What AI is

Technology, law, and practice. As of November 3, 2026.

Facts checked November 3, 2026

As of November 3, 2026

A plain-language handbook for public officials, staff, and organizations. Products are named only in Appendix B, as a dated reference list. Positions are attributed to their sources.

In one minute

In everyday government and vendor language, “AI” most often means generative systems—especially large language models—that predict tokens and produce text, images, audio, code, or video.
Training builds a model from data; inference is everyday use. Agents add tool use and multi-step actions on top of a model.
Documented limits include fabricated facts (“hallucinations”), training cut-offs, and measured performance gaps across groups and tasks.
As of Nov 3, 2026, U.S. rules are mostly state and sector-specific; the EU AI Act is phasing in by risk category; several countries use promotion laws, voluntary codes, or pending bills.
Public-sector decisions typically turn on data location, training-on-inputs terms, citation/retrieval behavior, records retention, and human sign-off—not on brand names.

What people usually mean by “AI”

When officials, vendors, and news coverage say “AI” in 2025–2026, they usually mean generative AI: systems that create new content rather than only classifying existing records. The most common office tools are built on large language models (LLMs). An LLM predicts the next token of text given prior tokens. Stacked at scale, that prediction can draft memos, summarize transcripts, translate documents, or write code. Related systems generate images, audio, or video. Computer vision, speech recognition, and classical machine-learning scoring systems remain in use for inspections, licensing queues, and fraud detection; they are AI under broader definitions, but they are not what most “AI assistant” procurements refer to.

Researchers and companies also use the labels AGI (artificial general intelligence) and ASI (artificial superintelligence). AGI usually means a system that could match or exceed human performance across most cognitive tasks. ASI usually means a system that substantially exceeds human ability across nearly all domains. As of Nov 2026, those terms describe research goals and policy debates, not a settled product category with shared technical criteria. This primer keeps the everyday name “AI” for deployed generative and predictive systems, and treats AGI/ASI as forward-looking vocabulary that appears in hearings and vendor materials.

Operational description used throughout this primer: a model maps inputs to outputs using patterns learned from training data and any tools connected at inference time. Predicting fluent language is not the same as verifying facts.

Models, data, compute, and agents

ModelThe trained artifact: billions of parameters (weights) that encode statistical patterns. Foundation models are trained broadly; they can be adapted with fine-tuning or prompting.

Training dataText, images, code, and other corpora used to set the weights. Composition affects capability, bias measurements, copyright disputes, and privacy risk.

InferenceRunning the trained model on a new prompt. Most day-to-day “using AI” is inference, billed by tokens, seats, or API calls.

ComputeGPUs, TPUs, and related accelerators used for training and inference. Training frontier-scale models requires large clusters; inference cost scales with usage.

Open vs closed weightsOpen-weight releases let an organization download and run a model under a license. Closed models are reached only as a service. Licenses and support differ from classic open-source software.

AgentsSystems that call tools, browse, write files, or take multi-step actions. They raise permission, logging, and oversight questions beyond a chat box.

Capabilities that drive adoption

Documented strengths of current generative systems, as described in vendor model cards and independent benchmarks, include drafting and revision of prose, translation, summarization of long text, code assistance, and question-answering when retrieval is available. Computer-vision systems continue to support document intake, license-plate or form extraction, and inspection workflows that predate generative AI. Speech-to-text systems feed those pipelines. The policy question for a legislature or city is rarely whether any of these tasks can be automated in part; it is which tasks may be automated under what disclosure, records, procurement, and human-review rules.

Why the topic reached governing bodies

Public bodies report three concurrent shifts: model capability rose; consumer interfaces removed the need for programming skill; and cloud pricing made pilots inexpensive to start. That combination moved AI questions from IT departments into procurement, records, workforce, elections, child-safety, and infrastructure dockets. Parallel debates over data-center load, training-data copyright, and synthetic media appear in statehouses and national legislatures at the same time (as of 2024–2026 legislative trackers summarized later in this primer).