Employers are not asking for less human judgment. They are asking for more of it, and they are saying so explicitly.
In 2026 employer surveys, the majority of companies rank critical thinking, judgment and decision-making, and creativity as having grown in importance over the past year, even as they simultaneously demand AI fluency and skills like workflow automation, data analysis, and AI governance.
Analytical thinking sits at the top of the World Economic Forum’s skills ranking through 2030, named essential by roughly seven in ten employers, which tells you something inconvenient for anyone selling AI literacy as the whole answer: the scarce skill was never operating the tool.
This matters because universities keep answering the wrong question. The question isn’t “how do we teach students to use AI.” Most eighteen-year-olds arriving on campus already use it daily. Undergraduate AI use rose from 92% in 2025 to 95% in 2026, and use of GenAI specifically for assessed coursework rose from 88% to 94% over the same period.
The real question is what happens to a mind that offloads its thinking to a tool before it has finished forming the habit of thinking independently. And here the evidence is not reassuring. A 2025 study by Gerlich, based on 666 participants, found a negative correlation between frequent AI usage and critical-thinking ability, with younger and more AI-dependent users scoring lower on critical-thinking assessments than older participants.
A separate Microsoft Research and Carnegie Mellon study of 319 knowledge workers found what researchers called an “automation irony”: AI removes routine cognitive tasks and, with them, the reps that build judgment, so people increasingly monitor AI systems rather than reason through problems themselves. Confidence in your own judgment, not confidence in the tool, was the variable that predicted whether someone kept thinking critically at all.
That is the actual argument for teaching resilience, empathy, and critical thinking hard, not as a nice-to-have humanities add-on, but as AI governance infrastructure. Responsible AI governance is not a policy document. It is a person, in a room, deciding whether to trust an output they didn’t generate, under time pressure, when the output looks plausible. That decision requires exactly the muscle that passive AI use atrophies.
If a university produces graduates who are fluent in prompting but have never practised holding a position under pressure, admitting they don’t know something, or noticing when a confident-sounding answer is wrong, it has produced people who will rubber-stamp whatever the model says. That is the opposite of governance.
One encouraging counterpoint from the same research literature: the effect isn’t uniform. Higher education itself acts as a partial buffer, and structured, deliberate AI use behaves differently from passive use. Which means the design of the learning experience, not the mere presence or absence of AI, is what determines the outcome.
That puts the responsibility back on the institution, not on the technology. This is exactly why at NEWTON University, we are running an MBA program on AI Transformation for the third year in a row now – and this year, a new subject, Critical Thinking in the Era of AI, is an absolute must, together with Change Management: these skills are increasingly a differentiator in the AI transformation.
So, what does integration actually look like, if the goal is graduates who can think rather than graduates who can prompt?
The direction several universities have already committed to is process over output. The University of Surrey redesigned its entire assessment policy, effective September 2026, around evaluating students’ process and their ability to critically check AI output rather than the final artefact, using worked examples like having engineering students verify AI-generated designs by hand calculation.
The University of Bath is replacing its AI traffic-light detection system with a simpler two-lane model distinguishing supervised, secure tasks from open tasks that require real-world reasoning.
Australia’s higher education regulator has pushed the same direction at the policy level, steering institutions toward oral, applied, and process-based evaluation precisely because it is harder to outsource to a model.
In practice, this means breaking one final deliverable into graded checkpoints (a proposal, an annotated source list, a draft with reflective commentary, a final version) and grading the evolution of the thinking, not just its polished endpoint. It means bringing back oral defence of ideas at scale, not as a nostalgia move but because articulating reasoning out loud is one of the only assessment formats a model cannot yet fake convincingly on someone’s behalf.
It is closer to admitting the old assessment model was already built on trust that no longer holds, and rebuilding it around what can still be verified: whether a student changed their mind somewhere in the process, whether they can defend a claim under a follow-up question, whether they know the difference between an AI output that looks right and one that is right.
That is a duller, harder job than “integrating AI into the classroom.” It is also the only version of the job that produces graduates employers actually need.
Sources:
- Undergraduate AI use 92%→95%, coursework use 88%→94%: — https://arxiv.org/pdf/2608.12351
- Gerlich study, — https://www.ie.edu/center-for-health-and-well-being/blog/ais-cognitive-implications-the-decline-of-our-thinking-skills/
- Microsoft Research + Carnegie Mellon, 319 knowledge workers, “automation irony”: https://www.microsoft.com/en-us/research/publication/the-impact-of-generative-ai-on-critical-thinking-self-reported-reductions-in-cognitive-effort-and-confidence-effects-from-a-survey-of-knowledge-workers/