms across the tech industry have struggled in the past two years as high interest rates and slowing growth made their devil-may-care spending in the early pandemic era look downright devil-may-care. "Programmers may be more likely than software developers to have more of their job replaced by generative AI, but the sharp decline cannot be attributed to generative AI alone," Shrivastava told us. She added that Indeed postings for tech jobs such as programmers and developers rose much faster than other job postings amid the 2022 labor market boom and thus had further to fall when the job market cooled. AI obviously is not a direct replacement for human programmers, given the precision demanded by the profession and the chatbots' predilection for fabrications. But in the hands of a software developer who knows the code well, the bots could pitch in on some of the grunt work. It makes us wonder if one day the programmer will go the way of the computer. For centuries, "computer" described a job done by human beings who performed complex computations. Now, it just refers to the machine humans use to handle the math they used to assign to other humans in the computing department. Greetings to friends and foes alike! The Department of Data's still craving queries. What's killing all the butterflies? Is artificial intelligence better at diagnosing than doctors are? How many federal workers are there, really? Just ask! If your question appears in a column, we'll send you an official Department of Data button and ID card.