What I Teach My Students About AI (That Courses Get Wrong)

Every week I stand in front of students who can smell theory from the back row. Undergraduates studying advertising and media at two universities in Karachi, plus school students learning entrepreneurship and financial literacy, and they all share one trait: if I say something about AI that I cannot back up with proof from real work, someone will call it out. Usually within seconds. That classroom has become the toughest filter my ideas ever pass through, and it taught me something most AI courses have not figured out yet. The gap between what we teach about AI and what the working world actually needs is wide, it is measurable, and it is getting worse.

The Classroom Is the Toughest Filter

Here is what people who do not teach tend to misunderstand. Students are not a polite audience waiting to be impressed. They are twenty-year-olds who use AI tools every single day, often more fluidly than the people lecturing them, and they have developed a finely tuned instinct for spotting a fraud. If I walk in and recite a list of tool features I read on a blog, I am finished before the projector warms up. They will test every claim in real time, on their phones, under the desk, and they will report the result out loud.

Teaching has made me a better practitioner, and practice has made me a better teacher. I run an AI practice alongside my teaching: brand strategy, AI video production, social media work, through Ailygence and Sagar Samy Ltd, across more than 250 brand engagements. When I tell students something, it is because a client paid for that lesson first. Students can tell the difference between a lecturer who tried the thing and one who read about it.

The data backs up what I see in the room. A nationally representative RAND Corporation survey of 1,214 students, published in March 2026, found the share using AI for homework rose from 48% to 62% between May and December 2025. In the UK, the Higher Education Policy Institute’s 2025 survey of 1,041 undergraduates found 92% use AI tools, up from 66% the year before, and 88% had used AI in assessments. Students are miles ahead of the syllabus: only 29% of UK students agreed their institution encourages AI use. Professor Michael Grove of the University of Birmingham put it bluntly on the HEPI website in 2025: students are using generative AI at scale, yet only a third say their institution has supported them to use these tools well. For many, he said, the message appears to be: you are on your own.

So the teacher who walks in with a theory-only lecture is not just boring. He is irrelevant, and the students know it before he reaches his second slide.

What Most AI Courses Get Wrong

Most AI courses I have reviewed, and most university modules being bolted together in a hurry right now, make the same three mistakes. They teach the tools, they police the cheating, and they quarantine the ethics. None of that is what the work actually requires.

Mistake one is tool-first teaching. The course becomes a tour of features: how to write a prompt, which model does what, a demo of image generation. The problem is that tools change every quarter, and a curriculum built on current tool features expires fast. Wharton’s Ethan Mollick has made the academic version of this point with his concept of the jagged frontier: AI completes some tasks with ease while failing at others of similar difficulty, and the line moves with every model release. In a study with BCG, Harvard and Wharton, consultants who used AI for tasks outside that frontier were 19 percentage points less likely to produce correct solutions. Using the tool badly is worse than not using it at all, and a course that teaches features without teaching where the frontier sits is manufacturing overconfidence.

Mistake two is teaching AI as a risk to be policed rather than a literacy to be built. Grove’s criticism is exact: the sector’s focus has largely been on mitigating risk, assessment guidance and misconduct policies, rather than teaching AI as a literacy. The result: students use the tools anyway (92% of them, per HEPI), but learn in the shadows, without guidance, while the institution pretends the policy is working.

Mistake three is the standalone ethics module, one quarantined week of responsible AI, after which nobody mentions judgment again. India’s national AI curriculum taskforce made this exact diagnosis at scale in May 2026, when Union Minister Ashwini Vaishnaw met NASSCOM and industry. The taskforce found lecture-based teaching, only 25 to 30% practical exposure in B.Tech computer science curricula, and responsible AI taught as a standalone module instead of being integrated across every semester. I read that finding and thought: that is not an India problem. That is a template of what is going wrong in classrooms everywhere.

This is not abstract for me. I teach in Pakistan, and in January 2026 the Khyber Pakhtunkhwa government announced AI would enter the school curriculum from grade 6, with implementation set from March 2026 and 35 to 50% AI-focused material. The ambition is real, but the fine print is familiar: 5,525 IT laboratories have yet to be established, and approval was given to recruit 7,555 IT teachers to make the plan work. Policy outpaces classroom reality, in my country and in yours.

