5 AI Safety Gaps Most Professionals Don't Know They Have (2026 Data)

Here is one number from 1,017 measured AI assessments that should change how organisations think about AI deployment:

36% of professionals have no functional AI safety practice.

Not "they do not follow best practices." They have no practice at all. No awareness of data boundaries. No process for adjusting scrutiny based on stakes. No consideration of what happens downstream when AI output is wrong.

These are not people avoiding AI. They are active users — professionals who have integrated ChatGPT, Copilot, or Claude into their daily workflow. They just have not thought about what can go wrong.

Here are the five measurable safety gaps the data reveals, ranked by how common they are.

1. No Verification Process for AI Output (20% of Professionals)

One in five professionals has no systematic process for checking AI output before using it. They accept what AI gives them, or check sporadically based on gut feeling.

At the top end, professionals scoring 7+ on Output Evaluation (T1) have built verification into their workflow architecture — cross-referencing with primary sources, automated checks, human-in-the-loop gates at decision points. At the bottom, there is no check at all. The same level of trust goes to a brainstorm draft and a client deliverable.

Why it matters: AI output that is confidently wrong is the most dangerous kind. Without a verification step, errors propagate into decisions, documents, and products with no friction.

2. Headline-Only Knowledge of AI Limitations (36% of Professionals)

More than a third of professionals know AI's limitations only from what they have read — hallucination, bias, confidentiality risks — not from anything they have personally tested, experienced, or built defences against.

Only 17% score Proficient or above on Limitation Awareness (T2), meaning they can predict failure before it happens and adjust their approach preemptively. The rest are reactive at best: they recognise an AI failure after it has occurred, if they recognise it at all.

Why it matters: Headline knowledge creates a false sense of safety. "I know AI can hallucinate" is not the same as "I know when AI is likely to hallucinate in my workflow and what I check for."

3. No Awareness of Data Boundaries (Part of the 36% With No Safety Practice)

The Safety & Responsibility criterion (S1) measures whether professionals think about what data should and should not be shared with AI tools. At the Developing level (score 3–4), professionals are aware of obvious risks but have not thought about subtler issues: what happens when they paste proprietary code into ChatGPT, share customer data with an AI summariser, or generate content that reflects training data biases.

Professionals scoring 1–2 on S1 use AI output without any scrutiny and have no awareness of data risks at all. Combined, 36% of professionals fall into this functional gap.

Why it matters: Data shared with AI tools may be logged, used for training (depending on the provider and tier), or reflected in outputs seen by other users. Without clear personal boundaries, sensitive data leaks are a matter of time.

4. Safety Gaps Are Worst in High-Exposure Roles

The role-by-role breakdown reveals that the professionals most likely to create AI content for external audiences are the ones with the weakest safety awareness:

Marketing & Content has the largest safety gap of any role — 15 points below their overall average. They use AI extensively, generate customer-facing content, and are the least aware of the risks.

Engineers have a meaningful gap too (11 points). They understand how the tools work technically, but that understanding has not translated into proportional caution.

Why it matters: Safety gaps in high-exposure roles have the highest blast radius. A marketing team publishing AI-generated content without safety awareness affects customers. An engineering team deploying AI features without safety checks affects users.

5. The Expert Paradox: Capability and Caution Are Correlated

The most capable AI users are also the most cautious — and the data shows this is not coincidence but causation.

Experts (score 92+) average 88/100 on Safety & Responsibility, compared to the population average of 45. That is nearly double. The same pattern holds for Technical Understanding: 87 vs 45.

Understanding how AI works (Technical Understanding) and knowing where it fails (Limitation Awareness) are prerequisites for responsible use. The safety gap is fundamentally a literacy gap — professionals who do not understand AI cannot be cautious about risks they cannot see.

This is why training programmes focused only on "AI safety dos and don'ts" are insufficient. Safety awareness requires understanding, and understanding requires AI literacy.

The Compliance Angle: EU AI Act Article 4

The EU AI Act is not hypothetical. Article 4 — requiring employers to ensure "sufficient AI literacy" for all staff interacting with AI systems — is already in force. Enforcement begins August 2026.

With 36% below a functional safety threshold and 44% unable to explain AI fundamentals, the compliance gap is measurable and specific. Organisations need evidence that their workforce understands AI risks — not a checkbox, not a survey, but demonstrated competence.

What Organisations Can Do

1. Measure your team's actual AI safety awareness. Not their confidence — their demonstrated practice. Self-reported surveys consistently overestimate capability. A team assessment gives criterion-level data for every employee.
2. Prioritise high-exposure roles. Marketing, customer support, and content teams generate AI output that reaches customers. Their safety gaps have the highest blast radius.
3. Make verification a workflow step, not a mindset. "Be careful with AI" is not actionable. "Before sharing any AI-generated content externally, verify X, Y, and Z" is. Build the check into the process, not the person.

Frequently Asked Questions

What percentage of professionals have no AI safety practice?

36% of professionals score below functional on AI Safety & Responsibility, based on 1,017 measured assessments. This means they have no awareness of data boundaries, no process for adjusting scrutiny based on stakes, and no consideration of downstream impact.

Which job role has the worst AI safety score?

Marketing & Content professionals have the lowest Safety score (37.2 out of 100) of any role — a 15-point gap below their overall AI score. This makes them the highest-risk group for unsupervised AI content generation. Full role breakdown: AI Skills by Job Role.

Does the EU AI Act require AI literacy?

Yes. Article 4 of the EU AI Act — already in force, enforcement beginning August 2026 — requires employers to ensure "sufficient AI literacy" for all staff interacting with AI systems. The Act does not define a specific score threshold, but AISA's data provides a measurable baseline for compliance.

Are AI safety skills correlated with overall AI ability?

Yes, strongly. Experts (score 92+) average 88 on Safety vs the population average of 45. The correlation runs through Technical Understanding — professionals who understand how AI works are better equipped to recognise its risks. Safety training without foundational AI literacy is less effective.

---

Related reading: The State of AI Literacy 2026 — the full report on AI literacy gaps. 44% of Professionals Can't Explain How AI Works — the knowledge deficit behind the safety gap. Top 10 AI Certifications in 2026 — how to prove and build your AI skills.