title: ‘Exploring the AI Act’ description: ‘We explore the AI Act and its pillars from the perspective of companies/organizations using artificial intelligence.’
On the 1st of August 2024 the European Union regulation on the use of artificial intelligence (AI), the AI Act, entered into force1. The regulation tries to promote the uptake of human-centric and trustworthy artificial intelligence (AI).
The regulation covers more in-depth high-risk general purpose AI systems such as the one provided by OpenAI, Mistral, Anthropic, Google, Microsoft or other generic Large Language Models.
What means high-risk can be defined based on the human involvement in the process and its application with respect to human rights and the depth of use in critical (national) infrastructure (biometric identification, emotion recognition, evaluation of natural person, access to essential services and benefits, cableway installations, personal protective equipment, medical devices, civil aviation, vehicle, marine, rail, … 2).
The more the human is in control of the process, i.e. revise and validate the output and take the final decision, the less the system is considered risky.
Let’s focus on consumers and users of AI (e.g., classification, recommendation, extraction, …). If you are creating a Large Language Model (like the ones from OpenAI, Google, …) you must read carefully all the AI Act.
When you use AI services hosted by third parties you are subject to transparency requirements while the provider must ensure that you get instructions of use of the provided AI models (e.g., GPT-4o System Card). Imagine that you pay the OpenAI subscription to allow your employees to summarize text. OpenAI is then responsible to provide all the information for a correct use of their models, while you (the company) are responsible to tell your employees that summaries are generated using AI techniques and should be double checked before use. From the organizational perspective you should probably think to establish processes around what is safe or not to do with the AI based on the instructions given by the provider.
Ask to the provider
Provide to end-users
When you create AI models or offer products that users AI models (e.g. OpenAI GPT models) you are considered a provider. Providers are subject to both transparency and instructions of use disclosure. From the transparency perspective your users must be informed when AI is used during his/her request. This time is your turn to provide instructions on how those models are best suitable for and the kind of interactions that understand.
While the AI Act focuses on high-risk systems (i.e. large language models), some of the recommendation applies to any kind of artificial intelligence (or machine learning) model.
Provide to users
The file contains a blog post on the European Regulation on Artificial Intelligence. From what is written, derive some conclusions to add to the blog post in a special “Key Takeaways” section. Certainly. Based on the content of the blog post, I’ll derive some key takeaways to add to the article. Here’s a “Key Takeaways” section we can add: […] can you make it shorter Certainly, I’ll create a more concise version of the key takeaways.
The key takeaways section was generated using Anthropic Claude 3.5 Sonnet3 via the chat interface.
This article does not constitute legal advice