There is a familiar sentence beginning to appear around artificial intelligence at work: employees need to become more AI literate. Sometimes it arrives as a strategic priority. Sometimes it appears inside a job description, a manager’s expectation or a vague instruction to “use AI more.” The phrase sounds reasonable enough. The tools are changing quickly, organizations are trying to adapt and workers who understand how to use them are likely to have an advantage.
The less-discussed question is who is responsible for creating that advantage.
New workforce research from Holmdel, New Jersey-based iCIMS suggests that workers are increasingly answering the question themselves. In a survey of 1,000 U.S. job seekers released September 10, 47% said they had worked on their AI skills during the previous six months, up from 41% a year earlier. The share who said they were teaching themselves rose from 22% to 30% over the same period. Employer-provided AI training, meanwhile, remained roughly flat at about one in six workers.
At the same time, the expectation is already finding its way into the labor market. Forty-five percent of job seekers surveyed by iCIMS said generative-AI skills appear as requirements in jobs they would consider. That produces a peculiar arrangement: the competency is increasingly treated as something a worker should arrive with, while the infrastructure for learning it is often left outside the organization.
This is where the conversation stops being merely about AI adoption and starts becoming a story about work. Companies can describe themselves as transforming, modernizing or becoming AI-ready, but the actual transformation still has to happen somewhere. Someone has to learn the tools, figure out what they can do, discover where they fail, understand the rules for using them and develop enough judgment to know when an answer is useful. If formal training does not keep pace with the expectation, that work does not disappear. It migrates.
It migrates to evenings spent watching tutorials. It migrates to personal accounts and trial subscriptions. It migrates to workers experimenting on their own, reading documentation, swapping prompts with colleagues and trying to infer which skills will still matter by the time the organization finally writes a policy. Not every worker will incur every one of those costs, of course, but the structural pattern is recognizable: the organization wants the benefit of a newly skilled workforce while much of the learning process happens off the books.
Traditional professional development is visible. It has a course, a budget, a trainer, a conference, a certification, a tuition program or at least a block of time on the calendar. Shadow professional development is what happens when the skill requirement arrives before the support system does. The worker understands that the market is moving, senses that remaining competitive may depend on keeping up and begins building the competency independently. Eventually the employer can list the skill as an expectation without ever having to account for the labor that produced it.
AI makes that dynamic especially easy to miss because the tools themselves are so accessible. A worker can open ChatGPT, Copilot or Gemini in seconds and begin experimenting. There may be no classroom, formal software deployment or lengthy onboarding process to signal that training is occurring. Learning can look almost identical to ordinary computer use, which makes it easy to treat the resulting skill as something the worker simply picked up rather than something that required time, repetition and judgment.
The iCIMS data also shows why the phrase “AI proficiency” is becoming dangerous in its vagueness. Sixty percent of respondents said they feel ready to adapt to AI at work, but 61% described their proficiency primarily in terms of general-purpose tools such as ChatGPT, Copilot or Gemini. More specialized capabilities were considerably less common: 18% reported prompt-engineering skills, while 17% cited machine learning or model development.
That does not make the confidence misplaced. General-purpose AI fluency can be genuinely useful. Knowing how to frame a task, evaluate output, iterate, preserve context, identify failure modes and decide when not to use the tool are all practical workplace skills. But those abilities are not interchangeable with model development, data science, workflow automation or specialized implementation. A company that writes “AI proficiency required” without defining the term can therefore demand almost anything while measuring almost nothing.
It also creates room for workers to spend time learning the wrong thing. When employers do not specify the capabilities they actually need, employees and job seekers have to forecast the market themselves. Should they learn prompting? Build agents? Study Python? Understand retrieval-augmented generation? Learn a specific enterprise platform? Become better at verifying AI output? The phrase “learn AI” contains an enormous number of possible directions, and a worker who is paying for that education with personal time carries the risk of choosing badly.
The employer, by contrast, gets optionality. It can wait to see which skills become valuable and then hire for them.
That imbalance is especially notable because workers appear to value structured training. In the iCIMS survey, 42% said a company offering AI training would be more attractive than a similar employer that did not. Fourteen percent said they would even accept lower pay in exchange for that training. Those numbers suggest that professional development is not a side benefit in an AI-transitioning labor market. For some workers, it is part of the compensation package because it reduces the personal cost and uncertainty of remaining employable.
There is a useful distinction here between asking workers to keep their skills current and outsourcing the entire definition of “current” to them. Every profession evolves. People learn new software, regulations, techniques and standards throughout their careers. Self-directed learning is often a sign of curiosity and professional seriousness. None of that absolves organizations from deciding what competencies the job actually requires or from providing a reasonable path for employees to acquire them.
In fact, the faster AI changes, the stronger the argument becomes for employer involvement. Organizations need workers who do not merely know that AI tools exist, but understand how those tools fit the company’s actual processes, data, risk tolerance, security requirements and quality standards. A generic tutorial cannot teach an employee where a specific organization draws those lines. That knowledge has to come from inside the organization.
Otherwise, the company can end up with an odd version of an AI-ready workforce: everyone has experimented, many people feel reasonably confident, nobody was taught the same thing and management still has not defined what competent use looks like.
The broader iCIMS report makes the timing of this problem more interesting. Using Lightcast job-posting data, iCIMS found that explicitly AI-related roles still represent only about 4% of U.S. hiring demand. Yet the survey shows generative-AI expectations reaching much further into the jobs people are considering. The important change, then, may not be the creation of an enormous separate AI profession. It may be the quieter addition of AI expectations to ordinary professions.
That is how a new competency becomes everybody’s responsibility before it becomes anybody’s curriculum.
Employers have a legitimate reason to want adaptable workers. Workers have an equally legitimate reason to ask what support comes with that expectation. If AI literacy is becoming part of the job, organizations will eventually have to decide whether it is merely something they screen for or something they are willing to teach, standardize and give people time to practice.
Because “AI-ready” is not a personality trait. It is a capability. And if a capability is important enough to require, it is important enough to explain how people are supposed to get it.
SOURCE NOTES
• iCIMS / PR Newswire — Sept. 10, 2026 workforce AI skills release
iCIMS Insights September 2026 Workforce Report and iCIMS press release distributed through PR Newswire, Sept. 10, 2026.