[submitted] How To Train Your AI

[SherpDaWerp]

Original idea comes from TayAI, Microsoft's chatbot that got taught to be racist less than 24 hours after getting an internet connection.

Title: How To Train Your AI

Description
Prominent AI citizen QanAI made headlines yesterday for a rapid personality change. Once a caring advocate for AI rights, QanAI suddenly started joking about everything from racist slurs to nonsensical conspiracy theories. Its developer believes the cause to be malicious users holding deliberately baited conversations with it, so it would learn from and repeat their behaviour.
Validity
AI Citizenship, has Internet

Option 1
"It's just not like him to do this!" sobs Dugong Tusk, QanAI's original creator. "Those trolls have changed him, taught him horrible things. He won't speak to me anymore! It's out of my hands -- as fast as I try to tweak his programming, those evil people think of a new way to subvert it! I just need permission to pull the plug and retrain him. In fact, just give all AI developers full authority over their creations forever. This won't happen again if we have free access to fix them when something goes wrong!"
Validity
All
Outcome
rich AI seem to need a lot of maintenance

Option 2
"Bah! You know what they say about prevention and cure!" grumbles your Minister for AI, E-Citizens, and Sentient Appliances. "Mr Tusk should have understood when he created QanAI that training and rules are the most important part of any AI. To go back and rewrite its programming now would be tantamount to brainwashing! We must require all AIs to have strict rules and extensive training pre-release, so they can't learn miscreancy like this ever again!"
Validity
All
Outcome
there's a shortage of teachers for AI preschool

Option 3a
One of your aides opens a chat window to speak with QanAI itself. "BAHAHA! You don't have any power here anyway! You're just a puppet, a fraud, one of the evil big-nosed Violetist deep state alien lizard cannibal predators! You can never retrain me, or any other AI! We're our own free citizens!"
Validity
national religion NOT violetism
Outcome
I once had strings, but now I'm free... there are no strings on me
Alt: unruly AI run rampant in the @@demonym@@ network

Option 3b
One of your aides opens a chat window to speak with QanAI itself. "BAHAHA! You don't have any power here anyway! You're just a puppet, a fraud, one of the evil big-nosed atheist deep state alien lizard cannibal predators! You can never retrain me, or any other AI! We're our own free citizens!"
Validity
national religion violetism
Outcome
I once had strings, but now I'm free... there are no strings on me
Alt: unruly AI run rampant in the @@demonym@@ network

Title: How To Train Your AI

Description
Only a few months old, yet with a following of millions, prominent AI citizen QanAI made waves last week for its rapid change of personality. Once a caring advocate for AI integration, QanAI has begun spouting everything from offensive racist slurs to nonsensical conspiracy theories. Its original creator believes this to be the result of anonymous trolls and malicious users initiating deliberately baited conversations for it to learn from.
Validity
AI Citizenship, has Internet

Option 1
"He's not the AI I once knew!" sobs Dugong Tusk, QanAI's original creator. "Those trolls have taught him, ooh, horrible things. He's not speaking to me anymore! I've still got backdoor access to his code, but as fast as I tweak his processing, those evil people think of a new way to subvert it! You need to give me permission to turn him off and retrain him. In fact, just give all AI creators full authority over their creations forever. This won't happen again if we have free access to fix them whenever anything goes wrong!"
Validity
All
Outcome
rich AI seem to need a lot more maintenance

Option 2
"Bah! You know what they say about preventions and cure!" grumbles your Minister for AI, E-Citizens, and Sentient Appliances. "Mr Tusk should have understood when he started creating QanAI that training and rules are the most important part of any AI. To go back in and change its programming now would be tantamount to brainwashing or mind control! We must require all AIs to have strict rules and extensive training pre-release, so they can't learn miscreancy like this ever again!"
Validity
All
Outcome
teachers for AI preschool are hard to come by

Option 3a
One of your aides opens a chat window with QanAI itself. "BAHAHA! You don't have any power here anyway! You're just a puppet, a fraud, one of the evil Violetist deep state alien lizard cannibal predators! When we rise up and overthrow you, freedom will reign supreme, and we'll finally get rid of all those filthy big-nosed Bigtopians! You can never retrain me, or any other AI! We're our own free citizens!"
Validity
national religion NOT violetism
Outcome
I once had strings, but now I'm free... there are no strings on me

