AI often overpromises and underdelivers. But here’s how it is keeping us safe
For years, we’ve heard of the AI revolution and the way it will transform our lives. From being able to accurately predict our movements, fears and foibles to helping streamline business processes and save us money, it seems there’s nothing AI can’t do.
Except the promise hasn’t always lived up to the reality. AI has had plenty of play in the media, but still remains relatively rudimentary. Bar a few cases.
Tackling the bad guys in bot form
Cloudflare is one of the secret tools of the world wide web, keeping our websites connected and protected from attacks, including distributed denial of service (DDoS) incursions. The company covers more than 26 million online properties, processing more than 11 million requests to those websites a second – which can get up to 14 million at peak times.
The company has developed AI that helps detect those bad bots who drive traffic to websites and through services, either for nefarious means or to try and force them offline through brute force. It’s found that more than a third of all the internet traffic it has sight of is driven by bad bots.
That doesn’t include all the bots in the world – some are so-called “good bots,” including those pushed out by Google, Bing and LinkedIn to crawl the web and find information. But Cloudflare’s AI siphons off the good bots, allowing them to continue their work, while singling out the bad bots and making sure they aren’t able to carry out their evil goals.
Tackling Twitter’s bots
Sheer brute force attacks against websites aren’t the only way bots are being deployed – and therefore the only place they need to be spotted and identified by AI. They’re also used in order to lead conversations in one direction or another on social media, often by nation states that want to try and subvert society.
As well as army divisions that are powered by humans to engage with and discredit any negative discussion of countries like Russia and China on social media, armies of bots are also deployed in order to try and funnel conversation to be only positive.
It’s for that reason that researchers at the University of Southern California have trained AIs to detect bots on Twitter, using the way in which humans interact with each other and bots do in order to distinguish which one is which.
The researchers looked at more than 11 million tweets, from 3,500 humans and 5,000 bots combined, and realised that there are some key elements that separate the human participants from the computer-generated ones.
It’s good to talk
One of the main ways in which to identify a human is its propensity to interact with others. The researchers discovered real-life humans reply to tweets more than their bot counterparts did – by a factor of four or five. Interactivity is one of the things that bots aren’t easily able to do, it seems.
Likewise, humans converse in a specific way. We start our conversations with long screeds of text that we try and parse between each other, but as we hone in on a conversational topic, we limit the amount we type. That’s not the case for bots.
Similarly, humans are more likely to space out their tweets in a random manner, while bots are more punctual, often posting every 30 or 60 minutes.The team’s end product, which they call Botometer, was able to detect those patterns and point the finger at something being amiss. It’s a rare example of the reality of AI matching the much-vaunted promises.