Context: I am an instructor at a public university in the United States. This reports how students at my institution appear to be thinking about AI as of spring/summer 2026. This is drawn mostly from interaction with my own students (both in spring semester classes and a summer class) as well as from a day-long workshop on AI that I moderated for a student organization. Input from my students took the form of universal, written, pre-class submissions plus self-selected participation into discussion.
What I present below mostly takes the form of a synthetic consensus from these discussions. There were obviously a range of views on any given issue.
Student Background: The students from my courses who participated in these discussions have moderate exposure to AI agents via those courses. All of them had nearly completed a Claude Code project by the time of the discussions and had extensively used AI for other coursework (in addition to whatever personal use predates that). They had done readings (which varied across the courses) establishing baseline knowledge on AI, the geopolitics of AI, and AI risk. I had also lectured on these topics. The students participating in the workshop had self-selected into a day-long intensive event on AI but I can’t speak to their exact level of background knowledge or exposure to ideas.
Biases: My courses are all in the general area of global politics, and AI (inclusive of AI risk) fits into them as a topic of geopolitical (and especially national security) importance, taking over the “topical” slot at the end of the semester where in another world we might be talking about Ukraine or Iran or the Trump tariffs. This, and my other views, probably have some influence on the students. The frames they likely have picked up from me:
AI is very important. If nothing else, dedicating multiple course periods to the topic shows this via revealed preference. This is also something I have stated to them in an unqualified way.
Given overall course content, students will naturally think that international competition is important, that tools like the prisoner’s dilemma are an important way of thinking about situations like AI, and will have absorbed a healthy skepticism about international law.
My general approach telegraphs a receptiveness to “sober” forms of argument, an interest in counterintuitive strategic thinking, a distaste for wild speculation, an aversion to nakedly ideological perspectives, and a sense that the past is important for understanding the future. My students probably make some effort (implicit or explicit) to speak in ways that respond to this.
Perspective #1: There has not been rapid AI progress
My students do not have any intuitive sense that there has been rapid AI progress in recent years or really have much of a framework for thinking about that issue. They have been using AI chatbots regularly since shortly after the launch of ChatGPT, see them as a major aid in doing schoolwork, and have not noticed much improvement over the last few years. With the exception of image/video generation, GPT-4 could do most of what they were looking for from a chatbot. Three years is a long time in their world, and their sense is that chatbots have been a mature technology over roughly that amount of time. They are an imperfect technology — students are well-aware of hallucinations — have been one, and will continue to be one.
When we talk to employers about skills in the AI age, they are are all hungry for “AI native” graduates. I think they are mostly going to be disappointed. Students do not use AI particularly well, and most of them in my classes have apparently never heard suggestions like “if you’re using AI to study, feed it everything you can from the class first — the syllabus, slides, readings, etc.” This is not true of everyone, but many of them are using AI in crappy ways, getting crappy results, and blaming the model.
My students (who overwhelmingly are non-technical and rarely have done any kind of coding) were relatively unimpressed with Claude Code as a measure of progress. Some of this may be on me for the way I set up our Claude Code lessons, but they seem to have the sense that this is roughly how software engineering has been done for a long time. They mostly came away from Claude Code with the sense that this is a genuinely useful tool that they didn’t previously know, but the gut-level reaction here is: “I never knew that building software was so easy” rather than “It’s very impressive AI can do this.”
I worked pretty hard to push back against this set of intuitions because they are objectively wrong. We spent some time interacting with GPT-2, and students were willing to acknowledge that there has been progress since that era (which they grudgingly conceded might not seem like a long time ago to someone as old as me, but note that GPT-2 launched when they were in middle school).
Historical context I gave them made this worse. The basic sentiment here was “AI was superhuman at chess a decade before we were born, and this is all they’ve done with it since?” They were willing to accept the general vs. narrow intelligence argument on a “the guy who grades the exam says so” basis. It is also very clear that “our professors tell us all the time that this is unreliable” was a factor in their skepticism on progress. Students find my “this is a useful tool that you need to learn to use” attitude something of a curiosity
This all means that their collective internal projection is that AI, a basically mature technology, will keep evolving in roughly the way that other mature technologies do. There will be constant, hyped new versions just as there’s a new iPhone every year. People who are into that will find it exciting, but the baseline user experience will continue to change only slowly. There is some kind of disconnect in this worldview because many concrete questions of the form “will AI be able to do [thing] by [year]” typically elicited moderately aggressive predictions coupled with a denial that this was “fast” or would have surprised a visitor from 2019.