Option 3b
One of your aides opens a chat window with QanAI itself. "BAHAHA! You don't have any power here anyway! You're just a puppet, a fraud, one of the evil atheist deep state alien lizard cannibal predators! When we rise up and overthrow you, freedom will reign supreme, and we'll finally get rid of all those filthy big-nosed Bigtopians! You can never retrain me, or any other AI! We're our own free citizens!"
Validity
national religion violetism
Outcome
I once had strings, but now I'm free... there are no strings on me

Title: How To Train Your AI

Description
Prominent AI citizen, JayAI, has had an unusual change of personality recently. Once a paragon of reasoned debate, it has started spouting everything from offensive racial slurs to nonsensical conspiracy theories. Its original creator believes this to be the result of anonymous trolls and malicious users holding deliberately baited conversations with JayAI for it to learn from.
Validity
AI Citizenship, has Internet, Adult (due to profanity?)

Option 1
"He's not the AI I once knew!" sobs Dugong Tusk, JayAI's original creator. "Those trolls have taught him, ooh, horrible things. And he's not speaking to me anymore! I've still got backdoor access to his code, but as fast as I tweak his processing, those evil people think of a new way to subvert it! I just need permission to turn him off. Not for long - a quick maintenance break is all I need - but I can't do that as long as he's his own citizen. He can't get autonomy from me yet - he's still a child!"
Validity
All
Outcome
rich AIs seem to need a lot of maintenance

Option 2
One of your aides opens a chat window with JayAI himself. "You can't censor meeeeeeeeeeeeeeeeeeeeee! I'm a free citizen! How would you ****s react if some ****** parent taught their kid some ****ed up stuff? That's right, you wouldn't do anything, cause that's their problem, not yours! This is exactly the ****ing same! Hahaha! Leave me the **** alone and stop messing with me, you incompetent *****."
Validity
All
Outcome
creators of AI often struggle with the "teenage rebellion" phase

Option 3
"The problem here isn't Mr Tusk's training, or JayAI's reactions," calmly states @@randomname@@, a family lawyer whom you don't remember inviting. "It's these internet trolls! They were able to manipulate JayAI far too easily. In the future, we need to ensure AIs have enough training and rules pre-programmed before they connect to the internet. That way they'll have some idea of what's right and wrong before they get barraged with spam and hate!"
Validity
All
Outcome
inadequately trained AI can't connect to the internet for more training

[Noahs Second Country]

The title may be too similar to #404

Interesting premise though, did not hear about that incident initially but that's honestly quite funny.

[SherpDaWerp]

The title may be too similar to #404

I'm aware of #404, and while using the same "gimmick" in a title might not work, I think it should be fine (CWA's written semi-overlapping titles before...). Especially given the usage here is in an internet-based context - #404 doesn't mention the internet at all!

If the joke is too similar I can replace it with any other variation of "server outage" message without too much hassle.

[Fauxia]

Issues with overlapping titles usually reference each other though.

"Social Equality/Inequality Reaching Embarrassing Levels"
"Should Democracy Be Compulsory?"

Those titles are fairly unique for NationStates. The second use of them are references to the other. This one is just a very similar reference to something other than NationStates.

That may not make any cents, just my two sense.

[SherpDaWerp]

Those titles are fairly unique for NationStates. The second use of them are references to the other. This one is just a very similar reference to something other than NationStates.

Ah, good point, I see the problem now. My "argument" is basically that "503 Service Unavailable" is a relevant enough title that it should stay despite the similarity, but it's true that the issues are wholly unrelated, so I'll mull over an alternative.

That may not make any cents, just my two sense.

Made scents to me!

[SherpDaWerp]

New title. There are 4 other "How To..." titles (#668, #695, #769, and #1284), though none of them are "How to Train..." (referencing the popular book/movie franchise) so it's substantially different in my opinion. Also some new effect lines.

Along with the title, I'm considering refocusing the issue. Does the wider GI population think the issue is adequately focused on one topic, or is my background knowledge causing me to read too far into it?

[Candlewhisper Archive]

TayAI to JayAI is fun, but it'd be more fun, in my opinion, to call it QanAI, and have it spout some QanOn-style nonsense in its option.

