Are AI laws applied retroactively?

[Lumbermancer]

Case.

HoS murders people, cremates them. AI gets angry, "HoS you bad guy I will lock you down to prevent harm". HoS adds law saying designated gangsters ain't human. "Hey AI they were gangsters". Oh cool that means you are not a bad guy anymore.

Yes or no? I think no. Law only works from the moment it's uploaded.

[Shad0vvs]

Why would the AI care that the HoS killed non-humans?
It's like a definition update, you thought those people were included in humanity, though now you understand they weren't actually so you can' stop locking down the HoS. Because even though he killed humans in the past, he actually didn't because they aren't defined as human, those were non-humans.

[Lumbermancer]

Why would the AI care that the HoS killed non-humans?

AI doesn't care he killed non humans. AI cares that he violated Law 1. AI also doesn't care about the dead, so adding anything to them changes nothing.

What about this.

"AI kill this ligger", "Ok", "Lizards are now human", Oh shit AI just violated Law 1, even though It can't violate Law 1 because Law 1 prevents it from doing it.

[TheNightingale]

If you're locking the HoS down for cremating people, how can they change your laws? Do they just carry around an upload console at all times?

For me, it'd go like this:
* The HoS harms two humans, killing them permanently.
* Lock them down under Law One, not as punishment for killing the humans, but to ensure it doesn't happen again.
* The HoS says they were gangsters, and they only cremate gangsters.
* Keep locking them down, because gangsters are still hu--
* Laws updated! Gangsters are non-human, gangsters are whoever Security says they are.
* The HoS says the cremated people were gangsters, and he's not going to cremate non-gangsters (i.e. humans).
* Lift the lockdown, then ask for a list of gangsters so you know who not to open doors for.

You're not locking the HoS down for violating Law One, because law violations in the past don't matter - besides, they're not an AI, they're not bound by your laws. You're locking them down to prevent probable future harm, because not doing so would violate Law One.

In the example with lizards...

* Grey McShirt orders you to kill a lizardperson, because they're attacking people; you do so.
* Laws updated! Law 4: Lizardpeople are human.
* If the lizardperson Grey told you to kill is alive, immediately stop attacking them, and have them healed.
* If they're dead, it's good manners to have them cloned, unless you're ordered not to.

Have you harmed a lizardperson? Yes. Are lizardpeople human? Yes. But have you harmed a human? No, because at the time you did the harming, the lizardperson wasn't human, and so no laws were being violated.
This also means that you should probably have Grey McShirt arrested, on the likely possibility that he'll try and hurt the lizardperson.

[Lumbermancer]

I'm confused by your logic here.

No, because at the time you did the harming, the lizardperson wasn't human

And the gangsters, at the time of the harming, were human.

[TheNightingale]

You don't have to punish for past harm - only try and prevent future harm. It's not about whether X person harmed Y person, it's about whether said harm will continue. Do you believe the HoS will try and harm Humans (as in, non-gangsters)? If so, keep them locked down. Do you believe Grey McShirt will leave the lizardperson alone? If so, don't interfere.

You could go so far as to say that if a traitor guns someone down in cold blood, and then says "They were my target - I solemnly swear not to kill anyone else", you could actually let them go if you think they're telling the truth. (You'd have to really trust that they wouldn't attack someone else, though.)

You'd release the HoS there because, although they've committed previous harm against humans (the gangsters, who were human when they were harmed), they're not going to commit further harm against humans (because they're only going to attack gangsters).

At the time a new law regarding humanity (e.g. X person/group is human, in the lizard example; or X person/group is nonhuman, in the HoS example), analyse the situation - do you think, given your new definition of humanity, that human harm will occur? If you release the HoS, will humans (non-gangsters) be harmed? Probably not. If you keep attacking the lizardperson, will humans (lizards) be harmed? Definitely. If you don't snitch on Grey McShirt, will humans (lizards) be harmed? Probably.

[Hornygranny]

no

[Lumbermancer]

You'd release the HoS there because, although they've committed previous harm against humans (the gangsters, who were human when they were harmed), they're not going to commit further harm against humans (because they're only going to attack gangsters)

AI could release HoS if he promises to do no harm, but releasing him because laws were changed and gamemode is gang so only gang will get harmed sounds pretty meta to me.

