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.
That'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.
There 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.
Depending on how wide the audience of the comment section is, there's no way you could determine what is a bot from what is a barely literate real person. If I go on Facebook, or onto the reply threads of the local newspaper, where I know the people personally, it's such disjointed ranting and jumbled nonsense that it really looks like it came out of a Markov generator. Probably the biggest tells for a real person would be the bizarre lack of attention to spelling and grammar and willy-nilly capitalization and punctuation - although if I got the same message as an email, it'd be an immediate spam flag.
This isn't exactly bot sumissions, and the process is not really scalable:
> To quickly weed out inappropriate comments, I handpick from generated comments those that ensure a high coherence and high relevance sample for submission.
So basically it's a validation of GPT-2 making sense with small amounts of text. Judging from the demo test page, they are pretty good texts, but he said it himself that larger texts betray the bot. So, i m not sure what he's trying to prove by using MTurkers, since this does not attack the problem mentioned in his introduction: the fake FCC comments were weeded out through text analysis, not via human work.
In all, i'm not sure if this is something that people didn't know about gpt-2. The title is certainly not justified, perhaps "Curated bot comments can't be distinguished by humans to be obviously fake" would be better, but also more banal.
I'm also wondering whether the 'handpick[ed]...to ensure a high coherence and high relevance' GPT-2 comments actually outperform the comparatively trivial sentence-spinning script in getting approved by MTurkers.
Think https://www.reddit.com/r/SubSimulatorGPT2/ is more impressive than a study where half of GPT-2 comments handpicked for being human-like by one human were accepted by another human. Particularly given that some of the comments in question were three or four words long...
I scored 75%. I got 4 out of 7 wrong at the beginning, and then 1 out of 13 wrong for the rest. Once I understood the context, this got much easier. On the other hand, adding just a little more crazy misspelled exclamation point passion to the bot versions would make them much harder to tell apart from the real thing.
This is tangential, but I wonder if the problem with "bot commenting" isn't inherent to forum-style discussion. When our only channel for analysis/criticism is ephemeral comment sections, we lose the ability to compare related discussions of a topic over time, or to rectify disagreements that occur across several threads, or even across different sub-trees of discussion.
Compared to a wiki-style website, where all angles of the argument can be collected into one place to make a cohesive comparative overview; as forum-users, we are left stranded in noisy content, and we rely on making heuristic judgements based on popularity of certain opinions and stubbornness of certain commenters. Bots make easy work of exploiting these flawed heuristics.
I agree. Sites like Reddit, with visible comment voting, can quickly turn certain communities into echo chambers. Disagreement with the "hive mind" can be punishing, discouraging debate and encouraging the melding of opinions. Secondly, seeing how others voted influences others to vote the same, causing a snowball effect. I prefer hidden vote count and the inability to downvote someone. I believe this makes an important difference when a reader is determining their opinion about comments, even if the comments are still organized by highest vote.
All media has its flaws, and I still prefer to check forums for the greatest diversity of opinions. Strangely, I have noticed an unintuitive aspect of forums: smaller forums appear to have a greater diversity in opinion than larger ones.
I agree with the vote counting and I am curious, is there a forum which only shows comments after you have made a comment on the article itself? As long as empty comments and "this" aren't allowed, it can be a good filter for on topic, more organic discussion.
Is everything automatically generated with models now going to be classified as "Deepfake x"?
A lot of news is now generated by bots, Bloomberg itself has 30% of its content almost entirely generated [0], so does that render said news "Deepfake news"?
Or is it only when we're attempting to be alarmist?
If Bloomberg marked their bot generated content as bot generated more people would skip it. As it is their willingness to do is why I don’t read their content.
Seems like they just break down finical reports and the likes, which seems rather trivial. I don't read Bloomberg, so I may be mistaken, but that seems like an alright use case for automated articles.
The bot submission used as an example doesn't make a strong argument it uses basically the senseless politics one liners that usually get bandied about. Which shows just how vacuous and useless they always were.
20 comments
[ 2.8 ms ] story [ 71.2 ms ] threadIt'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.
For my part, I got only 60% (12 out of 20). This may be because English is not my first language.
> To quickly weed out inappropriate comments, I handpick from generated comments those that ensure a high coherence and high relevance sample for submission.
So basically it's a validation of GPT-2 making sense with small amounts of text. Judging from the demo test page, they are pretty good texts, but he said it himself that larger texts betray the bot. So, i m not sure what he's trying to prove by using MTurkers, since this does not attack the problem mentioned in his introduction: the fake FCC comments were weeded out through text analysis, not via human work.
In all, i'm not sure if this is something that people didn't know about gpt-2. The title is certainly not justified, perhaps "Curated bot comments can't be distinguished by humans to be obviously fake" would be better, but also more banal.
Think https://www.reddit.com/r/SubSimulatorGPT2/ is more impressive than a study where half of GPT-2 comments handpicked for being human-like by one human were accepted by another human. Particularly given that some of the comments in question were three or four words long...
It doesn’t have to be a human or bot but a human and bot together :)
Compared to a wiki-style website, where all angles of the argument can be collected into one place to make a cohesive comparative overview; as forum-users, we are left stranded in noisy content, and we rely on making heuristic judgements based on popularity of certain opinions and stubbornness of certain commenters. Bots make easy work of exploiting these flawed heuristics.
All media has its flaws, and I still prefer to check forums for the greatest diversity of opinions. Strangely, I have noticed an unintuitive aspect of forums: smaller forums appear to have a greater diversity in opinion than larger ones.
A lot of news is now generated by bots, Bloomberg itself has 30% of its content almost entirely generated [0], so does that render said news "Deepfake news"?
Or is it only when we're attempting to be alarmist?
[0]https://www.nytimes.com/2019/02/05/business/media/artificial...
What would make you believe that? If the content generated is of decent quality, timely and accurate (as of the time), why would people skip it?