AI systems are safer to use when their risks are managed throughout the lifecycle—not when a single test, filter, or human reviewer is treated as a cure-all. Start by defining where and how the system will be used, identify likely harms and foreseeable misuse, match controls to those risks, test the controls, give qualified people authority to intervene, and monitor what happens after release. The right safeguards depend on the system, its users, and the people affected by its outputs.
How can AI systems be made safer?
Use a continuing risk-management process. A safeguard is useful only if it addresses a plausible failure in the real context of use, can be checked, and has an owner who can act when it fails. Risk management therefore continues after deployment: new uses, incidents, or performance changes can make earlier assumptions obsolete.
As an Amazon Associate I earn from qualifying purchases.
- Define the system and its intended use. Record what the AI does, who will use it, what decisions or actions depend on it, and the settings in which it will operate. Include important limits on its use.
- Identify harms, affected people, and foreseeable misuse. Consider how errors, biased outcomes, unsafe recommendations, privacy or security failures, or overreliance could affect users and others. Account for the groups likely to be affected, including minors or other vulnerable people where relevant.
- Choose controls that fit the risks. Use design changes to prevent or reduce risks where feasible; add safeguards such as access restrictions, warnings, review steps, or operational limits for risks that remain. Assign an accountable owner to each control.
- Test against defined criteria. Evaluate behavior in conditions that reflect the intended use and foreseeable misuse. Set relevant metrics and thresholds before testing, investigate failures, and revise the system or its controls when results fall short.
- Prepare people to oversee the system. Give designated operators the information, training, competence, and authority needed to notice problems, intervene, or stop use when appropriate.
- Monitor and update after release. Review performance and incidents, including whether the system is being used as intended. Feed what you learn into risk assessments, controls, and decisions about whether to continue, restrict, or withdraw use.
This sequence is iterative rather than a one-time checklist. Monitoring information should be able to change the risk assessment and the safeguards, not merely be collected.
What safeguards should AI systems have?
There is no universal safeguard set that makes every AI system safe. Controls should address distinct failure modes and work together. For example, better data does not by itself prevent cyberattacks, and a human review step does not compensate for unclear operating limits or poor testing.
#1 Best Overall
| Safeguard area | What it helps address | What to put in place |
|---|---|---|
| Data governance and quality | Errors or skewed outcomes associated with unsuitable, incomplete, or unrepresentative data | Define data practices and assess quality and relevant statistical properties, paying attention to groups that may be affected by bias. |
| Accuracy and robustness | Incorrect outputs or behavior that breaks down under expected variation or difficult conditions | Test against the system’s intended purpose and operating conditions; set acceptance criteria and address failures. |
| Cybersecurity | Attacks, unauthorized access, or interference with the system and its supporting infrastructure | Apply security controls appropriate to the system and its deployment, and reassess them as threats and system conditions change. |
| Transparency and documentation | Deployers or operators misunderstanding capabilities, limits, or how to use the system | Provide relevant information and instructions, and keep technical documentation and records appropriate to the system and applicable duties. |
| Human oversight | Problems that require judgment or intervention rather than unattended operation | Design oversight into the workflow; specify who may intervene, what they need to know, and how to pause or stop use. |
| Monitoring and incident response | Failures or emerging risks that appear after deployment or as use changes | Track relevant behavior and incidents, define escalation and response responsibilities, and use findings to update controls. |
These controls can reduce risks but do not prove that a system is fair, accurate, secure, or safe in every situation. Their effectiveness depends on the evidence collected and whether the controls are maintained in practice.
How do you manage risks from generative AI?
Apply the same lifecycle, but examine how generative AI creates or transforms content and how people may rely on that content. NIST’s Generative AI Profile says it defines “risks that are novel to or exacerbated by the use of GAI.” It is a cross-sector companion to NIST’s AI Risk Management Framework (AI RMF), published July 26, 2024, and proposes actions aligned with that framework.
The profile organizes risk work around four functions:
Do these 3 things before closing this tab:
1Repair Windows errors before they cause bigger problems2Fix the driver behind crashes, sound loss and screen glitches3Clear out junk files and repair common Windows errors- Govern: establish accountability, policies, and responsibilities for managing risk.
- Map: describe the intended context, affected people, and potential impacts.
- Measure: evaluate risks and system behavior using appropriate evidence.
- Manage: select, implement, and monitor mitigations, adjusting them as evidence changes.
For a generative system, make the use context specific: distinguish a tool that drafts material for a trained employee to review from one whose output is delivered directly to the public or informs a consequential decision. Test the relevant outputs and misuse scenarios, give users clear operating information, and define when output must be checked or rejected. The profile is guidance, not a certification or proof that a system is safe.
