Why AI Safety Benchmarks Are a Mess

In the fast-moving world of AI safety, benchmarks are multiplying like rabbits. But is more always better?

Proliferation Without Standardization

Since 2018, 195 AI safety benchmarks have emerged, but the measurement side? Not so much. We're seeing a surge of benchmarks but little coherence in how to evaluate them. Enter AISafetyBenchExplorer, a catalogue designed to organize these benchmarks with a detailed multi-sheet schema. It's like a GPS for navigating the benchmark jungle.

Here's where things get messy. Out of 195 benchmarks, 94 are medium-complexity, while just 7 have reached the Popular tier. That's a lot of benchmarks, with few standing out. English dominates the scene with 165 benchmarks. If you're looking for diversity, better luck next time. Most resources are evaluation-only and plenty of repositories are gathering dust on GitHub and Hugging Face. It seems we're drowning in options but lacking in quality.

Tangled Metrics

At the metric level, terms like accuracy and F1 score might sound familiar, but don't be fooled. Each benchmark packs its own judges and rules. It's like trying to compare apples and oranges with a banana scorecard. This fragmentation is the main failure mode of the field, too many disconnected benchmarks, not enough shared measurement language.

Why should you care? Well, without standardized metrics, comparing benchmarks is like trying to nail jelly to a wall. Researchers have plenty of artifacts to choose from, but lack the tools to pick wisely. AISafetyBenchExplorer wants to fix that by offering a traceable benchmark catalogue and a structured approach for discovery and comparison. Will it succeed?, but it's a start.

The Path Forward

Here's the kicker: despite the chaos, there's hope. AISafetyBenchExplorer aims to create a common measurement language and offer guidance for choosing benchmarks. It's a step towards untangling the web, providing a clearer path for AI safety evaluation. Is this the silver bullet? Probably not. But it's a solid move in the right direction.

So, what's the takeaway? The AI safety field is rich in resources but poor in organization. AISafetyBenchExplorer isn't perfect, but it's addressing a gap that's been ignored for too long. If you're diving into AI safety, this tool might just keep you from getting lost in the chaos.