Learn AI /Responsible AI, Human Rights and Safety /AI safety

AI safety

Starter
Responsible AI, Human Rights and Safety
Policy officers
Managers & front offices

Also called: safety.

Detail

In five seconds

What it is

AI safety means designing, testing and using AI systems so they are less likely to cause harm, especially in situations where people could be affected.

In plain words

AI safety means making sure an AI tool is less likely to harm people and that humans stay in control where the stakes are high.

An everyday example

A city engineering office uses satellite images to mark damaged roads after heavy rain. Before repair crews are sent out, a supervisor checks a sample of the marked locations, compares them with recent field notes, and flags doubtful cases for manual review. The office also sets a rule that no repair order is issued from the map alone in areas where schools, hospitals or bridges are nearby. This helps the team avoid wasting fuel, missing dangerous damage, or sending crews to the wrong place.

Where the story stops working

A road map is usually easier to inspect than many AI systems, and the risks are often more visible. Some AI systems generate text, rankings or recommendations where mistakes are harder to spot, can spread quickly, and may affect rights or safety in less obvious ways. Also, safety work includes design, testing and monitoring before and after deployment, not only checking results at the end.

Remember thisAI safety is the continuing work of finding possible harms, testing failure modes, setting limits, keeping appropriate human oversight and responding when problems emerge.

This term is contested

The term is used in more than one way. In many practical governance settings, it means reducing harms from current AI systems. In some research communities, it also refers to longer-term risks from highly capable systems.

Understand it

Why it matters

Non-specialists often meet AI as a convenient tool, but convenience can hide risk. A system can sound confident while being wrong, miss important cases, produce harmful advice, or push people to trust it too much. In low-risk tasks, that may waste time. In higher-risk tasks, it can affect safety, rights, services, reputation or access to help. AI safety helps organisations ask practical questions before relying on a system: What could go wrong? Who could be harmed? What checks are needed? When must a human make the final call? These questions support better decisions about whether to use AI, where to limit it, and when not to use it at all.

How AI safety reduces harm

After viewing this diagram, a reader should be able to describe AI safety as a process: identify risks, add controls, test, use with oversight, and monitor for harm.

01
1. Identify possible harms
List who could be affected, what could go wrong, and which cases are high risk.
Show a rectangular box with small icons for people, warning sign and checklist; include a short subtitle: 'Users, non-users, rights, safety, errors'.
02
2. Add limits and safeguards
Set guardrails, restrict risky uses, and define when humans must review or approve.
Show a box with sliders, a shield outline and a hand symbol; add a note: 'No automatic action in critical cases'.
03
3. Test normal and edge cases
Check performance on routine cases and difficult cases, including failures, bias and misuse attempts.
Show a box with three test cards labelled 'routine', 'edge', and 'misuse'; connect them into the same box.
04
4. Use with human oversight
Deploy the system only within approved conditions, with trained people able to question, stop or override it.
Show a box with a screen and a person beside it; add a visible stop button icon.
05
5. Monitor and improve
Track incidents, complaints and near misses, then update controls or stop use if harms appear.
Show a box with a trend line, alert bell and loop-back arrow to step 1.

A left-to-right five-step diagram showing how an AI system moves from idea to monitored use. Read each box in order. Each step adds a way to reduce the chance of harm, with special attention to high-risk cases and human review.

On hover or tap, each step reveals one example of a possible harm and one matching safety control.

Describe this without the diagram

This diagram presents AI safety as five ordered steps. First, identify possible harms by asking who could be affected and what could go wrong. Second, add safeguards such as guardrails, restricted uses and rules for human approval in risky cases. Third, test both ordinary and difficult situations, including misuse attempts and possible bias. Fourth, use the system only with human oversight, so a person can question, stop or override it. Fifth, monitor real-world use for incidents and near misses, then improve the system or stop using it if needed. The main message is that safety is an ongoing process, not a single test.

With reduced motion: A static five-box horizontal sequence with arrows between boxes, showing the same steps and labels without animation.

AI safety focuses on preventing harm from AI systems before, during and after use. Harm can include wrong answers in high-stakes settings, unsafe instructions, discriminatory outputs, privacy breaches, or people trusting the system too much. The aim is not only to make a system useful, but also to reduce the chance that it causes damage.

What goes in includes training data, design choices, rules for use, and the task given to the system. What happens includes testing, monitoring, and adding controls such as guardrails, which are limits or checks intended to reduce risky behaviour. What comes out can create risks that people need to manage, depending on the output, use and context.

Safety is often confused with AI security and AI ethics. AI security protects systems from attacks, theft or unauthorised access. AI safety asks whether a system could cause harm, including when there is no attack. AI ethics examines whether AI use is fair, safe, accountable and respectful of human rights. AI safety focuses more specifically on practical measures to identify and reduce harmful outcomes.

In practice, safety usually means keeping humans involved, testing edge cases, limiting what the system can do, and planning what to do if it fails. For higher-risk uses, safety is not a one-time check. It is an ongoing process of identifying risks, reducing them, watching for problems, and stopping use if needed.

How “Checking a road map before sending repair teams” maps onto the real thing

In the story	In the AI system
Satellite-based road markings	An AI system producing outputs from input data
Supervisor checking a sample	Human oversight and review of model outputs
Flagging doubtful cases	Escalating uncertain or high-risk results for manual assessment
Rule against acting on the map alone near critical sites	Safety guardrails and limits on automated action

A worked example

At a regional hub, an information management officer prepares a note for the file before a pilot of an AI assistant that routes public enquiries to the appropriate service.

