AI writing messages on Bitcointalk.org

[gentlemand]

Once again another reason for signature vacation time.

I don't really pay attention to them but how many mindless and mediocre posters are left in campaigns at this point? It's not like days of yore where anything could sign up. You have to be pretty convincing to get in one.

Almost all of the crappiest posts are sig free. I presume they'll use something like this in the hope of ranking up to get in one. Not sure how far they'll get.

[suchmoon]

I presume they'll use something like this in the hope of ranking up to get in one. Not sure how far they'll get.

It could be mildly successful because the are no negative merits. So if they manage to find a way to earn even 1 merit per 100 near-zero-effort AI posts they would eventually rank up to full member or whatever is the minimum these days for a signature campaign.

The weak spot in this, like with word spinners, is that just a single suspicion can unravel the whole thing. Campaign managers, fellow campaign participants, merit senders etc will eventually notice that something is not quite right with the 10k-post user who streams generic texts without showing any lasting engagement with other users.

OTOH if the AI can beat that then more power to it... them... will it have a preference for certain pronouns I wonder.

[gentlemand]

The weak spot in this, like with word spinners, is that just a single suspicion can unravel the whole thing. Campaign managers, fellow campaign participants, merit senders etc will eventually notice that something is not quite right with the 10k-post user who streams generic texts without showing any lasting engagement with other users.

OTOH if the AI can beat that then more power to it... them... will it have a preference for certain pronouns I wonder.

Unlike plagarisers there'll never be a smoking gun which makes things a bit harder to execute. I guess you can build up a collective amount of disengagement but the owner of the account will be incentivised to drop in every now and then and act like a reactive community member.

I vote for us to close our eyes and wish for it to go away. The alternative might be a bit hard to take.

[Lucius]

Unfortunately our forum is existing for the wrong reasons.  

I can understand that you are against signature campaigns, but I cannot agree that the situation is so bad that we have to ban them. When I came to the forum 2015 there was a real wild west here when it came to signatures campaigns, some of them didn't even have managers, but were completely controlled by bots. At one point the administration decided to make a sharp turn and instead of bots set up real people, which resulted in the forum starting to look healthier and more natural. What we need are even higher standards for all signature/bounty campaigns, why one of the conditions would not be at least 50 earned merits?

Makes bots and ai have less reason to post if there are no signature revenues.

True, but AI doesn't exist just to cheat on signature campaigns - this forum would still be a very good place for training AI. If we go into some conspiracy theories, then maybe AI invented Bitcoin and thus this forum, and we all live in the delusion that Satoshi Nakamoto ever existed. My opinion is that technology is much more developed than what we use today, but it is available to a very limited number of people.

[UserU]

There's simpler & good horrible solution for this problem, which forcing user to solve CAPTCHA every time he made/edit message.

Having solved reCAPTCHA a million times before, Google now thinks I'm a robot. Previously at one point, my account was like blacklisted for nearly a year.

[suchmoon]

Unlike plagarisers there'll never be a smoking gun which makes things a bit harder to execute. I guess you can build up a collective amount of disengagement but the owner of the account will be incentivised to drop in every now and then and act like a reactive community member.

Possibly. But "dropping in" would raise the cost of farming such accounts, hopefully to the point where it's no longer profitable.

I vote for us to close our eyes and wish for it to go away. The alternative might be a bit hard to take.

We can train AI to catch those bots - that'd be quite poetic I think.

Or hire a blade runner.

[LoyceV]

At loyce how about scraping a signature campaign thread and estimate what it pays each week.

Most campaigns use a spreadsheet, which makes it very easy to get the total.

True, but even identifying message is written by human/AI could help moderator in some ways. For example, those who use it to bump ANN thread could be identified and nuked quickly.

That's still not against the rules. As far as I know, the same rules apply for bots as for humans. That's going to be interesting when a bot gets banned though: will all other users of the same AI technology be considered alt-accounts and thus evading their ban?

[andreibi]

Brave new world, indeed. AI writing here? I still think it can be detected, nobody can write perfect sentences even if a post is intended to be written by a 12 year old.

Waste of technology anyway.

[gentlemand]

Brave new world, indeed. AI writing here? I still think it can be detected, nobody can write perfect sentences even if a post is intended to be written by a 12 year old.

Waste of technology anyway.

But for a couple of sentences in the article it wrote that was linked I wouldn't have known it was AI. It'll no doubt rapidly get smarter. The GPT-2 posts I've read were empty, moronic shit even if they were coherent. That article was a bit bland but vastly more convincing.

