Generative AI

Traditional AI systems recognize, sort and predict. They return a category ("spam"), a number (expected sales) or a recommendation. Generative AI returns something that did not exist before: a draft, an image, a summary, a piece of code. In its generative AI profile from July 2024, the US National Institute of Standards and Technology (NIST) describes it as the class of AI models that emulate the structure and characteristics of input data in order to generate derived synthetic content.

Technically, generative AI is part of machine learning and today almost always runs on deep learning, meaning large neural networks. The best-known examples are large language models (LLMs) for text. There are also models for images, audio, video and code, and more and more models handle several of these at once. In 2021, Stanford researchers coined the term foundation model for such broadly trained base models on which many applications are built. Generative AI reached a mass audience with ChatGPT at the end of 2022.

For executives, what matters most is the different failure mode. A traditional prediction is wrong in a measurable way. A generated output can be fluent, plausible and still wrong. Generative AI also raises its own legal questions: labeling under the EU AI Act, copyright in inputs and outputs, and data protection for everything employees type into a tool.

How a generative model is built.

It starts with pretraining on very large datasets. A language model learns to predict the next piece of text. Because the right answer is in the text itself, no human-labeled examples are needed, which is why this is called self-supervised learning. By the end, the model has tuned billions of internal weights so that it can reproduce language, knowledge and style to a high degree.

The model is then trained further so that it follows instructions and gives helpful answers. A common method is reinforcement learning from human feedback (RLHF): people rate answers, and the model learns to give preferred answers more often. OpenAI described this approach for InstructGPT in 2022.

In use, a language model builds its answer piece by piece, each time from the probability of the next token. Image models today usually work differently. Diffusion models, described in a widely cited form by Jonathan Ho and colleagues in 2020, start from noise and remove it step by step until an image emerges that matches the description. Older methods such as generative adversarial networks (GANs, 2014) play a smaller role today. What they all share: the model reproduces patterns; it does not look up facts.

Traditional AI and generative AI compared.

The reliable difference is the output. Traditional AI classifies or forecasts: this email is spam, this part is defective, next week we will probably sell 400 units. Generative AI writes, draws or codes. Both rest on machine learning, and many systems combine them, for example when a classifier decides which requests a language model may answer.

The difference calls for a different way of testing. With traditional AI there is usually one right answer. You measure how often the model is wrong on test data and decide whether that rate is acceptable. With generative AI there are many possible good answers. Whether a text is correct, complete and in the right tone can only be judged with review criteria, spot checks and human judgment.

The errors look different too. Traditional AI puts things in the wrong category. Generative AI invents content that sounds plausible: sources, quotes, figures, legal provisions. NIST calls this confabulation; in everyday use it is called hallucination. The rule of thumb that follows: the more an output goes outside the company or triggers decisions, the tighter the review before it is passed on.

Generative AI and LLMs are not the same thing.

A large language model is a generative model for text. Generative AI is the broader term and also covers image, audio, video and code models. Many current models are multimodal: they read text, images and documents and respond in text or create images. The lines between model types are blurring as a result.

Just as important is the line between model and application. ChatGPT, Claude and Gemini are applications, each backed by several models. The model itself has no interface, no memory beyond the conversation and no access to company data. The application builds those capabilities around it, for example by searching documents (retrieval-augmented generation, RAG) or by connecting tools.

AI agents go one step further. They use a generative model to break tasks into steps, call tools and react to the results. That means they also trigger actions, which calls for different safeguards. When people talk about generative AI in a company, it helps to clarify whether they mean a model, an application or an agent. Costs, risks and responsibilities differ considerably from one level to the next.

Where generative AI works in operations and where it doesn't.

Generative AI suits work where a good draft saves a lot of time and a person can check the result quickly. Typical cases are drafts and translations, summaries of long documents, suggested replies in customer service, searches across internal knowledge and code. What they have in common is that the model does the legwork and the judgment stays with a person.

Generative AI is weak where exact facts without a source, reliable arithmetic or identical results every time are required. A language model may answer the same question differently twice. For calculations, bookings or legally binding statements it is at most an aid, framed by fixed rules or conventional software.

In practice, one simple question per use case works well: how expensive is an error, and how quickly will it be noticed? An internal draft that someone will rework anyway tolerates more uncertainty than a customer email sent automatically. That tells you where human sign-off is mandatory and where spot checks are enough. Whoever passes a result on is accountable for it, regardless of who or what wrote it.

Labeling, copyright and data protection.

Article 50 of the EU AI Act sets transparency obligations that have applied since August 2, 2026. Providers of generative systems must mark outputs in a machine-readable format as artificially generated, for example through metadata or watermarks. For systems already on the market before that date, the amending Regulation (EU) 2026/1744 allows a transition period until December 2, 2026. Companies using generative AI must disclose deepfakes. AI-generated text published to inform the public on matters of public interest must also be disclosed, unless a person has reviewed it and someone holds editorial responsibility (as of November 2026).

Since August 2, 2025, providers of general-purpose AI models have had to maintain a copyright compliance policy and publish a summary of their training content, among other duties. For outputs, German copyright law protects only an author's own intellectual creation under Section 2(2) of the Copyright Act (UrhG). A purely machine-generated result is therefore generally not protected; the greater a person's creative contribution, the more likely protection becomes.

The GDPR applies to everything employees enter into prompts. Before rollout, it should be clear whether the vendor stores inputs or uses them for training and whether a data processing agreement is in place.

What good implementation looks like.

Good implementations start with access, not with the tool. Employees sign in with their company account, permissions follow their role, and there is an approved route to AI. Without that route, many people use personal accounts, and the company loses sight of which data goes where. This is known as shadow AI.

The second sign is that results can be checked. Answers drawn from internal knowledge cite their source. Recurring tasks have a list of real sample cases with known good answers, and every change is tested against it, whether a new model, a revised prompt or new documents. Judging a demo alone means the weaknesses show up only in production.

The third is control. Every model call is logged with its cost, so it stays visible which application causes which spend. The application is not locked to a single model, so it can switch without a rebuild when prices, quality or contract terms change. And training uses real tasks from each team's own work, because people who have seen the typical errors in their own process review results more effectively. These are the points worth raising with any vendor.

Where the term stops working

Generative AI produces plausible content, not verified facts, and it does not know whether its output is correct. It should be distinguished from traditional AI, which classifies and predicts, and from rule-based automation such as RPA, which creates nothing new. An LLM is one form of generative AI for text, not a synonym. AI agents use generative models but also act through tools.