AI evaluation frameworks are currently struggling with maturity and standardization.
Irene Solaiman of Hugging Face discusses the shift toward agentic system evaluations and why the lack of consensus on testing protocols complicates result reproducibility.
This perspective highlights the need for consistent standards as the industry evolves beyond simple model-level benchmarks.
The problem of comparing results
There is no consensus in the industry on how to compare the results of different tests. Each company follows its own approach, which makes it difficult to draw firm conclusions. Furthermore, one of the most controversial points remains the ability to repeat the same tests and get the same results.
The challenge of external audits
Due to ongoing security concerns, the idea of placing external reviewers within AI companies has gained popularity. Anthropic, for example, chose consulting firm Accenture for this role. The choice has drawn criticism in the tech industry, with some questioning whether such a company is the right fit to take on the review.
The need for overall control of systems
Solaiman emphasizes that audits need to change direction: it is no longer enough to examine just the AI model. Evaluations are shifting to examining the entire system, including the tools and software that support its operation. This is because a less powerful model can perform better than a more powerful one if it has a more effective support system.
READ ALSO
Microsoft: Artificial Intelligence is not completely safe and needs human control
However, such tests are considered expensive and difficult to implement. According to Solaiman, the industry needs more standardization in rules and evaluation standards so that these tests have a meaningful value for the safety and reliability of systems.
G
Hacks.gr on Google
Set it as a preferred source for cybersecurity updates.
Preferred source
Top comments (0)