4 ms·
Taking the test linked from the article, I got 75% correct. Most of my failures were for very short (3-5 word) sentences which could simply be memorized from ex
by inciampati 7y ago
Taking the test linked from the article, I got 75% correct. Most of my failures were for very short (3-5 word) sentences which could simply be memorized from existing texts. Human responses were largely more coherent, keeping on topic through multiple sentences, while bot ones didn't.
It's a very good idea to make sure submissions come from humans, but it's also slightly overstated and alarmist to state that model-generated text is reliably passing the Turing test.
- lgeorget 7y agoThat's a dataset of one, and due to your presence on HackerNews, you're likely more tech-savvy than the average reader of comments on news site, so more able to tell comments by bots apart from comments by humans. If a sufficiently large percentage of the population gets 50%, you can say that it passes the Turing test. For my part, I got only 60% (12 out of 20). This may be because English is not my first language.
- BiteCode_dev 7y agoBesides, you have only access to the text. A website like Amazon has a huge list of additional data points to make guesses.
- ses1984 7y agoThere is no single Turing test. A bot might fool me but not you. You don't have to fool everyone to have an impact. Also not everyone reads everything as though they are in the middle of a Turing test.
- sharemywin 7y agoPriming does make a difference.
- ses1984 7y agoAre we at the stage where a comments section needs an anti bot primer blastered over the top?