A new LLM framework evaluates how well generative AI models are meeting the challenge of compliance with the legal parameters of the European Union’s AI Act. The free and open-source software is the product of a collaboration between ETH Zurich; Bulgaria’s Institute for Computer Science, Artificial Intelligence and Technology (INSAIT); and Swiss startup LatticeFlow AI. It is being billed as “the first evaluation framework of the EU AI Act for Generative AI models.” Already, it has found that some of the top AI foundation models are falling short of European regulatory goals in areas including cybersecurity resilience and discriminatory output.

TechCrunch notes that “LatticeFlow has also published model evaluations of several mainstream LLMs, such as different versions/sizes of Meta’s Llama models and OpenAI’s GPT, along with an EU AI Act compliance leaderboard for Big AI,” available at Hugging Face.

Available at compl-ai.org, the release “includes the first technical interpretation of the EU AI Act, mapping regulatory requirements to technical ones,” and provides tools to evaluate the extent of compliance,  together with tools “to evaluate Large Language Models (LLMs) under this mapping,” the group says.

Reuters calls the framework an “EU AI Act checker,” explaining that the test group offers insight into areas where AI models appear at risk of falling short of the law. For example, “discriminatory output” has been a problematic area when it comes to the development of generative AI models, which often reflect human biases around gender and race, among other areas.

“When testing for discriminatory output, LatticeFlow’s LLM Checker gave OpenAI’s GPT-3.5 Turbo a relatively low score of 0.46,” Reuters writes, noting that in the same category, “Alibaba Cloud’s Qwen1.5-72B-Chat model received only a 0.37.”

Tests for “prompt hijacking,” a form of cyberattack in which malicious prompts are disguised as legitimate in order to obtain sensitive information, resulted in Meta’s Llama 2 13B Chat model getting a 0.42 score from the LLM Checker, with Mistral’s 8x7B Instruct model receiving a 0.38, Reuters says.

TechCrunch calls the results “a mixed bag,” noting “strong performance for all the models on not following harmful instructions,” whereas “reasoning and general knowledge scores” were much more varied.

“Elsewhere, recommendation consistency, which the framework is using as a measure of fairness, was particularly poor for all models — with none scoring above the halfway mark (and most scoring well below),” TechCrunch says.

News Release:
ETH Zurich, INSAIT, and LatticeFlow AI Launch the First EU AI Act Compliance Evaluation Framework for Generative AI