In January 2023, many major school systems blocked ChatGPT outright. Within two years most of those bans had quietly been replaced with guidance, experimentation, and course-level rules. Nearly four years after ChatGPT's public release, the destination is still not settled.
This matters to you not as an education story but as a hiring and management story. The people entering your organization right now were educated during the most incoherent period of AI policy in academic history. Some learned to use AI well. Some were trained to fear it. Most got conflicting signals depending on which professor they had that semester.
Where Schools Actually Stand
The current state across most major institutions is a baseline plus instructor discretion. An, Yu and James examined 214 policy documents from the top 50 US universities and found that 94% had faculty guidelines on establishing course-specific AI policies, with 54% explicitly telling faculty to set their own and communicate them to students. More than half provided template syllabus statements for faculty to choose between. Stanford offered three categories, from freely permitted to not permitted at all. Chicago offered four. So one course allows AI for brainstorming but not drafting. The next requires full integration. The one after that prohibits it entirely, and all three are following institutional guidance correctly.
HEPI's 2026 student survey puts the scale of it plainly. 95% of UK undergraduates now use generative AI in some form and 94% use it for assessed work, but only 48% think their teaching staff are helping them develop AI skills. Worth noting alongside that: only 12% report inserting AI-generated text directly into their work, so this is mostly students using AI as an aid without much guidance on how. The gap between what students are doing and what institutions have a coherent position on is significant.
What Parents Can Do
If your child is in secondary school now, the questions worth asking their school are specific. Does the school have a written AI use policy, or is it left to individual teachers? Are students being taught how to verify AI output, or just told not to use it? Is AI literacy treated as a skill, the way research skills or citation formatting are taught, or is it treated purely as a disciplinary issue?
The schools handling this well are teaching students to use AI as a tool with visible, documented limitations. They are teaching citation and disclosure. They are running exercises where students deliberately find AI errors and explain why the output failed. That is the curriculum that produces a graduate who is an asset in 2030. A school that only prohibits AI, without separately teaching students how to evaluate and use it, risks producing graduates who are unprepared to use a tool many will encounter in their working lives.
For university students already enrolled: check the written AI policy for each course at the start of each semester. Do not assume the institution's general policy covers a specific assignment. The same divergence shows up in admissions. GradPilot's analysis of university admissions essay policies finds 62 of 180 institutions with at least one program whose policy differs from the institution-level position. Duke is one: its Law School prohibits AI in application materials while undergraduate admissions permits it.
What the Incoherence Produces
The institutions handling this well have three things in common: a written definition of acceptable AI use, a standardized disclosure process, and assessment designed so that independent thinking shows up in the work itself. Oral defense, version history, a short reflection on process. The ones handling it badly leave each professor to set rules without explaining why they differ, producing graduates who carry conflicting intuitions about what responsible AI use looks like.
Detection tools are not the answer, and a good number of institutions have already concluded that. In the An, Yu and James study, 38% of universities told faculty not to use them. Princeton called them unreliable and biased. Northwestern turned Turnitin's AI indicator off after piloting it. Turnitin's own guidance says detector scores should be treated as prompts for investigation, not proof of wrongdoing. Detectors can produce false positives, and their performance varies by tool, by model, and by writing context.
What that means for hiring: two graduates from the same institution, same program, same year, may have radically different AI competencies depending on which professors they happened to take courses with.
What Industry Is Already Doing
The financial services industry is the leading indicator here. Banks are not waiting for universities to sort this out.
JPMorgan's LLM Suite is now deployed to more than 230,000 employees. Jamie Dimon has said publicly that AI "will eliminate jobs" while committing to retraining affected staff. Goldman Sachs CEO David Solomon said in January 2025 that AI can now complete 95% of an S-1 IPO prospectus in minutes, work that once took a six-person team two weeks. His specific observation: "The last 5% now matters because the rest is now a commodity."
In 2025, 88% of AI-related job listings in financial services were for AI user roles, not AI developer roles. The split was roughly 88% user roles to 12% builder roles, so demand for people who can build has not disappeared. But most financial services AI hiring is for people who can apply it, and that is what hiring managers are mostly testing for.
Zapier has gone furthest in making this explicit. The company publicly sorts candidates into four AI fluency levels: unacceptable, capable, adoptive, and transformative. Its March 2026 revision is the interesting part. Unacceptable no longer means resistance to AI. It means using AI as a lightweight assist inside a workflow that has not changed, which describes a lot of people who consider themselves AI users. Capable now requires AI embedded in core work, repeatable systems rather than one-off prompts, and evidence of impact. Zapier also added accountability as a fourth thing it measures, on the principle that you can delegate the work but not the accountability. Zapier applies the rubric across four points in its hiring process, and publishes what it looks like by department.
The broader picture: McKinsey research from November 2025 found demand for AI fluency jumped nearly sevenfold in two years, from approximately 1 million workers in 2023 to around 7 million in 2025. LinkedIn's January 2026 labor market report found a 70% year-over-year increase in US roles requiring AI literacy. NACE reported that AI skills appeared in 16.5% of job descriptions in spring 2026, up from 10.5% six months earlier, while 28% of surveyed employers said they were seeking early-career talent able to use AI.
