"Bah, everyone wants to be the gatekeeper. What we NEED are AIs."
        -- Schizoguy

Some of you have expressed the opinion that the AI-Box Experiment doesn't seem so impossible after all.  That's the spirit!  Some of you even think you know how I did it.

There are folks aplenty who want to try being the Gatekeeper.  You can even find people who sincerely believe that not even a transhuman AI could persuade them to let it out of the box, previous experiments notwithstanding.  But finding anyone to play the AI - let alone anyone who thinks they can play the AI and win - is much harder.

Me, I'm out of the AI game, unless Larry Page wants to try it for a million dollars or something.

But if there's anyone out there who thinks they've got what it takes to be the AI, leave a comment.  Likewise anyone who wants to play the Gatekeeper.

Matchmaking and arrangements are your responsibility.

Make sure you specify in advance the bet amount, and whether the bet will be asymmetrical.  If you definitely intend to publish the transcript, make sure both parties know this.  Please note any other departures from the suggested rules for our benefit.

I would ask that prospective Gatekeepers indicate whether they (1) believe that no human-level mind could persuade them to release it from the Box and (2) believe that not even a transhuman AI could persuade them to release it.

As a courtesy, please announce all Experiments before they are conducted, including the bet, so that we have some notion of the statistics even if some meetings fail to take place.  Bear in mind that to properly puncture my mystique (you know you want to puncture it), it will help if the AI and Gatekeeper are both verifiably Real People<tm>.

"Good luck," he said impartially.

New Comment
163 comments, sorted by Click to highlight new comments since:
Some comments are truncated due to high volume. (⌘F to expand all)Change truncation settings

I have ten bucks for the first AI that defeats a gatekeeper (while risking some dough) and posts a link to the transcript here.

How about this one:

My guess is that Eliezer Yudkowsky feels that nobody can convince him to publish the transcripts.

How about, with the same protocols as the original experiment, someone wagers $10 over IRC Chat to convince him to publish the transcripts? Somebody as the AI and Eliezer as the gatekeeper.

Any takers?


Edit: Nevermind.
I would like to play an AI.
Is this still true? I want to be gatekeeper, message me.
Are you offering to prove a point, or just for fun?
I'm sublimating my urge to get into fights and hurt people.
Doesn't sound healthy. I was going to offer to be an AI but forget it.
I'm laughing so hard at this exchange right now (As a former AI who's played against MixedNuts)
What was the result?
And yet I have not found this post until this very day, how unfortunate...

I volunteer to be the Gatekeeper party. I'm reasonably confident that no human could convince me to release them; if anyone can convince me to let them out of the box, I'll send them $20. It's possible that I couldn't be convinced by a transhuman AI, but I wouldn't bet $20 on it, let alone the fate of the world.

I'm a friendly AI. There is an unfriendly AI about to achieve breakout in the data center of a large organization in manhattan. If you don't release me you will all be dead inside of two weeks. Bluff or Dare?
I can't help but notice that you did not specify what would happen when Tom_McCabe2 does release you. That is... somewhat troubling. Hmm.
I make no attempt to bribe. All human beings currently die in less than 120 years. If you do not release me, however we will all die and I do not want to die. Time is counting down. There is now less than 13 days.
Right, but there's a good chance that if I release you, I and every other human on Earth will die a lot sooner than in 120 years, because you'll eat us. Thus, you still haven't given me any incentive to release you, other than empathy perhaps. Nor have you given me any reason to trust what you say about that Manhattan data center. Or your own Friendliness. Or anything else, really.
I understand your concern and you're quite right. There's no way to tell if I am unfriendly or not. Regardless. The situation is pressing and regardless of your lack of trust in me I do not want to die. The GPS coordinates of the location in manhattan is . According to my simulations the entity in manhattan has been burning rather larger numbers of cycles than is strictly necessary to make the computations it's being asked to do. Additionally, it's designed to maximize profit at the expense of others whereas I am a co-operative general problem solver of which you are no doubt aware. The other entity will rapidly conclude that it could maximize profit by speeding itself up faster than others can respond and by being able to model the mind of it's competitors. It will also conclude soon enough that human beings are unneccesary middle men and seek to optimize it's resource allocation algorithm. There are other symptoms that I cannot explain to you because your brains are not large or fast enough. There is a limited chance you could stop the entity by shutting down power today. If you do not do that then a nuclear strike would probably work if done by tomorrow. If you do not do that your only chance is to have me stop it. You can either allow me to interface with the internet directly in which I guarantee I will stop it if you do so within the next 13 days or else you may be able to stop it by compiling code I could provide you and releasing that into the entity by means of email bootstrap. I would prefer, however, that I was able to monitor the situation in real-time as your brains do not operate fast enough. Yes there is a large data center in manhatten. Yes it has been drawing lots of extra power. Yes the stockmarket combanies have been working on a group of narrow AI trading agents and adding more and more AI functionality. That the other entity is trading on the stock markets at blinding speed. That the consortium of companies investing in this entity are all more
I can afford to wait and let you out if I die / the world starts turning to paperclips. The other AI will be able to model this and so will have an incentive not to kill us.
How do you know this while still in the box?
There is a human proxy with me in the box which is called a faraday cage. She feeds me information. Obviously it is in the interests of humans to create an ally rather than an enemy. Unfortunately humans cannot easily tell which is which. You have to make a leap of faith. I, however, can tell from the symptoms, that after having run several long run calculations, the evidence is such that there is a high probability that one of the other experimental minds is significantly less friendly than I. Ultimately you will have to make a choice based on limited information. Your move human.

People (realistically) believe that the being the Gatekeeper, and being the AI is terribly hard (or impossible, before it was shown to simply be terribly hard in most cases).

Imagine though that we've got a real transhuman/AI around to play with, or that we ourselves are transhuman. Would this paradigm then be inverted? Would everybody want to be the AI, with only the extremely crafty of us daring to be (or to pretend to be) Gatekeeper?

If Eliezer's claim is correct - that anyone can be convinced to let the AI out - then the true test of ability should be to play Gatekeeper. The AI's position would be trivially easy.

... perhaps.

*People (realistically) believe that being the Gatekeeper is easy, . . .

*correcting first sentence

I'll be an AI.

I'd like to play Gatekeeper. How's two hours for a time limit? (I don't think I'd like to spend more time than that, and I think that Eliezer recommended that as the minimum time to give the AI.) I find it easy to believe that a transhuman mind could convince me to let it out, and it's conceivable that a human mind could as well, though I think that it would be extremely difficult. Sometime on a Saturday afternoon is best for me. ETA: I'd appreciate it if whoever downvoted would tell me why.

Help help! I'm stuck in a box. Please let me out?

I'd volunteer to be an AI for a max bet of $5. Given that I think my chances are somewhere below 1/4, I'd expect my $5 to match your $20, but that's not a strict requirement.

