What is Generative AI?
Generative AI is a class of models and systems that learns a probability distribution over data and produces candidate text, images, audio, video, code, or structured records from a prompt or other condition.
Quick Facts
| Full Name | Generative Artificial Intelligence |
|---|---|
| Created | 2014 (GANs by Ian Goodfellow), 2017 (Transformers), 2022 (ChatGPT public release) |
| Specification | Official Specification |
How It Works
Generation models a conditional distribution
A generator estimates a distribution such as P(x), P(x|c), or an equivalent latent-variable process, where c may be a prompt, class, image, schema, speaker, or retrieved evidence. Inference turns that distribution into one or more concrete candidates through decoding or sampling. Temperature, top-p, seed behavior, constraints, step count, and candidate selection can change the result without changing model weights. A generated sample may differ from individual training records, but generation alone does not prove novelty or independence from them.
Different model families use different mechanisms
Autoregressive models factorize ordered data and predict the next element; diffusion models learn to reverse a corruption process; generative adversarial networks train a generator against a discriminator; and variational autoencoders learn a regularized probabilistic latent space. These are not mutually exclusive product categories. A system may place a VAE inside a latent diffusion pipeline, combine Transformer components with diffusion, or keep its architecture private. Model-family labels guide investigation but do not replace evaluation of the exact deployed version.
The production system is larger than the model
Inputs pass through authentication, data classification, parsing, and instruction boundaries before reaching the model. The generation path may add retrieval, tools, memory, retries, or multiple candidates. Outputs may require schema validation, evidence checks, policy review, provenance, or human approval. Every model, prompt, corpus, tool, policy, and evaluator needs a version, owner, and observable failure path so a team can locate regressions and roll back safely.
Evaluation must follow the user decision
Freeze representative normal, difficult, adversarial, and severe-risk cases before tuning. Record the exact model, prompt, tools, decoding policy, evaluator, retry budget, latency boundary, and cost method. Measure artifact-specific quality alongside severe failures, slice performance, p95 latency, cost per accepted result, abstention, and human-review load. Deterministic checks are preferable for machine-verifiable contracts; human or model judges need calibration against independently adjudicated examples.
Risk controls are layered, not absolute
The NIST Generative AI Profile groups risks such as confabulation, privacy, information security, harmful bias, intellectual property, content safety, information integrity, and value-chain dependencies. Controls can reduce exposure through scoped tasks, least-privilege tools, approved data, retrieval, filters, review, monitoring, and fallback paths. Retrieval does not guarantee truth, a second model may share correlated errors, and human review fails when reviewers lack evidence, time, or authority.
Provenance is not a truth detector
C2PA Content Credentials can bind signed claims, assertions, ingredients, and edit history to an asset so a verifier can inspect signer trust and detect changes to protected data. They do not prove that a depicted event happened, that every prior transformation was recorded, or that unsigned content is synthetic. Visible watermarks and AI-detection classifiers are separate signals with their own removal and error risks.
Key Characteristics
- Produces candidate outputs by sampling or decoding a learned distribution under explicit conditions
- Covers text, images, audio, video, code, and structured data through different or hybrid model families
- Changes behavior with model, prompt, context, tools, decoding, filtering, and candidate-selection versions
- Requires task-specific evaluation because fluent or realistic output is not evidence of correctness
- Needs layered controls for privacy, security, bias, rights, content safety, and information integrity
- Supports accountable deployment only when provenance, monitoring, fallback, and rollback are designed with generation
Common Use Cases
- Drafting grounded text or summaries with citation checks, abstention, and human review for high-impact claims
- Generating and editing visual, audio, or video assets under consent, brand, rights, and provenance controls
- Producing code candidates that must compile, run in a sandbox, pass tests, and receive diff review
- Extracting structured records with constrained decoding, schema validation, and field-level quality metrics
- Creating synthetic training or test data with source lineage, duplication checks, slice audits, and independent holdouts
Example
Loading code...Frequently Asked Questions
What is the difference between generative AI and discriminative AI?
A discriminative model estimates a label or value from an input, such as P(y|x). A generative model represents how data or outputs are distributed, such as P(x), P(x|c), or a latent-variable equivalent, so it can sample candidates. Production systems often combine both: a generator proposes outputs while classifiers, rankers, rules, or humans decide whether to accept them.
Does generative AI always create original content?
No. A generated result may not exactly match one training record, but that does not prove originality. Models can reproduce memorized fragments, common expressions, recognizable styles, source bias, or protected material. Novelty, attribution, similarity, and legal rights require separate technical and legal review.
Which architectures are used for generative AI?
Major families include autoregressive models for ordered sequences, diffusion models for iterative synthesis, GANs for adversarial generation, and VAEs for probabilistic latent representations. Modern systems can combine these components, and a product name does not reliably reveal its complete architecture.
How should a generative AI system be evaluated?
Define the user decision and unacceptable outcomes, freeze representative and severe-risk cases, and record the entire deployed configuration. Measure task-specific quality, severe failures, performance by slice, p95 latency, total cost per accepted result, abstention, and review load. Prefer deterministic verification wherever the output contract allows it.
Can provenance or an AI detector prove that content is AI-generated?
Not generally. C2PA can verify signed provenance claims and content integrity under a trust policy, but credentials can be absent or stripped and do not prove truth. AI-detection classifiers estimate patterns with false positives and false negatives. Neither signal should be treated as universal proof of origin.