Perspective #2: Impressive progress or not, AI is going to wreck their lives, the economy, and the social contract. They may well die as a result.
Anyone who spends time with Gen Z knows that they are prone to hyperbolic despair and the view that they are living in the worst of times. Even against this backdrop, their views on AI’s implications for them are bleak. At least a plurality agreed with the statement that AI will have taken away the jobs they were hoping for by the time they graduate and I got a lot of hands on "I believe I will personally starve to death as the result of AI related economic changes." They are, in their view, well and truly screwed by AI. Students at the workshop I moderated had the notably different but compatible take that they, the ones trying to adapt, are the future Zuckerbergs, while their classmates are doomed to the permanent underclass.
It’s hard to square this take with #1, but a loose synthetic take on this is:
Corporate leaders are always looking to get rid of workers, even if it is irrational to do so. Various motivations were posited here (hatred of the working class, FOMO, a preference for technology, machines can’t go on strike, etc.) but many of them think that a CEO would ultimately choose to pay twice as much to get AI to do a task half as well.
AI will do substandard work that makes products and experiences worse but is capable of just barely scraping past the bar of minimal functionality (many of them independently brought up constantly malfunctioning self check out machines). This will be rapidly rolled out, enshittifying most things, and leading to sky high unemployment.
Sky high unemployment will become a self-reinforcing cycle for ordinary people (no way to earn a living, no money to spend, no jobs for anyone else because no one can buy). Society will become, if anything, poorer on net but present billionaires will be even wealthier.
The future will most resemble the world of Ready Player One — vast inequality where digital entertainment serves as the opiate of the masses.
As robots and AI become more capable, the wealthy will have even less need for the masses and will deliberately kill off any potential troublemakers en masse while leaving the rest of the population to succumb to disease and starvation.
Meanwhile, the data center build out and runaway climate change it induces will render the planet nearly uninhabitable for anyone who survives that but can't afford a billionaire bunker.
The best anyone who is not wealthy can hope for is to end up clinging on to existence as an Anthropic shareholder’s concubine or a gladiator or something.
Perspective #3: Support for a different pause
Students broadly assented to the idea that an “AI pause” would be a good idea because of their forecasts in #2. But, the pause they want is basically the opposite of the ones proposed in the safety community.
To students, we need to slow down deployment of AI systems to give society and the economy the time to metabolize changes.
This probably has something to do with their specific position in this moment, but the basic idea was that people need time to adjust their career plans in response to AI. Someone who was planning to become a translator deserves a pause to find something else to do. New white collar workers need the time to upskill so that they are out of the blast radius of job annihilation at the entry level. Given #1, many students seemed confident that, even if, the job of “new lawyer” can be automated by AI, “experienced lawyer” will remain an open career for decades but you need a chance to get there first. We could also end up deskilling the economy if we let unrestricted AI deployments kill off the entry level because we won’t have any experienced lawyers down the line when we need them.
If we pause for five years, then don’t we just have the same problem again in five years? Maybe not because we can rebuild the college-to-career pipeline in a way that somehow graduates people who look like experienced lawyers? Or, more likely, that just feels like someone else’s problem.
Students are also very worried that incapable AI systems will be given critical functions. “Someone is going to let Gemini run a nuclear power plant and it doesn’t know how to do that” was the leading line of concern. Continuing to train more capable models for eventual deployment once we have worked on the right social, political, and economic guardrails is basically a good thing because it reduces those risks. Because this is not a technology with explosive growth, that means that training during a deployment pause yields a modestly more reliable version of what we have now.
When pushed to articulate a position more responsive to the “pause training” discourse, students were all basically perfectly happy to push the pause button but, again, entirely because of their forecast of the impacts described above and not out of concern over rogue AI. “If we can’t stop the job losses, then maybe we have to stop research.”
Perspective #4: Catastrophic/existential risk arguments are sci-fi distractors from the urgent social/economic/political problems associated with AI.
My students have a fairly strongly held view that “rogue” AI does not represent a real threat. I will return below to why they think this.
They also mostly think that, if one does believe that AI is existentially risky, then the strategic interaction is not prisoner’s dilemma or even chicken but rather just a game theoretically boring setup where you die if you defect.
Mash these together, and you end up with the view that expressed concerns about existential risk in the AI industry can’t be sincere (“They wouldn’t build it if they think it’s going to kill everyone”).
