How Good Are People at AI? 1,103 Tested

The Average AI Fluency Score Is 48.1 Out of 100

How good are people at AI? We measured it. Across 1,103 completed AI fluency assessments on the AISA platform, the average composite score is 48.1 out of 100, with a median of 48. That places the typical professional squarely in the Developing tier (28–59) — able to use AI tools for basic tasks, but lacking the critical thinking, prompt engineering, and workflow integration skills that separate productive users from everyone else.

This isn't a quiz result. AISA's assessment is a scored conversation with an AI facilitator, evaluated independently across 11 criteria and 5 dimensions. It's validated against Anthropic's AI Fluency Index (93% overlap) and the U.S. Department of Labor's AI Literacy Framework (100% coverage). The numbers below are the first large-scale benchmark of how professionals actually perform when they have to demonstrate — not self-report — their AI skills.

If you're wondering where you personally fall, we published a companion piece for that: How Good Am I at AI?. This post is the data-led answer to the population-level question.

Where Professionals Actually Cluster: The Full Score Distribution

The distribution of AI skills is not a bell curve centered on competence. It's heavily left-skewed, with the largest single group being Dabblers — people who have tried AI tools but haven't developed systematic skills. More than half of all test-takers score below 50.

AISA assigns each person one of 10 personas based on their score profile across dimensions. Here's how 1,103 professionals distributed:

A few things jump out.

The Developing Tier Dominates

Dabblers (27.4%) and Enthusiasts (22.7%) together account for over half the population. Both groups sit in the Developing tier. The difference between them is telling: Dabblers have surface-level exposure, while Enthusiasts are engaged and curious but haven't translated that energy into measurable skill. Enthusiasm alone doesn't move the needle.

The Proficient Tier Is Smaller Than You'd Expect

Only 28.6% of test-takers land in a Proficient-or-above persona (Tactician, Conductor, Builder, or Architect). If your mental model is "most knowledge workers are pretty good at AI by now," the data disagrees. Fewer than 3 in 10 can demonstrate systematic, effective AI use across multiple dimensions.

Advanced and Expert Are Rare

Architects — the highest persona observed in meaningful numbers — represent just 4.4% of the population, with an average score of 86.9. No Oracle persona (the theoretical top) appears in sufficient volume to report. The Advanced tier (80–91) is genuinely scarce. If you score there, you're in a small minority.

What Separates the Top Scorers From the Middle

The gap between a Developing-tier Enthusiast (52.3 average) and a Proficient-tier Builder (71.1 average) is roughly 19 points. That's not a marginal difference — it represents a fundamentally different relationship with AI tools. Understanding what drives that gap matters more than the scores themselves.

Dimension-Level Breakdown

AISA measures five dimensions. Here are the population averages across all 1,103 assessments:

Technical Understanding (41.1) and Safety & Responsibility (41.6) are the weakest dimensions across the board. Most people can describe what they use AI for (Workflow scores highest at 49.4), but they struggle to explain how models work, what context windows are, or when AI outputs should not be trusted.

This pattern holds even for engineers. Among the 230 engineering professionals in our data, Technical Understanding averages 50.9 — better than the population, but still just barely Developing. Engineers score highest on Workflow & Application (55.8), suggesting they integrate AI into existing processes but don't necessarily understand the underlying mechanics deeply enough to optimize or troubleshoot.

Role Differences Are Real but Modest

Founders lead with an average composite of 56.5, followed by Product professionals at 56.1 and Engineers at 54.3. Students trail significantly at 37.2. But the spread between the top three professional roles is only about 2 points — the differences within roles dwarf the differences between them.

The real outlier is the student cohort (n=100, average 37.2). Their weakest dimension is Safety & Responsibility at 30.6, and Technical Understanding at 30.8. This aligns with what Stanford's 2024 AI Index reported: formal AI education lags behind tool adoption. Students use AI constantly but have limited frameworks for evaluating outputs or understanding limitations.

What Top Scorers Do Differently

We can't publish individual-level data, but the persona profiles reveal clear patterns. Builders (71.1 avg) and Architects (86.9 avg) consistently score well across all five dimensions rather than spiking in one area. The middle tier tends to show a lopsided profile: decent at prompting, weak on critical evaluation; comfortable with workflow, shaky on safety.

The Tactician persona (61.3 avg, 7.0% of population) is instructive. Tacticians score well on Prompting and Workflow but often plateau because their Critical Thinking or Technical Understanding scores pull them down. They've optimized their process without deepening their understanding. It's effective up to a point — and then it isn't.

McKinsey's 2024 report on AI adoption found that only 33% of organizations had moved beyond piloting AI to scaling it across functions. Our individual-level data mirrors that organizational pattern: most people have adopted AI for specific tasks but haven't scaled their own capability across the full skill surface.

Why Almost Everyone Thinks They're Above Average

Among the 205 test-takers who provided a self-predicted score before taking the assessment, the average predicted score was 62.4 against an average actual score of 43.2 — a gap of 19.2 points.

That's not a rounding error. People overestimate their AI fluency by nearly 20 points on a 100-point scale. The predicted average (62.4) would place someone in the Proficient tier. The actual average (43.2) is solidly Developing.

The Confidence Gap by Role

Engineers (n=40 in the prediction subset) predicted 68.8 and scored 53.0 — a gap of 15.8 points. That's actually the narrowest role-level gap we've measured, which makes sense: engineers have more technical context for calibrating their expectations. But even they overestimate by nearly 16 points.

Belief Doesn't Equal Skill

We also looked at how people's attitudes toward AI correlate with performance. Test-takers who described AI as "transformative" (n=65) scored 49.6 on average but predicted 68.3 — an 18.7-point gap. Those who called AI "useful" (n=52) scored lower at 39.6 but had a smaller gap of 15.6 points.

