Generative AI

Generative AI is software that makes something new on request: a text, an image, a voice recording or computer code. You type an instruction, such as "write a friendly reply to this complaint", and get a draft within seconds. ChatGPT is a well-known example. A business can use it for product texts and customer emails, but it sometimes states false things with confidence, so someone checks the result.

How does it differ from other AI?

McKinsey (April 2024) describes generative AI as algorithms, such as ChatGPT, that can create new content, including audio, code, images, text, simulations and videos. Before it, machine learning was largely limited to models that classify or predict, such as recognising a cat in a photo. A generative model can create a new picture or description of a cat on request. Google's machine learning course draws the same line: a discriminative model can tell a dog from a cat, a generative model can make new photos of animals that look like real ones (Google for Developers).

How does it work?

A generative model learns from a very large collection of examples what that material looks like, and then makes new material with the same patterns. In Google's formal definition, a generative model captures the probability distribution of the data itself. Three common families of models do this in different ways:

• Language models, such as the large language models behind ChatGPT, predict the next token, a piece of a word, and repeat that until the answer is complete (autoregressive generation).
• Diffusion models start from random noise and remove it step by step until an image appears (Ho et al., 2020). Stable Diffusion does this in a compressed latent space, which takes much less computing than working on the pixels directly (Rombach et al., 2022).
• A GAN (generative adversarial network) trains two networks together: a generator makes examples, and a discriminator estimates whether each one comes from the training data or from the generator (Goodfellow et al., 2014).

Multimodal models handle several kinds of input and output, such as text and images, in one model. Most generative models include an element of chance, so the same request can give a different result each time; in a language model, the temperature setting controls how much.

What do people use it for?

Researchers at OpenAI, Duke and Harvard analysed a representative sample of ChatGPT conversations (Chatterji et al., NBER, September 2025). By July 2025, ChatGPT had more than 700 million weekly users, around 10% of the world's adult population. At work, writing was the most common use, about 40% of work-related messages in June 2025, and about two-thirds of all writing requests asked ChatGPT to edit, critique or translate the user's own text instead of writing something new.

Where does it go wrong?

The US standards institute NIST lists twelve risks of generative AI in its Generative AI Profile (July 2024). Among them:

• Confabulation, better known as hallucination: false content stated with confidence. NIST calls it a natural result of how these models are designed, since their output approximates the statistical distribution of their training data. It is mostly a problem in text; in images, audio or video, invention can be exactly what you want.
• Harmful bias: output that amplifies existing biases or works less well for some groups or languages, for example because the training data was not representative.
• Intellectual property: copyrighted or trademarked content reproduced without permission.
• Data privacy: personal or sensitive data that leaks or is disclosed without authorisation.

McKinsey advises organisations to have a person check the output before it is published or used (human-in-the-loop), and to avoid generative AI for critical decisions, such as those involving significant resources or people's welfare.

Example

Suppose a webshop has 300 products and only the suppliers' spec sheets. A language model drafts a description for each product from its spec sheet, in the shop's tone of voice, and translates it into German. An employee reads every text before it goes live, because the model can add a feature the product does not have.