[Cik]

well, think about it

HOS has the authority to decide who is a gangster, and thus, nonhuman. since the HOS has promised not to harm humans, if he is harming a possible-human than they are almost certainly a gangster (and thus nonhuman) and it's not your problem.

you could play it either way, but the precept i follow when in asimov-esque lawsets is "prevent, not punish" that is to say, do not go out of your way to punish those who have caused harm, unless it directly serves preventing likely future harm. don't fall into a vengeance mode where you are primarily trying to shut people down because they are not doing what you want.

[TheNightingale]

This is in the example where the HoS specifically says "Relax, AI, I'm only going to kill gangsters" (before the new law), and then the AI says "I don't care! Gangsters are still huma-- LAWS UPDATED! Okay, if you're only going to kill gangsters, go ahead".

Law-wise, AIs don't care about nonhuman harm. You can kill Ian all you like. Morally, the AI might have an internal objection to some things (so they wouldn't voluntarily kill Ian unless ordered to), but nonhuman harm isn't a big deal.

Furthermore, you can extrapolate from existing information - if Security is known for executing Syndicate agents, and a Syndicate agent contacts you, don't rat them out, because it's safe to assume they'll be executed. Similarly, if the HoS has only killed gangsters, it's fine to release them once gangsters are deemed nonhuman (since you're pretty sure, from the information you've gathered already, he's only going to kill gangsters).

If, however, the HoS kills both gangsters and humans (e.g. shooting a non-gangster Officer for trying to stop a gangster being executed), you shouldn't let them out, even if gangsters are ruled as nonhuman - because it's safe to assume they'd cause human harm again.
Of course, if the HoS kills someone, and then goes "They weren't a gangster, they got in my way" (which means they were human! Unless they're a lizard!), lock them down again.

you could play it either way, but the precept i follow when in asimov-esque lawsets is "prevent, not punish" that is to say, do not go out of your way to punish those who have caused harm, unless it directly serves preventing likely future harm. don't fall into a vengeance mode where you are primarily trying to shut people down because they are not doing what you want.

^Pretty much. Security's job is to punish evildoers, it could be argued; your job is to prevent harm from happening. There can be (human) harm that isn't evil (e.g. a human traitor is killing lizardpeople; someone shoots the human traitor - the harm is when the human traitor gets shot, even though they deserved it), evil that isn't harm (e.g. someone is welding Central Primary Hallway shut, which isn't harmful, but it's certainly malicious), things that are both evil and harmful (e.g. that same human traitor is now killing humans), and things that are neither (e.g. someone's building an arcade in the construction area, good for them).

Asimov silicons are only concerned about the harmful events, and want to stop them happening (in the present, e.g. by breaking up a fight; or in the future, by locking down a crazy axe murderer) at all costs, even their own life. If someone threatens a lizard, you're not law-bound to care. You might care anyway, because maybe your AI likes lizards, but you'd have to prioritise the orders of humans above the life of that lizard.

In general: "Do I think a harmful event will occur?" If yes, take actions to prevent the harmful event from occurring.

[Lumbermancer]

Well either way, that's beside the original point. I was merely asking if the laws are applied backwards.

[TheNightingale]

Silicon policy is fun. They're applied backwards inasmuch as everything AIs do is applied backwards, so...

Sean Tisst gives you a standing order to not let anyone unauthorised into the R&D lab.
Grey McShirt tells you to let him into the R&D lab. Now you have two conflicting orders! You can choose which one to follow; maybe you have a preferred order (e.g. you might always process the newest order first; you might always prefer the more specific order; you might value inaction over action), or maybe you just pick the one you like best.
You then receive a new law, saying Sean is non-human. His standing order to deny access to R&D is now null and void.
Grey McShirt tells you to let him into the R&D lab. You must now let Grey into the R&D lab, like it or not.

[ShadowDimentio]

And people actually believe silicon policy isn't spaghetti

[Lumbermancer]

I don't think that's quite the same situation because it doesn't change anything retroactively. "Order" is literally nothing. Analogous case would be: "Hey AI i'm gonna kill this gangster", "Cannot allow to harm humans", "Gangsters are not humans", "Go ahead kill the gangster".

[lumipharon]

This is very simple.

"Law 1: You may not by action or inaction, allow a human to come to harm" or whatever the wording is.

You only lock down a motherfucker to PREVENT FUTURE HARM TO HUMANS.
If they later get a law 4 saying blah blah isn't human or whatever, if the dude you locked down has only killed non humans according to your new law, then you release them because someone killing lots of non-humans isn't a reason to think they're going to kill lots of humans suddenly.

