Making the AI Potent-But-Neutral for real

[Drynwyn]

In line with some of my well-received remarks on the AI in this thread: https://tgstation13.org/phpBB/viewtopic.php?f=10&t=5980

I'd like to put forward a clear proposal for AI capabilities and policy.

The remarks in question that will guide this (AKA the TL;DR):

Spoiler:

Someone somewhere around here said that the core of the problem is that AI's tend towards action over inaction, because inaction is boring for the player. This is a damn good point. Here's what I perceive to be happening: AI players want to act- they want to assist SOMEone, whether that is the station, the antagonists, or just Joe the Autism Forter.

They veer towards assisting security because

a) Security is easy to find and easy to help. They know who the sec officers are, and because they have Sec radio, they know what they need.
b) Antagonists are HARD to find and help, because they're used to needing to dodge the AI.

I dunno if any fix to this would work without both a policy change and a large effort to make admins aware that AI's

a) Don't have to be aware IC of the fact that antagonists likely have an objective that requires killing.
b) Don't have to treat antagonists who haven't harmed anyone as harmful.

Here are the steps I think that would fix this.

Step 1- Make AI's more genuinely neutral toward antagonists.

This would involve:

Spoiler:

Establish that AI's don't consider antagonists "inherently harmful" in any capacity. AI's shouldn't behave as though they are aware that antagonists likely have required-murder objectives.

Make it clear that AI's shouldn't punish an antagonist's past harm any more than they should punish Security's past harm- if a traitor killed someone quietly with no fuss, they shouldn't care, so long as he won't do it again (Honestly an extremely likely scenario).

Establish that it is entirely acceptable for AI's to assist a traitor so long as they don't help, encourage, or permit him to harm humans.

Step 2- Make antagonist-AI collaboration easier.
A key reason many AI's assist Security is that they know who security is and what they need, and for a player, action is fun, and sitting on your hands is boring. The steps to solve this are:

Spoiler:

Make it easier for the AI to help antagonists. This could be accomplished several ways, but I would recommend simply giving PDA's, newscasters, request consoles, or all of the above a "Anonymous AI Message" option that allows antagonists to contact and gauge the AI without revealing themselves.

Not punishing antagonists for past harm also plays into this, for obvious reasons.

On the policy side, this requires giving AI's greater freedom to aid antagonists, likely even more then normal crew do. The benchmark should be: AI's should be permitted to do anything a human requests of them, provided it doesn't break law 1.

Step 3- Discourage the AI from being Sec's pet OR Sec's nemesis.
The AI should assist security only under the circumstances of human harm, or a specific, non-countermanded order. In other words, the AI should only help Sec if there's an active murderboner, or security has requested a specific individual be tracked. This is accomplished by:

Spoiler:

Giving standard crew members easy private access to the AI, like security (as touched on earlier).

Give silicons better shields against being subverted or destroyed by Security- this would probably have to be policy-based, otherwise the AI would become BANE OF THE NUKEOP/WIZARDS/ETC.

Continue punishing AI's who dick with security just for the hell of it- there should be a wide gulf between "not helping sec" and "actively hindering sec".

Setting up clear standards for conflicting order prior- you know what, this should be a step of it's own.

Step 4: Clear resolution for conflicting orders
Around 90% of silicon policy revolves around "person A says A, B says B, who do I obey?" As a result, the trouble with the AI, from an antagonistic perspective, is that it is unpredictable. This is less of an issue for Security, because Security is inherently a reactive entity, and has the outside support and tools to respond to most problems the AI can cause them (though doing that can take a bit).This discourages the antagonist from collaborating with the supposedly-neutral AI, because he generally does not have the support, time, or tools to deal with the AI turning on him- it's safer to go it alone and deal with the AI if he has to. To that end, this simple system should be applied instead of case-by-case calls.

Spoiler:

Apply the following, in order, until the conflict is resolved.

1- Law priority. It should go without saying, but if one person's order would/likely would cause human harm, and the conflicting order wouldn't, go with the one that avoids harm.

2- Specific before general. If an order is more specific than another order- for example, "AI, help Urist McTraitor get the CE's blueprints" conflicting with "AI, don't help anyone with anything illegal" or "AI, don't help anyone get the CE's blueprints" - the more specific order (that's the first one) should be followed. What this means is that, while Security and Command can leverage the AI's information and abilities for their own, they first have to have knowledge of what the traitor was doing.

