What is ChatGPT and how does it work?

What exactly is ChatGPT?

ChatGPT is an application from the US company OpenAI, released in late 2022, which brought generative AI to a mass audience. The name combines "Chat" with "GPT" (Generative Pre-trained Transformer), the family of models working in the background. You type a question or instruction (a "prompt") and receive an answer in natural language.

The distinction matters: ChatGPT is the product (the chat interface and apps), while GPT is the underlying language model. The same model can also be embedded into your own software through an application programming interface (API). ChatGPT is therefore not a search engine or a database, but a tool that produces language.

How does ChatGPT work?

The language model was trained on very large amounts of text, learning the statistical patterns of language. It breaks input into small units ("tokens") and predicts, step by step, the most likely next token. That is how an answer is built word by word. It does not "understand" content like a human; it computes plausible continuations.

After base training, the model is refined with human feedback (Reinforcement Learning from Human Feedback) to make answers more helpful and safer. Modern versions are multimodal: alongside text, they process images, speech and files. The exact feature set depends on the version and subscription tier.

What can ChatGPT do? Typical capabilities

• Write and edit text: emails, reports, proposals, blog posts, job ads, including tone and style adjustments.
• Summarize and structure: condense long documents, minutes or transcripts to the essentials.
• Translate and localize: fast translation across many languages, useful for multilingual markets like Switzerland.
• Assist with coding: explain, generate and debug code, and draft tests.
• Ideas and analysis: brainstorming, pros-and-cons lists, strategy drafts, interview questions.
• Multimodal tasks: describe images, generate visuals (via DALL-E), understand and read out speech, analyze files, depending on the version.

Business applications in companies

In companies, ChatGPT is used mainly as a productivity and assistance tool. Value appears wherever a lot of language is processed and a human checks the output. For sensitive data, providers offer business and enterprise options with stricter data-protection and administration features; check the exact terms with the provider.

• Customer service: draft replies, FAQ drafts and summaries of enquiries (with human sign-off).
• Marketing and content: first drafts for campaigns, social posts and product copy.
• Knowledge work: structure meeting notes, bundle research, prepare presentations.
• IT and development: speed up prototypes, code reviews and documentation.
• HR and internal communication: templates, onboarding materials and translations for multilingual teams.

Limits and risks

• Hallucinations: ChatGPT can present false statements, invented sources or quotes convincingly. Always verify facts independently.
• Knowledge cutoff: the model knows events only up to a training cutoff; without web search it lacks up-to-date information.
• Bias: training data can contain prejudices that surface in answers.
• Data protection: do not put trade secrets, health or personal data into standard prompts unless processing is contractually covered.
• No substitute for expert judgement: for legal, medical, financial or safety matters, qualified professionals remain essential.

ChatGPT in the Swiss and DACH context

In Switzerland, handling personal data is governed by the revised Data Protection Act (revDSG/nFADP), supervised by the EDOEB. Anyone using ChatGPT commercially with personal data must clarify the legal basis, transparency and data-processing agreements, especially because data is often processed outside Switzerland. For Swiss matters, the "GDPR" is not the governing law; the revDSG applies.

The EU AI Act also has extraterritorial effect: Swiss companies that offer AI systems in the EU, or whose outputs are used there, can fall within its scope. Internal AI policies, training and, where data sovereignty matters, a look at alternatives such as the open model Apertus developed at ETH and EPFL are advisable.