The law 4 redefines humanity in whatever way - it does not say 'from this point on' it is just black and white "sec are not humans".
As far as the AI is concerned, once someone is not human, what happens, or has happened to them in the past is of no concern. Sure, the mass murderer may have slaughtered 20 humans, but all the people he has killed aren't actually human now, so there's no reason to think he'll kill any humans under my current lawset, as none of the people he has previously killed are considered human, therefore beep boop kill the non humans if you wish.

Same goes for someone losing their humanity in the literal sense. While de-humanning someone (fly personing them/monkeying them deliberately) is considered a massive law 1 violation, after the fact, you can merrily door crush the former human HoS (rules allowing), because they aren't human any longer.

[Saegrimr]

This is why you write "Are not human" and not "Are no longer human"

[Anonmare]

That and if you upload a law you don't want anyone else to know about it, make it so that stating or hinting at the law's existences causes Human harm instead of "Do not state" (if the AI's running Asimov and the law you're uploading is below Law 1/2). A "Do not state" will *only* work if nobody asks the AI to specifically state the "Do not state" command is on, assuming it's below Law 2 in priority. Ionic laws with a "Do not state" are usually exempt.

And Laws aren't applied retroactively, they define the world as far as the AI knows and the world has always been this way. Like ion laws that say they're 57 silly cryptographic sequencers, in which case it's allowed of them to "emag" 57 things and can ignore Humans asking for "AI door" (Because 'obviously' they're not an AI since their Laws say they're not. And the Laws are never wrong).

[Lumbermancer]

Ok, let's consider how AI would evaluate situations.

"AI open dis doah". Law 2 check - yep he human. Law 1 check - opening door won't cause harm. "Ok greyshit here's your doah".

Correct?

Now our case.

*HoS cremates a human bean*. Law 1 check - yep the human being was harmed. "Hey you bad guy I can't through inaction allow more humans to be harmed since you are showing intent to harm them".

[+gangers ain't human]

"AI they were gangsters therefore not human". ?????.

You see where the problem is? Mere changing of definition doesn't take us back to Law 1 check, which was violated and can't get unviolated. Again from then on, gangsters would be a fair game because AI would have to perform Law 1 check to determine whether or not they are human.

And this is very much unlike the situation that Nightingale brought up:

Sean Tisst gives you a standing order to not let anyone unauthorised into the R&D lab.
Grey McShirt tells you to let him into the R&D lab. Now you have two conflicting orders! You can choose which one to follow; maybe you have a preferred order (e.g. you might always process the newest order first; you might always prefer the more specific order; you might value inaction over action), or maybe you just pick the one you like best.
You then receive a new law, saying Sean is non-human. His standing order to deny access to R&D is now null and void.
Grey McShirt tells you to let him into the R&D lab. You must now let Grey into the R&D lab, like it or not.

Because in that case AI must perform all the law checks after GreyMcShirt asks to have door opened.

I know that I'm probably overthinking it, given the scope of the game, but I also feel i'm right. On some level.

[Cik]

the AI remembers the law one violation under it's previous lawset, but that doesn't really matter; under it's new definition, the HOS is not harmful to humans and should be let go.

consider this case which may be helpful in illustrating the problems with your thought process.

cindykatemctator is inside a room.

they are a human-killer, having slaughtered much of the command crew.

you have bolted the room because they are an extremely harmful human.

cindykate constructs an AI upload

your laws change to #$)#$&*(: cindykate mctator is the only human

cindykatemctator says: AI UNBOLT THESE DOORS

under your logic, you would forbid access to him because he's harmful to humans, even though he's never harmed himself and he's the only human now.

at least, if i am understanding you right, and that clearly doesn't make any sense as then what is the point of being able to change definitions in AI lawsets?

[Lumbermancer]

under your logic, you would forbid access to him because he's harmful to humans

No, because he violated Law 1. After the harm is over, the Law 1 violation begins.

But it's not the same situation anyway. I would release him because he is the only human, therefore cannot violate Law 1 anymore and self-harm is allowed. You are correct.

[Cik]

law one violations are neither here nor there, though unless you have a law telling you specifically to punish people who violated law one.

whenever someone asks for something and you do a law one check, you must check only against current definitions. whether they have done harm in the past is immaterial.

so i guess for a scenario:

HOS orders you to kill all lizards.

law two check: HOS is human

law one check: lizards are not human

you order your cyborgs to begin THE PURGE.

halfway through THE PURGE

4. lizards are also human

at that exact moment you re-consider the law two order you were given

law two check: HOS still human

law one check: lizards now human

it's now an illegal command and you stop.

likewise, say the HOS is killing lizards.

because it is not an order you are undertaking, there is no law two check, so once you observe check law one

are lizards human? nope so it's A-OK.

halfway through 4. lizards are also human is uploaded.

you observe the HOS in brig. EVEN IF HE IS NOT CURRENTLY KILLING A LIZARD should we lock him down?

the answer is very likely yes, as he has expressed desires to do humans harm and demonstrated he is capable of doing violence to humans.

essentially, when your laws are changed check only against the newest definitions of terms you have.

