iTechGuides is reader-supported. When you buy through links on our site, we may earn an affiliate commission. As an Amazon Associate I earn from qualifying purchases. Learn more
AI harms are negative consequences an AI system may cause or contribute to—not simply mistakes made by a model. They can affect people who never use the system, as well as groups, organizations, communities, society, ecosystems and the planet. Reducing those harms requires assessing risk before design, testing in the intended context, monitoring after deployment and stopping when the remaining risk is unacceptable.
What counts as an AI harm?
A model error becomes a harm question when you ask what consequence follows, who bears it and how serious it could be. A faulty output might be inconsequential in one setting and consequential in another—for example, if it informs a decision affecting someone’s access to an opportunity or service. The relevant assessment is about the system in context, not accuracy in isolation.
The NIST AI Risk Management Framework (AI RMF) treats impacts as potentially positive, negative or both. For a potential negative event, it frames risk in terms of both the likelihood of occurrence and the magnitude of the resulting harm. Neither dimension alone tells the whole story: an unlikely event may still warrant attention if its consequences could be severe, while a frequent but minor failure calls for a different response.
Potentially affected parties extend well beyond a system’s direct users. They can include individuals, groups, communities, organizations, society, ecosystems and the planet. The system’s developers, deployers and evaluators also have roles in understanding and managing these effects.
#1 Best Overall
Which kinds of harm should an assessment consider?
Start with the intended use and its setting, then consider how effects could arise directly or indirectly. These risk dimensions can overlap; they are not mutually exclusive categories.
- Fairness and harmful bias: A system may produce unequal outcomes or representations across groups, even if its overall performance appears acceptable.
- Privacy: Risks may arise from the data used to train or operate a system, as well as from the system’s outputs and use.
- Safety and reliability: Invalid, inconsistent or unsafe outputs can have different consequences depending on the task and how people rely on them.
- Security and resilience: Consider threats to confidentiality, integrity and availability, including risks in the system’s software and hardware.
- Accountability and transparency: Consider whether people can understand who is responsible for a system’s use and how its outputs inform decisions.
- Environmental effects: Training, maintaining and operating AI systems can consume resources, with impacts that vary by model activity, modality, hardware and task.
NIST’s trustworthiness considerations also include explainability and interpretability. These concerns can intersect: for example, privacy risks may accompany training data, while a system’s limitations may make it harder to explain or challenge an outcome.
How should an organization assess risk?
Use a lifecycle process rather than treating a launch review as a one-time sign-off. The AI RMF is intended for developers, users and evaluators, and it emphasizes accounting for system limitations and reassessing risk as uses and conditions change.
Free tools Windows power users keep installed
One-click scans. No signup required.
- Set context before design. Define the intended use, the task or decision the system supports, affected people and communities, foreseeable misuse, and the organization’s tolerance for negative impacts. Record what is out of scope as well as what the system is meant to do.
- Identify plausible harms and assess them. Consider direct and indirect consequences, their likelihood and their magnitude. Seek input from people who may be affected. Document assumptions, evidence and uncertainty; some harms and their severity are difficult to predict, especially as applications evolve.
- Test for the actual use. Evaluate relevant failure modes and performance in the intended context, including across populations and languages that matter to that use. NIST’s AI Resource Center provides technical materials to support testing, evaluation, verification and validation.
- Choose safeguards and record residual risk. Match mitigations to the harms identified. Document which risks remain and who is accountable for accepting or addressing them.
- Monitor and reassess. Watch for changes in use, operating conditions and outcomes after deployment. Revisit the assessment when those changes could alter the likelihood or severity of harm.
- Stop safely if risk is unacceptable. NIST’s 2023 AI RMF says development and deployment should cease in a safe manner until risks can be sufficiently managed when unacceptable negative risk levels are present, such as imminent significant impacts, severe harms occurring or catastrophic risks.
These steps organize risk management; they do not guarantee that every harm will be predicted or eliminated. The framework provides a way to structure decisions and accountability, not a universal score that makes every trade-off comparable.
Rank #3
What changes when the system is generative AI?
NIST’s cross-sector Generative AI Profile, NIST AI 600-1, released July 26, 2024, complements the AI RMF with generative-AI-specific risks and suggested actions. It highlights the potential for generative systems to increase the speed and scale at which harmful bias manifests. Representational harms and differences in performance across demographic groups and languages can contribute to unequal outcomes or reinforce existing social biases.
That makes evaluation choices important: a single overall performance measure may conceal differences that matter in a particular use. Test the system with populations, languages and contexts relevant to the task, and assess what consequences a failure could have there.
The profile also identifies environmental impacts as a concern. Training, maintaining and operating generative AI can be resource-intensive, but the impact depends on factors such as the system’s activity, modality, hardware and task. NIST says there is currently no agreed method to estimate generative AI’s environmental impacts, so a single emissions figure should not be treated as a universal measure.
What’s actually slowing this PC down?
Pick the symptom - the matching free tool is one click away.
For context only, the 2024 profile reports an estimate that training a single transformer large language model can emit as much carbon as 300 round-trip flights between San Francisco and New York. This is a comparison reported by NIST, not a universal emissions factor; the profile’s measurement caveat still applies.
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.When is the NIST AI RMF useful, and what does it not do?
NIST released AI RMF 1.0 on January 26, 2023, after an 18-month development effort involving more than 240 contributing organizations, according to NIST’s AI Resource Center. NIST says the framework is voluntary and its current framework page indicates it is being revised. The Generative AI Profile adds guidance for generative systems; it does not turn the framework into a mandatory standard.
The framework can help developers, users and evaluators organize risk work around trustworthiness considerations: validity and reliability, safety, security and resilience, accountability and transparency, explainability and interpretability, privacy, and fairness with harmful bias managed. It is a useful management framework, not a guarantee of safety or a substitute for applicable law. Legal duties and sector requirements depend on jurisdiction and use case.
Use the framework to make risks, assumptions, mitigations and residual concerns visible throughout a system’s lifecycle. Do not treat completion of a checklist as proof that a system is safe: outcomes can be difficult to assess, and risks may change as the system or its use changes.
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 errorsQuick 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.
