Which task is a generative AI task? One rule and 12 worked examples

A task is a generative AI task when the thing the model hands back is new content: a sentence, a picture, a sound, a clip or a program that did not exist before the request. If the model hands back a label, a score, a number, an order or a document that already existed, the task is not generative, whatever model is running underneath. Those two sentences settle the quiz. The rest of this page is the rule written out once, twelve tasks run through it, and the one case (translation) where answer keys disagree.

I have marked this question on intro course quizzes for three years, and the wrong answers cluster in one place: people judge the task by how impressive the model is, not by what comes out of it. A fraud model at a bank is enormous, expensive and clever, and it is not generative. A tiny script that turns a line of text into a spoken sentence is generative. The output decides, never the size.

What is a generative AI task?

A generative AI task is any task whose required output is new content, meaning text, an image, audio, video or code that is produced rather than selected, scored or predicted. That sentence is the standalone definition; everything below is evidence for it and practice with it.

The vendors and the regulator agree on the shape of that definition even where their wording differs. AWS writes that generative AI "can create new content and ideas, including conversations, stories, images, videos, and music," and calls it "a subset that generates new content" inside the broader field of AI. IBM, in a page updated 18 March 2026, defines it as AI "that can create original content such as text, images, video, audio or software code in response to a user's prompt or request." The NIST Generative AI Profile (NIST AI 600-1, July 2024) adopts the US Executive Order 14110 wording: "the class of AI models that emulate the structure and characteristics of input data" so as to "generate derived synthetic content. This can include images, videos, audio, text, and other digital content." The Google machine learning glossary, last updated 10 April 2026, is the most candid of the four: it calls generative AI "an emerging powerful field with no formal definition" and then says most experts agree the output has to be "complex", "coherent" and "original."

Four sources, one common noun: content. Nobody defines generative AI by the model or the dataset. They define it by what comes out.

What is the one-line rule for deciding?

Ask whether the output is content that did not exist before the request. If yes, the task is generative. Text, image, audio, video, code: if the model produced one of those and it is new, you have your answer. If the model produced a label, a rank, a probability, a forecast or a pointer to an existing item, you have the other answer.

Two checks catch the tricky cases. First, "new" means the artifact is new, not that the information in it is new. A summary contains nothing you could not find in the source, and it is still a fresh paragraph nobody wrote before, so it is generated. Second, "content" means something a person would consume as content. A probability of 0.93 is not content, even though the model produced it. A four-word caption is content, even though it is short.

The same glossary gives the technical version of the rule. A discriminative model is one that predicts labels from the inputs it is given, and "the vast majority of supervised learning models, including classification and regression models, are discriminative models." A generative model, by contrast, "creates (generates) new examples from the training dataset," such as writing poetry after training on poems. Discriminative models answer "which?" and "how much?" Generative models answer "make one."

Which tasks count as generative? Six examples with the reason

Every task below passes the rule because its output is new text, image, audio, video or code. The reason is the one-line rule applied to that task.

1. Writing a product description from five bullet points. The output is a paragraph of text that did not exist. The bullets are the prompt, the paragraph is new content, and this is the canonical yes answer.

1. Producing an image from the prompt "a lighthouse in a snowstorm, watercolor." The output is an image file whose pixels were sampled by the model. No lighthouse photo was retrieved; one was made.

1. Reading a script aloud in a synthetic voice. The output is an audio waveform. Text-to-speech is generative because the sound did not exist before; the model produced every sample of it.

1. Turning a still image and a text instruction into a six-second clip. The output is a set of video frames that were never filmed.

1. Writing a function from a docstring that says "return the median of a list, ignoring None values." The output is source code. Code is text with a compiler, and the function is new content, so code generation is generative.

1. Replying in character in a roleplay chat. The output is the character's next line, which is new text conditioned on the persona and the conversation so far. A companion character that answers "you remembered my sister's name" is doing generation on every turn. This is the task that platforms built around characters run all day, and the fields that shape that output (backstory, traits and greeting) are described in what an AI character creator actually needs from you.

Notice that none of those six depends on how good the output is. A bad poem is still generated. A wrong function is still generated. Quality is a separate question from category.

Which tasks do not count as generative? Six examples with the reason

Every task below fails the rule for the same underlying reason: its output is a label, an order, a flag, a number, a pointer to an existing item, or a grouping. None of those is new content. The model may be large. It may even be a language model. The task still is not generative.

1. Classifying an email as spam or not spam. The output is one of two labels. The Google glossary uses exactly this example for a discriminative model, one that predicts whether an email is spam from its measurable properties and the model weights. An AI-text detector is the same task with different labels, which is why the tools in AI humanizers and detectors are classifiers, not generators.

1. Ranking search results for a query. The output is an ordering of pages that already exist. The glossary defines ranking as "a type of supervised learning whose objective is to order a list of items." Nothing new gets written. Something existing gets sorted.

1. Flagging a credit card transaction as fraud. The output is a flag and usually a score. This is detection, which the glossary files under anomaly detection, "the process of identifying outliers." The transaction and the account both existed before the model ran.

1. Forecasting next week's demand for a product. The output is a number, or a curve of numbers. Forecasting is regression over time, and regression models are discriminative by the glossary's own sentence. A forecast is a prediction about content (sales), not content itself.

