Which task is a generative AI task?
The generative AI task is the one whose output is new content: a poem, an email draft, an image, a block of code. Tasks that return a label, a number or a reordered list are not generative. Google's machine learning course draws the line in two sentences: "Generative models can generate new data instances." and "Discriminative models discriminate between different kinds of data instances." The course page is the plain version of the whole topic.

Use this test on any quiz. When the task finishes, is there something on the screen that didn't exist before, or is there a label, a number or a shuffle of what was already there? A new thing means generative. In a typical four option question, three choices count, detect or sort, and one writes, draws or composes. That one is the answer.
One real example: a study site lists the question with four options, which are calculating the total word count in an article, detecting credit card fraud, identifying the main topic of a news article, and writing a poem based on a given theme. Its answer is writing a poem, because the poem is new text. The other three each return a count, a flag or a label.
Which of the following is a generative AI task? 18 tasks sorted
Here are 18 tasks, each marked generative or not, with the reason. The reason column is the part to learn, because a quiz changes the task but not the logic. The sorting is this page's own, built on the output test above and on the definitions from Google and Wikipedia.
| Task | Generative? | Reason |
|---|---|---|
| Writing a poem from a theme | Yes | The poem is new text |
| Drafting an email from a short brief | Yes | The draft is new text |
| Making an image from a text prompt | Yes | The picture is new content |
| Writing code from a description | Yes | The program is new text that runs |
| Summarizing a report | Yes | The summary is newly written sentences |
| Composing a short tune in a given style | Yes | The melody is new audio |
| Reading text aloud in a synthetic voice | Yes | The audio waveform is generated |
| Suggesting a new product slogan | Yes | The slogans are new text |
| Translating a paragraph | Borderline | The output is new text, though it follows the input closely |
| Calculating the word count of an article | No | The output is a count |
| Detecting credit card fraud | No | The output is a flag on a transaction |
| Identifying the main topic of an article | No | The output is a label from a set |
| Marking an email as spam or not spam | No | The output is one of two labels |
| Predicting a house price | No | The output is a single number |
| Sorting records by date | No | The output is the same records reordered |
| Recognizing a face in a photo | No | The output is a match to a known person |
| Recommending the next video to watch | No | The output is a pick from existing items |
| Finding an existing document by keyword | No | The output is retrieval of what is already stored |
Count the rows: eight yes, one borderline, nine no, which makes 18. The pattern holds across the nine no rows. Each one returns something that was always inside the input or a short list of options, so a discriminative or plain computing method finishes the job. Generative models do something different, and Wikipedia describes it as learning the patterns of training data and using them to generate new data in response to input.
Is summarizing, translating or classifying a generative AI task?
Summarizing and translating are generative, and classifying usually is not. A summary is a fresh piece of writing that you never typed, and a translation is a fresh paragraph in another language. Both pass the new content test, though translation sits closest to the line because it follows the source word for word in places.
Classifying is the trap. A chatbot can answer "spam" when you paste an email, so the tool is generative, but the task is classification, because the answer is one word from a fixed list. Quiz writers mean the task, not the tool. Lists on the web are not consistent here, so when two options both look possible, take the one that produces content over the one that produces a label.
The same logic covers sentiment: calling a review positive returns a label, so it isn't a generative task, while writing a polite reply to that review is. For the wider map of what counts as AI at all, types of AI sorts the families, and what is machine learning covers the prediction side that this page keeps on the not generative list.
Which task is best suited for generative AI?
Generative AI suits tasks where you want a first version fast and can judge the result yourself. Drafts, rewrites, summaries, outlines, example code and image concepts fit, because the output is something a person reads or looks at before using. These are the five outputs Wikipedia names: text, images, videos, audio and code.
It suits exact answers badly. A model writes likely text, so for a sum of 4,000 cells, a total or a date difference, a spreadsheet formula or a calculator is the right tool, and the generative option is the wrong one even when it sounds confident. Why AI gets things wrong explains the mechanism, and what is an AI hallucination shows what the failure looks like.
That is also why a quiz question about the most reliable generative task has no fixed answer. Reliability depends on whether you can check the output, and drafts you can read in a minute are easier to check than a long chain of numbers.
How do generative models make new content, and where do businesses use them?
TechTarget names three model families behind the new content. Generative adversarial networks pair a generator with a discriminator that tries to spot generated content. Transformers made it possible to train ever larger models and added attention, which tracks links across longer text. Diffusion models are used in image generation and turn data into a latent form, then reverse the process to build an image.
The same article lists use cases that all pass the new content test: customer service chatbots, writing email responses, resumes and business reports, suggesting new drug compounds to test, and generating code. It also lists analyzing and summarizing events from security and IT logs, which is generative only in the summary, because the log events themselves already exist.
