How Good Am I at AI Safety? [2026 Data]

If you're wondering about your AI safety skills, you're already ahead of most people — because most people never ask. Across 2,470 completed AI fluency assessments, Safety & Responsibility is the lowest-scoring dimension for nearly every role. The average safety score is 41.5 out of 100, trailing even Technical Understanding (39.2 on its own scale, but safety underperforms relative to expectations). This post gives you the data, a quick self-check, and a clear path forward.

Why AI Safety Is the Skill Nobody Thinks About

AI safety skills are the most neglected competency in professional AI use because they don't produce visible output. Nobody sees the data leak you prevented or the bias you caught before it shipped. The work is invisible — until it isn't.

Three forces drive this blind spot:

Safety Doesn't Feel Like a "Skill"

Prompting produces a draft. Workflow automation saves hours. Critical thinking catches a hallucination. Each of these has a tangible before-and-after. Responsible AI use, by contrast, is about what you don't do: you don't paste client PII into a public model, you don't ship AI-generated legal text without review, you don't deploy a customer-facing agent without a fallback. The absence of a mistake doesn't feel like competence.

Training Programs Skip It

Most AI upskilling programs focus on prompting and tool adoption. A 2024 UNESCO report found that only 15% of corporate AI training programs included a dedicated module on responsible AI use. The emphasis is on productivity gains, not risk management. When safety is covered, it's often a single slide about "being careful" — not a structured framework for data classification, PII handling, or stakes-based review.

Overconfidence Masks the Gap

Across 1,560 AISA candidates who predicted their own scores before taking the assessment, the average prediction gap was 17.7 points — people expected a 62.3 but scored 44.6. Students showed the widest gap at 31.4 points. When you think you're already competent, you don't seek out training on the dimension you're weakest in. The Dunning-Kruger effect is alive and well in AI safety.

The DOJ's newly launched Super Intelligence Task Force (November 2026) signals that regulators are paying attention to AI misuse — even as most professionals haven't built the habits to avoid it. Goldman Sachs recently pulled AI coding tools from traders, citing risk concerns. The gap between institutional caution and individual skill is widening.

What the Average AI Safety Score Actually Is

The average safety score across 2,470 AISA candidates is 41.5 out of 100 — the lowest of all five assessed dimensions. Here's how it compares:

Technical Understanding scores lower in absolute terms (39.2), but that dimension tests knowledge of model architectures, tokenization, and inference — topics many non-technical professionals wouldn't claim to know. Safety is different. Most people believe they handle AI responsibly. The data says otherwise.

What Safety & Responsibility Actually Measures

AISA's published rubric evaluates safety across specific behaviours: whether you can identify when data shouldn't be shared with an AI system, whether you understand the downstream consequences of AI-generated outputs, whether you apply appropriate review processes based on the stakes of a task, and whether you can articulate the ethical boundaries of AI use in your context. It's not a quiz about AI policy — it's a conversation where the evaluator looks for evidence that you practice these things.

Why 41.5 Matters

A score of 41.5 places the average candidate squarely in the Developing tier (28–59). That means most professionals can identify obvious safety concerns — "don't paste passwords into ChatGPT" — but struggle with nuanced scenarios: Should you use AI to draft a performance review? What's the right level of disclosure when AI contributed to a client deliverable? When does a "low-stakes" task become high-stakes because of who reads the output?

These are the questions that separate a Competent from a Proficient safety practitioner.

AI Safety Scores by Role: Who's Worst, Who's Best

Designers score lowest on AI safety at 35.8, and Students trail at 30.9, while Founders lead at 50.2 — still barely above the midpoint. No role group clears the Proficient threshold.

Every single role scores lower on safety than on their overall composite. The gap column tells the real story.

Why Designers Have the Widest Safety Gap

Designers show a 12.0-point gap between their safety score (35.8) and their overall composite (47.8) — the largest of any role. Design workflows tend to emphasise speed and iteration: generate concepts, refine visuals, test with users. The safety questions — "What happens if this AI-generated image contains copyrighted elements?" or "Should you disclose AI involvement in client-facing design work?" — sit outside the typical design feedback loop. For a deeper analysis, see Designers' AI Skills: The Safety Gap.

Why Founders Score Highest (But Still Not High Enough)

Founders lead at 50.2, likely because they carry liability. When you're the one signing the terms of service, approving the privacy policy, and answering to investors about risk, safety stops being abstract. But 50.2 is still Developing-tier. Founders tend to understand that safety matters without always knowing how to implement it — the difference between policy awareness and operational practice. The Founders vs Engineers comparison explores this dynamic further.

The Student Problem

Students score 30.9 on safety — barely above the Emerging tier boundary of 27. This aligns with the broader pattern: students show the largest prediction gap of any role (31.4 points), suggesting they haven't yet encountered the real-world consequences that build safety intuition. Most academic AI use happens in low-stakes contexts — essays, research summaries, study aids — where the cost of a safety failure is a bad grade, not a data breach.

5 Quick Self-Check Questions for AI Safety

These five questions map to the core behaviours AISA evaluates in the Safety & Responsibility dimension. Answer honestly — nobody's scoring you yet.