The honest summary of the situation looks like this:

AI Literacy Is Not AI Fluency

Every semester, a student asks me some version of the same question: if the AI can do the assignment, what am I actually supposed to learn? It is a fair question, and it exposes the central confusion in how we talk about AI education. We keep treating AI literacy and AI fluency as the same thing. They are not.

Literacy is knowing the tools exist and roughly what they do. Fluency is being able to verify an output, judge its quality, catch its lies, and take responsibility for publishing it. Literacy is a tourist with a phrasebook. Fluency is a professional who can argue in the language.

Pearson and AWS surveyed 2,711 stakeholders across six countries in April 2026, and the numbers are stark. Only 14% of current graduates report a high level of proficiency in applying AI tools to a professional workflow. And 58% of employers rated a graduate’s ability to critically verify AI outputs as their weakest competency. Verification. The single most important skill in an AI-saturated workplace is the one courses barely teach.

Students already feel this, which is why the RAND findings matter. In that same 2026 survey, 67% of students said using AI for schoolwork harms their critical thinking, up from 54% earlier in the year, even as their own usage kept climbing. Usage rose and concern rose together. They know the tool is doing their thinking for them. What they lack, as Grove said, is anyone teaching them the alternative. RAND researchers gave us the precise vocabulary: cognitive offloading, where AI does the thinking, versus cognitive augmentation, where AI pushes the student to think harder. The difference is not the tool. It is the assignment design. It is the teacher.

Fluency has one non-negotiable core: never ship an AI output you cannot defend. If you cannot explain why the answer is right, you do not know it. Eberly Barnes, who coordinates UC San Diego’s writing program, said it bluntly in 2025: you are not going to learn to write if you are just going to machine. The same applies to strategy, to design, to media planning. The tool can draft. Only the human can own.

What Employers Are Actually Asking For

Now the employer side, because this is where the story gets uncomfortable for universities. Pearson and AWS found that 53% of employers say their primary challenge is finding graduates with the right AI skills. Meanwhile 78% of higher education leaders believe they are meeting employer expectations. Both numbers come from the same 2026 research. That 25-point perception gap is the whole story in one statistic: institutions think they are delivering; employers say the graduates are not ready.

Kim Majerus, vice president of global education at AWS, named the actual opportunity in the research: helping translate AI tool engagement into real workplace capability. Tool engagement is what most courses produce. Workplace capability is what employers pay for. They are different products, and most of higher education is manufacturing the wrong one.

The hiring data keeps piling up. A Western Governors University survey of 3,128 US hiring professionals, published in September 2026, found that 60% say AI is making it harder to evaluate candidates’ real skills. The share of employers still figuring out how to evaluate AI skills doubled from 16% in 2025 to 32% in 2026. Nearly one in five cited difficulty confirming whether they were interviewing a human rather than an AI as a top challenge. WGU president Scott Pulsipher’s verdict: higher education and employers have to build a better proof system, credentials that carry meaning, assessments that verify what a person can do, and real-world experiences that show how they think, adapt, and use AI responsibly.

Microsoft’s 2025 Work Trend Index, based on 31,000 professionals across 31 countries plus LinkedIn data, adds the demand side. 78% of leaders were considering hiring for AI-specific roles (95% among what Microsoft calls Frontier Firms), and 67% of leaders versus 40% of employees reported familiarity with AI agents. The World Economic Forum’s Future of Jobs Report 2025, surveying 1,000+ employers representing 14+ million workers across 55 economies, projected that 39% of workers’ core skills will change or become outdated by 2030, with 63% of employers identifying skills gaps as a major barrier to business transformation. And Randstad’s analysis of 126 million postings found entry-level postings down 29% since January 2024. The jobs students are preparing for are being reshaped while they sit in the lecture hall.

What I Changed in My Own Teaching

I did not arrive at this from theory. Students kept breaking my assignments, and I kept rebuilding them better. Here is what changed.

First, I stopped grading the artifact and started grading the judgment. In advertising and media, the deliverable used to be the campaign concept or the content calendar. Now it includes the student’s written account of what the AI produced, what was wrong with it, and what they changed and why. The output matters less than the audit trail.

Second, I design assignments the AI can help with but cannot finish. If the task is generate ten taglines, the AI wins and the student learns nothing. If the task is produce the taglines, test them against three real audience insights, kill seven, and defend the surviving three with reasoning, then the AI is a sparring partner and the student does the actual learning. Real briefs, real constraints, real defense of choices.