Three real cases worked through
Case 1, the quiz item. Question: which task is a generative AI task? Options: count the words in an article, detect credit card fraud, identify the main topic, write a poem from a theme. Work it: count gives a number, fraud gives a flag, topic gives a label, poem gives new text. Answer: the poem. A good explanation names the output of all four, not only the winner.
Case 2, the school question. Another study site asks what would be an appropriate task for using generative AI, and answers with brainstorming topics, developing a draft, proofreading, getting feedback and using an AI draft in class discussion. Each of those produces new text, so each is generative. The same answer warns that a school's AI policy decides what is allowed, so read your own before you use it on graded work.
Case 3, the work question. A manager pastes a three page report and asks two things: the total of the cost column, and a five line summary for the team. The summary is generative, because the five lines are new sentences. The total is not, because it is a figure that arithmetic gives exactly. The right move is a formula for the total and a model for the summary.
These cases are examples written for this page, not outputs from a tested model. The logic, not the tool, carries the answer.
What are examples of generative AI tasks by kind of output?
Generative AI tasks fall into five kinds, one for each output. Wikipedia lists them in its first sentence: text, images, videos, audio, computer code or other forms of digital data. Here is a quick set of examples for each.
- Text: drafting an email, writing a story, summarizing an article, rewriting a paragraph in a friendlier tone, answering a question in full sentences.
- Images: a poster concept from a prompt, a product mockup, a cartoon character, an edited version of a photo you attach.
- Audio: a voice reading a script, a short jingle, a sound effect.
- Code: a function from a description, a test for existing code, a script that renames files.
- Video: a short clip from a prompt, which Wikipedia notes can now be photorealistic.
For the history and the models behind these, what is generative AI covers how they work. The idea that a model could make new examples at all goes back to the 2014 paper on generative adversarial networks by Goodfellow and colleagues, and the public caught on when OpenAI released ChatGPT on 30 November 2022.
What can you ask an AI to check your understanding?
The fastest way to learn the line is to test it against a model. Paste any of these into Claude, ChatGPT, Gemini or Whizi, and read the reasons it gives, not only the verdicts.
- Sort a list: "Here are ten tasks: [paste]. Mark each one generative or not, give a one sentence reason, and flag any borderline case. Use the rule that a generative task produces new content."
- Make me a quiz: "Write five multiple choice questions in the style of 'which task is a generative AI task'. Use four options each, one correct, and give the reason for every wrong option."
- Argue the edge: "Is [task, such as translation or sentiment analysis] a generative AI task? Give the best argument for yes and for no, then say which one a course quiz would expect."
- Rewrite a task: "Take this classification task: [paste]. Rewrite it as a generative task that uses the same input, and say what the new output is."
- Check my answer: "I think [task] is generative because [my reason]. Find the weakest step in that reasoning and stop there."
Treat the output as study help, not as an answer key. If your course publishes its own definition, the course's wording wins over any model.
Where does Whizi fit when you study this?
Whizi is one app instead of a pile of single purpose tools, and it suits this topic because you can ask Claude, GPT and Gemini the same question and see where they differ. Pick a borderline task from the table, put it to two models, and compare the reasons. Side by side Compare is a Pro feature, and its first week is $1.99.
The free plan gives you 10 messages a day on Whizi AI, Auto and Claude Haiku, with no card. Web search is on by default in each chat, which helps when you want a course's own definition checked against a source. A dedicated study site is quicker for a single quiz answer, and a free tool with no sign up beats an account for one lookup.
Where does this test break down?
The new content test is a study rule, not a law. Some real systems blend the two jobs: a model that rates a loan application and writes the reasons is classifying and generating at once, and a tool that finds a document and writes a summary of it does both. When a quiz asks for one answer, choose by the main output.
The wider risks of generative tools, from made up facts to copied style, have their own page: generative AI risks. For a school, college or course, the policy of that institution decides what use is allowed, and a model can't tell you what your instructor will accept.
- Name the output of the task before judging it
- Mark a task generative only when the output is new content
- Treat a label, a number, a flag or a sorted list as not generative
- Count summaries and translations as generative
- Pick the content creating option when two answers look alike
Frequently asked questions
Is classifying an email as spam a generative AI task?
No, it isn't. The output is one of two labels, so it is classification. A chatbot could answer the question, but the task still returns a label from a fixed list, and quiz writers judge the task and not the tool.
Is summarizing a text a generative AI task?
Yes, it is. A summary is newly written text, so it passes the new content test, even though it rests on the original. Pulling out a keyword or counting words does not pass, because those return an item or a number.
Is writing code a generative AI task?
Yes. Wikipedia lists computer code among the outputs of generative AI, and a model can write a program from a prompt. A person still has to run and test the result, because a program that looks right can fail.
What is the difference between generative and predictive AI?
Generative AI makes new content, and predictive AI returns a label or a number about data it was given. Google's machine learning course frames a close cousin of that split: generative models create new data instances, while discriminative models tell kinds of data apart.