AI Skills for Students: 222 Assessed [2026]

Students are the lowest-scoring group in the AISA dataset — and the most overconfident. Across 222 assessed students, the average composite score is 36.4 out of 100, placing the typical student squarely in the Developing tier. That's 9.9 points below the population average of 46.3. When students predict their own scores before the assessment, they estimate 66.7 — a 31.4-point gap between expectation and reality. No other role group comes close.

This post breaks down the full AI skills assessment data for students: where they score, where they struggle, why the overconfidence gap matters, and what educators and hiring managers should take from it.

Student Sample Overview

The student cohort in the AISA dataset includes 222 completed assessments with scored results across all five dimensions. Of those, 145 provided a self-predicted score before beginning the assessment, enabling the overconfidence analysis below.

Students in this sample are self-selecting — they chose to take AISA, which means they likely have at least some interest in AI. This isn't a random cross-section of every university student. It's a group that was curious enough, or motivated enough, to sit a 20-minute conversational assessment. That makes the results more striking, not less: even the students who show up to be measured are scoring well below the professional average.

How AISA Measures Students

AISA uses a conversational format — candidates talk to an AI facilitator, and a separate AI evaluator scores independently against a published rubric. Every score requires evidence in the candidate's own words. There are no multiple-choice shortcuts. This matters for students because it eliminates the ability to pattern-match or guess. You either demonstrate AI fluency or you don't.

Who Are These Students?

The dataset doesn't break down by university, major, or year of study. What we know is that they self-identified as students when taking the assessment. Some may be undergraduates, others postgraduates. The sample is large enough (n=222) to draw meaningful comparisons against other role groups.

Composite Score: Students vs. the Population

The student average composite of 36.4 sits 9.9 points below the overall population average of 46.3. To put that in context, the population average itself isn't high — it falls in the Developing tier (28–59). Students are in the lower half of that tier.

Here's how students compare to every other role group in the dataset:

Every professional group outscores students by at least 11 points. Product managers and founders lead by nearly 19 points. Even designers — the next-lowest group — score 11.4 points higher.

This gap isn't explained by age alone. Career changers, who are often early in their professional journeys, score 36.8 on average (per the motivation data for career transitions). That's comparable to students, which suggests that work experience with AI tools, not just years of experience, is the differentiator. Professionals who use AI in their jobs — even imperfectly — build fluency that students haven't yet had the chance to develop.

For a deeper look at how AI fluency levels map to real capability, the tier breakdown is worth reviewing.

Dimension Breakdown: Where Students Are Weakest

AISA scores across five dimensions, each weighted differently. Students underperform the population average in every single dimension, but the gaps aren't uniform. The weakest areas — technical understanding and safety — reveal specific blind spots.

Technical Understanding: 30.9

This is the students' lowest dimension. A score of 30.9 maps to the upper end of the Novice–Developing boundary at the criterion level. It means the typical student struggles to explain how large language models work, what tokens are, why context windows matter, or how temperature settings affect output. They use AI tools without understanding the mechanics underneath.

This isn't surprising. Most university curricula haven't caught up. A Stanford HAI report from 2024 found that fewer than 10% of U.S. universities had formal AI literacy requirements for non-CS students. Unless a student is in a computer science or data science program, they're unlikely to encounter these concepts in coursework.

For students looking to close this gap, the How AI Works guide covers the foundational concepts AISA tests.

Safety & Responsibility: 31.0

Nearly as low as technical understanding. Students score 10.5 points below the population average on safety. This dimension covers data privacy, AI bias awareness, responsible deployment, and understanding when AI shouldn't be used.

The pattern we observe is that students often treat AI tools as neutral utilities — like a calculator or search engine. They don't think about what data they're sharing, whether the output could cause harm, or what the ethical implications of their use case might be. This is consistent with findings from our broader safety data analysis, where safety is the second-lowest dimension across the entire population.

Workflow & Application: 35.2

The largest absolute gap (−12.0) is in workflow. This dimension measures whether someone can integrate AI into real tasks: decomposing problems, chaining tools, evaluating when AI adds value versus when it doesn't. Students score 35.2 here.

