A useful working definition
Artificial intelligence describes machine-based systems that infer how to produce outputs—such as predictions, recommendations, generated content or decisions—from the inputs they receive. The output may influence a digital or physical environment. This broad definition covers a spam filter, a product recommender, a vision system and a generative assistant without pretending they work in the same way.
The word ‘intelligence’ can mislead because it invites comparison with a complete human mind. Most deployed AI is specialized. A system may recognize objects and fail at arithmetic, write persuasive prose and misread a chart, or optimize delivery while knowing nothing about the customer’s actual goal. Capability is always tied to a task and operating context.
Rules, machine learning and foundation models
Some systems follow human-written rules. Machine-learning systems instead fit patterns from examples, using an objective that rewards better predictions on training data. Deep learning uses layered neural networks to learn complex representations, while foundation models are trained broadly and adapted to many downstream tasks through prompting, tools or additional training.
These categories overlap in real products. A generative model may draft an answer, retrieval software may supply documents, rules may block prohibited actions and a conventional classifier may score risk. Calling the whole product ‘the AI’ hides those boundaries, even though each component can fail differently and may require a different owner.
Training is not the same as understanding
During training, a model adjusts internal parameters to reduce error across many examples. A language model learns statistical structure that supports useful completion, reasoning-like behavior and tool selection. That capability can be striking, but training does not guarantee grounded knowledge, stable beliefs or awareness of the real-world consequences of an answer.
This is why a model can produce a strong explanation and an invented source in the same tone. Fluency is evidence of learned language structure, not proof of truth. Systems that require current or private facts need retrieval, source tracking and verification outside the model’s generated text.
The full AI system
An operational system includes data collection, preprocessing, the model, prompts or policies, tools, a user interface, logging, monitoring and people who approve or act on the output. Performance can collapse even when the model is unchanged: a data feed drifts, a search index becomes stale or the interface encourages users to trust uncertain output.
Evaluation should mirror that system. Offline benchmark scores are useful but incomplete. Teams also need scenario tests, security review, latency and cost measurements, accessibility checks, monitoring and a process for correcting harmful or inaccurate outcomes after launch.
What trustworthy AI requires
NIST organizes trustworthy AI around characteristics including validity, safety, security, accountability, transparency, explainability, privacy and fairness. These qualities can conflict. A more explainable system may lose accuracy; collecting more monitoring data may create privacy risk. Responsible deployment is the practice of making those tradeoffs visible and justified.
The first question should therefore be practical: what decision or experience will this system change, for whom, and what happens when it is wrong? From that point, teams can define evidence, permissions and escalation. AI is most useful when its role is precise enough to evaluate—and limited enough that people remain able to intervene.
Sources & further reading
Social-media activity is treated as a signal of attention, not proof. Product claims are attributed to the linked publisher or announcement.