7 AI Limitations and Responsible Use
Learn what AI systems can and cannot do, how their risks interact, and how to evaluate and manage those risks in context.
Capabilities depend on context
AI systems can recognize patterns, classify information, make predictions, recommend actions, and generate text or images. Their performance depends on the task, training data, system design, and conditions of use.
A system that performs well on familiar examples may be less reliable when the data or setting changes. Therefore, performance should be assessed in the context where the system will actually be used.
Interpreting AI outputs
AI output should not automatically be treated as fact. A generative system may produce a fluent answer that contains false details or invented sources; this is sometimes called a or confabulation. The confidence or polish of an answer does not prove it is correct, so important claims should be checked against reliable evidence.
matter across tasks and groups: an AI system can make mistakes, and its accuracy may vary. Testing in the intended setting, tracking uncertainty and errors, and having a way to pause or fall back to another process can help when results are unreliable.
Risks to people and information
AI risks can arise from training data, labels, design choices, and how people use a system. A hiring model trained on past hiring records, for example, may reproduce patterns in those records. Evaluate performance for relevant groups, involve affected people, and address harmful disparities; one overall accuracy score may hide unequal errors.
AI systems may process personal or sensitive information. Collect only what is needed, protect access, and consider throughout design and use. Avoid entering confidential information into a tool unless its data practices and your authorization make that appropriate.
Errors, misuse, or attacks can cause harm, especially in consequential settings such as health, education, finance, or public services. Match safeguards to possible harm by testing before deployment, monitoring for problems, limiting access where needed, and establishing a safe way to override, repair, or stop the system.
Transparency, oversight, and tradeoffs
People should know when AI meaningfully contributes to a decision that affects them. Organizations should document what a system is for, its limits, who is responsible for it, and how errors can be challenged or corrected. must be meaningful, not merely a person approving an output without the ability or information to assess it.
Risks can interact. Collecting more personal data might improve a model’s performance but increase risk. Trustworthiness therefore involves balancing reliability, safety, security, fairness, , explainability, and accountability in the system’s specific context.
A cycle for responsible use
Responsible AI is an ongoing practice, not a one-time test or a guarantee that a system is risk-free. The cycle offers a practical way to address risks:
Govern: Set policies, roles, and oversight. Identify who is accountable for the system and its effects.
Map: Define the system’s purpose, users, affected people, and context. Identify likely benefits, limitations, and harms before deciding whether to use it.
Measure: Test performance, uncertainty, fairness, , security, and safety. Include realistic conditions and relevant groups, document findings, and reassess as the system changes.
Manage: Prioritize risks and take action. Add safeguards, restrict or change the use, provide human review and ways to correct errors, or do not deploy the system if important risks cannot be acceptably addressed.
Applying safeguards to hiring
Before using AI to help screen job applications, an employer should establish that the tool is suitable for that purpose, examine how errors differ among applicant groups, protect applicant data, tell people how the tool is used, and ensure a responsible person can review and correct outcomes.
These checks should continue after launch because data, systems, and uses can change. The higher the potential impact on people, the stronger the safeguards should be.