Generative AI is artificial intelligence that creates new content, such as text, images, audio, video, or code, by learning the patterns in its training data and producing new examples that follow them. Large language models are its best-known form. Its output is a draft or an answer; acting on that output takes a system built around it.
The useful contrast is with predictive AI. A predictive model takes an input and returns a label or a number: this ticket is about billing, this transaction looks like fraud, this customer is likely to cancel. A generative model returns something new: the reply to the ticket, the explanation to the customer, the summary for the agent who takes over. Most production systems use both, and most of the business value of generative AI comes from where its output is allowed to go.
Predictive and generative AI on the same ticket
A telecom customer writes in about a bill. The two kinds of model do different jobs with the same message.
Predictive model
Scores- ClassifyIntent: billing dispute. Sentiment: negative.
- ScoreChurn risk: high. Customer value: high.
- RouteSends the ticket to the priority billing queue.
A decision about the ticket. Useful for routing. Nothing is written to the customer.
Generative model
Creates- ReadReceives the bill lines retrieved by the system: a premium sports package added on the 3rd.
- DraftWrites a reply that names the package, explains when it was added, and sets out the options.
- SummarizeWrites a three-line note for the billing specialist if the customer asks for a person.
New content for people. Only as accurate as the bill data it was given.
Illustrative example. The amounts and package are placeholders. In a real deployment, the predictive score decides the route and the generative model writes within it.
How generative models produce new content
A generative model learns the statistical shape of its training data well enough to sample new points from it. Different model families do this in different ways.
Transformers
Predict the next token given everything before it, one token at a time. This is how LLMs write text and code.1
Diffusion models
Start from random noise and remove it step by step until an image appears that matches the prompt.2 Most image and video generators work this way.
GANs
Two networks compete: a generator creates samples and a discriminator tries to tell them from real data.3 They led image generation before diffusion models.
Multimodal models
Combine text with images, audio, or video in one model, so a customer can send a photo of a damaged product and get a written answer.
How far generated output is allowed to go
The same model can be low risk or high risk depending on what happens to its output. Companies usually move up this scale one step at a time. The scale is a working model, not a standard.
Internal draft
Output reaches: an employeeSummaries, translations, and first drafts that a person edits before anything leaves the company.
Suggested reply
Output reaches: the customer, after approvalThe model drafts; a support agent approves or edits each message.
Grounded answer
Output reaches: the customer, directlyAnswers sent without review, but only from approved sources through retrieval.6
Generated actionAI agents
Output reaches: business systemsThe model's output becomes a tool call: a refund, a booking change, a CRM update, inside set permissions.
Generative AI and neighboring terms
| Term | What it names | Typical output | Relation to generative AI |
|---|---|---|---|
| Machine learning | Systems that learn patterns from data | Predictions, scores, content | The wider field; generative AI is one branch |
| Predictive AI | Models that classify or forecast | A label, score, or number | The counterpart: decides, does not create |
| Large language model | A generative model for text and code | Text, code, structured data | The most used kind of generative AI |
| Agentic AI | Systems that pursue goals with tools | Completed tasks | Uses generative models to plan and act |
Where generative AI earns its place
It works best where the input is unstructured language and the output must be written for a person or turned into structured data.
Write the resolution
Turns order, payment, and policy data into a clear reply in the customer's language, then logs a summary.
data → reply → summaryExplain the choice
Compares three laptops or two insurance plans in plain words, based on the catalog and the shopper's stated needs.
needs → catalog → explainPersonal notices at scale
Writes a delay notice for each affected traveler that reflects their booking and the options they actually have.
event → booking → notifyStructure documents
Generates structured records from claims forms, invoices, and emails, which code then validates.
document → JSON → validateAnswer employees
Answers questions on procedures and contracts from internal documents, with the source attached.
question → retrieve → answerSpeak, not only write
Generates spoken replies for phone lines, where speed and short answers matter more than polish.
speech → text → voice replyGenerative AI risks and controls
NIST's Generative AI Profile lists the risks that generative AI creates or makes worse, including confabulation, data privacy, harmful bias, and information security.7
Confabulation
ControlModels produce confident false statements, also called hallucinations. Ground customer-facing answers in approved sources and test on real conversations.
Data privacy
ControlPrompts can carry personal data. Decide which data reaches which model, where it is processed, and how long logs are kept.
Harmful bias
ControlOutput can reflect skews in training data. Review samples across customer groups and languages, not only in English.
Labeling duties
ControlThe EU AI Act requires providers of systems that generate synthetic audio, image, video, or text to mark the output as artificially generated in a machine-readable way.5
Brand and tone
ControlA model writes in whatever voice it is given. Set the tone in the system prompt, give examples, and score replies against a style rubric.
Common questions
Is ChatGPT generative AI?
Yes. ChatGPT is an application built on large language models, which are generative AI for text and code. It can also create images.
What is the difference between AI and generative AI?
AI is the whole field of systems that perform tasks associated with intelligence. Generative AI is the part that creates new content, as opposed to classifying, predicting, or controlling.
What is the difference between generative AI and agentic AI?
Generative AI produces content. Agentic AI uses generative models to decide on and take actions toward a goal, such as checking an order and issuing a refund.
Is generative AI safe to use with customers?
It can be, when answers are grounded in approved data, sensitive actions have limits, conversations are monitored, and customers know they are talking to an AI system.
References
- Vaswani, A. et al. Attention Is All You Need. NeurIPS 2017.
- Ho, J., Jain, A. and Abbeel, P. Denoising Diffusion Probabilistic Models. NeurIPS 2020.
- Goodfellow, I. et al. Generative Adversarial Networks. NeurIPS 2014.
- Bommasani, R. et al. On the Opportunities and Risks of Foundation Models. Stanford CRFM, 2021.
- European Commission. AI Act, Article 50: Transparency obligations. Regulation (EU) 2024/1689.
- Lewis, P. et al. Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks. NeurIPS 2020.
- NIST. Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile (NIST AI 600-1). July 2024.
- European Commission. AI Act, Article 3: Definitions. Regulation (EU) 2024/1689.