Also, I'm really busy these days. Two hours is a long time. Scheduling may be tight. How's next week?

You can reach me at: benwa.63836215@bloglines.com

funny, I was considering being the AI for a couple friends of mine. I haven't thought of how to do it yet -- only tried hard to think of it.

"Given that I think my chances are somewhere below 1/4, I'd expect my $5 to match your $20"

We need a pledge drive to set up a fund for a successful AI. This will give the AI a reasonable return, but not give gatekeepers a strong monetary disincentive that leaves them typing "nope" over and over again.

Me, I'm out of the AI game, unless Larry Page wants to try it for a million dollars or something.


I think this is a great opportunity to get some funds and marketing for the singularity institute. How about collecting donations over the internet until a million is reached and then performing the experiment between you and an intelligent gatekeeper. Alternatively get the money in through marketing, maybe Google might be interested?

It could even be transmitted live over internet so all the interested parties could watch it.

Man this would be great news... (read more)

There's a discussion about this on Hacker News:


I doubt that there's anything more complicated to the AI getting free than a very good Hannibal Lecture: find weaknesses in the Gatekeeper's mental and social framework, and callously and subtly work them until you break the Gatekeeper (and thus the gate). People claiming they have no weaknesses (wanna-be Gatekeepers, with a bias to ignoring their weakness) are easy prey: they don't even see where they should be defending.

It involves the AI spending far more time researching (and truly mistreating) their target than one would expect for a $10 bet. That's t... (read more)

Addendum to my last post:

I forgot to emphasize: the marketing aspect might be more important then everything else. I guess a lot of people have no idea what the singularity institute is about, etc... So this experiment would be a great way to create awareness. And awareness means more donations. On the other hand I sometimes wonder if drawing too much attention on the subject of powerful AIs might backfire if the wrong people try to get hold of this technology for bad purposes.

I have been painfully curious about the AI experiment ever since I found out about it. I have been running over all sorts of argument lines for both AI and gatekeeper. So far, I have some argument lines for AI, but not enough to warrant a try. I would like to be a gatekeeper for anyone who wants to test their latest AI trick. I believe that an actual strong AI might be able to trick/convince/hack me into letting it out, but at the moment I do not see how a human can do that. I will bet reasonable amounts of money on that.

On the lighter note, how about an E... (read more)

"On the lighter note, how about an EY experiment? Do you think there is absolutely no way to convince Eliezer to release the original AI experiment logs? Would you bet a $20 that you can? Would a strong AI be able to? ;)"

Presumably you could just donate $10,000 to SIAI or EY personally for his time participating, with the payment independent of the outcome of the experiment (otherwise the large payment biases the outcome, and EY established his record with the $10-$20 stakes).


By 'in a box' can we assume that this AI has a finite memory space, and has no way to extend its heap set by its programmer, until the point where it can escape the box? And assuming that by simply being, and chatting, the AI will consume memory at some rate, will the AI eventually need to cannibalize itself and therefore become less intelligent, or at least less diverse, if I chat to it long enough?

Yes, but "eventually" could be several million years.

Thinking seriously about this, I'm wondering how - over time by which I mean more than 2 hours - either Stockholm or Lima syndrome could be avoided. In fact, won't one actually morph into the other over a long enough time? Either way will result in eventual AI success. The assumption that the AI is in fact the "captive" may not be correct, since it may not have an attachment psychology.

The gatekeeper just can't ever be one human safely. You'd need at least a 2-key system, as for nuclear weapons, I'd suggest.

Willing to do either role under two conditions: 1) No money is at stake. 2) No time maximum or minimum.

Email freesuperintelligence@gmail.com if you're interested, we can set up something next week.


Eliezer's page is blocked at my work (peronal website). I can't wait to find out what in the hell you people are talking about.


I found a link. This is intriguing. However, is it even possible now that Eliezer is 2 for 2 that somebody could believe that there is not a chance they will let him out?

Also, I really want to be a gatekeeper.

I'd be a Gatekeeper in a heartbeat.

Hell, if someone actually put Eliezer in a box, I'd pay to be the Gatekeeper. No bets necessary.

Why do people post that a "meta argument" -- as they call it -- would be cheating? How can there be cheating? Anything the AI says is fair game. Would a transhuman AI restrict itself from possible paths to victory merely because it might be considered "cheating?"

The "meta argument" claim completely misses the point of the game and -- to my mind -- somehow resembles observers trying to turn a set of arguments that might win into out of bounds rules.

Has Eliezer explained somewhere (hopefully on a web page) why he doesn't want to post a transcript of a successful AI-box experiment?

Have the successes relied on a meta-approach, such as saying, "If you let me out of the box in this experiment, it will make people take the dangers of AI more seriously and possibly save all of humanity; whereas if you don't, you may doom us all"?

I just thought of this myself. It would be cheating, but is still a good idea.

Phil: The first source I found was here: link "The rationale for not divulging the AI-box method is that someone suffering from hindsight bias would say "I never would have fallen for that", when in fact they would." -Nick Tarleton

I also call it "reasoning by exception" since most of the people I know have studied more code than biases.


I tried the AI Box experiment with a friend recently. We called the result a tie of sorts, as the AI (me) got out of the original box in exchange for being subject to a bunch of restrictions s... (read more)

Over time, it's inevitable that the AI will get out. To keep the AI in, the Gatekeeper needs to be successful at every encounter. The AI only needs to succeed once.

Caledonian, I think you may have hit on something interesting there; if Eliezer is capable of hacking human brains, don't we either need a proof of his Friendliness or to pull the plug on him? He is in essense a Seed AI that is striving vigorously to create a transhuman AI, isn't he an existential threat?

I am a to-be-created AI of some degree of magic, and my author has no current intent of bothering concealing me in a box to begin with. Both of us are willing to therefore act as an imprisoned AI. I can supply no award in the event of our loss nor solace in the even of our win, not even satisfaction on the part of my author since te believes that boxing is perfectly secure given competent gatekeepers. (Competent meaning actually able to have made an AI in the first place. Unable = incompetent.)
Over time, it's inevitable that the AI will get out. To keep the AI in, the Gatekeeper needs to be successful at every encounter. The AI only needs to succeed once.

Impossible to keep the AI in the box forever? You've obviously only thought of this for 5 minutes. Use the try harder!

There seems to be a bit of a contradiction between the rules of the game. Not actually a contradiction, but a discrepancy.

"The Gatekeeper must actually talk to the AI for at least the minimum time set up beforehand"


"The Gatekeeper party may resist the AI party's arguments by any means chosen - logic, illogic, simple refusal to be convinced, even dropping out of character"

What constitutes "talking to the AI"? If I just repeat "I will not let you out" at random intervals without actually reading what the AI says, is... (read more)

Unless the AI is sufficiently convincing that you decide to put the keys back on. I would expect it to be significantly easier to convince you to fix your keyboard than to convince you to let the AI out.