What does effective human oversight require?
“A human in the loop” is not an effective safeguard if that person lacks time, relevant information, training, or authority. Oversight should be part of the system’s operating design, with responsibility assigned to people able to recognize when the system is outside its intended use or not working as expected.
- Define which person or role is responsible for oversight and what decisions they can make.
- Provide information and training sufficient to interpret system outputs and recognize limitations or warning signs.
- Make intervention practical: identify when and how to correct, override, pause, or stop the system.
- Where appropriate to the context, set operating constraints that the AI system itself cannot override.
The EU AI Act’s recitals describe oversight as enabling people to ensure intended use and address impacts across the system lifecycle. The appropriate mechanism depends on the application; the underlying test is whether a person can make an informed intervention when one is needed.
Rank #3
What does the EU AI Act require, and who is covered?
The EU AI Act does not impose the same duties on every AI use. Its obligations depend on the category of system or model and the role of the relevant actor. In particular, Article 9 sets out a documented, iterative risk-management system for high-risk AI systems; it is not a universal rule that every AI product is legally high-risk.
What’s actually slowing this PC down?
Pick the symptom - the matching free tool is one click away.
For high-risk systems in scope, Article 9 calls for identifying and analyzing known and reasonably foreseeable risks to health, safety, or fundamental rights under intended use; estimating risks under intended use and reasonably foreseeable misuse; considering post-market monitoring information; and taking targeted measures for identified risks. Potential adverse impacts on minors and, where appropriate, other vulnerable groups must also be considered.
Article 9 also calls for testing as appropriate throughout development and, in any event, before the system is placed on the market or put into service. Testing is to use predefined metrics and thresholds appropriate to the system’s purpose. The regulation calls for eliminating or reducing risks as far as technically feasible through design and development, then applying mitigation and control measures where risks cannot be eliminated. It also links controls with information and, where appropriate, training for deployers.
Other high-risk requirements include data governance and management, technical documentation and record-keeping, transparency and information for deployers, human oversight, robustness, accuracy, and cybersecurity. These are complementary duties, not a guarantee that a system will never cause harm.
A separate category also matters: Article 55 adds duties for general-purpose AI models with systemic risk. Those duties include model evaluation using protocols and tools reflecting the state of the art, documented adversarial testing, systemic-risk assessment and mitigation, serious-incident reporting, and adequate cybersecurity for the model and its physical infrastructure. They should not be generalized to every general-purpose model or every AI system.
How is NIST guidance different from EU law?
| Framework or rule | Status and scope | How it helps |
|---|---|---|
| NIST AI Risk Management Framework (AI RMF) | Voluntary guidance; NIST says it is intended for voluntary use. AI RMF 1.0 was released January 26, 2023. | Helps organizations incorporate trustworthiness considerations into AI design, development, use, and evaluation through a risk-management approach. |
| NIST Generative AI Profile, AI 600-1 | Published July 26, 2024; cross-sector companion guidance for generative AI risks. | Proposes actions aligned with the AI RMF functions Govern, Map, Measure, and Manage. |
| EU AI Act | Binding EU regulation with duties that apply to defined categories and actors, including high-risk AI systems and general-purpose AI models with systemic risk. | Sets legal requirements for systems and models within its scope, including documented risk management and specified technical and organizational safeguards. |
NIST’s framework page says the AI RMF is “intended for voluntary use” to improve the ability to incorporate trustworthiness considerations into AI products, services, and systems. NIST’s overview reports that AI RMF 1.0 is being revised; that status can change. The framework is not a certification and does not itself establish compliance with the EU AI Act. Conversely, a legal obligation under the Act depends on whether the system, model, actor, and use fall within the relevant provisions.
Best Value
How can an organization tell whether its safeguards are working?
Do not judge safeguards only by whether they exist on paper. Connect each significant risk to a control, an accountable owner, and evidence that the control works for the intended use.
- Risk coverage: Can the organization explain which foreseeable harms and affected groups each control addresses?
- Test evidence: Were the system and controls tested against predefined criteria that reflect the use context? Were failures followed by changes or a justified decision to limit use?
- Operational readiness: Can the people assigned to oversee the system identify a problem and intervene with the authority and information they need?
- Post-release learning: Are incidents and changes in system behavior reviewed, escalated, and used to revisit assumptions and mitigations?
- Scope and accountability: Is it clear who owns each control, which rules apply, and what action follows if the control is ineffective?
A documented process makes those questions answerable and supports improvement; documentation alone is not evidence that risk has been controlled.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
