What Does AI Do?
AI can recognize patterns, generate content, answer questions, and support decisions. Learn what these systems do and where their capabilities have limits.
Overview
Artificial intelligence (AI) is a broad class of engineered systems that use data, rules, or learned models to produce outputs such as predictions, recommendations, generated content, or decisions. In practical terms, AI classifies images, finds patterns in records, forecasts outcomes, ranks options, generates text or images, and can control machines within defined limits. What an AI does depends on its design, data, objective, and deployment context; the term does not describe one single capability. NIST defines an AI system as machine-based and able, for stated objectives, to generate outputs that influence real or virtual environments, with varying degrees of autonomy (Artificial Intelligence Risk Management Framework (AI RMF 1.0)).
What does AI do?
AI systems mainly turn inputs into useful outputs. Inputs may include photographs, spoken questions, sensor readings, transaction records, or text. Outputs may be labels, scores, forecasts, recommendations, actions, or generated material. A system may operate entirely in software, inform a person, or control a physical device.
| Task | Typical output | Example |
|---|---|---|
| Classification | Category or label | Flagging a message as likely spam |
| Detection | Location or presence of a pattern | Identifying a feature in a medical image |
| Prediction | Estimate of an unknown or future value | Forecasting demand from historical data |
| Ranking and recommendation | Ordered options | Selecting which search results to show first |
| Generation | New text, images, audio, or code | Drafting a response to a request |
| Control and planning | Proposed or executed action | Navigating a rover around obstacles |
These functions overlap and are not an exhaustive taxonomy. A medical system, for example, may detect image features and produce information supporting clinical assessment. The FDA says AI-enabled medical devices can analyze complex data, identify patterns, and support disease detection, diagnosis, treatment, and other care functions (Artificial Intelligence-Enabled Medical Devices | FDA).
What is AI and what does it do in everyday services?
People usually encounter AI as a component inside a larger product. Search and recommendation systems rank material; voice interfaces convert speech to text and interpret commands; navigation tools estimate routes; fraud systems flag transactions; and writing assistants generate or revise content. The visible service combines the AI component with software, data, user interfaces, policies, and often human oversight. An AI estimate does not by itself guarantee that the whole service is accurate or appropriate.
Generative AI is a subset of AI that creates outputs such as text, images, or audio in response to a prompt or other context. Large language models commonly generate text by predicting successive tokens from context, rather than by checking each statement against a database of verified facts. NIST describes confidently presented false or inconsistent generated content as confabulation, often called hallucination (Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile).
How AI works
Many current AI systems use machine learning. Developers provide examples or other data and specify a task or objective. A learning procedure adjusts a model so its outputs fit examples or optimize a defined measure. The trained model is then applied to new inputs. Results depend on the data and objective used in development and on whether new inputs resemble the conditions represented in training and evaluation.
This is a common lifecycle, not a requirement that every AI system learns from data. Some systems rely substantially on explicit rules; others combine learned models with conventional software or human judgment. NIST emphasizes that risks and benefits depend on the system lifecycle and the context in which people use it, not just on a model in isolation (Artificial Intelligence Risk Management Framework (AI RMF 1.0)).
Machine learning, deep learning, and generative AI are related but distinct terms. Machine learning fits patterns from data; deep learning uses multi-layer neural networks; generative AI produces new outputs. Some AI classifies or predicts without generating content, while software that appears intelligent may rely on hand-written rules rather than machine learning (AI vs. Machine Learning vs. Deep Learning vs. Neural Networks | IBM).
Examples across fields
Health care
AI-enabled medical devices can analyze images and other health data, identify patterns, and provide information supporting detection, diagnosis, or treatment. FDA examples include software that detects diabetic retinopathy from retinal images, tools that sharpen images, and systems that automate insulin dosing from continuous glucose-monitor readings. These examples describe functions, not a guarantee that every system improves outcomes. In the United States, the FDA regulates medical devices, including AI-enabled devices, according to intended use and technical characteristics, and considers safety and effectiveness across the product lifecycle (Artificial Intelligence-Enabled Medical Devices | FDA).