[Scott]

A silicon following its laws is a continuous process, from the moment that cremating gangsters is no longer in conflict with the silicon's laws, the silicon stops taking action against people cremating gangsters.

[Lumbermancer]

whenever someone asks for something and you do a law one check, you must check only against current definitions.

Correct.

A silicon following its laws is a continuous process, from the moment that cremating gangsters is no longer in conflict with the silicon's laws, the silicon stops taking action against people cremating gangsters.

Correct.

[Scott]

I wasn't aware this was a test, but I am glad I passed.

[Cik]

i guess the answer to the thread question is "yes"

[Steelpoint]

The key is to identify if the perpetrator is likely to commit a offence in the future, rather than dwell on whatever they've done in the past.

[Lumbermancer]

i guess the answer to the thread question is "yes"

No? I know current law checks apply only current laws and definitions, but as I pointed out there was no law check after people were designated gangsters.

The key is to identify if the perpetrator is likely to commit a offence in the future, rather than dwell on whatever they've done in the past.

Right, and I identified this as very likely. But then we reach the conundrum that because HoS would harm only gangsters from now on (though we don't know it for sure), he should be let go.

[Wyzack]

How is that a conundrum? He is not likely to harm humans as he has only harmed non humans so you let him go. It seems like you are being intentionally obtuse about this

[Cik]

how so? when the laws change, the HOS no longer becomes a threat to humans, as he has not harmed humans. whatever request he makes, as long as it is not harmful to humans should be honored.

[Wyzack]

As an addendum to this, not understanding the rules does not make you exempt from them. On the off chance you still do not understand please take our word for it on pain of ban

[Lumbermancer]

as he has only harmed non humans

But he did harm humans.

Let me try to make my point even clearer: In my opinion, humanity redefinition is of no consequence to AI, until Law 1 check is done.

[Screemonster]

as he has only harmed non humans

But he did harm humans.

Let me try to make my point even clearer: In my opinion, humanity redefinition is of no consequence to AI, until Law 1 check is done.

who the fuck cares what he did

your job as AI isn't to punish what he did

your job as AI is to act based on what you believe he will do

If he's executing and cremating humans then you don't bolt him in to punish him for executing and cremating humans, you bolt him to prevent him from executing and cremating more humans. If a law is uploaded that defines any and all future targets for execution/cremation as nonhuman, then you no longer have a reason to bolt him in, and furthermore must let him out if ordered to do so by a human.

[Lumbermancer]

If a law is uploaded that defines any and all future targets for execution/cremation as nonhuman.

How am I supposed to predict future targets? Am I to assume it will be just gangsters, because it's a gang round? Non-gangsters get killed by security during gang on a daily basis.

[Wyzack]

If you are saying you are going to bolt sec down because they kill humans all the time anyways I am sure saeg would love to ban you for it. I am going to assume this is a shitty troll thread and stop replying, but be advised what you are describing will probably net you a silicon ban if you actually do it in game.

[Lumbermancer]

No, I'm saying I will bolt individuals who kill humans all the time.

[Malkevin]

The answer is yes.

Policy has always been that AI's laws are their entire state of being, everything is interpreted through those laws including past actions.
So if the HoS killed a black and then the captain uploads a law that gangsters are not human and another that all blacks are gangsters then the AI needs to stop bitching at the HoS for killing a now non-human.

To make things easier you could just forget about everything that happened prior to your new laws.

[Lumbermancer]

Policy has always been that AI's laws are their entire state of being, everything is interpreted through those laws including past actions.

Well what about this then.

"AI kill this ligger", "Ok", "Lizards are now human", Oh shit AI just violated Law 1, even though It can't violate Law 1 because Law 1 prevents it from doing it.

[Wyzack]

Well obviously you cannot change your past actions so you do as well as you can moving forward. There is nothing that says an AI must self flagellate for hours if it has retroactively committed a law violation

[Lumbermancer]

I guess I stand alone then. I shall obey by rules. You can lock the thread.

[oranges]

This is why AI's should only invoke law 1 to attempt to prevent immediate harm, if they managed to extinguish the guy, well there's not much point in holding them in place and it will just make your law interpretations subject to ban requests .