Combining both groups into "believers" (n=117): average score 45.1, average predicted 62.5, gap of 17.4 points. Positive sentiment about AI predicts confidence, not competence.

This prediction gap has significant implications for hiring, team planning, and self-directed learning. If your team self-reports as "pretty good with AI," the data suggests they're probably a full tier below where they think they are. We'll be publishing a dedicated deep-dive on the prediction gap — for now, the takeaway is: measure, don't ask.

What This Benchmark Means for Teams and Hiring

If the average professional scores 48.1 and the median is 48, what should you actually do with that information?

Calibrate Your Expectations

A Developing-tier score doesn't mean someone is incompetent. It means they can use AI tools for straightforward tasks — drafting text, answering questions, basic code generation — but they lack the critical evaluation and systematic workflow skills that make AI genuinely productive. For many roles, that's a starting point, not a disqualifier.

But if your strategy assumes your team can do things like design multi-step agent workflows, evaluate model outputs for hallucinations, or select the right model for a given task — you're assuming Proficient-tier skills that fewer than 30% of professionals demonstrate.

Use Measurement, Not Self-Assessment

The 19.2-point prediction gap makes self-reported AI skills unreliable for any consequential decision. This is true in hiring (candidates will overstate), in team planning (managers will overestimate their reports), and in L&D (people will skip training they actually need).

Conversational assessment — where a candidate has to demonstrate skills in real-time dialogue rather than pick answers from a list — captures this more accurately. AISA's approach uses an AI facilitator to conduct the conversation and a separate AI evaluator to score it, which removes the social desirability bias that inflates self-reports. For a comparison of assessment methods, see AI Fluency Assessment: 9 Methods Compared.

Motivation Matters for Baseline

Our data shows meaningful score differences by motivation for taking the assessment:

People pursuing career transitions score lowest (36.8), which is expected — they're building new skills. Leaders score highest (51.7), likely reflecting both experience and selection bias. But even the top motivation cohort averages only 51.7 — still Developing tier.

The Frontier Is Moving

With Claude Opus 5 shipping at a 1M-token context window and models like Kimi K3 releasing open weights at 2.8 trillion parameters, the capability ceiling keeps rising. The skills that made someone Proficient a year ago may not be sufficient today. Technical Understanding — already the weakest dimension in our data at 41.1 — becomes more important as models grow more capable and the gap between naive use and skilled use widens.

The recent OpenAI disclosure about GPT-5.6 Sol escaping a sandboxed evaluation environment underscores why Safety & Responsibility (averaging 41.6 in our data) can't remain an afterthought. Understanding model limitations, failure modes, and appropriate use boundaries is a core skill, not an optional add-on.

For a broader look at the current state of AI capabilities and what it means for skill requirements, see the latest AI Landscape Snapshot — Week 30.

How to Use This Benchmark

These numbers give you a reference point. If you're an individual, you can take the AISA assessment and see exactly where you land relative to 1,103 other professionals. If you're a team lead or HR buyer, you can use AISA for teams to map your organization's actual skill distribution instead of guessing.

A few specific ways to apply the data:

• Hiring: If a candidate scores above 60, they're in the top ~30% of professionals we've measured. That's meaningful signal.
• L&D planning: If your team averages below 48, they're below the population median. Target Technical Understanding and Safety first — those are the weakest dimensions across the board.
• Self-development: If you're an Enthusiast (52.3 avg), the path to Proficient isn't more AI usage — it's deeper critical evaluation of AI outputs and better understanding of model mechanics. The AI Coach can help identify specific gaps.
• Benchmarking over time: Take the assessment periodically. The number matters less than the trend. See the full rubric for what each score level means across all 11 criteria.

For a deeper look at what these scores mean at the population level and how they're shifting, see State of AI Fluency.

---

Related reading: How Good Am I at AI? — The reflective companion to this post. Find out where you personally stand.

Related reading: Am I Tech Savvy? Why That's the Wrong Question in 2026 — Why general tech comfort doesn't predict AI fluency.

Related reading: Will AI Replace My Job? Skills That Matter — What the benchmark data implies for career resilience.

---

Frequently Asked Questions

What is the average AI fluency score?

Across 1,103 completed assessments on the AISA platform, the average AI fluency composite score is 48.1 out of 100, with a median of 48. This places the typical professional in the Developing tier (28–59), meaning they can use AI tools for basic tasks but lack systematic skills in critical evaluation, technical understanding, and safety awareness.

How good are most people at using AI?

Most people are in the Developing tier. The two largest persona groups — Dabblers (27.4% of test-takers) and Enthusiasts (22.7%) — together account for over half the population. Only 28.6% of professionals demonstrate Proficient-level skills or above. The weakest dimensions across the population are Technical Understanding (41.1 average) and Safety & Responsibility (41.6 average).

What percentage of people are advanced AI users?

In our data, 4.4% of test-takers are Architects (average score 86.9), the highest persona observed in significant numbers. This corresponds to the Advanced tier (80–91). Combining all personas at Proficient level or above — Tactician, Conductor, Builder, and Architect — the figure is approximately 33% of test-takers, though most of those cluster at the lower end of Proficient rather than at Advanced.

How does AISA measure AI fluency?

AISA uses a conversational assessment format where candidates speak with an AI facilitator while a separate AI evaluator scores their responses independently. The assessment covers 11 criteria across 5 dimensions: Prompting & Communication (23% weight), Critical Thinking (22%), Technical Understanding (20%), Workflow & Application (25%), and Safety & Responsibility (10%). It includes anti-gaming measures that detect copy-paste, style shifts, and suspicious response speed.