Space and environmental data
NASA describes using AI to analyze satellite imagery, support disaster-response information, and assist autonomous navigation. Its account of the Perseverance rover explains that the rover analyzes camera images to identify hazards and navigate terrain while out of contact with Earth (Artificial Intelligence - NASA). This is task-specific autonomy, not general-purpose human judgment.
Business and public services
Organizations use AI to extract information from documents, identify anomalies, forecast quantities, sort requests, and assist customer interactions. Such systems can help prioritize work or automate repetitive steps. Their practical value depends on error rates, data quality, workflow design, and the cost of correcting mistakes. A score or recommendation is not the same as a justified decision, particularly when access to employment, credit, education, or public services is affected.
What AI does not reliably do
AI is not one uniform agent with human-like understanding. A system may perform well on a narrow task but fail on unfamiliar inputs, changed conditions, or questions outside its intended use. Generative systems can state false claims confidently. Predictive systems estimate patterns in data; they do not establish that a correlation is causal. Outputs should be interpreted in light of purpose, validation, and operating conditions.
Performance can change after deployment. Data may shift as populations, equipment, user behavior, or processes change. NIST notes that changes in data can affect system functionality and trustworthiness, while complex systems can make failures difficult to detect (Artificial Intelligence Risk Management Framework (AI RMF 1.0)). Evaluation and monitoring are therefore part of responsible operation.
Benefits, risks, and trade-offs
AI can process large volumes of data quickly, apply a procedure repeatedly, and identify patterns that are difficult to detect manually. Those properties may support faster analysis or automate parts of a workflow, but they do not guarantee better results. A system can reproduce errors or biases in data, produce false positives or omissions, or fail when used in a different context. Automation can also shift work: people may still need to check outputs, handle exceptions, and maintain the system.
NIST identifies validity and reliability, safety, security, transparency, explainability, privacy, and fairness as important characteristics of trustworthy AI. It describes AI systems as socio-technical: effects arise from the interaction of technical design, organizational practice, users, and social context. A system that performs acceptably in testing can still cause harm if it is deployed for an unsuitable purpose or its outputs are treated as unquestionable (Artificial Intelligence Risk Management Framework (AI RMF 1.0)).
| Potential benefit | Corresponding limitation or risk | Practical safeguard |
|---|---|---|
| Faster screening or analysis | False positives, missed cases, or automation bias | Validate for intended use; review consequential outputs |
| Personalization | Privacy loss or unequal performance | Limit data collection and test relevant groups |
| Automated content creation | False claims or misleading media | Verify claims and provenance before use |
| Autonomous operation | Unexpected behavior outside tested conditions | Set operating boundaries and fallback procedures |
| Repeated decision support | Difficult-to-explain or contest decisions | Document purpose, limits, and routes for appeal |
Safeguards should match the use. A low-consequence suggestion may need clear labeling and easy correction. Medical or other high-impact uses warrant stronger validation, monitoring, traceability, and human accountability. NIST's voluntary risk framework organizes organizational activities around governing, mapping, measuring, and managing risks throughout the AI lifecycle (AI Risk Management Framework | NIST).
Regulation and accountability
Legal treatment depends on jurisdiction and application. The European Union's AI Act entered into force on 1 August 2024 and establishes a risk-based framework. The European Commission describes categories from minimal-risk uses to high-risk systems subject to requirements such as risk mitigation, data quality, user information, and human oversight; certain practices are prohibited. The Act's obligations apply on a staged timetable, so entry into force should not be confused with the date every rule applies (AI Act enters into force).
In the United States, the FDA's approach illustrates sector-specific regulation: it regulates medical devices that incorporate AI, rather than treating every AI application as a medical device. Requirements depend on intended function and applicable device pathway (Artificial Intelligence-Enabled Medical Devices | FDA). Responsibility remains with the people and organizations that develop, deploy, operate, and rely on a system; an AI output does not remove the need for accountable decisions.
Direct answer
AI takes data or other inputs and uses rules or learned models to produce predictions, classifications, recommendations, generated content, or actions. It is used to automate defined tasks and support human work in health care, transport, business, research, and public services. Its usefulness is conditional: a system can be effective within a validated purpose but may be wrong, biased, insecure, privacy-invasive, or unsuitable outside its design conditions. Consequential uses require evaluation, appropriate oversight, and continuing monitoring.