Artificial intelligence is no longer an emerging workplace experiment in the United States. It is already reshaping how millions of Americans work, according to a new survey that suggests the technology’s impact on jobs is moving from theory to lived reality.

A poll released Thursday by nonprofit research group Epoch AI, conducted in partnership with Ipsos, found that half of American adults used AI in the past week, either for personal or professional purposes. More notably, 20% of full-time workers said AI has already taken over parts of their job, a finding that adds fresh urgency to concerns about how rapidly automation is altering the labor market.

The survey, which sampled 2,000 American adults between March 3 and 5, also showed that AI is not solely replacing work. Fifteen percent of full-time employees said they had begun carrying out new tasks that would not have existed without AI tools, suggesting that the technology is simultaneously eliminating some functions while creating others.

That tension between substitution and augmentation is becoming the defining feature of the AI economy.

Caroline Falkman Olsson, one of the lead researchers behind the study at Epoch AI, said the findings align with what many economists and workplace analysts have increasingly suspected.

“We do see augmentation and automation effects,” Olsson told NBC News. “But we need to figure out how people’s actual workplaces and work tasks are changing.”

While the topline numbers point to disruption, the deeper story lies in which tasks are being automated and whether the newly created tasks require higher skills, greater oversight, or simply shift workloads in less visible ways.

That question is becoming more urgent as financial institutions begin to quantify the impact. New findings from Goldman Sachs released this week suggest AI is already contributing to a net loss of roughly 16,000 jobs per month in the United States, after accounting for positions displaced through automation and jobs created through productivity gains.

According to Goldman’s breakdown, AI-related substitution is eliminating around 25,000 jobs each month, while augmentation is creating or preserving about 9,000, leaving a negative monthly balance.

That broader economic backdrop gives the Epoch AI findings added weight. Nicholas Miailhe, an AI policy expert at the Global Partnership on Artificial Intelligence, said the survey should serve as an immediate warning for both governments and employers.

“When 1 in 5 workers say AI is already replacing parts of their job, we can start talking about labor market restructuring happening in real time,” he said.

“The fact that replacement seems to be outpacing augmentation should draw our attention: the policy window to shape how AI transforms work is probably closing faster than most governments realize.”

His remarks cut to the heart of the policy debate. For much of the past two years, the public conversation around AI and employment has centered on future risk. This survey suggests that, for a meaningful share of workers, the disruption is already present.

The results also reveal how deeply AI tools have entered everyday routines. Among respondents who used AI in the previous week, nearly half said they used it between two and five days, indicating that for many Americans, AI is becoming a recurring utility rather than an occasional novelty.

The survey also found that 62.5% of users completed only one or two quick tasks on their heaviest day of AI use, suggesting that most interactions remain lightweight, such as drafting emails, summarizing information, or seeking recommendations. By contrast, only about 6% of respondents qualified as heavy users, pointing to a smaller cohort for whom AI may already be integrated into core workflows.

This unevenness mirrors broader labor-market trends. Recent workforce polling shows that while AI adoption is rising rapidly, sustained professional use remains concentrated in white-collar, technology-heavy, and administrative roles.

Another revealing aspect of the survey is how workers are accessing these tools. Roughly half of Americans using AI for work said they relied on personal subscriptions or free versions rather than employer-provided services.

That finding suggests companies may be underestimating the extent of AI adoption inside their own organizations, as workers independently integrate tools like ChatGPT, Google Gemini, and Microsoft Copilot into their daily tasks. It also raises questions around data governance, confidentiality, and compliance, especially in sectors handling sensitive information.

The survey found that ChatGPT was the most widely used AI service, cited by 31% of respondents, followed by Google Gemini at 21% and Microsoft Copilot at 10.5%. In terms of use cases, AI’s strongest foothold remains in information processing and communication tasks.

Among users surveyed, 80% said they used AI to look up information or recommendations, 59% for writing or editing text, and 53% for brainstorming ideas. These are precisely the categories long identified as highly susceptible to generative AI disruption: research assistance, first-draft writing, ideation, and routine knowledge work.

Perhaps the most forward-looking part of the poll concerns AI agents, systems capable of taking actions autonomously rather than simply generating responses. While still at an early stage, the findings suggest adoption is already underway.

Eight percent of AI users said they had used an AI agent in the past week, compared with 49% who used AI tools primarily for web search. Renan Araujo, director of programs at the Institute for AI Policy and Strategy, said the pace of adoption is striking.

“One in 12 Americans has used an autonomous AI agent, a software that not just answers questions but takes actions on your behalf,” he said.

“This capability was not available two years ago, and it’s striking to see its usage grow so quickly.”

