When it comes to AI features, it can sometimes feel like there are approximately five million of the things buzzing about right now. Every phone seems to have a dozen to play with, but there's a slight issue when it comes to how useful they actually are.
Superzoom and image creation tools are well and good, but I'm still a little suspicious of how many people actually use them particularly often. The biggest standout success in this area has been the proliferation of magic eraser-style options to let you get rid of annoyances in photos, but few features can match that utility.
At MWC this year, though, Honor had a more in-depth demonstration of the AI deepfake detection tech that it announced six months ago, and I was struck by how potentially useful it could be. The feature works easily, on the surface – when active, it can screen videos that you're watching, whether live or recorded, and give you an indication of whether they're likely to be genuine or involve a face-swap.
It takes a few seconds to process, before you're given a result that summarises how confident the AI is that you're being duped as a percentage, and seems like a complete slam-dunk in security terms. After all, we all hear a bunch about how scammers are starting to use deepfakes and vocal imitations to get one over on people.
A feature to mitigate that risk is a no-brainer, even if there are still questions about whether it'll always be active, or instead remain as a toggle that people have to use when they're suspicious of a video.
There are also still question marks in my head when it comes to reliability and messaging. In a pool interview I took part in, Honor's Eric Fang indicated that the feature currently has a 70% success rate, which is one of those numbers that both impresses and doesn't at the same time. A 70% chance of being saved from a fake video – that's ideal. A 30% chance that your phone tells you a video's fine when it's not – that feels sub-optimal.
Still, it wouldn't be the first AI feature that gets better as time goes on, as its model gets more sophisticated. Plus, the fact of the matter remains that I would actually use deepfake detection, which is a surprisingly high bar. Hopefully it's the first of a cavalcade of non-novelty features in the same vein.
