That AI Score Everyone Trusts? It's Easy to Fake

If you've ever downloaded a smaller version of an AI model to run on your laptop or phone, you might have seen a score called KLD (Kullback-Leibler Divergence). It's often used to rank these compressed models, but it's not the reliable quality measure people think. KLD compares the outputs of a compressed model to the original, but it's sensitive to the data used for testing. If the test data is similar to what the model was trained on, the score can look artificially good.

Even when test data is supposedly separate, it might still be too similar to the training data, making the score misleading. Some developers might even tweak their models to get a better KLD score without actually improving performance on real tasks. So a low KLD doesn't mean the model will work well for your needs—it just means it's good at mimicking the original on that specific test.

For regular users, this means you shouldn't rely on KLD alone when choosing an AI model. Instead, look for real-world benchmarks or try the model yourself on tasks you care about. KLD can be a useful diagnostic for experts, but it's not a universal leaderboard. The next time you see a model advertised with a great KLD score, take it with a grain of salt—it might not be the best choice for you.