Hacker News Reader

Writing an automated service that classifies if a twitter account is a political bot. No word on their error rates.

It's impossible to determine the error rate - there is no way of knowing for sure if an account is a bot or not. If there was, this model would be useless. I'd be interested to know how they trained it though, given they didn't have any true data to check against

They could check it against known bots though.

That's only valid if the known bots are a random sample of all bots.

They can't test against a large array of public bots because they're only detecting political bots, not everything automated. So they'd have to train/test on "accounts which are definitely known to be bots, but trying to hide it". Meaning, presumably, the least-convincing bots or bots specific to a previously-exposed network.

There is a way, doxx every account. It's not nice but it would work 100% of the time.

Similarly, just because you can't dox an account doesn't make it a bot.

Also, there are APIs for acquiring real-world mailing addresses inside of "subdivided" warehouses. The only way to verify that a person isn't a bot is to make babies with them.

> The only way to verify that a person isn't a bot is to make babies with them.

For some reason I feel like there are other ways......

In the article they claim 93.5% accuracy.

Reply on news.ycombinator.com