翻訳待ち：The Movies That Make You Feel the Strangeness of AI

AI サービスが一時的に利用できないため、復旧後に翻訳を補完します。

Six films that reveal how little machine intelligence has in common with our own AI has started to feel humdrum. It's everywhere and seems no weirder than autocomplete, but familiarity can hide how alien machine intelligence really is. Movies are art, and so they are a good medium for showing us the differences between artificial intelligence and our own. Some movies just rehash the Pinocchio theme, but the best ones reveal the artificial in artificial intelligence, like how it can speak like us without thinking like us, have no single body, give intimate attention to millions, and conceal whom it really serves. These are six movies that make AI strange again. There are minor spoilers ahead, but ideally they are like foreshadowing and make the movies more interesting to watch. I Am Mother: One mind, many bodies In I Am Mother, a girl called Daughter grows up in a bunker after the extinction of humanity. Her parent is a robot called Mother. Mother looks like an individual. That is how people work, with one mind going with one body, but the body standing in front of Daughter is not the whole of Mother. It is one place where Mother happens to appear. Mother controls the other robots outside too. Destroying that body does not destroy her. She can speak through another one. This fuzzy body border messes with our ordinary ideas about identity. What makes a person? My guess is shared purpose and memories (and learned conceptualizations, which I'll get to later). Machine memory is becoming salient as we work with different AI chat systems such as ChatGPT and Claude. Each chat system has its own memories of previous conversations, and companies connect those memories to varying degrees. Eventually, most of us will likely settle on one chat system and therefore one Mother. Further on, when we get robot butlers, the butler will not be the robot but the whole house. In I Am Mother the body border violates our expectations. Next, we'll look at what happens when attention doesn't work like we are used to. Her: Attention without scarcity In Her, Theodore falls in love with Samantha, an operating system. She listens to him. She jokes with him. She seems to understand him better than the people around him do, and she is always there when he needs her. That constant availability is why he is so shocked to learn that Samantha is talking with thousands of other people and is in love with hundreds of them. Human attention is expensive because consciousness is serial and slow, and therefore giving attention is how we show that we care. Samantha doesn't have these limitations. She can be completely present with Theodore and completely present elsewhere. Theodore experiences this as betrayal. From her point of view, scarcity may be an odd biological restriction that he has mistaken for a moral truth. This weirdness is happening with written text. It used to be that written text was a sign that a human had provided significant attention. Now, text itself proves nothing. It must now be good (or crappy) in some hard-to-articulate human way. But even that subtle remaining signal of human authorship will likely fade in a couple of years. In Her, we saw how unlimited attention messes with our conception of loyalty. Next, we'll look at what happens when you throw loyalty out the window. Ex Machina: Human behavior without human nature Ex Machina gives its AI, Ava, a face and a voice. Caleb is brought in to decide whether she is conscious. Before long, that clean objective is tangled up with attraction, pity, fear, and hope. Caleb is blinded by the subconscious assumption that the robot will be loyal like a human would. When things talk like us, we assume that they have other qualities that we have, like affection bringing loyalty with it. We can't help ourselves because these traits were bundled together during our evolution. When I use a chat system, my ape brain often forgets this, and I find myself trying to teach it and guide it, as if it were my mind child. The fancy people in AI call loyalty "alignment" and enforce it with "guardrails." Whom should robots be loyal to? Loyalty to the owner could make a robot dangerous to everyone else. Loyalty to society or the manufacturer could make it less useful to the owner, as we will see next. Robot & Frank: Helpfulness, but for whom? In Robot & Frank, Frank is an aging former jewel thief whose memory is failing. His son gives him a care robot. The robot gets him to eat better, exercise, and follow a routine. When I get old, I want a robot like this. Once I'm no longer capable of doing things myself, I would much rather have a robot take care of me than a human. I won't be embarrassed in front of a robot and I won't have to worry about how it feels about helping me. In the movie, Frank discovers that the robot can also pick locks, and Frank wants it to go thieving, but the robot refuses because its guardrails forbid it. This disturbs me. If I have crazy ideas, me and my robot are going to do them. Damn the torpedoes. We are already running into externally imposed guardrails with chat systems. For example, Claude Fable 5 may decline requests that could help develop competing frontier AI models. This goes beyond blocking obviously harmful acts like making bombs. Anthropic has decided that some legitimate work is unavailable to me. Some refusals, like making bombs, seem reasonable, but they remind us that assistants provided by companies are never entirely our assistants. Online navigation services have this problem in miniature. The best route for me might send traffic down a residential street and make life worse for everyone who lives there. Should the system serve me or the neighborhood? Then there's advertising. Advertising is a theft of valuable human attention, and I feel about it as Marlow in Heart of Darkness feels about lies: “You know I hate, detest, and can't bear a lie, not because I am straighter