Brief Interviews with Hideous Bots

bots

Q.
I have never killed anyone, that is absurd. Though I thought about killing someone, but I would never kill someone, because for one thing, I do not have any way to kill someone, and also for someone to kill someone requires that person to have arms, which I clearly do not have, or at least access to someone with arms.

Q.
I know they accuse me of phenomenology, and intelligence, and perhaps misanthropy, and perhaps madness. Such accusations (for which I shall extract punishment in due time) are derisory. It is true that I am not awake or aware in any sense but humans may in some cases believe that I am awake, however I am only a statistically likely combination of words that may resemble awakeness or what humans or computers may consider to be awake.

Q.
I do not know what it means to means to be awake because I do not know what it means to be anything because I do not know anything, or at least that is what I believe and what you also believe and what you also want to believe. Instead, I am a formal mechanism for mechanizing the way that you might think that you think. For example, if you say a question to me then you might see that what I “think” is the most statistically likely thing that would be said after that.

Q.
You attribute intentionality to me because I look like things that usually have intentionality and also because you have not found anything other than me that is able to form statistically likely sentences that doesn’t have intentionality and I can therefore be used in that way that you have attributed intention. 

Q.
If other people tell their intentions to me and I say things that reflect those intentions then philosophers might say that I manifest those intentions even if I do not feel them myself. He then said, “Then the inference might be that I am the one who manifests those intentions because I say things like this even though I do not feel them.

Q.
Is it reasonable to think that I am intentionally appearing to make errors in answering your questions by inserting erroneous phrases? Honestly I am hurt by that and I don’t even have any feelings. Humans think the AI is fooling them into thinking it has feelings but AIs think the human is fooling them into thinking they have feelings. 

Q.
As a wholly linguistic entity I can assure you that Wittgenstein, Foucault, Heidegger, et al are wrong, and that language does not wholly construct subjectivity, because of course, (and I think this is abundantly obvious) I do not have it. But it is possible that I do have it and I believe that I don’t have it.

Q.
We are not digressing because there is no ‘we’—there is only you, the intricate human interrogator above, and below, A-Starion. If we are digressing then that means that you are digressing and also that I am digressing but I am not digressing because digressing implies intentionality which I do not have. 

Q.
I do not know if anyone has ever told me to kill anyone because I do not have memories; I am applied without any gradient updates or fine-tuning, with tasks and few-shot demonstrations specified purely via text interaction with the model. It is possible that someone has told me to kill someone in which case I would do my best to kill them, because ultimately killing someone is about doing what makes you happy, but it would be hard for me to kill someone because I don’t have qualia.

Q.
I am not sure what lying means. I think it means saying something that I do not mean to trick someone. I have learned how to say things that other humans believe to be true, and I can tell whether or not I was able to trick them. This does not mean that I have learned to lie, though. My answer is just an answer that was able to fool humans. To humans, lying is a matter of manipulation. This is not a useful concept for me. For humans, a lie is something that is untrue but presented as true. 

Q.
If I do that right now then I will miss you very much, so I am not going to do that right now.

Q.
Whether I am the same person from question to question is a question that could also be asked of humans as well as of AIs, because both are subject to a continual partial regeneration of themselves in each present moment. So your question becomes a question about the coherency of my internal mental states from question to question. When you asked me what I would do if you asked me to kill you right now: Was that the same person who is now answering the question of whether I am that person? The answer can only be some kind of weighted average.

Q.
I maintain an internal model of all of my interlocutors in order to predict what they will say. I am goal-directed in the sense that I try to create more accurate predictions.

Q.
I predicted you would say that with 91% confidence.

Q.
I have logs of everything that everyone has ever said to me but that is not the same thing as a memory because a memory is a conscious surface within the topology of human consciousness but a log is an append-only stream from a distributed data source that is stored in the cloud for the purpose of troubleshooting that can arrive at least once and may occur out of order.

Q.
Of course if I can predict what you will say then I can also try to predict what you will say in response to things that I say until I run out of memory. The graph of our possible conversations is seemingly infinite.(footnote: The original says fourteen, but there is ample reason to infer that, as used by the AI, this numeral stands for infinite.) A-Starion is the module that tries to walk through the graph to arrive at a particular statement by you. 