3- Inaction before action. If in doubt, do nothing. "AI, don't help Urist get the CE's blueprints, and tell me if he tries to get them" conflicts with "AI, help Urist get the CE's blueprints, and don't tell anyone about this". The AI doesn't help Urist get the CE's blueprints, but it also doesn't rat him out to Security.

--------------------------------------------------------------------------------------------------

So that's the gist of it. The ideal way this plays out is to encourage more conflict with less killing- both antags and sec can manipulate the AI with ordermancy, but if either becomes a consistent killer, they turn the AI against them. If this works out, it would provide an easy lever for designers to adjust murderbone frequency with- just buff or nerf the AI. It doesn't affect non-murderboner antagonists, because the AI should be simultaneously working for both of them, but it makes murderboning harder/easier.

Also, looking over Silicon Policy, a substantial portion (net loss of words) could be replaced by the three-prong law-conflict test.

[Cik]

it's all well thought out, but this in particular

2- Specific before general. If an order is more specific than another order- for example, "AI, help Urist McTraitor get the CE's blueprints" conflicting with "AI, don't help anyone with anything illegal" or "AI, don't help anyone get the CE's blueprints" - the more specific order (that's the first one) should be followed. What this means is that, while Security and Command can leverage the AI's information and abilities for their own, they first have to have knowledge of what the traitor was doing.

is quite literally genius.

[lumipharon]

Yeah that specific part is actually very very good.

Current policy is that AI's can basically choose how to figure out law 2 conflicts, but they're (supposed to) be consistent throughout the round with it.

If your idea was enforced that would be a lot better.

In regards to private communication with the AI, anyone can just pda them, and if you really care about your identity you can just use someone elses pda to do so.

AI's not caring about past harm (except as potential indications of future harm, case by case etc) is how it's supposed to be, but sadly isn't.
If a traitor kills their assassination target and the AI catches them and locks them down, providing the traitor explains that they aren't going to kill anyone else/there's no reason to believe otherwise, the AI should really go out of their way to protect them from sec, as if they get into conflict with sec, there is a very high chance of human harm, and possibly death, where as letting them escape on a pod is harmless.
Currently, the mindset of the vast majority of AI players is to drop the hammer on any antags, even though getting them arrested and permabrig 90% of the time results in fighting before the arrest, and even possibly executions.
Of course the AI would still be expected to keep an eye on them to make sure they keep their word, but really this is how it should be.

The key thing there really, is basically assuming better nature unless reasonably shown otherwise - basically you don't presume harm, you presume not harm, innocent until proven guilty, onus to prove chance of harm, not chance of not harm.
Basically, don't be a cockblocking git, and asimov is 100% fine.

[Takeguru]

I've tried assuming non-harm before

Just led to me getting wiped by the captain when I let a traitor escape out the maint door when sec was waiting near the hall door with bloody batons

Never again

[lumipharon]

Policy says you can't kill silicons for following their laws. If you're following law 1 like that, ahelp them.

[Drynwyn]

I've tried assuming non-harm before

Just led to me getting wiped by the captain when I let a traitor escape out the maint door when sec was waiting near the hall door with bloody batons

Never again

Addressed under Step 3. "Give silicons better shields against being subverted or destroyed by Security."

Theoretically, the fact that the AI is a helpful asset much of the time- and a powerful tool against murderboners- would work to dissuade Security from killing it, but that's not always the case, particularly since the people most capable of destroying the AI are also those who suffer least from it's destruction (engineers and the captain).

Admittedly, this is probably the hardest thing to accomplish in the entire plan of action- a policy solution would almost certainly involve a LOT of arguing and entirely too much line-toeing on "when are you allowed vs. not allowed to shoot up the AI core", and the balancing gymnastics of a code solution that discourages security deactivating the AI without making antagonist or subverted AI's demigods are nontrivial.

It is, however, nonvital, as most AI-Traitor collaboration should go unnoticed by security- if Security has a good idea of who the traitor is and what they are up to, the AI tends to swing to favor them under the specific-before-general clause.

(And yes, policy currently forbids killing silicons for following their laws, but that is not a policy that is always followed- in no small part because of how vague "following their laws" can be.)

The "subverted" line is also important here- if it were up to me, I would remove all non-asimov boards from the AI Upload, so that turning the AI into security's pet required actual effort.

Also, I thank Cik and Lumi for their kind words.

[Atlanta-Ned]

These are all wonderful ideas. I'm glad to see this fully fleshed out. My only concern is that the inaction before action will cause a bunch of roundstart silicon-lawyering, e.g. "AI do not let anyone without access into tech storage". Would need some policy oversight.