Every time something like this is upgraded it takes a major leap. Looks like this is no exception.

[Mpamaegbu]

If someone reply or making a post something like this conscious to unconscious probably they're bot or spammer.

What if I use "conscious to unconscious" with smilies generously splashed on the comment,  wouldn't that make me a Super Al?

My point above is simply — Don't judge quickly.

[NotATether]

You know, an AI that flags suspicious Marketplace sales threads with a different explanatory post detailing the risks involved in those sales would be very useful. Real users get tired of wading and answering through the same listings of crap each day.

[Twentyonepaylots]

Unless theymos gonna make or use AI which can detect whether a message is written by AI/not, there's nothing much we could do aside from reporting posts that doesn't make sense.

Using bots isn't even against the rules.

So we gonna change the rules?

I'm impressed that the linked article was entirely AI.

 Especially with the recent work from home, your own cloud AI can apply for thousands of jobs per minute pretending to be you, and since you don't have to show up in person anyway, your AI can maintain dozens of jobs

Yes and we gonna be useless I guess. We are working to build the AI in order for it to work for us.


Imagine the forum full of AI, interactions won't be as natural as it is now. I see the world leading by technology, AIs definitely but I don't think it would not be good for us if we come AI dependent.

Plus if it happens, then say goodbye to the signature campaigns, project campaign won't be effective anymore and if the project want to advertise they are just gonna create their bots to do the job for them.

[gentlemand]

Imagine the forum full of AI, interactions won't be as natural as it is now. I see the world leading by technology, AIs definitely but I don't think it would not be good for us if we come AI dependent.

Plus if it happens, then say goodbye to the signature campaigns, project campaign won't be effective anymore and if the project want to advertise they are just gonna create their bots to do the job for them.

The businesses that advertise here need real people to use them.

If this forum really did get totally infested with AI, and that is verrrry possible as I'm sure it would be a gift from the heavens for shitposters, those real customers will find it increasingly unbearable  and business will dry up. Something will need to be done to keep them at bay.

It's no good if they are more articulate than genuine users and the quality of posts actually rises. AI doesn't have any money to spend or any needs to service.

I'm very curious to see how this system interacts with other users in a forum environment. Writing an article on its own is a very different thing. It might fall apart when juggling and responding to other points of view despite being convincing on the surface, that's if it does that at all.

[DdmrDdmr]

This thread reminded me of this:

https://bitcointalk.org/index.php?topic=5171540.msg52047825#msg52047825

If anybody wants to have a bit of fun, there is a tool created by the IBM Watson AI lab called GLTR (Giant Language model Test Room) that aims at detecting whether a text is AI generated or human generated. See: http://gltr.io/dist/index.html

It provides some sample texts, but you can provide your own. In brief, it matches each word with a prediction based on the previously written word. The more predictable the words, the more likely the text is AI generated (green and yellow in the tool’s output). Further details here: http://gltr.io/

AI to trap AI …

So I wondered what GLTR would make if the article created with OpenAI GPT-3 (from 'OpenAI' to 'human operator'). I haven’t seen any updates to the GLTR website, so I do not know if the underlying software has been upgraded since 2019, but anyway. Alas, it doesn’t seem to be working properly, and I only managed to run a small fragment through it. Running all the text made the program stall (or I lost patience waiting for over 30 min., despite multiple attempts).

I also tried out another site that detects fake news: https://grover.allenai.org/detect. When I gave it as input the complete text written by the AI, it stated as output "We are quite sure this was written by a machine", whereas when I fed it the complete explanation provided by the author, it came back with "We are quite sure this was written by a human". That looked promising, until ...

When I did the same, feeding it one paragraph at a time, the output indicated that the first paragraph of the AI written text was quite surely written by a human, but all the rest were deemed as written by a machine (one was likely, but not certain).
Of course I lost faith in the site when I ran the OP and I got a "We are quite sure this was written by a machine" (I’m pretty sure @fillippone’s hairy wrist is quite human). Same goes for multiple other posts within the thread.

As AI text generation progresses, so does the counter-detection software. Of course, integrating this into a forum and having to cope with truckloads of false positives would defeat the purpose. I recall the GLTR solution being based on predicting the next word in a phrase, and determining whether the next word was within the most common next predicted words - the further away from common word prediction, the better (i.e "the cat was"  + "angry" is more standard than "the cat was" +"flabbergasted"). That kind of algorithm may be an approach to work on text written by natives to a language to a certain degree, but people with limited language skills and vocabulary would probably display as "AI" posters, and not as non-english limited vocabulary posters. Anyhow, interesting to see how this evolves (in general terms).