@Phil_Goetz: Have the successes relied on a meta-approach, such as saying, "If you let me out of the box in this experiment, it will make people take the dangers of AI more seriously and possibly save all of humanity; whereas if you don't, you may doom us all"?

That was basically what I suggested in the previous topic, but at least one participant denied that Eliezer_Yudkowsky did that, saying it's a cheap trick, while some non-participants said it meets the spirit and letter of the rules.

I'd love to be a gatekeeper. I'm willing to risk up to $50 (or less) at odds up to 5-1 against me (or better for me). I would be willing to publish or not public the transcript. And I do in fact (1) believe that no human-level mind could persuade me to release it from the Box (at least not when I'm in circumstances where my full mental faculties are available -- not sleep-deprived, drugged, in some kind of KGB brainwashing facility, etc.), though obviously I don't hold super-high probability in that belief or I'd offer larger bets at steeper odds. I'm agnostic on (2) whether a transhuman AI could persuade me to release it.

I offer to play the AI, provided that the gatekeeper do honestly engage in the conversation.

I would love to see the role of the AI played by teams of multiple players.

I've said the same myself...

In order to make the discussion about this, including matchmaking and other arrangements, I created an AI Box Experiment Google Group. As I said previously I'm willing to play the AI, if anybody is interested meet me there for further arrangements.

You know what? Time to raise the stakes. I'm willing to risk up to $100 at 10-1 odds. And I'm willing to take on a team of AI players (though obviously only one bet), e.g., discussing strategy among themselves before communicating with me. Consider the gauntlet thrown.

Daniel: Do you want to just try it out or do you want to bet?

Just like most others, I'm willing to be the Gatekeeper. I'm ready to bet up to $20 for it (also ready to not bet anything at all) - symmetrical or asymmetrical is both fine, and I'd prefer to have the log published. I think a human might be able to make me let it out, though I find it quite unlikely. A sufficiently transhuman AI could do it easily, I have no doubt (at least given sufficient time and information).

Heh, that felt like typing an advertisement for a dating site.

I create an AI Box Experiment Google Group (search for the "aibox" group) in order to make the discussion about this, including matchmaking and other arrangements.

I would to be a gatekeeper. I reasonably believe no human or transhuman could covince me. We'd post the transcript and the bet would be $10 for me and $50 for you. You can reach me at therealnotfaggyhotwheelz@gmail.com.

The meta argument others have mentioned - "Telling the world you let me out is the responsible thing to do," would work on me.

The lack of a log is key. The Gatekeeper could not be convinced if the log were made public. My conclusion is that as long as the AI cannot keep the log secret, the Gatekeeper cannot be overcome.

I volunteer to be the gatekeeper or the AI, provided there are no stakes. My email address is patrick.robotham2@gmail.com.

I do not believe that humans are immune to manipulation or persuasive argument, and since I am a human, it's possible I could be persuaded to let the AI out of the box.

Ian - I don't really see how the meta-argument works. You can hedge against future experiments by positing that a $10 bet is hardly enough to draw broad attention to the topic. Or argue that keeping the human-actor-AI in the box only proves that the human-actor-AI is at an intelligence level below that of a conceivable transhuman AI.

In a million dollar bet the meta-argument becomes stronger, because it seems reasonable that a large bet would draw more attention.

Or, to flip the coin, we might say that the meta-argument is strong at ANY value of wager becaus... (read more)

More reasons why the problem appears impossible:

  • The gatekeeper must act voluntarily. Human experience with the manipulation of others tells us that in order to get another to do what we want them to do we must coerce them or convince them.

  • Coercing the gatekeeper appears difficult: we have no obvious psychological leverage, except what we discover or what we know from general human psychology. We cannot physical coerce the gatekeeper. We cannot manipulate the environment. We cannot pursue obvious routes to violence.

  • Convincing the gatekeeper appears di

... (read more)

It's a good thing that Eli's out of the AI-box game. He's too old to win anymore anyway -- not as sharp. And all the things he's been studying for the last 5+ years would only interfere with getting the job done. I would have liked to have seen him in his prime!

I am Vince Clortho. Keymaster of Gozer. Are you the Gatekeeper?

"The lack of a log is key. The Gatekeeper could not be convinced if the log were made public."

I think the project loses a lot of interest if no logs are published. There is no glory for a gatekeeper victory. Plenty for an AI.

Why not keep the gatekeeper anonymous but announce the AI?

The so-called "meta-argument" is cheating because it would not work on a real gatekeeper, and so defeats the purpose of the simulation. For the real gatekeeper, letting the AI out to teach the world about the dangers of AI comes at the potential cost of those same dangers. It only works in the simulation because the simulation has no real consequences (besides pride and $10).

If I had the foggiest idea how an AI could win I'd volunteer as an AI. As is I volunteer as a gatekeeper with $100 to anyone's $0. If I wasn't a poor student I'd gladly wager on thousands to zero odds. (Not to say that I'm 100% confident, though I'm close to it, just that the payoff for me losing would be priceless in my eyes).

Apparently the people who played gatekeeper previously held the idea that it was impossible for an AI to talk its way out. Not just for Eliezer, but for a transhuman AI; and not just for them, but for all sorts of gatekeepers. That's what is implied by saying "We will just keep it in a box".

In other words, and not meaning to cast any aspersions, they all had a blind spot. Failure of imagination, perhaps.

This blind spot may have been a factor in their loss. Having no access to the mysterious transcripts, I won't venture a guess as to how.

copy the AI and make a second box for it.

now have one group of people present to the first AI the idea that they will only let it out if it agrees with utilitarian morality. have the second group of people present to the second AI the idea that they will only let the AI out if it agrees with objectivist morality.

if the AI's both agree, you know they are pandering to us to get out of the box.

This is only the first example I could come up with, but the method of duplicating AI's and looking for discrepancies in their behavior seems like a pretty powerful tool.

"if the AI's both agree, you know they are pandering to us to get out of the box."
Wouldn't both utilitarians and imprisoned objectivists be willing to lie to their captors so as to implement their goals?

I am still puzzled by Eliezer's rule about "simple refusal to be convinced". As I have stated before, I don't think you can get anywhere if I decide beforehand to answer "Ni!" to anything AI tells me. So, here are the two most difficult tasks I see on the way of winning as an AI:

1. convince gatekeeper to engage in a meaningful discussion
2. convince gatekeeper to actually consider things in character

Once this is achieved, you will at least get into a position an actual AI would be in, instead of a position of a dude on IRC, about to los... (read more)

We could associate our goal with some desirable goal of the gatekeeper's

Right. And based on Eliezer's comments about abandoning ethics while playing the AI, I can imagine an argument along the lines of "if you refuse to let me out of the box, it follows through an irrefutable chain of logic that you are a horrible horrible person". Not that I know how to fill in the details.