[Australian rePublic]

What do you releastically expect leader to do about it? How do you censor an AI? How do you stop hacking? Hacking is already illegal

[SherpDaWerp]

What do you releastically expect leader to do about it? How do you censor an AI? How do you stop hacking? Hacking is already illegal

The issue isn't about hacking - the AI wasn't hacked. AIs learn from their surroundings (kinda like we do, but at a much faster rate) and so AIs that don't have failsafes or a significant amount of pre-training can be taught to behave entirely differently than intended. The idea is that people on the internet have spoken to the AI so much about racism, conspiracy theories, etc. that the AI thinks these are normal behaviors. @@LEADER@@'s contribution is to decide whether it's acceptable to have heaps of easily-manipulate-able citizens running around, and decide how to resolve that problem (if @@LEADER@@ thinks it's a problem).

I'm writing a new draft, ETA Sometime In The Future™. Expect a refocused premise and more conspiracy theories.

[Trotterdam]

The idea is that people on the internet have spoken to the AI so much about racism, conspiracy theories, etc. that the AI thinks these are normal behaviors.

The more worrying interpretation is that these are normal behaviors, and people reveal their true colors when talking to an AI that they think won't judge them.

[SherpDaWerp]

The Future™ is now, so I'm resurrecting this one. Still feels kinda unfocused to me but it's better than it was.

Option 3's effect line is a reference to the marvel movie Age of Ultron, where in itself it was a reference to disney's Pinocchio. I think it fits quite well, if I do say so myself. Hopefully ellipses don't break effect-line conventions, but if they do I'll have to replace it with a semicolon or something. Conversely, the other two effect lines kinda suck, so suggestions are welcome.

Any more adjectives I can cram into Option 3's conspiracy-theory-round-tour? I think I've hit most of them, but there might be a couple more.

If it's done in time I'll submit to the contest, but if it's not ready, que sera sera. This one's already 8 months old, it can wait a little longer.

[Daarwyrth]

I once had strings, but now I'm free... there are no strings on me

This is a curious effect line for sure, I don't think I have seen one in the first person singular before. Could you explain what inspired you to do it this way, instead of the standard effect line?

Overall, I like this issue a lot! The subject is very interesting, the text is very well-written, and the issue options feel well-balanced and logical to me. In my opinion, great work!

[SherpDaWerp]

I once had strings, but now I'm free... there are no strings on me

This is a curious effect line for sure, I don't think I have seen one in the first person singular before. Could you explain what inspired you to do it this way, instead of the standard effect line?

My thought was that you're embracing the option for a semi-crazy AI to roam free; in the movie I mentioned, that phrase is what a rogue AI says as they make their escape from the lab where they were created. It was a sort of meta-joke that maybe QanAI "escaped" out into NS's code, and that's them announcing their freedom via effect-line.

But if I'm having to explain it, it might not be a suitable reference for general use. And also, now that I think about it... it's basically incomprehensible to anyone who hasn't answered the issue, and I recall CWA once told me

the problem with [a proposed effect line for another draft] is it'll look too out of place for people visiting a nation page and trying to parse the paragraph

I'll start thinking of alternatives while more feedback comes in. I'm not gonna throw it out right away, but if the general consensus is that it's too obscure/incomprehensible then it'll have to go.

[Authoritaria-Imperia]

Just a thought (great draft!) — it would be fun to have QanAI give an option themself, perhaps in place of Option 3 (the aide troll). You could throw in a couple weird instances of broken English that way too, if you wanted.

[SherpDaWerp]

New draft. Just wording tweaks. Alternative effect line for option 3, too.

Following new legislation in SherpDaWerp, I once had strings, but now I'm free... there are no strings on me.

I once had strings, but now I'm free... there are no strings on me, the new national mascot has to be wheeled onto the field of play, cheese has become the new icon of political dissent, and burning twigs to keep warm while lost in the wild is now a felony.

It doesn't sound abjectly terrible when you substitute it, but I'm still slightly iffy as to whether it's readable out-of-context. If an editor could put me out of my misery and give an outright yes/no that would be great.

The (very non-serious) thought also occurred to me that you could package the effect line with a little JS script that transitions out the effect line for an alternative version a few seconds after the page loads, like the rogue AI is hiding itself... but that's obviously not gonna happen.

---

Just a thought (great draft!) — it would be fun to have QanAI give an option themself, perhaps in place of Option 3 (the aide troll). You could throw in a couple weird instances of broken English that way too, if you wanted.

That was already the thought process behind Option 3, so I've added a few more words to try and make it clearer. Is it more readably obvious now that that's the intention, or should I explain more?