A generative AI task is any task where the system creates new content: writing text, generating an image from a description, composing code, drafting a summary, producing audio or video. Tasks that label, rank, predict, or retrieve existing information (classifying an email as spam, forecasting next quarter's sales, detecting a fraudulent transaction) are not generative, even when they use sophisticated AI. If you arrived here from a certification course or an interview prep list, that one paragraph is the answer; the rest of this page gives you the test to apply to any task you have never seen before, a worked examples table, and the reasoning behind the exam variants, so the distinction sticks rather than being memorized.
The direct answer
Generation means producing an artifact. The tasks below are generative because in each case the model composes something that did not exist until it ran:
- Writing an email, article, product description, or social post
- Generating an image, illustration, or design from a text description
- Composing, completing, or refactoring code
- Summarizing a document (the summary is new text)
- Translating between languages, since the target-language text is newly composed even though the meaning is not
- Creating audio, music, voice, or video
- Drafting answers in a conversation
And these common AI tasks are not generative, because their output is a decision about existing data rather than a new artifact: classifying emails as spam, predicting sales or churn, detecting fraud or defects, ranking search results, recommending products, and transcribing speech to text (transcription records what was said; it does not invent content).
The one-question test
Every borderline case resolves with the same question: is the output new content, or a judgment about existing content? New content means text, imagery, code, or audio that the model composed. A judgment means a label, score, category, prediction, or retrieved item.
Apply it to the two cases that trip people up most:
- Summarization feels like extraction but is generation. The model does not copy sentences out; it writes new ones that express the source's meaning. New text = generative.
- Recommendation feels creative but is ranking. "You might like this" selects from items that already exist. No new artifact = not generative, no matter how uncanny the suggestions.
A few genuinely borderline cases, resolved with the same question. Paraphrasing and rewriting are generative: the new phrasing is composed text. Autocomplete is generative in miniature; the suggested continuation is new content, just short. Data augmentation, creating synthetic training examples, is generative by definition. Sentiment analysis is not: "positive" is a label. Entity extraction is not: it points at spans that already exist in the text. Clustering is not: grouping items creates no artifact, only structure over existing data.
Examples: which tasks are generative AI tasks, and why
The table below covers the tasks that appear most often in courses, exams, and real product decisions. The "why" column is the part worth internalizing; the verdicts follow mechanically from it.
| Task | Generative? | Why |
|---|---|---|
| Writing a product description | Yes | New text is composed |
| Creating an image from a text prompt | Yes | New image is synthesized |
| Summarizing a report | Yes | The summary is newly written text |
| Writing or completing code | Yes | New code is composed |
| Translating a paragraph | Yes | The target-language text is newly generated |
| Classifying email as spam | No | Assigns an existing label to an existing item |
| Forecasting sales | No | Outputs a number about existing patterns |
| Detecting fraud | No | Flags existing transactions; creates nothing |
| Ranking search results | No | Orders existing documents |
| Recommending products | No | Selects from an existing catalog |
| Transcribing a meeting | No | Records existing speech as text |
| Answering a question in a chat | Yes | The answer is newly composed text |
The exam-question variants, worked through
This question appears in AI fundamentals courses and certification exams because it tests whether you can separate the generative family from the broader machine-learning toolbox. The multiple-choice version nearly always pairs one creation task against three judgment tasks. Two representative shapes:
- "Which task is a generative AI task: forecasting demand, classifying images, writing a product description, or detecting anomalies?" Answer: writing a product description. The other three output predictions, labels, and flags about existing data.
- "Which of the following is NOT a generative AI task: composing an email, summarizing a document, ranking search results, or generating an image?" Answer: ranking search results. It orders existing items; the other three create new artifacts.
If an unfamiliar option appears, apply the one-question test rather than pattern-matching keywords. "Generating a forecast" is still prediction (the output is a number), and "creating a spam filter" is still classification; exam writers enjoy planting creation verbs in front of judgment tasks.
Generative vs discriminative, in one minute
The formal version of this distinction: discriminative models learn boundaries between categories, so given an input they answer "which one?" or "how much?". Generative models learn the structure of the data deeply enough to sample new examples from it, so they answer "here is one". IBM's and Google Cloud's primers both draw the line the same way: generative AI is defined by producing novel content, in contrast to systems that classify or predict over existing data.
In practice, the two are teammates. A support product might use a discriminative model to route a ticket (classification), then a generative model to draft the reply (generation). The same handoff runs inside Gravity's own agents: a judgment step decides what the task needs, a generative step produces the draft, and the user only ever sees the finished result. Neither replaces the other; they hand off.
Beyond generation: when tasks chain into agents
Most real work is not one task. "Handle my overdue invoices" involves reading the ledger (retrieval), deciding which are overdue (classification), writing reminder emails (generation), sending them (action), and checking replies (judgment). Software that chains those steps toward the goal and knows when to pause for a human is an AI agent, and generation is just its creative engine. That layered relationship (generation inside, pursuit of a goal outside) is the subject of agentic AI vs generative AI, with the conversational sibling covered in AI agent vs chatbot.
If the exam question brought you here but the underlying interest is what this technology can actually do for your work, start with what is an AI agent?, browse concrete AI agent examples, and calibrate expectations with what can an AI agent actually do?. When you want the buying view rather than the theory, the category guide is what is an AI agent platform?.
Frequently asked questions
Which task is a generative AI task?
Any task where the AI produces new content is a generative AI task: writing an email or article, generating an image from a description, composing code, drafting a summary, creating audio or video. Tasks that only label, rank, predict, or retrieve existing information, like classifying spam or forecasting sales, are not generative.
Is summarizing a document a generative AI task?
Yes. A summary is new text the model composes; it did not exist in the source in that form. This regularly trips people up because summarizing feels like extraction, but the model writes original sentences, which makes it generation.
Is classifying emails as spam a generative AI task?
No. Spam filtering assigns an existing label to an existing item; nothing new is created. It is a classification task, the textbook example of discriminative rather than generative AI, even when a large language model performs it.
Is predicting or forecasting a generative AI task?
No. Forecasting sales, predicting churn, or estimating a delivery time outputs a number or probability about existing patterns, not new content. Prediction is classic machine learning; generation creates artifacts like text, images, code, or audio.
What is the difference between generative and discriminative AI?
Discriminative models draw boundaries: given an input, they decide which category it belongs to or what value to predict. Generative models learn the underlying patterns well enough to produce new samples: new sentences, new images, new code. Ask whether the output is a label or a creation.
Is using an AI agent a generative AI task?
It is more than one. An agent chains many tasks together: some generative (drafting the email), some not (deciding which tool to call, checking a result). Generation is the creative engine inside; the agent adds planning, tool use, and follow-through until the work is finished.
Sources
- IBM, "What is generative AI?", ibm.com, backs the definition of generative AI as systems that create novel content (text, images, audio, code) in contrast to classification and prediction systems.
- Google Cloud, "What is generative AI?", cloud.google.com, backs the content-creation framing and the task examples used in the comparison table.
- IBM, "What is agentic AI?", ibm.com, backs the description of agents as systems that chain generative and non-generative steps toward a goal.