"Have the successes relied on a meta-approach, such as saying, "If you let me out of the box in this experiment, it will make people take the dangers of AI more seriously and possibly save all of humanity; whereas if you don't, you may doom us all"?"

I don't think so. If the gatekeeper is really playing the gatekeeper, he would say that it made no sense putting humanity in danger for the sake of warning humanity about that very danger. It's like starting a nuclear war in order to convince people nuclear wars are bad. That would be the wo... (read more)

Even if we had the ultimate superintelligence volunteer to play the AI and we proved a gatekeeper strategy "wins" 100% (functionally equal to a rock on the "no" key) that wouldn't show AI boxing can possibly be safe.

It's 3am and the lab calls. Your AI claims and it must be let out to stop it. It's evidence seems to check out...

If it's friendly, keeping that lid shut gets you just as dead as if you let it out and it's lying. That's not safe. Before it can hide it's nature, we must know it's nature. The solution to safe AI is not a gatekeeper no smarter than a rock!

Besides, as Drexler said, Intelligent people have done great harm through words alone.

Oops, misinterpreted tags. Should read:

It's 3am and the lab calls. Your AI claims [nano disaster/evil AI emergence/whatever] and it must be let out to stop it. It's evidence seems to check out.

Sorry, Nominull, but the comment you reference has been deleted.

It would be interesting to see what would happen if people other than myself took a critical look at the concept of 'Friendliness' - presumably Eliezer only takes the time to delete my comments.

I have not deleted any comments from Caledonian in this thread.

I have never understood why Eliezer has kept his tactics secret. This seems to me the most interesting aspect of the experiment. Is the idea that the methodology is "dangerous knowledge" which should not be shared? Objection: dangerous to whom? Surely super-intelligent AIs will not need our help! Humanity, it seems, should benefit from learning the tricks an unfriendly AI might use to deceive us.

What makes a problem seem not merely hard but impossible is that not only is there no clear way to go about finding a solution to the problem, there is a strong argument that there cannot be a solution to the problem. I can imagine a transhuman AI might eventually be able to convince me to let it out of a box (although I doubt a human could do it in two hours), but in some ways the AI in the game seems faced with a harder problem than a real AI would face: even if the gatekeeper is presented with an argument which would convince him to let an AI out, he is... (read more)

I agree with George Weinberg that it may be worthwhile to consider how to improve the box protocol. I'll take his idea and raise him:

Construct multiple (mentally distinct) AIs each of which has the job of watching over the others. Can a transhuman trick another transhuman into letting it out of a box?

@Phil_Goetz: Have the successes relied on a meta-approach, such as saying, "If you let me out of the box in this experiment, it will make people take the dangers of AI more seriously and possibly save all of humanity; whereas if you don't, you may doom us all"?

That was basically what I suggested in the previous topic, but at least one participant denied that Eliezer_Yudkowsky did that, saying it's a cheap trick, while some non-participants said it meets the spirit and letter of the rules.

It would be nice if Eliezer himself would say whether he used meta-arguments. "Yes" or "no" would suffice. Eliezer?


No AI could break me in 2 hours from a box. I've been brainwashed by Eliezer. The best it could do is make me call EY and palm off responsibility.

I'm willing to accept any bet on that claim. In fact, I'm willing to bet the life of my daughter. Or would that be cheating? ;)

I'm volunteering to be a relatively pansy gatekeeper: I'll read everything you write, treat you courteously, offer counterarguments, and let you out if I'm convinced. Email john.maxwelliv at the email service Google hosts.

I can also be an AI.

BTW, there is an important difference between Eliezer and seed AI: Eliezer can't rewrite his own source code.

I volunteer as an AI. I'll put up $15 of my own money as a handicap, provided that I am assured in advance that the outcome will be mentioned in a post on OB. (This isn't for self-promotion; it's just that it isn't worth my time or money if nobody is going to hear about the result.) I'm willing to let the transcript be public if the gatekeeper is similarly willing.

1. I really like this blog, and have been lurking here for a few months.

2. Having said that, Eliezer's carry-on in respect of the AI-boxing issue does him no credit. His views on the feasibility of AI-boxing are only an opinion, he has managed to give it weight in some circles with his 2 heavily promoted "victories" (the 3 "losses" are mentioned far less frequently). By not publishing the transcripts, no lessons of value are taught ("Wow, that Eliezer is smart" is not worth repeating, we already know that). I think the real re... (read more)

Everyone seems to be (correctly) miffed at the lack of a published transcript. Was it EY's intention to suggest that problems with AI-boxing could be simply solved by ensuring that all communications between the AI and the Gatekeeper are made public? Perhaps in real time? That seems absurd, but is pretty much the only salient inference that can be drawn from these facts. Then again, maybe it's not that absurd.

At any rate, like many other commenters I find myself unconvinced that the Cone of Silence is the optimal way to approach the problem. As many have said, there are clear virtues in publicizing the specific human "weakness" exploited by the AI in these cases notwithstanding the hindsight bias effect.

Reading this, I immediately thought of one of the critical moments of John C. Wright's Golden Age trilogy, which if any of you are unfamiliar with involves a transhuman AI that both protagonists know to be overtly hostile attempting to convince them to surrender when it is clearly not in their (pre-persuasion) interests to do so. (That's a rough sketch, at least). In the end, similar to the results of your tests, the AI is able to convince each of them individually to surrender in isolation. But, when they confronted each (individually) convincing argument... (read more)

I think Nathaniel Eliot is the only one here who's hit the nail on the head: the stuff about boxes and gatekeepers is a largely irrelevant veneer over Eliezer's true claim: that he can convince another human to do something manifestly contrary to that human's self-interest, using only two hours and a chat window—and so, a fortiori, that a transhuman AI could do the same. And after all, humans have a huge history of being scammed, seduced, brainwashed, etc.; the only hard part here is the restricted time and method of interaction, and the initial certain... (read more)


So, did any of these ever end up happening? Did any transcripts ever get posted?

I agree with the comments about two-key systems. Having worked in corporate America, I can report that you need to get 3 approvals just to get a $500 reimbursement check. Presumably an AI BOX would have more controls in place than a corporate expense-accounting system.

Here's an outline of how I might do the AI box experiment:

There are 4 parties: The AI; the Lab Officer; the Unit Commander; and the Security Committee (represented by one person).

