Artificial intelligence covers many techniques: recognizing an image, detecting fraud, recommending a product, forecasting demand. Generative AI is a particular branch of it. It produces new content, text, images, sound, video or code, from a request written in everyday language. Conversational assistants such as ChatGPT, Gemini or Claude are the best-known examples.
Where generative AI comes from
Generative techniques have existed for a long time, but three steps made their large-scale use possible:
- 2017: Google researchers publish "Attention Is All You Need", which describes the Transformer architecture. It makes it possible to train models efficiently on huge amounts of text and underpins most of today's language models
- 2020: OpenAI presents GPT-3, a model able to write coherent text on almost any subject from a few lines of instruction
- 30 November 2022: OpenAI launches ChatGPT, which makes these capabilities available to everyone in a conversational interface
Since then, several providers have offered competing models, and capabilities improve with each generation: longer texts, several languages, reading images and documents, using tools.
How it works, in one sentence
A generative text model predicts, piece by piece, the most likely continuation of a text, based on what it learned during training and on what it is given to read at the time of the request. These pieces are called tokens. This principle explains both the fluency of the answers and their main flaw: a likely text is not necessarily a true one.
What generative AI does well
- Understanding a freely worded question, with its typos, detours and implications
- Rephrasing information in another register, length or language
- Summarizing and structuring a long document or a conversation
- Writing a first draft: an e-mail, a product sheet, a template reply
- Extracting information from free text: a name, a date, a stated need
What it does badly, or not alone
- Guaranteeing that a fact is accurate. Without a reliable source, a model can confidently make things up; this is called hallucination
- Knowing your business. It knows nothing about your rates, opening hours or terms unless you provide them
- Knowing current events. Its general knowledge stops at the end date of its training
- Making a binding decision. A discount, a refund or a diagnosis belong to a person who takes responsibility for them
Generative AI in a business assistant
A business AI assistant uses generative AI within a narrow frame. It is given a prompt that sets its role and limits, the organization's documents as the source of truth, and rules for handing over to the team. The model then serves mostly to understand the question and write a clear answer from the documents, rather than drawing on its general knowledge.
This frame changes the nature of the risk. The question is no longer "does the model know?" but "do the documents say the right thing, and does the model stick to them?". That is why the quality of the knowledge base, source citation and reviewing conversations matter more than choosing the latest model.
What European law says
The European regulation on artificial intelligence, the AI Act, governs generative AI at two levels. Providers of general-purpose AI models, such as those developing large language models, have documentation and transparency obligations. Organizations that deploy an AI system in contact with the public must, among other things, inform people that they are interacting with an AI. Generated content must also, in some cases, be identifiable as such.
Personal data processed in conversations remains subject to the GDPR, as with any other processing.
Generative AI and "classic" AI
The two are not opposed. Many systems combine generative building blocks with older ones: a keyword search to find a document, a fixed rule to block a topic, a classification model to spot an emergency. A good assistant uses generative AI where it adds something, understanding and wording, and deterministic mechanisms where predictability matters.
Generative AI in KernChat
KernChat uses generative models available on the market (Google Gemini, Anthropic Claude or OpenAI) to understand visitors' questions and write answers from the organization's documents, in French, Dutch and English. Every answer that relies on a document cites its source, the visitor is told they are talking to an assistant, and reserved topics go to the team.