“AI-generated” content is sometimes just unaltered training data

https://twitter.com/eevee/status/1298932940697550848

by weswpg • 6 years ago

3 2 6 years ago

weswpg• 6 years ago

it appears this was "produced" by an AI, but it also exists verbatim elsewhere as written by a human, so it was probably picked up in the training data and then spat back out unaltered
so if it decides to start with A, and it's only ever seen A next to B, and it's only ever seen B next to C, and it's only ever seen C next to D... then you might get out ABCDEFG and go "wow amazing it came up with the alphabet all by itself"
but no, it didn't do that. it saw someone else do that and it couldn't figure out anything else to do it. it's not a sign of intelligence; it's exactly the opposite. and the messier your training data is, the less similarities within it, the more likely this is to happen
yeah you can get an essay out of that one text generator. cool. but do keep in mind 1. it had //millions// of pages of input from formal sources like newspapers and encyclopedias, all written to a similar style. 2. how much of its output is truly original? have you checked?
the funny thing is that we already saw a lot of this stuff happen 20+ years ago with markov chain bots in irc. sometimes the bot would say something remarkably poignant and, oh, no, it's basically repeating something a human said earlier because it started with a rare word

MrStonedOne• 6 years ago

that one copyright case everybody in AI is depending on to protect them is not likely to be as protective as they are assuming.

“AI-generated” content is sometimes just unaltered training data

Related Stories