This makes sense. Workflow fluency comes from doing real work with AI — shipping code, writing reports, managing projects, analyzing data. Students have fewer opportunities for this kind of applied practice. They may use ChatGPT for essays or homework, but that's a narrow slice of what workflow fluency actually means.

Prompting and Critical Thinking

Prompting (36.6) and critical thinking (34.7) are also below average, but the gaps are slightly smaller. Students can generally form a prompt and get a response. What they lack is the ability to iterate, constrain, and evaluate. They accept first-pass outputs without checking for hallucinations, don't structure multi-step prompts, and rarely verify claims against external sources.

The Critical Thinking with AI guide covers the verification habits that separate Developing from Proficient scorers.

The Overconfidence Problem: Predicted 66.7, Actual 35.3

This is the most important finding in the student data. Of the 145 students who predicted their score before taking AISA, the average prediction was 66.7 — solidly in the Proficient tier. The average actual score was 35.3. That's a 31.4-point gap.

To understand how extreme this is, compare it to every other role group:

Students' gap is nearly double the population average of 17.7 and more than four times the gap for founders (7.3). We've written about this pattern in detail in our overconfidence analysis.

Why Students Are So Overconfident

Several factors contribute:

Exposure ≠ fluency. Students are heavy users of AI tools. A 2024 UNESCO survey of 4,700 students across 74 countries found that over 80% had used generative AI tools. But using ChatGPT to draft an essay is not the same as understanding how to evaluate, iterate, and responsibly deploy AI outputs. Frequency of use creates a false sense of mastery.

No feedback loop. In professional settings, bad AI outputs have consequences — a flawed analysis gets challenged in a review, a hallucinated code snippet breaks in production. Students rarely face this kind of feedback. If ChatGPT gives them a plausible-sounding answer, they often have no mechanism to discover it was wrong.

The Dunning-Kruger dynamic. This is a textbook case of confidence calibration failure. Students don't know enough to know what they don't know. They haven't been exposed to the full scope of what AI fluency entails — the technical understanding, the safety considerations, the workflow integration — so they anchor their self-assessment on the narrow slice they've experienced.

Social signaling. Being "good at AI" carries social currency among students. There's pressure to appear fluent, which inflates self-perception.

Why This Matters

Overconfidence isn't just an ego problem. It's a learning problem. A student who believes they're Proficient (66.7) has no motivation to improve. They think they've already arrived. The 31.4-point gap means students are systematically misjudging where they need to invest effort — and universities that don't measure AI skills have no way to correct this.

For hiring managers, the implication is direct: a candidate's self-reported AI proficiency on a resume or in an interview is unreliable. The data shows that students overestimate by a factor that would place them two full tiers above their actual performance.

What This Means for Educators and Hiring Managers

The student data carries concrete implications for two audiences: universities designing curricula and employers evaluating early-career candidates.

For Universities and Educators

Measurement comes first. You can't improve what you don't measure. The student overconfidence gap suggests that self-assessment and informal observation are insufficient. A structured AI fluency assessment gives educators a baseline and students a reality check. Our data on AI skills for teachers and educators shows that educators themselves are still building fluency, which makes objective measurement even more important.

Technical understanding and safety need explicit instruction. These are the two lowest dimensions for students, and they're unlikely to improve through casual tool use alone. Universities need to teach how models work, what their limitations are, and how to use them responsibly — not just in CS departments, but across disciplines.

Workflow practice requires real tasks. The 12-point workflow gap won't close with lectures. Students need project-based assignments that require them to integrate AI into multi-step work: research, analysis, drafting, verification, iteration. The Using AI at Work guide outlines the task decomposition and tool-chaining skills that define workflow fluency.

A McKinsey Global Institute report from 2024 estimated that 70% of work activities could be automated or augmented by AI by 2030. Students entering the workforce without workflow fluency will be at a measurable disadvantage.

For Hiring Managers

Don't trust self-reported AI skills. The 31.4-point prediction gap is the single strongest argument for objective assessment in early-career hiring. A student who claims to be "proficient with AI tools" may genuinely believe it — and be wrong by 31 points.