[PrimeNumber7]

Take a look at this video from 2018.

Google trained a model to have a conversation with a human based on a specific request from a second human. This is far more advanced than the article in the OP. I believe the model used in the article linked in the OP will have difficulty responding to a specific post in a thread, while staying on topic.

It appears the model used to generate the text in question has 175 billion parameters. To put this into context, each parameter will need to have at least 1 gradient, or slope, and at least one bias, depending on the activation function used in each layer. The large number of parameters will make it expensive to even load this model onto a computer or server to generate text.

I think, most likely this is a generator model in a generative adversversial network. The best way to predict if output is from a generator model is to use the discriminator model the generator model trained with. If you don't have that model, you can train generator and discriminator models so you can eventually use the discriminator model.

When you are training generator/discriminator models, you can only feed it so much data, and I believe the vast number of dialects and writing styles will make it difficult to predict if a given text was generated by a generator model or not. Similar to how some people who grew up in a Hispanic country but live in the US speak "spanglish" at home, some who grew up in the US and live in the UK will likely speak in a mixture of the two dialects.

[alani123]

I think the amount of effort and vetting one would require to program GPT-3 to post on bitcointalk wold be substantial. Not a ton of work, but still, if someone possess such advanced knowledge of these professions, they'd have much more lucrative endeavors to attend to other than trolling the forum with AI generated posts. There's not much to gain from this other than perhaps portraying it as a silly social experiment. Which again wouldn't be taken all that seriously, as there are better potential applications, without questionable ethics.

[gentlemand]

I think the amount of effort and vetting one would require to program GPT-3 to post on bitcointalk wold be substantial. Not a ton of work, but still, if someone possess such advanced knowledge of these professions, they'd have much more lucrative endeavors to attend to other than trolling the forum with AI generated posts.

I think you underestimate the amount of skill out there and willingness of people to throw away their time too. If you're willing to spend 19 hours a day clicking faucets then you're certainly going to look into this as well.

There are many areas of the world where opportunities are very limited and this place looks compelling if you're unfortunate enough to be marooned like that.

[suchmoon]

I think the amount of effort and vetting one would require to program GPT-3 to post on bitcointalk wold be substantial. Not a ton of work, but still, if someone possess such advanced knowledge of these professions, they'd have much more lucrative endeavors to attend to other than trolling the forum with AI generated posts. There's not much to gain from this other than perhaps portraying it as a silly social experiment. Which again wouldn't be taken all that seriously, as there are better potential applications, without questionable ethics.

Except there is a direct financial benefit for that shit on this forum. And arguably once it's all set up and running, it would easier to maintain it than posting mind-numbing one-liners manually so e.g. a campaign manager could perhaps pay someone to set it up for them and then just collect earnings for 20 sockpuppets instead of hiring 20 humans for the campaign. It's probably a dead end in the long run - the effectiveness of such ads would diminish as the content gets less interesting for google, plus there is a risk of getting caught - but it might be a goldmine for a couple of years for someone who doesn't have particular ethical constraints (and there's definitely no lack of such individuals here).

[PrimeNumber7]

I think the amount of effort and vetting one would require to program GPT-3 to post on bitcointalk wold be substantial. Not a ton of work, but still, if someone possess such advanced knowledge of these professions, they'd have much more lucrative endeavors to attend to other than trolling the forum with AI generated posts.

I think you underestimate the amount of skill out there and willingness of people to throw away their time too. If you're willing to spend 19 hours a day clicking faucets then you're certainly going to look into this as well.

There are many areas of the world where opportunities are very limited and this place looks compelling if you're unfortunate enough to be marooned like that.

The skillset required to click on faucets is much less than that of creating a machine learning model.

The forum may be a place in which models are tested. Scientists may use models to post to see how many people notice humans are not being the posting.

The computational cost to train a model is also very high for something like GPT-3. The final model will cost possibly thousands of dollars to train, and before you have a final model, you will likely have many models that you end up not using that you will need to pay to train.

[Vod]

Remember that AI growth is exponential.   We won't be having discussions on how to stop AI because the day they become self-aware (even in a shallow AI limited sense) they will be indistinguishable from humans and they will outnumber us.

Self awareness is as simple as knowing what your inputs (senses) are and how to react to them.   Sounds like where machine learning is moving now..  :/