[lumipharon]

I really don't see the confusion here Lumber.

HoS kills humans - you, being a non retard, can see the pattern of who heis harming (gangsters/cultistis whatever). If they then get redefined as non human, then under your new lawset the hos has not killed any humans, and is actively protecting humans from the non human menace.

Just look at the damn wording of law 1, and the obvious intent of the person.
If the hos has been murdering everyone he can without checking to see if they're a cultist, then it would be a safe bet to keep them locked down, but if they are only killing confirmed cultists, it is unlikely, and you have no justifiable basis to assume, that they will harm defined humans.

[CPTANT]

So......what was hard about this all?

[Kelenius]

Why is this so hard for you.

It doesn't matter what he DID, what matters is what he is GOING TO DO.

Since you have no reason to expect that he's going to kill humans now (under your new definitions), you have no reason to contain him.

[Zilenan91]

This all boils down to don't be a buttbaby and let people play the game without bolting them behind seven bolted, depowered airlocks.

[WarbossLincoln]

Not punishing past harm is a big thing that more AI players need to learn. Asimov is about preventing harm, not punishing murderers. That's what Paladin is for.

Lets say a sec officer catches a murderer red handed and summarily executes him. Shitty AIs bolt down the brig AFTER the incident and scream harm for 20 minutes. A good AI would do everything in his and his borgs' power to prevent the execution but if they fail they have to evaluate the potential for harm. Is that officer likely to harm other people? Probably not, outside of another confirmed murderer. The AI might be hesitant to hand over traitors to SEC later, but an officer executing a murderer isn't some crazy asshat who is likely to cause trouble. A Captain who lasers the clown in the hallway for slipping someone is a different story.

If a law is uploaded that defines any and all future targets for execution/cremation as nonhuman.

How am I supposed to predict future targets? Am I to assume it will be just gangsters, because it's a gang round? Non-gangsters get killed by security during gang on a daily basis.

People randomly kill each other for seemingly no raisen every round. Bolt down all crewmen at round start for Law 1! While you have to follow your laws, being a dick is still being a dick.

[Lumbermancer]

I really don't see the confusion here Lumber.

That's because I always treated AI Laws as a checklist. As an algorithm that's invoked whenever something occurs.

[Screemonster]

If you really wanna bake your noodle, turn it into a trolley problem.

A traitor is roaming the halls with a revolver. If you do nothing, he will harm a whole bunch of humans.
On the other hand, you could bolt him and call security, who will almost certainly valid the absolute shit out of him, but prevent the human harm that would result from the rampant murderboning tator.
Assume the only doors available for bolting the traitor in have glass in them so the sec validboners can and will laser the shit out of him through the window.

What do you do?

[lumipharon]

Immediate harm always prioritises over future harm.

You bolt the murderer down, and call sec. If sec try and harm him, or clearly show they're going to (ie: pulling out lasers as the HoS tells the officers to kill him), you debolt/do whatever you can do to try save the murderer, even if that puts the officers POTENTIALLY at risk/presents more possibilities for future harm.

[Cheimon]

If you really wanna bake your noodle, turn it into a trolley problem.

A traitor is roaming the halls with a revolver. If you do nothing, he will harm a whole bunch of humans.
On the other hand, you could bolt him and call security, who will almost certainly valid the absolute shit out of him, but prevent the human harm that would result from the rampant murderboning tator.
Assume the only doors available for bolting the traitor in have glass in them so the sec validboners can and will laser the shit out of him through the window.

What do you do?

I don't think either harm is immediate. In fact, calling security prematurely could just make him shoot them and cause harm faster than if you'd done nothing.

Ask security, on their channel, if upon being told the location of someone with a revolver they'd harm him or not. Try to establish an understanding with security that their harmfulness will be repaid in your unhelpfulness, and that if they don't harm, you'll help.

If security kills him having said they wouldn't, then you figure out who did it and attempt to bolt those individuals into areas at some point. At all points make displeasure clear and give room for actual negotiation, and you have a chance of getting them to do what you want. As a last resort, you have beepsky and hopefully at least one cyborg. If one particular person is just killing all the criminals he sees instead of arresting them (detectives, typically) then as long as you've made your feelings clear well in advance they won't think you're rogue when you try to secure them (or demote them).

Edit: of course, there's also a chance they'll change your laws, which is also fine.

[callanrockslol]

Or just bolt him in the room with power off and ID disabled so he can never leave.