What Is Generative AI?

In its introduction to machine learning, Google describes generative AI as a class of models that creates content from user input. It can take a variety of inputs and create a variety of outputs, such as text, images, audio and video. The input is called a prompt; how to write one is covered in the prompt engineering entry.

What sets it apart from other machine learning is the output. A spam filter assigns an email to a class, and a bidding system estimates a probability. A generative model returns something new that did not exist in that form in the training data.

Recital 99 of the EU AI Act names large generative AI models as a typical example of general-purpose AI models, because they generate content flexibly, as text, audio, images or video, and can handle a wide range of distinct tasks. Germany's Federal Office for Information Security (BSI) also classes generative AI models as a subset of these general-purpose models.

During training, a generative model learns how the data it is trained on is structured. It then creates new examples that follow that structure. Three papers stand for the approaches most widely used today.

In 2014, Ian Goodfellow and colleagues introduced generative adversarial networks. Two models are trained at the same time: one generates data, and the other estimates whether an example came from the training data or from the generator.

In 2017, Ashish Vaswani and colleagues introduced the Transformer, an architecture based solely on attention mechanisms. They tested it on machine translation. How a language model generates text from there, token by token, is explained in the LLM entry.

In 2020, Jonathan Ho, Ajay Jain and Pieter Abbeel showed high-quality image synthesis with diffusion models. These models learn to remove noise from an image step by step, and in this way create a new image from pure noise.

The same principle applies to audio and video, with different data. Some models combine several types: they take an image and a text as input, for example, and generate a video.

In marketing, generative AI produces drafts of copy, image variants, translations, summaries and spoken versions. Google's guidance on generative AI content says it can be particularly useful when researching a topic and to add structure to original content.

For search, Google draws the line at the purpose, not the tool. In February 2023 it wrote that not all use of automation, including AI generation, is spam. Its spam policies define scaled content abuse as generating many pages for the primary purpose of manipulating search rankings, and the first example they give is using generative AI tools to generate many pages without adding value for users.

Generative models produce plausible outputs, not verified ones. The BSI calls hallucinations a major problem because the outputs usually seem credible, down to made-up references.

Two legal questions come on top. Under Article 50(2) of the AI Act, providers of AI systems that generate synthetic audio, image, video or text content must mark the outputs in a machine-readable format; deadlines and the duties of users are covered in the AI content labelling entry. And anyone publishing content online can reserve text and data mining by machine-readable means under Article 4(3) of Directive (EU) 2019/790. Providers of general-purpose AI models must identify and comply with such a reservation under Article 53(1)(c) of the AI Act.

• Use generative AI for drafts, variants and structure, and check every factual claim against its source before publishing.
• Never copy figures, quotes or references from an AI output without checking them.
• Add your own experience, data and examples, which the model cannot know.
• Keep a record of which content was made with which tool, so you can meet labelling obligations.
• Before use, clarify whether inputs contain confidential or personal data and what the provider does with them.
• Decide your own position on text and data mining and, if you want, declare a reservation by machine-readable means.

• Mistaking generative AI for a search engine and treating its answers as verified facts.
• Generating many pages with no added value automatically, assuming Google will not notice.
• Equating generative AI with language models and overlooking image, audio and video models.
• Publishing generated images of real people or places without checking the labelling rules.
• Entering customer data or contract details into a public tool without knowing its terms of use.