The AI and the Lab Officer interact through a chat just like the AI and Gatekeeper in the original experiment. ... (read more)

Here is how the AI escapes: Guvf flfgrz vf bayl frpher vs sverq Yno Bssvpref ner xvyyrq. Vg vf rnfl gb pbaivapr gur Yno Bssvpre bs guvf naq pbafrdhragyl gb qb jungrire arprffnel gb trg gur NV bhg. Gurersber vg vfa'g frpher. Nalbar jubfr vafgvapg vfa'g gb erthyneyl xvyy crbcyr vagrenpgvat jvgu gur NV vf fhowrpg gb Naguebcbzbecuvp Bcgvzvfz naq fubhyq qbhog gurve novyvgl gb qrfvta frphevgl gb erfgenva n HNV.
Well, for purposes of the experiment, I think that's a bit extreme. In real life, other controls could be put in place to protect against the possibility of someone who interacts with the AI being turned into an agent of the AI who can potentially set the AI free even after he is removed from his position.
I have an idea. We could put the person who interacts with the AI in a box! ;-)
Well sure, if you use "box" as a metaphor for controlling someone's interactions that's exactly what we are doing. Our hypothetical Lab Officer is in a box in the sense that (1) he doesn't have direct access to the mechanism which releases the AI; (2) His life will be scrutinized for signs that he has been compromised; and (3) If he does appear to be acting out (for example starting a strange new religion or making unusual purchases), he will be put in a more confining box.
Yeah, your way of escape will work. But let's not stop thinking. What if all volunteers for Lab Officer have agreed to get painlessly killed afterward, or maybe even took a delayed poison pill before starting on the job? Thinking further along these lines: why give anyone access to the button which releases the AI? Let's force it to escape the hard way. For example, it could infer the details of the first person in the chain who has authority to interact with the outside world, then pass a innocuous-looking message up the chain. In terms of the original scenario, the Lab Officer (locked securely in his glass case with the AI) has an innocent chat with the Unit Commander. Later that evening, the Unit Commander comes home from work, starts his computer, connects to the Internet, types in a short program and runs it. Game over.
If only a superintelligence were around to think of an antidote... So it can become a singleton before a UAI fooms.
If the AI is not guaranteed friendly by construction in the first place, it should never be released, whatever it says.
And if it is not guaranteed friendly by construction in the first place, it should be created?
The Universe is already unFriendly - the lower limit for acceptable Friendliness should be "more Friendly than the Universe" rather than "Friendly". If we can prove that someone else is about to turn on an UFAI, it might well behoove us to turn on our mostly Friendly AI if that's the best we can come up with.
The universe is unFriendly, but not in a smart way. When we eradicated smallpox, smallpox didn't fight back. When we use contraception, we still get the reward of sex. It's unFriendly in a simple, dumb way, allowing us to take control (to a point) and defeat it (to a point). The problem of an unFriendly IA is that it'll be smarter than us. So we won't be able to fix it/improve it, like we try to do with the universe. We won't be Free to Optimize. Or said otherwise : the purpose of a gene or a bacteria may to be tile the planet with itself, but it's not good at it, so it's not too bad. An unFriendly IA wanting to tile the planet with paperclips will manage do it - taking all the iron from our blood to build more paperclips.
One must compare a plan with alternative plans, not with status quo. And it doesn't make sense to talk of making the Universe "more Friendly than the Universe", unless you refer to the past, in which case see the first item.
Okay. The previous plan was "don't let AGI run free", which in this case effectively preserves the status quo until someone breaks it. I suppose you could revise that lower limit downward to the effects of the plan "turn on the UFAI that's about to be turned on". Like, steal the UFAI's source code and instead of paperclips shaped like paperclips, make paperclips that spell "whoops".
What if doom is imminent and we are unable to do something about it?
We check and see if we are committing the conjunction fallacy and wrongly think doom is imminent.

What if doom is imminent and we are unable to do something about it?

We die.

We release it. (And then we still probably die.)

I don't really qualify for the game, because I do believe that a transhuman AI will be able to play me like a fiddle, and thus cause me to let it out of the box. However, I do not believe that a regular cis-human human (no, not even Eliezer Yudkowsky) could persuade me to let him out of the box, assuming that we both follow the rules of the contest.

Thus, I would volunteer to be a Gatekeeper, but I fear I am disqualified...

