Hacker News Reader

I haven't read the directive/law, but the EU should also invite EVERYONE to submit well documented and reproducible examples of AI/LLM failures. It's only through public exposure that improvement will be assured.

The AI vendors will NEVER fix any system flaws that can be ignored or hidden. Only a public database can force these into the open.

View on HN

The AI Act classifies systems depending on their capabilities and their intended use, and the risk that they pose on the people using or being exposed to them. If a tool ends up being classified as a high-risk AI system or as a model with systemic risk there are strict requirements in terms of reporting incidents. For certain types of incidents it is as little as two days, for others upwards of 15 days.

What is a failure?

When an AI gives you a link that it made up. (This is easily verifiable)

Telling lies basically.

In AI-speak, a "hallucination".

Reply on news.ycombinator.com