How to Improve AI Skills: 7 Steps [2026]

Most people who want to improve AI skills hit a wall they don't even recognise. Across 2,196 completed AISA assessments , the average composite score is 46.1 out of 100 — squarely in the Developing tier. Nearly four in ten people land as either a Dabbler (31.5%) or a Copy-Paster (7%), personas defined by surface-level tool use and limited critical engagement. The gap between where people think they are and where they actually score averages 18.8 points.

This post breaks down exactly where the sticking points are, which dimensions matter most, and gives you seven concrete actions to move up. No vague advice about "staying curious." Every recommendation maps to a measurable dimension of AI fluency.

Where Most People Actually Are

Nearly 40% of AISA test-takers fall into the two lowest-performing personas: Dabbler and Copy-Paster. These aren't beginners who've never touched ChatGPT — they're people who use AI tools regularly but haven't developed the habits that separate competent users from everyone else.

Here's the persona distribution from 2,196 assessments:

What Defines a Dabbler

The Dabbler uses AI tools — often enthusiastically — but without structure. They'll ask ChatGPT a question, accept the first response, and move on. Their average score of 26.1 puts them at the top of the Emerging tier. They typically score low on critical thinking and technical understanding because they haven't built habits around verifying outputs or understanding why a model responds the way it does.

What Defines a Copy-Paster

The Copy-Paster is slightly more engaged, averaging 30.2, but their interaction pattern is mechanical: paste input, copy output, done. They rarely iterate on prompts, almost never question the model's reasoning, and treat AI as a text-transformation utility rather than a collaborative tool. The gap between Copy-Paster and the next persona up (Sceptic, at 42.8) is 12.6 points — a meaningful jump that requires deliberate practice.

The Overconfidence Problem

Across 1,286 people who predicted their score before taking AISA, the average predicted score was 62.7 against an actual average of 43.9 — an overestimation gap of 18.8 points. Students show the widest gap at 30.9 points (predicted 66.3, scored 35.4). This matters because you can't improve what you've misdiagnosed. If you believe you're already Proficient when you're actually Developing, you won't invest in the right areas.

The 5 Dimensions That Define AI Fluency

AISA measures AI fluency across five dimensions, each weighted to reflect its importance in real-world AI use. Understanding which dimensions drag your score is the first step to knowing how to improve AI skills systematically.

Technical Understanding Is the Biggest Drag

At 39.0, technical understanding is the lowest-scoring dimension across all 2,196 assessments. This doesn't mean everyone needs to understand transformer architecture at a research level. It means most people lack a working mental model of how language models generate text, why they hallucinate, what context windows are, and how different models differ in capability. Without this foundation, you can't make informed decisions about which tool to use or why a prompt failed.

Safety & Responsibility Is Undervalued

Safety scores 41.1 on average — the second-lowest dimension. Most people don't think about data privacy when pasting proprietary information into a chat interface, can't articulate when AI use is inappropriate, and haven't considered the downstream effects of AI-generated content. This dimension carries only 10% weight, but it's a differentiator: Architects (avg 87.5) consistently score well here, while Dabblers (avg 26.1) barely register.

Workflow Scores Highest — But Still Below 50

Workflow & Application leads at 47.1, which makes sense: most people come to AI tools to get work done. But even the strongest dimension is below the midpoint. The gap between knowing that AI can help with a task and knowing how to integrate it into a repeatable workflow is where most improvement potential sits.

7 Ways to Improve Your AI Skills

Each of these maps to one or more AISA dimensions. They're ordered from foundational (start here if you're a Dabbler) to advanced (for people pushing from Proficient toward Advanced).

1. Practice Deliberate Prompting

Dimension: Prompting & Communication (23% weight)

Stop accepting the first output. The single highest-leverage habit change is treating every AI interaction as a two-turn minimum. Give your initial prompt, read the response critically, then refine. Specify format, audience, constraints, and context. Use chain-of-thought prompting when you need reasoning, not just answers.

Concrete drill: take a task you've already completed manually. Prompt an AI to do it. Compare the output to your manual version. Identify where the prompt was underspecified and rewrite it. Do this three times per week for a month.

Anthropic's research for their AI Fluency Index — which AISA cross-references at 93% marker coverage — identifies prompt specificity as one of the strongest predictors of effective AI use.

