Botsight says it uses artificial intelligence to decide if there is a human or a bot behind a Twitter account. The software was developed by NortonLifeLock, which was formerly part of Symantec.
The goal is to help fight disinformation campaigns. It’s hard to argue with the sentiment behind this.
Botsight in a browser
You install Botsight as a browser extension. NortonLifeLock says it works with the major browsers. It turns out that mainly means Chrome. There’s no support for Safari and when I first tested the Firefox version that wasn’t delivering. These things happen with beta software. It’s no big deal.
Then, when Twitter is running in your browser, Botsight flags whether an account is likely to be human or a bot. You have to use the office Twitter website. A green flag shows an account that is likely to be human, red tells users to be wary.
The flags also show percentages. In my case the score is 80 percent, that’s enough for alarm bells to ring.
At Botsight says, I’m “almost certainly a bot”.
The developers say they collected terabytes of data then looked at a number of features to determine if an account is human or not. The software uses 20 factors to make this decision.
More AI nonsense
NortonLifeLock says its AI model detects bots with a high degree of accuracy. It’s a typical AI claim and like many of them, doesn’t stand up too well when tested in the real world.
No doubt a lot of Botsight readers who encounter my Twitter wit and wisdom will assume the worst.
It’s not going to happen, but that could be grounds for a defamation action. Sooner or later someone is going to sue a bot for character assassination.
Like it says at the top of the story I’m on the wrong side of this equation.
I asked NortonLifeLock how come I’m identified as a bot. Daniel Kats, the principal researcher at NortonLifeLock Research Group says there are three main reasons.
The first is my Twitter handle: @billbennettnz.
“The reporter’s handle is quite long, and contains many “bigrams” (groups of two characters) that are uncommon together. This is a sign of auto-generated handles (ex. lb, tn, nz). It’s also quite a long handle, which in our experience is common of bots.”
I didn’t have much choice here. My given name includes that tricky LB combination. I doubt changing Bill to William would have made any difference.
There are a lot of other Bill Bennetts in the world. Others got to the obvious Twitter handles first. Mine tells people I’m in New Zealand. Trust me, the alternatives look more bot-like.
The only practical way to change this is to kill the account and start Twitter again from scratch. It is an option.
Following too many
Botsight’s second alarm is triggered by my follow to follower ratio. It turns out that following 2888 people is ’an usually high number, especially in relation to the number of followers”. Kats says it is no common for a human to follow that many others.
Well, that’s partly because I use Twitter to follow people who might be news sources.
The idea of letting bots or AI bot detectors dictate behaviour bothers me. Yet, if Botsight thinks I’m a bot, it’s possible other researchers and analytical tools looking at my account think so too. We can’t have that. Perhaps I should cull my follow list.
So, please don’t take offence if you’re unfollowed. I need to look more human. Only up to a point. On one level I don’t care what a piece of software thinks about me. On another, I get a fair bit of work come to me via the Twitter account so it may need a bit more care and attention.
Not enough likes
The third sign that I’m a bot is that my number of favourite is low. Favourite is the official Twitter terms for liking a tweet. Apparently I don’t do this as much as other humans.
On the other hand, I link to a lot of web posts. Linking lots and not favouriting much is, apparently, a sign of a bot.
The Botsight software could take note that I often get involved in discussion threads on Twitter. That’s something that a human would do, but would be beyond most bot accounts.
From the bot’s mouth:
Well, there you have it. I’m a bot. Perhaps that means I should put my freelance rates up.
Of course, any AI model is only as good as the assumptions that are fed into it. This is where lots of them fall down. We’ve all heard stories of AI recruitment tools or bank loan tools that discriminate against women or minorities. Bias is hard coded.
This is nothing like as bad. On a personal level I’m not unduly worried or offended by Botsight. Yet it does give an insight into the power and potential misuse or misinterpretation of AI analysis.