Third, I teach the jagged frontier explicitly. Students need to know where these tools are strong and where they quietly fall apart, and the only way to learn that is supervised failure. We run the tool into its weak spots on purpose: hallucinated citations, confident nonsense, the plausible-but-wrong statistic. Then we dissect it. I would rather my students discover those limits in my classroom, where the cost is a grade, than in a client meeting, where the cost is a career.

Fourth, I brought my client work into the room. Not as war stories, as material. Real briefs with real constraints, anonymized where needed. When students work on the kind of problem a brand actually paid to solve, the theory question answers itself. The World Economic Forum’s 2026 education readiness report made the institutional case: the surrounding system determines whether AI adoption helps or harms. A classroom connected to real work is that system.

My Advice to Students, and to the People Hiring Them

If you are a student

Learn the work before you learn the tool. The students who thrive with AI are the ones who could do the task without it, slowly and imperfectly, because they understand what good looks like. AI multiplies judgment; it cannot create it. A strategist who cannot think will produce faster bad strategy. A writer who cannot write will produce faster bad writing. Build the underlying skill first, then let the tool accelerate it.

Second, make verification your signature skill. Employers told Pearson and AWS that this is the scarcest competency in graduates, and scarcity is opportunity. The graduate who can say here is what the model produced, here is what I checked, here is what I changed, and here is why I stand behind it, is the graduate who gets hired.

Third, build a portfolio of judged work, not generated work. Show the before and after: the AI draft and your improvements with reasoning. Hiring managers cannot evaluate real skills easily anymore (60% admit it, per WGU), so your proof system has to do the talking. Give them the audit trail, not just the artifact.

If you are hiring them

Stop screening for tool familiarity. Familiarity is cheap now; 92% of undergraduates have it. Screen for verification habits instead. Ask candidates to critique an AI output in the interview. You will learn more in ten minutes than from any certificate.

And stop expecting universities to deliver finished AI-fluent professionals. Only 24% of stakeholders in the Pearson and AWS research believe universities are keeping pace with AI, so companies that invest in internal AI training gain a real advantage.

That last point is where my teaching and my practice converge. The same judgment-first approach I use with undergraduates in Karachi is what I now run as workshops for company teams: less tool tourism, more supervised practice, more verification, more real briefs. The tools change faster than the syllabus, and the only durable skill is knowing how to think with the machine instead of letting it think for you.

Frequently Asked Questions

What should students actually learn about AI in university?

Students should learn judgment first and tool features second. The scarcest graduate skill, according to Pearson and AWS research from April 2026, is the ability to critically verify AI outputs, which 58% of employers rate as graduates’ weakest competency. Courses should teach where AI is reliable and where it fails, how to check its work, and how to take responsibility for the final output.

Is it cheating if students use AI for their assignments?

The better question is whether the assignment still teaches anything. A 2026 RAND survey found 62% of students use AI for homework while 67% believe it harms their critical thinking. RAND researchers distinguish cognitive offloading, where AI does the thinking, from cognitive augmentation, where AI pushes the student to think harder. The difference comes down to assignment design: tasks the AI can help with but cannot finish.

Why do employers say graduates are not AI-ready?

There is a 25-point perception gap. Pearson and AWS found in April 2026 that 78% of higher education leaders believe they meet employer expectations, while 53% of employers say their primary challenge is finding graduates with the right AI skills. Employers report that graduates can engage with AI tools but cannot translate that into real workplace capability, especially verifying outputs and applying AI to professional workflows.

What is the difference between AI literacy and AI fluency?

Literacy is knowing AI tools exist and roughly what they do. Fluency is being able to verify an output, judge its quality, catch its errors, and take responsibility for publishing it. Only 14% of current graduates report a high level of proficiency in applying AI tools to a professional workflow, according to Pearson and AWS, which shows how wide the fluency gap really is.

How can companies train their teams to use AI well?

Stop screening for tool familiarity, which 92% of undergraduates already have, and start building verification habits. Practical AI workshops that use real briefs, supervised practice, and explicit training on where AI fails produce fluency faster than tool tours. Only 24% of stakeholders believe universities are keeping pace with AI, so companies that invest in internal AI training gain a real advantage.

Do you offer AI workshops for companies and teams?

Yes, and as a corporate trainer and lecturer at Indus University and Ilma University, I run practical AI workshops for company teams focused on judgment-first AI use: verifying outputs, knowing the limits of the tools, and applying AI to real business workflows. Contact me through the contact page to discuss a workshop for your team.