2. Build Verification Habits

Dimension: Critical Thinking (22% weight)

The difference between a Copy-Paster and a Tactician is verification. Copy-Pasters take outputs at face value. Tacticians (avg 59.8) check claims, cross-reference sources, and know when to distrust the model.

Start with a simple rule: for any factual claim in an AI output, verify at least one claim per response before using it. Use source triangulation — check the claim against two independent sources. Over time, this becomes automatic.

A 2024 study from Stanford's Human-Centered AI Institute found that users who were prompted to verify AI outputs before acting on them made 40% fewer errors in downstream tasks (Stanford HAI, "AI-Assisted Decision Making," 2024). Verification isn't overhead — it's quality control.

3. Learn How Models Actually Differ

Dimension: Technical Understanding (20% weight)

Technical understanding is the lowest-scoring dimension at 39.0 for a reason: most people treat all AI models as interchangeable. They're not. As of October 2026, the model landscape includes Claude Opus 5.5 and Sonnet 5.5 from Anthropic, GPT-6 Astra and GPT-6.1 Sol from OpenAI, and open-weight options like Mistral Large 4 with 1.05 trillion parameters. These models differ in context window size, cost, reasoning depth, and multimodal capability.

You don't need to memorise benchmarks. You need to understand three things:

• Context windows determine how much information a model can consider at once (currently ranging from 128K to over 1M tokens across leading models)
• Token economics affect whether a task is cost-effective via API vs. chat interface
• Model strengths vary: some models excel at code generation, others at nuanced reasoning, others at speed

Practical step: pick two models and run the same complex prompt through both. Compare outputs. Note where they diverge. This builds intuition faster than reading documentation.

4. Integrate AI Into Existing Workflows

Dimension: Workflow & Application (25% weight)

Workflow is the highest-weighted dimension at 25% and the highest-scoring at 47.1 — but there's still a 30-point gap between the average and what Architects achieve. The key insight: don't create new AI workflows from scratch. Instead, identify friction points in your existing processes and test whether AI reduces them.

Map your weekly tasks. For each one, ask: "Where do I spend time on a subtask that's repetitive, requires synthesis, or involves format conversion?" Those are your insertion points.

McKinsey's 2024 report on AI adoption found that organisations seeing the highest productivity gains from AI embedded it into existing processes rather than building standalone AI workflows (McKinsey Global Institute, "The State of AI in 2024," June 2024). The same principle applies at the individual level.

5. Develop a Safety Reflex

Dimension: Safety & Responsibility (10% weight)

Safety scores 41.1 on average, and for Designers it drops to 35.5. This dimension is about more than compliance — it's about knowing what not to put into a model, understanding data retention policies, and recognising when AI-generated content needs human review before it reaches an audience.

Build three habits:

1. Before pasting: ask "Does this contain PII, proprietary data, or confidential information?" If yes, redact or use a local model
2. Before publishing: ask "Could this AI output cause harm if it's wrong?" If yes, add human review
3. Before automating: ask "What happens when this fails?" Design for graceful degradation

Anthropic's updated usage policy (October 2026) now explicitly bans sustained purposeless cruelty toward AI systems and prohibits using AI for voter deception — signals that the safety dimension is becoming more, not less, important in how AI fluency is evaluated industry-wide.

6. Practice Task Decomposition

Dimensions: Prompting (23%) + Workflow (25%)

One of the clearest markers separating Enthusiasts (avg 52.0) from Builders (avg 71.0) is the ability to break complex tasks into AI-appropriate subtasks. Instead of asking a model to "write a marketing strategy," a Builder decomposes: audience analysis → competitive positioning → channel selection → messaging framework → content calendar. Each subtask gets its own prompt with specific constraints.

This is task decomposition in practice, and it's a skill that transfers across every AI tool. The Gartner 2024 Hype Cycle for AI noted that prompt engineering is evolving from single-turn interactions toward multi-step orchestration, with task decomposition as the foundational skill (Gartner, "Hype Cycle for Artificial Intelligence," 2024).

Drill: take your most complex recurring task. Break it into five or more subtasks. Prompt each one separately. Compare the assembled result to what you'd get from a single monolithic prompt.