1. Do You Know Your Organisation's AI Data Policy?

Not "does one exist" — can you describe what data categories are approved for use with external AI tools? If you can't name the boundary between what's allowed and what isn't, you're operating on instinct rather than policy. Anthropic's 2025 Responsible Scaling Policy notes that organisational data governance is the first line of defence against AI-related data exposure.

2. Have You Ever Checked Whether an AI Output Could Cause Harm Downstream?

This goes beyond factual accuracy. A technically correct AI-generated email could still cause harm if it's sent to the wrong audience, uses an inappropriate tone for a sensitive situation, or makes commitments the organisation can't keep. Stakes-based review means adjusting your verification effort to match the potential consequences.

3. Can You Explain When AI Disclosure Is Appropriate?

Some contexts require it (EU AI Act mandates disclosure for certain AI-generated content). Others leave it to professional judgement. If your answer is "I always disclose" or "I never disclose," you're probably not thinking about it with enough nuance. The right answer depends on the audience, the stakes, and the norms of your industry.

4. Do You Treat AI-Generated Content Differently Based on Who Will See It?

An internal brainstorming doc has different safety requirements than a customer-facing report. A draft for your own review is different from a draft that goes straight to a client. If you apply the same level of scrutiny to everything — or worse, the same level of trust — your safety practice is likely underdeveloped.

5. Have You Ever Deliberately Tested an AI Tool's Limits?

Not to break it for fun, but to understand where it fails. Do you know what happens when your preferred model encounters ambiguous ethical scenarios? Have you tried feeding it edge cases relevant to your work? According to Stanford's 2024 AI Index Report, only 22% of organisations conduct any form of red-teaming or adversarial testing on the AI tools they deploy. Individual practitioners do it even less.

Scoring yourself:

• 5 confident "yes" answers: You're likely in the Proficient range (47+). Take the actual assessment to confirm.
• 3-4 "yes" answers: Developing range. You have awareness but gaps in practice.
• 0-2 "yes" answers: Emerging to low Developing. The good news: safety skills are highly trainable.

What to Do If Your AI Safety Skills Are Low

The fastest path to improving AI safety skills is building three specific habits, not completing a course. Safety is a practice, not a knowledge domain.

Start With Data Classification

Before you use any AI tool, spend 10 seconds asking: "What category of data am I about to share?" Public information, internal-only, confidential, or restricted? This single habit eliminates the most common safety failure — accidentally exposing sensitive data to a third-party model. AISA's responsible AI use guide walks through a practical classification framework.

Build a Personal Review Threshold

Decide in advance which outputs get extra scrutiny. A useful rule: if the output will be seen by someone outside your immediate team, or if it involves numbers that someone might act on, it gets a manual review. Write this rule down. Tape it to your monitor. The 36% of professionals with no safety practice at all aren't careless — they just never established a threshold.

Learn One Framework

You don't need to memorise the entire EU AI Act or NIST AI RMF. Pick one framework relevant to your role and learn its core principles. For most professionals, the NIST AI Risk Management Framework's four functions — Govern, Map, Measure, Manage — provide enough structure to move from instinct-based safety to systematic practice.

Measure Where You Actually Stand

Self-assessment is a starting point, but the prediction gap data shows it's unreliable. Across AISA candidates, people overestimate their abilities by an average of 17.7 points. The only way to know your actual safety score is to have it measured against a published rubric with evidence-linked scoring. AISA's conversational assessment evaluates safety through your own words — not multiple choice — and the free report breaks down exactly where your safety gaps are.

The assessment takes about 25 minutes. Your safety score is reported alongside the other four dimensions, so you can see precisely how it compares to your prompting, critical thinking, technical understanding, and workflow skills. If you land in the Developing tier, you'll know exactly which safety behaviours to build. If you're already Proficient, you'll have the evidence to prove it.

For a deeper look at why safety consistently underperforms, read AI Safety Skills: The Lowest Score.

---

Related reading: Am I Falling Behind on AI? [2026 Data] — see how your overall AI skills compare to 2,470 professionals.

Related reading: AI Certification Worth It? Seekers Score 45 — what certification-motivated candidates actually score.

Related reading: Should I Learn AI Now or Wait? [2026] — the cost of waiting, measured in real data.

---

Frequently Asked Questions

What does AI safety actually test?

AISA's Safety & Responsibility dimension evaluates whether you can identify appropriate data boundaries for AI tools, apply review processes proportional to the stakes of a task, recognise downstream consequences of AI-generated outputs, and articulate when and how to disclose AI involvement. It accounts for 10% of the overall composite score and is assessed through a conversation — not a multiple-choice quiz — where every score requires evidence in your own words.

Why do designers score lowest on AI safety?

Designers score 35.8 on safety, a 12.0-point gap below their overall composite of 47.8 — the widest gap of any role. Design workflows prioritise rapid iteration and visual output, which means safety considerations like copyright risk in AI-generated images, disclosure norms for AI-assisted design, and data handling for user research inputs often fall outside the standard feedback loop. The skill isn't absent; it's just not reinforced by typical design practice.

Is AI safety the same as AI ethics?

They overlap but aren't identical. AI ethics is a broader field covering fairness, accountability, transparency, and societal impact — often at the policy or organisational level. AI safety skills, as measured by AISA, focus on the individual practitioner's behaviours: data handling, stakes-based review, disclosure practices, and consequence awareness. You can have strong ethical convictions and still score low on safety if you haven't translated those convictions into consistent working habits.