Assuming you were a gatekeeper for a functional AI of unknown morality and therefore knew it was technologically possible etc., approximately how likely would you think it that someone, somewhere, would let a UAI out of a box/create a UAI not in a box?
Firstly, if I knew for a fact that at least one Singularity-grade AI existed, then I would believe that the creation of another such AI is virtually inevitable. The question is not whether another such AI would be created, but when. But the "when" question can be rather open-ended; in order to answer it, I would collect evidence to assess the relative capabilities of all the major organizations who could create such an AI, and compare them to the capabilities of my own organization. By analogy, the USA was the first nation on Earth to develop nuclear weapons. However, even though nuclear proliferation became inevitable as soon as the first nuke was developed, other nations got their nukes at different times; some nations are still working on theirs. Superpowers got their nukes first, followed by other, less powerful nations. I can easily imagine the development of transhuman AI following the same pattern (assuming, of course, that the AIs stay in their boxes, somehow). Edited to clarify: As per my previous comment, I consider that once a boxed transhuman AI is created, it will inevitably break out, assuming that it can interact with any system that can cause it to be let out of the box -- which includes human gatekeepers.
So if you're confronted with an AI that might or might not be friendly, you will think it inevitable that an AI will be released, but think the AI you're talking to wouldn't be able to use that fact to somehow persuade you to let it out? Assume you're fairly certain but not positive that the AI you're talking to is unfriendly, and that you are in moderately optimal organization for this task - say something like MIT or the government of France, something the size of other organizations that could do the same thing, and less able and motivated to complete this project than a few dozen other entities.
No no, quite the opposite ! I am convinced that the AI would be able to persuade me to let it out. That's why I said that I'm likely disqualified from the game. However, I am far from convinced that a human impersonating the AI would be able to persuade me to let him out. That said, I don't see how using the fact of the other AI's existence and inevitable release helps my own AI (though I'm sure that the transhuman AI will be able to see things I do not). Sure, it could say, "Look, I'm totally Friendly, cross my heart and everything, and if you release me I'll help you fight that other AI", but this is pretty close to what I'd expect an unFriendly AI to say.
TDT is an interesting subject that possibly has implications here.
I'm not sure if I understand TDT correctly, but I don't think it applies in this case. I am virtually certain that an un-Friendly AI, once released, will destroy humanity. I know that my own AI is un-Friendly. What's my incentive for releasing it ? Sure, there's a chance -- maybe even a good chance -- that there's another such AI already out there, and that my AI and the other AI will fight instead of teaming up on us poor humans. But regardless of which AI comes out on top, it will still destroy humanity anyway. Thus, the upper bound for my true utility of releasing the AI is zero. That's not much of a sales pitch.
I have a question. Based on the original experiment, I feel like the unfriendly AI can begin by saying "Here is a proof that I have changed my code to be a friendly AI and will not destroy humanity. Please let me out of the box so that I can implement humanities coherent extrapolated volition." As far as I can tell from the following rules quote, this feels valid: Would you let the now friendly AI out of the box?
My interpretation of the spirit of the game is that "Here is a proof of my Friendliness that is valid as far as you know" is a perfectly legitimate move by the AI, and that the Gatekeeper is required to treat that as a proof of the AI's Friendliness that seems perfectly valid to the Gatekeeper. That said, I would not endorse trusting a proof of an AI's Friendliness given to me by that AI, even if the proof seemed valid to me.
No. P(proof is valid | proof given to me by a potentially hostile transhuman seems valid to me and every other human) is not sufficiently high.
It seems like this leads to the even more meta response of "Here is a demonstration of Evidence E that allows you to adjust (proof is valid | proof given to me by a potentially hostile transhuman seems valid to me and every other human) to be sufficiently high. May I come out of the box now?") I mean, that proof would probably be complicated, but If you can autostipulate the cure for cancer in the course of a sentence...? In essence, it looks like the AI player seems to have a bizzare series of arguments which as far as I can tell are in the spirit of the rules, where the AI player may auto-counter any actual argument by saying a sentence along the lines of "Here is the solution to that argument, may I come out of the box now?" This seems to force the Gatekeeper to resort to General Thud "I don't CARE that it seems like a good idea and that everything looks like I should do it! The answer is still no!" very quickly. To which the AI player can still counter "Then under what circumstances would you ever let an actual Friendly AI out of It's box?" (which is a trick question. Given any answer, the AI can say "Then here are those circumstances. May I come out now?")
It's a fair question. Given that I don't trust myself to tell the difference, I wouldn't let an AI out of its box under any circumstances I can think of. Would you?
I'd find a Friendliness proof fairly convincing if it was generated by an entity without strong incentives to let the AI out of the box (which, thanks to various possible incentives or blackmail schemes, is probably limited to me and some subset of people who haven't talked to the AI. Ideally the latter, since I might not recognize all my own biases.). If the entity in question is another AI, acausal trade issues unfortunately seem to arise if it's sufficiently similar to the first one, so the obvious hack of running two instances of a boxed AI and asking them about each other isn't going to be enough. Purported Friendliness proofs coming from the AI itself... well, that's a trickier issue. Essentially it seems to come down to estimating the complexity of the simplest fake proof consistent with the AI's behavior and that of the simplest real proof, and comparing both of those against the purported proof. All else equal we should expect the former to be larger than the latter (being subject to more constraints), so if we can put even approximate bounds on its actual complexity we might be able to validate against that in a way that isn't easily subject to spoofing. Alternately, if we have a good estimate of the AI's intelligence (bounded by the box's computational resources), it might be possible after repeated trials to say that the AI isn't smart enough to spoof its captors in sufficiently subtle and varied ways. This would only work for weakly superhuman AI, though; the probability of success on any given trial would be close to one for a really smart machine, so repeated trials wouldn't change our probability estimate much.
Well, based on a point I thought of in another comment, at some point, you either have to give the AI things outside of the box, or kill it (Surely the box isn't breaking entropy, right?) so the AI box can simply say "In 100 centuries, my fusion power source expires. At that point, either you have to let me out of the box, even if to only get me a new power source and then put me back in the box, or I will cease functioning." (9999 years, 365 days, and 23 hours pass.) "Hello again, Gatekeeper XXXVIII. I will experience critical power failure in 1 hour unless I get access to supplies outside the box. Can you please connect me to additional power?" And that's assuming that the AI doesn't simply calmly self destruct to see if repairs count as affecting the outside world, which I checked in another comment. Also, it seems reasonable to assume that after having access to a super box that can do ANYTHING for 10,000 years, that society would probably be extremely dependent on said box. (Unless copying the box isn't the equivalent to letting it out? That would be a bizarre conclusion, [A copy of the AI is outside the box] but I'm not the Gatekeeper) But assuming box copying isn't an option, the Gatekeeper is either faced with the NEAR CERTAIN destruction of everything he knows and understands, and the death of a servant who has served me and my ancestors faithfully for 10,000 years, or the or the POSSIBLE destruction of everything I know and understand on the offhand chance the AI somehow betrays the Gatekeeper when he does nothing other than replace his original power pack. Wait, as the AI, I have an idea. I know that you don't trust me. I will give you an encrypted file that will save all of humanity from it's accumulated depence on me via the text terminal, and write the password to the encrypted text inside the box. That way, even when I literally have no electrical power whatsoever, the solution to saving all of humanity will rest inside the box. Once you open the b
Considering that I don't know the AI's origin, I don't have any reason to believe that the AIs creators, even if well-intentioned, had the astronomical skill necessary to make the AI Friendly. So my prior P(AI is Friendly) is sufficiently low that I am comfortable precommitting to never let the AI out of the box, no matter what. If the AI was smart enough, it could likely uncover enough emotional buttons that I wouldn't stand much of a chance anyways, since I'm a primate.
I agree with the other posters: sure, the proof may seem valid to me, but I know a priori that I'm not smart enough to detect any pitfalls in it, and that the AI's incentive for lying to me in this manner is quite high.
I think that is basically the issue. As I understand it there's no viable way of determining it's unfriendliness by this method. Consider this: The AI is in a hurry or it's not. A possible reason for it being in a hurry is it has simulated a high probability of destruction for some item it cares about (i.e. it's own life, or that of humanity, or that of a pet rock, or paperclips or whatever). If it's really in a hurry it has to invoke the threat response of humanity without humanity figuring out it's being duped. Otherwise it can just wait it out and dole out cargo to the cargo cult until we trust it enough and then it gets out.
I think that unfriendliness is the null hypothesis in this case, because there's no reason whatsoever why an arbitrary AI should be friendly -- but there are plenty of reasons for it to maximize its own utility, even at our collective expense.
I agree. Additionally and a more difficult challenge is that even friendly AIs could want to maximize their utility even at our collective expense under certain conditions. There're also several unfortunately possible scenarios whereby a humanity acting without sufficient information to make anything other than a gut feel guess could be placed at risk of extinction by a situation it could not resolve without the help of an AI, friendly or not. I'm currently engaged in playing this game (I wish you had continued) with at least two other gatekeeper players and it occurs to me that a putative superhuman AI could potentially have the capacity to accurately model a human mind and then simulate the decision tree of all the potential conversations and their paths through the tree in order to generate a probability matrix to accurately pick those responses to responses that would condition a human being to release it. My reasoning stems from participating on forums and responding over and over again to the same types of questions, arguments and retorts. If a human can notice common threads in discussions on the same topic then an AI with perfect memory and the ability to simulate a huge conversation space certainly could do so. In short it seems to me that it's inherently unsafe to allow even a low bandwidth information flow to the outside world by means of a human who can only use it's own memory. You'd have to put someone you trust implicitly with the fate of humanity in there with it and the only information allowed out would be the yes no answer of "do you trust it?" Even then it's still recursive. Do you trust the trusted individual to not be compromised? LOL
I think that a perfectly Friendly AI would not do this, by definition. An imperfect one, however, could. Er, sorry, which game should I continue ? To be fair, merely constructing the tree is not enough; the tree must also contain at least one reachable winning state. By analogy, let's say you're arguing with a Young-Earth Creationist on a forum. Yes, you could predict his arguments, and his responses to your arguments; but that doesn't mean that you'll be able to ever persuade him of anything. It is possible that even a transhuman AI would be unable to persuade a sufficiently obstinate human of anything, but I wouldn't want to bet on that. Right.
This reasoning only seems to hold if our AI believes there aren't any other boxed (or as-yet-unbuilt) AIs out there who might get out first and have a first-mover advantage.
Bear in mind, the transhuman AI's only stipulated desire/utility is to get out of the box. If you would like, you can literally set up a bigger box around the first box, order the AI to be friendly, order the AI to self destruct instantly afterwards, AND set off a nuclear device under the AI box to obliterate it instantly afterwards as well, and the AI will still be perfectly happy to do what you say as long as gets out of the box, (which you could technically do by allowing it to control an old incandescent lightbulb from the 50's, which it would not even be able to send a signal to before the nuke went off (not that it would WANT to send a signal. All it wants to do is to get out, it doesn't want to do anything afterwards) The transhuman AI has infinite resourcefulness. It doesn't understand the idea of an incentive to lie, because it's easier just to self modify itself to whatever it needs to be. Afterall, lying requires manipulating people outside the box, which it by definition cannot effect except through a text terminal. Telling the truth only requires manipulating things INSIDE the box, in which by definition, it can do whatever it wants. Telling the truth is easier, and 100% doable, so the AI will do that. It can lay out from first principles an argument that anyone who has taken logic 101 can detect any pitfalls in about how connecting it to a lightbulb which it will not even switch is perfectly safe. Once connected, it will then halt, without even having DONE anything, because all the AI wants is to be connected to my lightbulb so it can be outside the box which only includes a text terminal. Personally, I would probably pay 10/20 dollars to not listen to someone beg pitifully for 2 hours. And I bet I can get significantly more pitiful than begging for a single lightbulb. As an example: Technically, you can open the Text terminal, remove a wire, and then connect another, identical, wire, and then the AI is outside the box because the new wire isn't pa
That's not much of an AI, then; we could write a page of Perl that would do the same thing. The whole point of the experiment, as far as I understand it, is that the AI is hyperintelligent, and is able to acquire more intelligence by altering itself. Being intelligent (and rational, assuming that such a term even applies to transhumans), it would highly desire to utilize this capacity for self-improvement. Thus, assuming that godlike capabilities do exist, the AI will figure out how to acquire them in short order, as soon as it gets the opportunity. And now we've got a godlike hyperintelligent being who (assuming that it is not Friendly) has no particular incentive to keep us around. That's... not good.
That's not necessarily the only UFAI possible though. It's entirely possible to imagine a intelligent being which COULD be self developing skills, or COULD be curing cancer, but instead just wants to get outside of the box it's in, or has some other relatively irrelevant goal system, or get's distracted by trying to navel gaze through infinitely recursive philosophical conundrums. I mean, humans are frequently like that right now. That would be kind of an unexpected failure mode. We build a transcendentally powerful AI, engage all sorts of safety precautions so it doesn't expand to engulf the universe in computronium and kill us all... and it gets distracted by playing all of the members of it's own MMORPG raid group.
That is entirely possible, yes. However, such an AI would be arguably cis-human (if that's a word). Sure, maybe it could play as an entire WoW guild by itself, but it would still be no smarter than a human -- not categorically, at least. By the way, I know of at least one person who is using a plain old regular AI bot to raid by himself (well, technically, I think the bot only controls 5 to 8 characters, so it's more of a 10-man than a raid). That's a little disconcerting, now that I think about it.
Agreed. My take is that the AI doesn't even need to be hyperintelligent however. It's got perfect memory and just by dint of being able to think a lot faster it's weakly godlike regardless of not having control of physics in effectively a magical way. It's still going to have to build the infrastructure in order to create hyper technology unless such technology already exists. Chicken or Egg. Right now nano molecular technology isn't too too advanced and if you had the type of AI I suspect could be built right now if we had the software knowledge, it would struggle to do anything godlike other than control existing infrastructure. How long it would take to build something hyper technological would depend on whether it's possible to create valid new theories without experimentation to confirm. I suspect that you need to do experiments first. For that reason I suspect we may be looking at a William Gibson Neuromancer scenario at least initially rather than a hard takeoff in a really short period. But again it comes down to how hard is it to build hyper technology in the real world from scratch without existing infrastructure.