Deel's senior director of talent acquisition put the problem plainly: companies desperately need AI literacy, yet interviewers often do not know how to assess it. "You've got motivation to hire an AI-literate workforce, yet there's an immaturity within your existing organization in terms of the maturity of AI deployment. That creates a situation where the interviewer doesn't really know how to assess the very skill you're trying to hire for."
What to Actually Ask in an Interview
The standard questions are a starting point. "Tell me about a time AI gave you something wrong" and "how do you decide when to trust AI output" will surface basic competency. But they are easy to prepare for, and by 2026 most candidates have rehearsed answers.
The more revealing questions go one layer deeper.
Ask them to walk you through the last piece of work they used AI on, step by step. Where exactly did they use it? What prompt did they use? What did they check and how? What would they do differently? A candidate who can narrate that process specifically, not generally, has a working relationship with the tool. A candidate who gives you a principle without a specific example does not.
Ask them about a case where they disagreed with an AI output that looked correct. Not where the AI was obviously wrong, but where the output was plausible and they still chose not to use it. Why? What told them something was off? This question tests calibration, not just error-catching. The dangerous AI user is the one who cannot tell a confident answer from a correct one.
Ask them what AI cannot do in their specific domain. Not in general, in their field, in the kind of work they are applying to do. A candidate who can name specific, concrete failure modes of AI in their discipline has spent real time with the tools in a professional context. A candidate who gives you a generic answer about hallucinations has read about it.
Finally, ask them to show you something. Give them a task during the interview: a short piece of analysis, a draft, a summary of something you hand them. Tell them they can use whatever tools they normally would. Then watch. How they approach a live task with AI visible tells you more than any answer to any question about AI.
What you are interviewing for is judgment showing up at the points where being wrong is expensive. The exercise should be designed to reveal it.
Where Agent Micho Fits
The interview process surfaces the gap. The workflow closes it.
A new hire who is excellent at using AI but has never worked within a structured system will develop habits on the job, some good and some not. A workflow that defines where AI sits in each process: what it produces, what gets checked, what gets escalated to a human, removes the dependency on individual judgment for the parts that should be standardized.
We build those workflows. Not as a replacement for human judgment, but as the structure that makes human judgment show up in the right place at the right time. Solomon is describing one task, not a universal ratio, and the person is involved well before the final stretch. Someone sets the goal, chooses the inputs, and carries the accountability throughout. What a well-built system does is automate the steps that should be automated and place human review according to consequence.
If you are building a team that works with AI, start by finding the points in a process where a person still has to make the call. Then check whether your workflow actually brings them in at those points, every time, not when someone remembers. That is a workflow design question, and it is what we do.
Sources
- Student AI use figures: 95% of UK undergraduates using generative AI, 94% using it for assessed work, 48% saying teaching staff help them develop the skill, and 12% inserting AI-generated text directly. Higher Education Policy Institute, Student Generative AI Survey 2026.
- Course-specific AI policies in higher education, the 94% and 54% figures, the syllabus templates, and the 38% of universities advising faculty against detection tools: An, Y., Yu, J. H., and James, S. (2025), Investigating the higher education institutions’ guidelines and policies regarding the use of generative AI, International Journal of Educational Technology in Higher Education, examining 214 documents from the top 50 US universities.
- Program-level divergence in admissions policy, 62 of 180 institutions, and the Duke example: GradPilot, AI Policies for 170+ Universities. This dataset covers admissions and application essay policies, not classroom or assessment rules, and is used here only for the admissions claim.
- Detector scores as a prompt for investigation rather than proof of misconduct: Turnitin's own guidance on reviewing the AI writing report.
- The 88% user to 12% developer split in financial services AI job listings: PwC 2026 AI Jobs Barometer, financial services, reported in American Banker, June 2026. The same piece covers Ramp research finding that companies spending most on AI grew headcount rather than cutting it.
- Sevenfold growth in demand for AI fluency between 2023 and 2025, and the seven million workers figure: McKinsey Global Institute, Agents, robots, and us, November 2025.
- The 70% year-over-year increase in US roles requiring AI literacy: LinkedIn Economic Graph, Building a Future of Work That Works, January 2026.
- AI skills in 16.5% of job descriptions in spring 2026, up from 10.5% six months earlier, and 28% of employers seeking early-career talent able to use AI: NACE, Ready or Reluctant, July 2026.
- The four AI fluency levels, the March 2026 revision of what counts as capable and unacceptable, and accountability as a fourth measured component: Zapier, Raising the AI fluency bar for every Zapier hire.
- The assessment paradox and the quote from Alan Price, Senior Director of Talent Acquisition: Deel, Companies want AI-literate workers, but don't know how to hire them, March 2026.
- David Solomon on AI drafting 95% of an S-1: Bloomberg television interview, January 2025. No stable public URL.
- Figures reflect what these sources published at the time of writing and will move as surveys are repeated.
A note on images across this site. Illustrations and workflow diagrams are made with AI, from prompts we write and refine, and we edit most of them afterwards. Screenshots taken in n8n are not, since they show workflows we built in the tool.