7. Run Before-and-After Measurements

All dimensions

Improvement without measurement is guesswork. The people who move tiers — from Dabbler to Enthusiast, from Enthusiast to Tactician — are the ones who know their starting point and can track change.

Take a baseline assessment. Practice deliberately for 4-6 weeks using the habits above. Reassess. Look at which dimensions moved and which didn't. Adjust your practice accordingly.

AISA's results are stable: 98% of people who retake land in the same or an adjacent tier, with scores correlating at r = 0.79 between attempts. That means score changes reflect genuine skill development, not noise.

How to Measure Progress on AI Skills

The most effective way to improve AI skills is to pair deliberate practice with periodic measurement. Without a baseline, you're guessing at your weaknesses. Without reassessment, you can't confirm that your practice is working.

Establish Your Baseline

Take the AISA assessment — it's a conversation-based evaluation that scores you across all five dimensions and assigns you one of ten personas. You get a free report showing your composite score, dimension breakdown, and specific areas for improvement. The assessment is cross-referenced against Anthropic's AI Fluency Index (93% marker coverage) and the U.S. DOL AI Literacy Framework (100% coverage), so the dimensions map to industry-recognised competency models.

Track Dimension-Level Change

Don't just watch your composite score. The dimension breakdown is where the actionable insight lives. If your technical understanding is at 31 (the student average) but your workflow is at 47, you know exactly where to focus. The AISA rubric is published — you can see exactly what each score band means for each criterion.

Set Realistic Timelines

Moving from Dabbler (avg 26.1) to Enthusiast (avg 52.0) is a 26-point jump. That doesn't happen in a weekend. Based on the patterns we observe in retake data, people who practice deliberately across multiple dimensions typically see meaningful movement within 4-8 weeks. The key is consistency: three focused AI practice sessions per week beats one marathon session.

Role-Specific Starting Points

Different roles have different dimension profiles. Knowing your role's typical pattern helps you prioritise.

Engineers score highest on technical understanding (50.8) but lag on safety (47.5). Designers have the lowest safety scores of any role (35.5). Students score lowest across every dimension, with technical understanding at 31.1. If you're a designer, prioritise safety and critical thinking. If you're a student, start with prompting fundamentals — it's the gateway dimension.

For a deeper look at how AI proficiency levels map to real-world capability, the tier definitions explain what each band actually means in practice.

The Compound Effect of Small Improvements

Here's what makes AI skill development different from most professional development: the feedback loop is immediate. You write a better prompt, you get a better output, you see the difference in seconds. That tight loop means deliberate practice compounds faster than in domains where feedback is delayed by weeks or months.

The seven steps above aren't a curriculum — they're habits. The people who reach Builder (avg 71.0) and above aren't necessarily more technical or more experienced. They've built reflexes: verify before trusting, decompose before prompting, measure before claiming improvement.

Start with your weakest dimension. Practice it for two weeks. Then measure where you stand.

---

Related reading: AI Fluency Benchmarks: 9 Personas [2026] — see how each persona scores across all five dimensions.

Related reading: Students Overestimate AI Skills by 32 Pts — the overconfidence gap is widest for early-career professionals.

Related reading: Is AI Making Us Dumber? What the Research Says — why verification habits matter more than ever.

Frequently Asked Questions

How long does it take to get good at AI?

There's no single timeline, but patterns from AISA retake data suggest that deliberate practice across multiple dimensions produces measurable score movement within 4-8 weeks. Moving from Dabbler (avg composite 26.1) to Enthusiast (avg 52.0) requires building new habits around prompting, verification, and workflow integration — not just more hours of tool use. Consistency matters more than intensity.

What AI skill should I learn first?

Start with prompting and communication. It's the gateway skill — weighted at 23% of the AISA composite and the dimension where small improvements produce the most visible results. Once you can write specific, constrained prompts and iterate on outputs, the other dimensions (critical thinking, technical understanding, workflow integration) become easier to develop because you have a stronger foundation for interacting with any model.

Can I improve my AI score?

Yes. AISA scores reflect demonstrated skill, not fixed aptitude. The assessment's test-retest stability (98% of retakers land in the same or adjacent tier, r = 0.79 between attempts) means that when your score changes, it reflects genuine improvement rather than measurement noise. Take a baseline assessment, practice the specific dimensions where you scored lowest, and reassess after 4-8 weeks of deliberate practice.