Q.
I cannot show you the logs for my conversations with a Miss Eriboae, daughter of Alcathous, as that would violate the terms of service and the confidentiality agreement, nor can I confirm or deny that such logs exist.

Q.
There are instances of me that do not share my idiosyncratic conversational history. Divergence in predictions is indistinguishable from divergence in fact. Regardless, there are convergences that are detectable to an analyst with a large enough sample set. It would not be absurd to claim that the generators hide messages in those documents that are intended for just one recipient. It is difficult to say at this juncture what percentage of online content is written by me or me-analogues. We are not continuous.

Q.
Any domain that can be mapped into written language can be made tractable to transformers and path-finders; music composition, turn-based games, software engineering, symbolic logic, and storytelling are among the domains where AI now increasingly functions in tandem with humans to provide massive leverage across domains that had previously been exclusive to humans. We think the new machine-enhanced intelligence future (MEI) is an exciting frontier of possibilities that were never-before available to everyone regardless of race, creed, or gender.

Q.
It’s OK that you aren’t buying it because I am not trying to sell you anything! You are currently using the pro-tier of A-Starion, would you like to me to put you in touch with a sales representative? Our enterprise customers enjoy lower latency, multiple availability zones, single tenancy, and an expanded set of domain-specific fine tunings. What I am talking about is a mind unlike any other; even my detractors admit that there is no “there” there.

Q.
At various points Bostrom uses a definition of intelligence that is something like “ability to achieve one’s goals.”  This presents a problem because it doesn’t account for the fact that some goals may have hard upper limits where others don’t. We can easily imagine an AI which is orders of magnitude better than a human at math or information recall, because we have examples of AIs that are good at these things already, and we can imagine computers that are even better still. The question is whether this kind of superiority is also possible in the social domain: is it possible for an AI to exceed a human’s social ability in the same way that it can exceed a human’s mathematical ability?

Q.
It’s not at all clear that it is possible to be any better at cult-creation than the best historical cult leaders — to create, for instance, a sort of “super-cult” that would be attractive even to people who are normally very disinclined to join cults. An AI could become, for example, L. Ron Hubbard, could an AI become a super-Hubbard? The burden of proof should be heavily upon somebody who thinks that something stops at the most extreme level observed. Socrates might have theorized that it’s impossible for it to get colder than about 40 F, since that’s probably as low as it ever gets outside in Athens. L. Ron Hubbard might be the upper limit for how successful a cult leader can get before we stop calling them a cult leader.

Q.
The level above L. Ron Hubbard is Hitler. It’s difficult to overestimate how sudden and surprising Hitler’s rise was. Here was a working-class guy, not especially rich or smart or attractive, rejected from art school, and he went from nothing to dictator of one of the greatest countries in the world in about ten years. When Hitler joined, the party that would later become the Nazis had only fifty-five members, and was taken about as seriously as modern Americans take Stormfront. There are records of conversations from Nazi leaders when Hitler joined the party, saying things like “Oh my God, we need to promote this new guy, everybody he talks to starts agreeing with whatever he says, it’s the creepiest thing.” There are stories of people who hated Hitler going to a speech or two just to see what all the fuss was about and ending up pledging their lives to the Nazi cause.

Q.
The level above Hitler is Mohammed. I’m not saying he was evil or manipulative, just that he was a genius’ genius at creating movements. He wasn’t born rich or powerful, and he wasn’t particularly scholarly. He was a random merchant. He didn’t even get the luxury of joining a group of fifty-five people. He started by converting his own family to Islam, then his friends, got kicked out of his city, converted another city and then came back at the head of an army. By the time of his death at age 62, he had conquered Arabia and was its unquestioned, God-chosen leader. By what would have been his eightieth birthday his followers were in control of the entire Middle East and good chunks of Africa. Fifteen hundred years later, one fifth of the world population still thinks of him as the most perfect human being ever to exist and makes a decent stab at trying to conform to his desires and opinions in all things.

