Our AI processes run using the JPG preview image that is embedded within the RAW file. If you shoot using a camera that has very small preview images (like some Sony .ARW files for example), then this could lead to our focus assessments thinking some faces are more in focus than might appear in the RAW image.
One way to improve this in select is to shoot using RAW + JPG as the AI processes will then have a much large JPG file to assess and you'll get more accurate focus assessments.
If you're shooting RAW + JPG, you might have 3 files per image in your source folder (your RAW, your JPG, and your XMP file). Select is designed to present you the JPG version for culling, and will ship the associated RAW image file to your editing software when you've finished your selection.