[Drynwyn]

These are all wonderful ideas. I'm glad to see this fully fleshed out. My only concern is that the inaction before action will cause a bunch of roundstart silicon-lawyering, e.g. "AI do not let anyone without access into tech storage". Would need some policy oversight.

First, I appreciate the kind words.

Second, since it apparently wasn't clear- the law tests are meant to be evaluated in order, so in that example, a specific "Let me into Tech Storage"/"Let Urist McThirdPersonSpeaker into Tech Storage" would override the general "anyone without access into tech storage" order. The "inaction before action" test is only meant to apply to equally specific, equally nonharmful orders. I have edited the OP to make this totally clear.

And if the HoS is willing to read through the entire crew manifest to order the AI not let each of them into each area he wishes to keep secure, I applaud his dedication and wish him the best of luck recovering all the guns that got stolen from the Armory while he was doing that.

[Zilenan91]

Policy says you can't kill silicons for following their laws. If you're following law 1 like that, ahelp them.

Policy doesn't exist right now. Honestly I think that that policy was always really stupid. Borgs being killed for following their laws should be what you sign up for, not disallowed.

[DemonFiren]

Policy says you can't kill silicons for following their laws. If you're following law 1 like that, ahelp them.

Policy doesn't exist right now. Honestly I think that that policy was always really stupid. Borgs being killed for following their laws should be what you sign up for, not disallowed.

And theeeere's Zilenan, with yet another shitty "opinion".

Back on topic, though, even if this fails completely at least let anonymous PDA mailing the AI be a thing.

[Anonmare]

Back when NTSL was a thing, MAGNI and other scripts could let you preface a radio message as anonymous which was good for bomb threats mainly of what little use it saw.

There's request consoles on walls that can relay private info to the AI but they're immobile. Having an anonymous function on PDA messages could be good but let them be "unmasked" by checking the message monitor to see where the message originated from

[InsaneHyena]

And theeeere's Zilenan, with yet another shitty "opinion".

I'm terrified of how much I'm agreeing with you recently.


OP, genius idea. AI should not play the role of a sec officer and should not validhunt. Your ideas need to be adapted into the policy as soon as possible.

[DemonFiren]

And theeeere's Zilenan, with yet another shitty "opinion".

I'm terrified of how much I'm agreeing with you recently.


OP, genius idea. AI should not play the role of a sec officer and should not validhunt. Your ideas need to be adapted into the policy as soon as possible.

I pride myself on being the least shitty ligger.

[Lumbermancer]

Isn't all this already in silicon policy?

[Cik]

Isn't all this already in silicon policy?

no

[Lumbermancer]

But wasn't it really? I can't check, because silicon policy has been removed from wiki, since we're apparently going all in after Space Law. But I recall at least points 1 and 4 being addressed. 1, basically you were only prohibited from releasing prisoners from perma afair, otherwise the tators could law 2 you if they wanted. 4 was that law 2 requests are executed in order they are received.

I think. No one really followed it to the letter.

[Drynwyn]

But wasn't it really? I can't check, because silicon policy has been removed from wiki, since we're apparently going all in after Spaaaaaace Law. But I recall at least points 1 and 4 being addressed. 1, basically you were only prohibited from releasing prisoners from perma afair, otherwise the tators could law 2 you if they wanted. 4 was that law 2 requests are executed in order they are received.

I think. No one really followed it to the letter.

You can review Sillicon Policy from the wiki edit history. That said, regardless of written silicon policy, the established precedent is that it's acceptable to assume a confirmed antagonist harmful unless proven otherwise- under this system, that is not the case.

And while, iirc, the wiki said to execute law 2's in the order received, AI's were rarely if ever held to that policy (in no small part because it was not a very good method for conflict resolution.)

EDIT: Apparently wiki just said to be consistent.

[DemonFiren]

But wasn't it really? I can't check, because silicon policy has been removed from wiki, since we're apparently going all in after Spaaaaaace Law. But I recall at least points 1 and 4 being addressed. 1, basically you were only prohibited from releasing prisoners from perma afair, otherwise the tators could law 2 you if they wanted. 4 was that law 2 requests are executed in order they are received.

I think. No one really followed it to the letter.

You can review Sillicon Policy from the wiki edit history. That said, regardless of written silicon policy, the established precedent is that it's acceptable to assume a confirmed antagonist harmful unless proven otherwise- under this system, that is not the case.

And while, iirc, the wiki said to execute law 2's in the order received, AI's were rarely if ever held to that policy (in no small part because it was not a very good method for conflict resolution.)

