AI Masterclass: Ensuring Safe and Compliant Use of Artificial Intelligence!

Praxitelis hands the course to coach Michael Kraus, who presents ZM-I KI's 3 part portfolio: AI consulting, software and data products, and university research. Kraus then outlines the course's 5 lectures, from AI fundamentals to application competencies, EU regulatory frameworks and a closing summary.

Kraus places AI on the Gartner hype cycle, splits AI model from AI system using ChatGPT, and separates SPAI from GPAI. Labels on bridge images divide supervised from unsupervised learning, a house price example walks the prediction steps, and he ends on buying an AI tool versus building a proprietary one.

Kraus surveys Gen AI vendors (OpenAI, Google, Anthropic, Mistral) and shows a transformer giving 3 different answers as temperature varies. He splits a prompt into instruction, context, input and output indicator, uses a tennis balls sum for chain of thought, and ends on a reasoning LLM answering in 37 seconds.

Completes the chain of thought prompting recipe (task, format, warnings, context) via a 260 square meter house example, then covers DALL-E bias and a Dlubal agent turning a steel sketch into an RFEM model. Closes with checklists on hallucination and bias for users, developers and decision makers.

Opens on the legal duty of AI competence at work, then German law: data protection under the DSGVO, copyright that excludes machine made content, and trade secrets. Liability runs on contract, tort, delict and the Produkthaftungsgesetz, where the steel fabricator using the AI is liable, closing on humans in the loop.

Dates the EU AI Act from 1 August 2024 and splits the provider role from the operator role. Sorts cases from a price tool to a credit check into minimal, limited, high and unacceptable risk. Ends on the duties per class and a 10 million EUR fine or 2% of turnover for a chatbot that hides that content is AI generated.

The lesson sets out the 4 compliance steps and the 3 training groups (end users, developers, decision makers), then recaps the AI and generative AI definitions and bias. The legal classification follows: data protection, copyright, trade secrets, liability, and the EU AI Act operator and provider transparency duties.

The closing session takes the questions asked most often: what a company must do after buying ChatGPT or Copilot (operator role, risk classification, transparency), what an AI policy document defines, and where to find help. Kraus points to the Dlubal Community Channel QR code and to email.