1. Finding the three policy documents most relevant to a question. The output is three pointers to existing files. This is retrieval, and it's worth separating from what happens next. In retrieval-augmented generation, the retrieval step fetches documents "retrieved after the model was trained" and the generation step writes an answer from them. The fetch is not generative. The answer written afterwards is. A quiz that names only the retrieval half wants "not generative."

1. Grouping 10,000 support tickets into topics. The output is a set of cluster assignments. Clustering produces a partition of existing items and no new item, so it is not generative even though it runs unsupervised, and even though a language model may embed the tickets first.

The pattern across all six: the input already contains every artifact that comes out the other end. The model adds a judgement about those artifacts. Generation adds an artifact.

Is translation a generative AI task?

Translation is generative by mechanism and contested on answer keys. For a quiz, the honest answer is "yes, unless the question contrasts it with something like writing a story." A translation model reads a source sentence and writes a target sentence token by token, exactly the way a chatbot writes a reply. The architecture behind every large text generator was introduced as a translation system: the Transformer paper (Vaswani and colleagues, 2017) reports "experiments on two machine translation tasks," with a score of 28.4 BLEU on the WMT 2014 English-to-German task. If translation is not generative, then the architecture behind every text generator was not doing generative AI on the day it was published.

The case against runs like this: the meaning already existed, so the output is a transformation of given content rather than "original" content, and the Google glossary's word "original" is doing the work. Some answer keys sort translation with summarization and paraphrase as transformation and reserve "generative" for open-ended creation. That is a vocabulary choice, and my opinion, which you may argue with, is that it teaches the wrong thing. A student who learns that translation is not generative will later be surprised when a translator invents a clause that was not in the source. It invents because it generates, and that failure is the same one image generators have.

How to answer on a quiz: apply the one-line rule to every option. If translation is the only option whose output is new text, pick it. If the options include both translation and something like "write a cover letter," the examiner is drawing the "original versus transformed" line, and the cover letter is the intended answer. Summarization and transcription sit in the same spot. Speech-to-text writes new text that carries old information, and most keys I have seen count it as not generative for that reason, while abstractive summarization usually gets counted as generative. Neither key is wrong about the mechanism. They disagree about what "new" means.

Why does the word "generative" mean two different things?

The statistics meaning of "generative model" is older than generative AI, and it names a way of modelling data rather than a kind of output. That's why naive Bayes, a spam classifier, gets called a generative classifier. Ng and Jordan's 2001 NeurIPS paper compares "discriminative and generative learning as typified by logistic regression and naive Bayes," and both of those are classifiers producing labels. In that sense, "generative" means the model learns the joint distribution of inputs and labels. AWS puts it the same way: generative modelling "calculates the probability of x and y occurring together," while a discriminative model predicts a label from the inputs alone.

The course meaning, the one behind the quiz, is about the task. A naive Bayes spam filter is a generative model doing a non-generative task. A large language model asked to output only "yes" or "no" is a generative model doing a classification task. The quiz is asking about tasks, so the output rule applies and the modelling vocabulary can be set aside. If an exam question ever says "generative model" and lists naive Bayes among the options, it's testing the statistics meaning, and naive Bayes is a correct answer there.

Key takeaways

• A task is generative when its output is new content in one of five forms: text, image, audio, video or code.
• Labels, rankings, flags, forecasts, retrieved items and clusters are outputs of prediction tasks, not generation, no matter how large the model.
• Translation, summarization and transcription generate new text from old information. Answer keys split on whether that counts, and the mechanism is generative either way.
• "Generative model" in statistics describes how a model learns, and a generative model can do a non-generative task. The quiz is asking about the task.

Frequently asked questions

Which of the following is a generative AI task?

The option whose output is new text, image, audio, video or code. Run each option through that test. The options that produce a label, a ranking, a score, a forecast or a retrieved item are the distractors, and in a typical four-option version, three of the four are prediction tasks.

Is spam detection a generative AI task?

No. Spam detection outputs one of two labels for an email that already exists, and the Google machine learning glossary uses the spam filter as its example of a discriminative model.

Is machine translation generative AI?

Yes by mechanism: a translation model writes the target sentence token by token, and the Transformer architecture behind today's text generators was published in 2017 as a translation model. Some course keys count translation as a "transformation" rather than generation because the meaning already existed, so read the other options before answering.

Is summarizing a document generative AI?

Yes, for abstractive summarization, where the model writes a new paragraph in its own words. Extractive summarization, which selects existing sentences and returns them, is a selection task and not generative.

What is the difference between generative AI and predictive AI, with examples?

Generative AI produces content (a product description, an image, a spoken sentence, a video clip, a function). Predictive AI produces a judgement about existing data (spam or not, fraud or not, next week's demand, the best-matching document). AWS describes traditional machine learning models as "discriminative or focused on classifying data points," and generative models as ones that try to reconstruct the inputs that go with a label.

Is a chatbot replying in character generative AI?

Yes. Every reply is new text conditioned on the character's persona and the conversation so far, which is the same task as writing a story one line at a time. The general kinds of assistant that do this are sorted in what an AI assistant is and the four kinds.

A character reply on YourCrush.AI is the cleanest example: the words did not exist until the model wrote them in that character's voice.

The rule leaves one thing unresolved on purpose. Under it, a model that writes "yes" is doing classification and a model that writes "yes, because the invoice total does not match the line items" is doing generation, and the boundary between those two outputs is a handful of tokens. Anyone who says the line is sharp has not marked enough quizzes.

Editorial Team, Content Strategist. Last updated: 2026-09-21