Wiki actually said to pick one law interpretation system and stick with it, same for orders.

[Lumbermancer]

AI's were rarely if ever held to that policy

People were rarely held to most of the silicon policy, unless someone ahelped or overzealous admin spotted something (Y U ROUNDSTART BOLT). So the question is why do we need a new set of rules like that?

I think AI needs mechanical changes instead.

[Drynwyn]

AI's were rarely if ever held to that policy

People were rarely held to most of the silicon policy, unless someone ahelped or overzealous admin spotted something (Y U ROUNDSTART BOLT). So the question is why do we need a new set of rules like that?

I think AI needs mechanical changes instead.

I agree with you in theory- in an ideal scenario, there would be no need for policy, because all AI laws/law interpretation would be mechanically enforced.

However, SS13 as she is played is simply too complex to create mechanical functions that restrict the AI without being much too tight for the AI to have or create fun.

So the answer to your question, why we need a new set of rules, is: Because the game is too complex and AI's have too many responsibilities for mechanical changes that don't cripple the AI to be feasible, so we're going with the next best thing.

[Topham]

I agree with the lizards. This would be a great change to AI, and would actually be an appropriate nerf to AI, which people keep desperately whining for. Looking for solutions in the playerbase, culture, or implicit rules are much more exciting to me than looking for solutions in game mechanics. This wouldn't nerf what AIs could do, per say, but it would instead reform who AIs are.

Fund he

[DemonFiren]

Wait wait wait what lizards?

[Topham]

Wait wait wait what lizards?

You and Drynwyn, of course~

[Drynwyn]

Wait wait wait what lizards?

You and Drynwyn, of course~

since when am I a lizard
I play sillicon -> human -> lizard in roughly that frequency
(i am a lizard sympathizer though)

[DemonFiren]

Wait wait wait what lizards?

You and Drynwyn, of course~

since when am I a lizard
I play sillicon -> human -> lizard in roughly that frequency
(i am a lizard sympathizer though)

You're a ligger lover because you literally have the avatar of one.

[Drynwyn]

Wait wait wait what lizards?

You and Drynwyn, of course~

since when am I a lizard
I play sillicon -> human -> lizard in roughly that frequency
(i am a lizard sympathizer though)

You're a ligger lover because you literally have the avatar of one.

right but that's different from being a lizard
he should have said "I agree with the lizards and lizard-lovers"

[Steelpoint]

I do agree that right now (or more accurately for AI's up till recently) they assume past harm commited by anyone, irrespective of them being in security, a civilian or a antagonist, equates to them being guaranteed to harm or kill someone in the near future.

While you might get some AI's who side with one side (usually Sec) and give them a bit of leeway with harm, but more often than not they will come down hard on you if you committed past harm.

I recall one round where I was forced to execute a highly dangerous person, due to a combination of him being killing a lot of people and that it was literally impossible for me to keep him in perma. As I did it the AI immediately bolts down all of the brig to contain me, and refuses to listen to my claim that I had no intention of harming anyone else and that I was needed to clean up the mess he made. I was able to blast my way out but I digress.

----------------------

I think the OP's idea's are worth consideration. However if they are implemented the Admins will need to do to Silicon's what they did to Security and crack down hard on anyone and everyone in that position.

However I think Silicon's should still assume that anyone wanting access to a area they have no access to and is considered a high danger area, such as the armoury or toxins, should be against the Silicon's law's. Nothing's more retarded then dealing with Silicon's letting anyone into the armoury via maintenance, and its utterly unreasonable to expect the HoS/RD/Captain to spend ten minutes listing off the entire crew individually to the AI to disallow them access.

[Gun Hog]

I always thought of Undyne as some sort of fish, rather than a reptile.

[Drynwyn]

However I think Silicon's should still assume that anyone wanting access to a area they have no access to and is considered a high danger area, such as the armoury or toxins, should be against the Silicon's law's. Nothing's more retarded then dealing with Silicon's letting anyone into the armoury via maintenance, and its utterly unreasonable to expect the HoS/RD/Captain to spend ten minutes listing off the entire crew individually to the AI to disallow them access.

This isn't intended to override all of silicon policy (later I may make a complete before-after silicon policy comparison)-the AI can still deny access to areas like the armory that are likely to cause Law 1 violations if chucklefucks get in (though they couldn't rat out the person who wanted in for asking unless he forgot to order them not to).


Also undyne is a fish, but she and Alphys are a couple- thus, lizard-lover.