As AI creeps into every aspect of our lives the lines of reality are blurring. It doesn’t always matter whether we are aware something is AI assisted or generated but when it does it’s important that we can tell the difference. That’s where AI detection tools, like AI Aware, come into the picture as they are far more effective at spotting AI than mere mortals but they are only as good as the people training them. At the heart of every impressive AI detection tool is a bunch of sceptical, mischievous, right-brain thinkers doing their best to trip it up in every way possible.
As part of a group AI detection red teamers, that’s my weird job, a world where we spend all day training AI Aware how to recognise the difference between a human and a machine. Our aim is not to break our detection tool but make it better, more reliable and safer. We throw as many curve balls at it as possible and shout ‘Catch!!’ Usually it can, but, unlike our developers, I admit to a certain satisfaction when it can’t.
Admittedly, stress-testing AI detection is an odd job. We spend all day feeding the tool everything from poetry, academic papers, haikus and prayers to kid’s stories, recipes and jokes, some AI generated and some human. We mess with the grammar. syntax and vocabulary in a way we know AI detection is not keen on, use colloquialisms, sarcasm, dialects, Gen Z slang even emojis intentionally trying to make the tool fail and when it does, we have a team of super smart developers that makes it better. Red teamers prod, probe and poke, we take old and new text, flawed and perfect to see how our model handles complexity in text and attempt to find blind spots to mitigate risks. Algorithms can generate text, simulate a conversation, but they can’t predict human creativity especially when it comes to breaking rules.
Automated adversarial testing can help, but it can’t replicate the ingenuity of a human trying to trip up a model. The messiest, most creative, most unpredictable thing on Earth is still a human and I definitely tick a few of those boxes. Red teaming isn’t just about poking holes in AI. It’s about building reliable and resilient systems. Without human testers continually improving detection, AI would be rolling out to an unprepared and unprotected world.
So, next time your doom scrolling and wondering if a video is genuine or trying to figure out if an article or comment is human or AI generated rest assured there’s a bunch of oddball humans out there making sure the truth is more likely to be found and respected.