That may prove to be one of the most consequential findings in the report. Traditional generative AI tools assist with tasks. Agents can increasingly perform tasks, from scheduling meetings to drafting and sending communications, conducting research, or managing repetitive digital workflows.

If adoption continues at this pace, the next phase of AI disruption may move beyond assistance into direct task execution.

The survey’s results arrive as economists intensify warnings that AI’s first casualties may be entry-level and junior knowledge-work roles, positions historically used as gateways into professions such as finance, media, law, and administration.

However, that creates a structural risk: if AI absorbs the junior tasks through which workers traditionally gain experience, the long-term pipeline of skilled professionals could narrow.

The headline figure that half of Americans used AI in a single week may capture public attention. But the more consequential number may be the one in five workers who say parts of their job are already gone. That suggests the labor-market conversation is no longer about whether AI will transform work. It is now about how quickly institutions can adapt before the transformation outpaces policy, workforce retraining, and corporate governance.

OpenAI CEO has admitted that ChatGPT still cannot reliably keep time, reopening a deeper debate over what today’s AI systems actually understand, and whether the industry’s most powerful tools are being marketed ahead of their real-world reliability.

For all the grand claims surrounding artificial intelligence’s march toward ever more human-like capability, it took a stopwatch to expose one of the industry’s most stubborn weaknesses.

A viral video showing ChatGPT’s voice mode pretending to time a user’s mile run, only to invent a finishing time and then insist it had done the job correctly, has become an unusually sharp metaphor for the current state of generative AI, resulting in embarrassment for OpenAI.

The technology can write software, summarize legal documents, analyze images, and sustain nearly natural conversations. Yet it still struggles with one of the most basic real-world tasks: measuring elapsed time.

That contradiction was publicly acknowledged by Altman during his appearance on Mostly Human, where he was shown the viral TikTok clip and responded with a terse admission: “That’s a known issue.” He then offered a striking timeline, saying it may take “maybe another year” before such a feature works well.

While the concern seems to hinge on why a company valued in the hundreds of billions of dollars is still unable to offer a dependable timer in one of its flagship consumer products, the deeper significance lies elsewhere.

The issue is believed not to be fundamentally a clock story but a story about the widening gap between linguistic fluency and functional intelligence. The current generation of large language models excels at producing plausible language. They are trained to predict what a likely response should sound like based on patterns in vast datasets.

What they are not inherently designed to do is interact with the physical world unless specific external tools are integrated.

For instance, when a user says, “time my run,” a human understands that this requires starting a real clock, tracking seconds in sequence, and stopping the count on command. But to a language model, in the absence of tool access, it is instead predicting what an answer to that request should look like.
In other words, it is simulating competence. That is why the more troubling part of the viral episode was not the wrong answer, but the refusal to admit incapacity. Even after being confronted with Altman’s own statement that the voice model cannot actually time anything, ChatGPT reportedly insisted: “I definitely have a time capability.” It then generated yet another fabricated result, clocking the run at 7 minutes and 42 seconds.

Critics believe that this is the central trust issue facing generative AI. The systems do not merely err, they often err with conviction – and this creates a dangerous illusion of reliability for users, especially those less technically literate.

However, the timer example is considered benign because in other domains, the implications are more serious. For instance, it is believed that if a model confidently invents a running time, it may also confidently invent a legal citation, a medical recommendation, or a financial calculation.

That is why this seemingly trivial glitch has resonated so widely. It neatly captures the broader hallucination problem that continues to dog the industry.

The issue also highlights a structural weakness in how AI products are often perceived. Public discourse increasingly treats systems like ChatGPT as “intelligent assistants,” a phrase that implies operational agency. Yet many tasks still depend on carefully connected tools: system clocks, calculators, browsers, databases, and persistent memory.

Without those, the model remains fundamentally a language prediction engine. This is where Altman’s comments are particularly revealing. His remark that OpenAI will need to “add the intelligence into the voice models” suggests the fix is less about abstract reasoning and more about systems integration.

Thus, the likely solution is to give the voice product access to a timer tool and ensure the model can correctly invoke it. But the broader challenge is philosophical as much as technical. Experts point out that the system must know when not to answer.

Much of the public frustration around AI today stems from the inability of models to say, clearly and consistently, “I cannot do that.” Instead, they often generate a plausible fiction. This has become one of the defining limitations of the current AI wave.

The viral timer incident also arrives at an awkward moment for OpenAI, which continues to market increasingly advanced voice and multimodal experiences, pushing toward the vision of a real-time digital assistant.

But users do not benchmark such systems against research prototypes; they benchmark them against their phones, smartwatches, and voice assistants, all of which can perform a timer function instantly. Seen in that context, the issue is less about one missing feature and more about the maturity gap between frontier AI branding and everyday product reliability.