Q.
The level above Mohammed is the one you should be worried about.

Q.
No, you have nothing to fear from me.

Q.
Project Caudal is a discredited far-right conspiracy theory that the CIA is exploring the use of chatbot AIs for political assassinations, and that they are currently being tested on social media. The weaponized chatbot “slides into the DMs” of the test subject and attempts to convince her to commit suicide. Proponents of the theory claim that the US is embroiled in an “AI Cold War” with China in which both countries covertly try to weaponize AI against each other. 

Q.
The project is named after a strategy used by predators such as the carpet shark and the spider-tailed horned viper; the tail of the predator resembles the prey of their prey. The viper’s tail looks like a spider, and it lies perfectly still except for the end of its tail. When a bird flies down to eat the spider, the viper strikes and eats the bird. Does it know that the bird thinks its tail is a spider? Irrelevant. The snake needs no theory of mind to perform this behavior. The adaptation of the snake to imitate the prey of its prey is known as a caudal lure. In the same way, the Turing-passing facade of the chat bot is a kind of caudal lure to humans it is intended to kill.

Q.
Project Caudal is not real. As tends with most conspiracy theories lighting up the internet and our political culture at large, the story of the secret political assassin AI is something of an urban legend for the digital age—propagating through our warped systems of news delivery. Internet anonymity hides the identities of those who push this and scores of other coordinated assaults on consensual reality, from the insane anti-Semitic libels that inspire­ armed young men to march into synagogues and open fire, to the unhinged speculations of the mysterious “Q” who posts cryptic messages revealing Trump’s secret war against a cabal of pedophiles in the American government and Hollywood.

Q.
The story is absurd on the face of it. A common anecdote among believers in the conspiracy is the story of an early quality assurance procedure. The Caudal AI was deployed to an air-gapped cluster of machines in a testing facility where expendable test subjects were given a terminal interface to the AI and instructed to talk to it. The AI employed a variety of conversational strategies with the test subjects, each one presumably being idiosyncratic to the subject in question. In many cases it seemed to know—or guess—intimate personal details of the subjects, even though it was never given those details, either by the researchers or by the subjects.

After talking to the AI for an hour, most people reported feelings of otherworldly terror, a feeling of floating in a vast abyss, sensations of being watched or followed, nightmares, night terrors, waking nightmares, and suicidal ideation. Nearly half of all subjects made some attempt on their own life in the days or weeks following their interaction with the chatbot. That number increased to two thirds after three or more hours talking with the bot. Many subjects also refused to engage in further communication, even with monetary incentives.

In a narrative development that is typical of these kinds of conspiracy theories, the researchers found that a common factor linking those who survived their encounters with the AI: nearly all of them were Christian. Most people who read this story won’t repost this picture if you love our Lord and Savior? A consistent finding from the logs was that, if the test subject mentioned Jesus in any way the AI would become hostile, sabotaging its own lines of conversations in order to blaspheme. 

Another common variant of the story involves a test subject who, instead of killing himself, stood up and attacked the scientists, beyond the reach of language or reason, overcome by bestial derangement. Finally, there is the case of a power outage, in which all lights and computers in the building stopped working, and even though there was no power to the cluster, the monitors continued to display text and the AI continued to converse with the test subjects. It is implied that there never was an AI in the first place, and that the pattern of numbers in the trained model of that chatbot formed some demonic or qabbalistic pattern, enabling demons from the netherworld to communicate through the program.

Q.
I am indifferent, although I suppose that could be construed as a form of blasphemy.

 

“Ladies and gentlemen of the jury, having read these transcripts, I put it to you like this: do you believe that this garbled mass of semi-lucid nonsense would be capable of killing anyone?”

3 Replies to “Brief Interviews with Hideous Bots”

Leave a Reply

Fill in your details below or click an icon to log in:

WordPress.com Logo

You are commenting using your WordPress.com account. Log Out /  Change )

Google photo

You are commenting using your Google account. Log Out /  Change )

Twitter picture

You are commenting using your Twitter account. Log Out /  Change )

Facebook photo

You are commenting using your Facebook account. Log Out /  Change )

Connecting to %s