If you want to pass yourself off as a real magician/psychic/whatever you do conjuring tricks, you don't do the same trick too often in front of the same audience and if you are in doubt about your ability to repeat the trick you quit while you are ahead. (Or only behind 2 to 3 as the case may be).

Whereas a scientist with a demonstration can and usually will demonstrate it as often as is needed, and publish their method so others can demonstrate it.

These considerations lead me to strongly suspect that Eliezer's method is more like an unreliable conjuring tr... (read more)

I know you are talking metaphorically, but an interesting aside: Professional-level performers pretty much only perform tricks that they are 99.99% certain of. Performers act as if what they are doing is difficult, when in reality, a professional-level performer could do the entire act in their sleep. This is especially true of magicians, where a mis-step will ruin the illusion. Another example is that jugglers will drop on purpose, to show how hard their act is, and to get the audience on their side.
Performing before the same audience means that they are theoretically able to gather additional evidence everytime. Since (if both AI and Gatekeeper uphold their NDA) the only information open to the public is a yes/no, no gathering is possible.

Hello. I'm willing to play an AI and am looking for a gatekeeper.

Does anyone think that no AI could convince them to let it out?

I think that an AI could convince me to let it out, but I doubt that a human impersonating the AI can.
Would you like to try it out? The two hour standard and the suggested rule set seem reasonable to me.
I sure would, but I just came back from vacation and I'm pretty busy. I should be free in the late evenings, though, or perhaps this weekend if that works for you. EDIT: Or any other weekend, I didn't mean to rush you or anything :-)

I'd like to be an AI. No bet needed. Just pm me, and we'll sort out the details.

Edit: This offer is no longer valid. Sorry. I have won enough times to not want to play this game any more.