Starting from: The proposed workflow, approved public information and a set of synthetic test enquiries covering routine, ambiguous, urgent and deliberately misleading requests.

The officer maps foreseeable harms, including urgent requests being treated as routine, incorrect referrals, unsafe generated advice and poorer performance in some languages.
The pilot is tested with synthetic cases, including misspellings, incomplete information, unusual phrasing and requests that fall outside the system's intended scope.
The workflow blocks the assistant from giving case-specific advice, limits it to approved information and sends urgent, uncertain or out-of-scope enquiries to a duty officer.
Before launch, the office assigns responsibility for reviewing flagged cases, recording incidents and suspending the assistant if a serious failure appears.
After deployment, staff review error patterns and user complaints, then adjust the routing rules and test the changes before release.

Result: The assistant may reduce routine sorting work, but the safety comes from identifying hazards, restricting what the system can do, testing failure modes, retaining human oversight and monitoring what happens after launch.

In a United Nations workplace

In an evaluation office, a budget officer prepares talking points for the chief of section before informals on a proposal to procure an AI tool that would sort evaluation recommendations and estimate follow-up costs. The vendor has not shown how it handles unusual cases or how errors would be reported after deployment.

AI safety requires more than checking whether the tool works on typical files. The proposal should address foreseeable errors and misuse, testing on ordinary and edge cases, access limits, human review of consequential outputs, incident reporting and a way to suspend the tool if monitoring reveals harm.

Before you try thisEvaluation records may contain personal, confidential or procurement-sensitive material. They must not be entered into a tool that has not been approved for that material. Cost estimates or classifications produced by the system should not determine budget decisions without review by the responsible officer and the usual clearance.

Where you may hear it

A vendor demo before an ICT governance board
A procurement unit's technical evaluation meeting
A go or no-go review before a country office pilot
An internal audit discussion of controls after an incident
An inter-agency working group reviewing safeguards for a shared tool

Avoid confusion

Often confused with

Side-by-side

AI ethics vs AI safety vs AI governance
Three committees, three vocabularies, one topic.

Common misconceptions

Common misconception: If an AI system works well most of the time, it is safe.
Correction: Safety is about the consequences of failure, not only average performance. A system that is usually correct may still be unsafe if it fails in high-risk cases, fails unpredictably, or encourages people to rely on it without proper checks.
Common misconception: AI safety only matters for future advanced systems, not for tools used today.
Correction: Current systems already create safety issues through harmful advice, false information, discriminatory outputs, privacy risks and over-reliance by users. Everyday deployment choices, limits and review processes are part of AI safety now.
Common misconception: Safety and security are the same thing.
Correction: They overlap but are not identical. Security focuses on protection against attacks, theft and unauthorised access. Safety asks whether the system can cause harm even if no one attacks it, for example through errors, misuse, poor design or lack of oversight.

Quick check

Which action best reflects AI safety in practice?

Go deeper

The simple version

AI safety means making sure an AI tool is less likely to harm people and that humans stay in control where the stakes are high.

The technical version

AI safety concerns the prevention, reduction and management of harmful outcomes from AI systems across the lifecycle. This includes identifying hazards, evaluating foreseeable misuse and failure modes, adding safeguards such as guardrails and access limits, testing on ordinary and edge cases, monitoring incidents after deployment, and ensuring human oversight where impacts may be significant. Depending on context, harms may involve physical safety, psychological harm, rights impacts, discriminatory effects, privacy breaches, unsafe instructions, operational disruption or automation bias. The term is often used broadly, and in some communities it also refers to longer-term risks from highly capable systems; in everyday governance and deployment work, it usually refers to practical measures that make current systems safer to use.

Reviewed external resources

Primary sourceRecommendation on the Ethics of Artificial Intelligence (opens in a new tab)UNESCO
This UNESCO recommendation gives a policy and governance framework for ethical AI and suits policy, leadership and public-sector readers.
Intermediatelink checked 2026-07-31
Primary sourceB-Tech Project (opens in a new tab)UN Human Rights Office (OHCHR)
This OHCHR B-Tech material helps readers connect technology governance with human rights responsibilities and suits policy and accountability audiences.
Intermediatelink checked 2026-07-31
EncyclopaediaAI safety (opens in a new tab)Wikipedia
This Wikipedia page gives a broad introductory overview of AI safety and suits readers who want a quick first orientation.
Intermediatelink checked 2026-07-31
VideoVideo explanations: AI safety (opens in a new tab)YouTube search
This YouTube search results page helps readers find a range of video explanations on AI safety and suits readers who prefer audio-visual introductions, but it is a search page rather than one vetted lesson.
search results, not one pagelink checked 2026-07-31
DiscussionCommunity discussion: AI safety (opens in a new tab)Reddit search
This Reddit search results page helps readers see how people discuss AI safety in practice and suits readers exploring public debate, but it is a search page rather than an authoritative source.
search results, not one pagelink checked 2026-07-31

Where this came from

Drafted with a language model through an offline pipeline, checked against this site's schema, and every external link above fetched to see whether it answers. Treat it as a starting point and check anything that matters against the primary sources. How this site is built and what that means.

The one thing to remember

AI safety is the continuing work of finding possible harms, testing failure modes, setting limits, keeping appropriate human oversight and responding when problems emerge.