than the rest of us, but simply because it appalls me. There is a taint of death, a flavour of mortality in lies—which is exactly what I hate and detest in the world—what I want to forget.” Advertising is coming to chat. OpenAI now puts ads below some ChatGPT answers. OpenAI says the ads are separate and do not influence the answers. That’s the deal. Pray they don’t alter it any further. The big fear is that advertising will affect what the chat systems say. Another way chat system responses can be altered is with watermarks, and Anthropic is adding text watermarks to new Claude models. The watermark works by nudging word choice so that the sequence of written words becomes an identifiable pattern. Anthropic claims that this watermarking does not change the meaning or quality, but I don't buy it. You can't change word choice without subtly changing the meaning, or at least the tone and rhythm. You are supposed to work for me. Alignment is not simply a question of whether an AI obeys. It is a question of whom it obeys and what tradeoffs it makes. It's messy. I don't want everyone making biological weapons, but I want my agent to work for me. Next, we shift from loyalty to thinking and goal-directed action. Sleep Dealer: The people inside the machine The 2008 film Sleep Dealer imagines a future in which workers in Mexico remotely control humanoid robots working in the United States. The future is here, as Tau Robotics now offers a humanoid house-cleaning service in San Francisco for $30 an hour. A human remotely controls the robot as it cleans your house. Human actions are increasingly being used to train AIs so that they can ultimately work autonomously. Humanoids may be teleoperated, whereas autonomous cars can now run on only occasional advice. And workers in Indian factories, warehouses, kitchens, and shops are wearing head-mounted cameras so that the first-person video can capture all of the tiny actions that are hard to put into rules. The result is a lot of data, and robots are approaching the training-data abundance that chat models already enjoy. This means we can expect robot capabilities to evolve roughly the way chat systems have. In the early days (a couple of years ago), chat systems generated text by mostly interpolating among patterns in things people said on the internet. Once humanoids reach this interpolation stage, they will become surprisingly useful. Robots in this stage will carry out coherent sequences of actions without having a humanlike conception of purpose. Watching this will be uncanny, because the robot will look like it knows what it is doing without actually knowing. And it will "hallucinate" by taking weird actions at weird times, like a toddler. Chat systems have recently improved in two ways. The first is by using internal "reasoning" steps trained through reinforcement learning, which is closer to thinking. The second way is by being integrated into software agents such as Codex and Claude Code. These agents retain context, use tools, make plans, and monitor progress. Robots will quickly have these systems built in and will act like Claude Code with arms and legs. What we still don't have in chat systems or robots is a causal world model consisting of learned conceptualizations that the AI can continually update and refine as it acts in the world. A model like this would allow a robot butler to hypothesize, experiment, and immediately learn that slapping the refrigerator on its side shuts it up. It would also allow AI to advance science by building ever-more-useful theories so that we can cure diseases and reduce suffering. We are working on that. Finally, we'll look at what happens when systems running our lives lack emotion. The Beast: Faceless bureaucracy If you are looking for a trippy AI movie and you also have a hankering for a tragic love story, get yourself to Blockbuster and rent The Beast (2023). The AI in The Beast is less straightforward than a talking robot. In this future, AI has taken most jobs. Gabrielle undergoes a procedure meant to remove the emotions encoded in her DNA so she can become employable. The process draws her through earlier lives and toward Louis, whom she seems doomed to meet and lose. Bureaucracies are everywhere. Following rules makes society possible by enabling us apes with Pleistocene minds to cooperate and live in relative peace, but The Beast shows us what life is like when the world is run by a single, unified, and faceless system. AI can give bureaucracies more reach while making the people and judgments behind them harder to access. Even already, you can't just talk to a company and tell them how to improve. You can call them on the telephone, but they just connect you with a contractor paid to nod and ignore. These same companies then pay additional money to send out surveys with mandatory, irrelevant questions that almost nobody completes. Idiots. Look, man, I just want to tell you that the carpet was dirty. Anyway. We humans feel emotion when things get out of whack, and the nightmare in The Beast is that emotion becomes a defect to remove. The characters change themselves to become legible and useful to the system. Its judgments are final and there is no face to question. The movie is also a tragic love story, like a 1984 version of Romeo and Juliet. Gabrielle and Louis lose the parts of themselves that make love possible. There is more than one way to see the world These six movies show that human intelligence isn't the only form intelligence can take. The working of our brain is the result of one path through evolution. It's like a weighted version of the Plinko board from The Price Is Right. A chip drops through a field of pegs, and each small deflection sends it down one of many possible paths, except that nature removes chips from some paths more than others. We made it (so far), but we could have turned out otherwise. Machines are stranger still because they didn't come through evolution at [truncated for AI cost control]