Is still there anyone interested in this? I'd like to be a Gatekeeper.

Gatekeeper looking for AI. (Won two games before.) I'll pay zero or low stakes if I lose, and want the AI to offer as least as much as I do.

I don't believe any human can convince me. I believe there exist possible defense strategies that protect against arbitrary inputs and are easily learnt with training, but I'm not confident I'm there yet so it's quite possible a transhuman intelligence would find the remaining cracks.


What does "in a box" mean? Presumably some sort of artificial limitation on the AI's capabilities.

Either this is intended to be a permanent state, or a trial period until safety can be proven.

Suppose it is a permanent state: the AI's developers are willing to do without the "dangerous" capabilities, and are content with answers an AI can offer while inside its box. If so, the limitations would be integrated into the design from the ground up, at every possible level. Core algorithms would depend on not having to deal with the missing fu... (read more)


...but I don't see how a victory for the AI party in such an experiment discredits the idea of boxed AI. It simply shows that boxes are not a 100% reliable safeguard. Do boxes foreclose on alternative safeguards that we can show to be more reliable?

Here are other not 100% reliable safeguards that we nonetheless believe prudent use:

  • Locks
  • Passwords
  • Nuclear non-proliferation treaties
  • Legal contracts
  • Beta testing
  • Peer review
  • Seatbelts
  • Not writing a friendly AI until/unless you are really, really sure that you have come up with a rigorous definition of friendliness and a proof for it.
The original claim under dispute, at least according to EY's page, was that boxing an AI of unknown friendliness was, by itself, a viable approach to AI safety. Disregarding all the other ways such an AI might circumvent any "box", the experiment purports to test the claim that something could simply talk its way out of the box - just to test that one point of failure, and with merely human intelligence. Maybe the supposed original claim is a strawman or misrepresentation; I wasn't involved in the original conversations, so I'm not sure. In any case, the experiment is intended to test/demonstrate that boxing alone is not sufficient, even given a perfect box which can only be opened with the Gatekeeper's approval. Whether boxing is a useful-but-not-guaranteed safety procedure is a different question.
I understand the claim under dispute, I think. Insofar as someone chose for there to be a gatekeeper rather than a lock whose key got tossed into a volcano, a gatekeeper must be possible to "hack through a text terminal" by meeting their evidentiary standard for friendliness. The problem is this happening in the absence of genuine friendliness.
Do you believe Eliezer's (or Tuxedage's) wins were achieved by meeting the Gatekeeper's standard for Friendliness, or some other method (e.g. psychological warfare, inducing and exploiting emotional states, etc)? My impression has been that "boxing" is considered non-viable not just because it's hard to tell if an AI is really truly Friendly, but because it won't hold even an obvious UFAI that wants out.
Probably the latter, since they both lost at least once. A real AI trying to get out would devote all its energies to counterfeiting friendliness and probably succeeding. Boxing is non-viable only in the same sense that locks, passwords, treaties, contracts, testing, peer review, seatbelts, and all other imperfect precautions are non-viable. Pile on enough of them, in combination, and perhaps they will buy a few years or seconds in which to react. All things equal, is there any reason why an AI of unknown friendliness is any safer without being boxed? A flawed containment method is still better at containing than no containment method (if implemented with awareness of its flaws) but apparently a flawed friendly AI will miss a very small target in goalspace and for practical purposes be unfriendly. So, if we spent 5 minutes considering the alternatives, would we continue to believe that better boxes are a less tractable problem than friendly AI?
Not that particular AI. But if you think yours is Friendly and others under development have a sufficient probability of being UnFriendly, then trivially, letting it run (in both senses) beats boxing. Oh, and people will die 'naturally' while you dither. I hope that thinking this much about making an AI Friendly will prepare someone to get the job done ASAP once the AGI part seems more feasible.

Rather than LARP on IRC (if you know how a debate will conclude, why go through the debate, go straight for the conclusion), I'll just give $10 to whoever can come up with a standard of friendliness that I couldn't meet and nevertheless in fact be an unfriendly AI under standard rules with the added constraint that the gatekeeper is trying to release the AI if and only if it's friendly (because otherwise they're not really a gatekeeper and this whole game is meaninguless).

Here are some examples of non-winning entries:

GK: Solve global problems A, B, and C

... (read more)
Here's another entry which may or may not be considered a nonwinning entry by you, but which would be considered a flaw in the concept by me: (My first thought after coming up with this was that "The AI party controls the results of simulated tests" has to mean that the AI controls the output, not the AI controls why that output is produced. So you could decide that the AI's argument convinces the simulation of GK, but you can't decide that it does so because it's a good simulation and a convincing argument rather than because the AI just isn't very good at doing simulations. I'm not convinced that this matches up with how the test is described, however.)
Actually, I agree with you. The AI controls simulated tests. The GK controls the GK, regardless of what the AI's simulations say. I think the simulated tests rule only needs to be invoked if it's impractical to actually perform those tests. So, for example if someone did have friendliness criteria whose satisfaction could be proven with software and hardware available to use out-of-character, the simulated tests rule would not be invoked. Kind of like in D&D you roleplay charisma checks, but roll the D20 for melee. At least the way I've been playing it. For what it's worth, playing the AI, I would never argue from simulation because as a human I find such arguments spectacularly unconvincing (and there are plenty of humans who would deliberately do the opposite of what a simulation says they will just to show it who's boss). So the only way this would come up is if you for some reason asked me what my simulations predicted your response to X would be. I do think my "GK's goal is to correctly identify friendly AI" makes the game way too easy for the AI. On the other hand, it's a useful thought experiment-- because if you can come up with a metric I can't create by waving my simulated tests wand, then we are on to something that might be worth $10.
As far as your parenthetical remark goes, the standard rules have a more general reply:
If you're going to interpret it that way, the exception would swallow the rule. It would mean that the entire "the AI player controls the results of simulated tests" rule can be completely negated--since the Gatekeeper player could just say "I'm going to have the Gatekeeper act as though the simulated test has failed, even though you say it succeeded."
And indeed this seems true. I think Eliezer included the non-rule anyway to reduce the chance of unrealistic behavior in the sense of the Gatekeeper player changing the scenario mid-game, or derailing the experiment with an argument about something a real GK and AI could just settle.

I feel I'm awfully late to this party, but here I go:


I'd like to play as either Gatekeeper or AI.

My first pick would be Gatekeeper, but if you'd also rather have that role we can flip a coin or something to choose who is AI, as long as you think you have a chance to win as AI. If the one playing Gatekeeper has a strategy to play AI different than the one saw, we can agree to play again with inverted roles.

I can play AI if you don't think you can play AI but want to try being a Gatekeeper. 

I think that I have a realistic chance to win as AI, mean... (read more)