Use AI fluency as a differentiator, not a filter. Most students score in the Developing tier. Filtering for Proficient or above would eliminate the vast majority of candidates. Instead, use assessment data to identify candidates who are self-aware about their gaps and motivated to improve. A student who scores 40 but accurately predicted 42 is a better bet than one who scores 35 but predicted 70.

Invest in onboarding. If your early-career hires are arriving with Developing-tier AI skills, your onboarding program needs to account for that. The AI certification pathway gives new hires a structured way to build and demonstrate fluency in their first months.

For a comparison of how other professional groups perform, the founders vs. engineers breakdown shows what experienced professionals look like on the same rubric.

3 Steps Students Can Take Now

The data is clear about where students fall short. Here's what to do about it.

1. Get an Objective Baseline

Take an AI fluency assessment and find out where you actually stand. The AISA report breaks your score into five dimensions and eleven criteria, so you'll know exactly which areas need work. The free report includes your persona assignment and a comparison to the population. Knowing your real score — not your predicted score — is the prerequisite for everything else.

2. Close the Technical Understanding Gap

This is the lowest-scoring dimension for students (30.9). Start with the fundamentals: how language models generate text, what tokens are, how context windows work, what temperature and top-p settings do. You don't need to become an ML engineer. You need to understand the tool well enough to use it effectively and recognize when it's failing. The How AI Works guide is a free starting point.

3. Build Verification Habits

The critical thinking and safety gaps (34.7 and 31.0) both stem from the same root problem: accepting AI output without scrutiny. Start treating every AI response as a draft, not a final answer. Check claims against primary sources. Ask the model to show its reasoning. Look for hedging language that signals uncertainty. Run the same prompt through a second model and compare outputs. These habits are what separate a Copy-Paster from a Tactician.

For students wondering whether to start building these skills now or wait, the timing analysis makes the case with data.

How Students Compare to Career Changers

One useful comparison: career changers — people taking AISA because they're transitioning into a new field — score 36.8 on average. That's only 0.4 points above the student average of 36.4. Both groups are in the lower Developing tier.

The similarity suggests that lack of applied AI experience is the common factor, not age or education level. Career changers and students are both groups that haven't yet had sustained, professional-context exposure to AI tools. The difference is that career changers tend to have more realistic self-assessments — the student prediction gap of 31.4 is far larger than what we observe in other low-scoring groups.

This is encouraging for students in one sense: the gap is closable. Professionals who start using AI in their work move up the scoring distribution. The question is whether students will get that applied experience before or after they enter the workforce.

---

Related reading: Students Overestimate AI Skills: 31-Point Gap — the full overconfidence analysis with prediction data.

Related reading: AI Skills for Teachers & Educators [Data] — how the people teaching students score on the same rubric.

Related reading: Career Changers AI Skills [Data] — the closest comparison group to students, with similar composite scores.

Frequently Asked Questions

Do students know less about AI than professionals?

Yes, based on assessed performance. Students average a composite score of 36.4 across 222 assessments, compared to the population average of 46.3 across 2,475 assessments. Every professional role group in the AISA dataset — engineering (54.5), product (55.3), founders (55.3), data (50.6), and design (47.8) — outscores students. The gap is largest in workflow and application, where professionals benefit from using AI in real work contexts.

Which AI skill do students struggle with most?

Technical understanding is the weakest dimension for students, with an average score of 30.9 out of 100. This covers knowledge of how AI models work, token mechanics, context windows, and model selection. Safety and responsibility is nearly as low at 31.0. Both dimensions require conceptual knowledge that casual tool use doesn't build — they need deliberate study and structured learning.

Should universities test AI fluency?

The data makes a strong case for it. Students predict their AI skills at 66.7 on average but score 35.3 — a 31.4-point overconfidence gap, the largest of any role group. Without objective measurement, neither students nor educators know where the real gaps are. A structured AI fluency assessment gives universities a baseline for curriculum design and gives students the feedback they need to direct their learning effectively.