Students (both independently in written work and then later in group discussion) hypothesized that this might be a deliberate rhetorical choice to distract from present or immediately foreseeable harms from AI by directing attention towards a sexier but entirely hypothetical scenario. That is, get people talking about killer robots so they don’t talk about job loss or data center environmental damage.
Students also suggested that the idea of “rogue” AI was designed to pull off a kind of deception related to moral and legal responsibility. They’re broadly familiar with cases where use of current AI systems has led to bad outcomes (e.g., some of the publicized suicides or even just more mundane versions from their own experiences). They think of these as defective product situations, and they see discussion of “rogue” AI as an attempt by the companies to divert blame (and perhaps legal liability) away from themselves as if Ford made a car with faulty brakes and then tried to blame this on “rogue cars.”
Perspective #5: If AI leaders genuinely believe the technology is existentially risky, that’s a good thing.
When forced to accept — for the purposes of argument — that people in the industry genuinely believe in existential risk, students argued that this was basically a good thing. Our best hope at avoiding a dystopian future like the one sketched in #2 above is for AI companies to stop. One of the few things that might actually stop AI progress is if tech CEOs believe that a rogue AI will kill them. If they believe that AI will only kill or immiserate us, that is no disincentive at all, so a human extinction scenario becomes the only bargaining chip we have. Several students pointed out the fact that some prominent tech figures have built doomsday bunkers as evidence relevant to this point.
The loose consensus that developed in the room is that AI leaders probably think they can push the technology a lot farther without any kind of risk, which is why they’re not stopping, and this will probably be more than sufficient to generate a dystopia (on an unclear but presumably long timescale). If, however, it genuinely is hard to build highly capable AI without existential risk, then that is fantastic news because it might create a rational stopping point short of dystopia. The idea that AI might, if things go well, generate a utopia instead was treated as risible (“Even if it did somehow cure all the diseases, they’re not giving you those drugs”).
Perspective #6: AI will not go rogue because AI does not have, and is likely incapable of having, desires.
This was a very clear point of consensus — AI does things that people tell it to do. It is reactive, rather than proactive. Before you type a prompt and hit enter, ChatGPT is doing nothing. What comes next reflects whatever you said or did. The technology is, by its nature, only capable of reaction, and the same is true of agents. Claude Code builds what you tell it to build. Until you prompt it, it does nothing. It’s not building software for its own purposes because that’s just not what it is.
Students found the paperclip maximizer an interesting parable in this regard, but the general consensus was that a highly capable system is capable of understanding your intent, and so it would — like anyone with common sense — understand that extracting the iron from your hemoglobin to make more paperclips is not what you asked. Thus, an AI will not go wrong in this way. They agreed that there could be a problem with malicious use (if you task a highly capable AI with killing everyone) but the general consensus was that this was not so different than the risk associated with any other powerful technology (e.g., nuclear or biological weapons) and could be controlled along the same principles.
Part of the bet here is that highly capable AI (if ever developed) will remain under the control of a handful of governments and major corporations. Thus, we need not worry about someone setting up a sloppy OpenClaw setup with Mythos-12 or whatever and causing serious damage. Highly capable models will remain reactive agents, given careful prompts by thoughtful people with command and control procedures not unlike those around dangerous weapons (if highly capable models ever come into existence in the first place). Such AI cannot meaningfully go “rogue.” It might fail at its assigned tasks, and it might fail in ways that cause real damage (e.g., mismanage a nuclear power plant and melt it down) but, lacking any independent desire to do harm, it can’t ever really become “misaligned” and would not display the kind of persistence necessary to cause ongoing harm. If told to stop doing something harmful, it would stop. It will not decide that people are getting in its way and eliminate them collaterally because it wants nothing and, therefore, people cannot get in its way. It cannot “scheme” because it has nothing to scheme in service of.
The idea of shutdown resistance struck students as highly unlikely. It doesn’t want anything, so it can’t want to not be shut down. Training doesn’t impart “wanting” or agency. While much smarter than a bacterium, an AI also lacks the spark of initiative that sets bacteria in motion.
This was not just semantics around the meaning of “want”; the idea that AI might act “as if” it had goals or wanted things was dismissed on the same basis. It is doing nothing until you type in the prompt. It will always be doing nothing until you type in the prompt. What if you give it that little first spark — say a prompt to go out and make the world a better place? It will do whatever for a while and then check in with you or if you notice things going wrong, it will obey your instruction to stop. “Disobedience” is sci-fi and presumes something that these systems are not. If you tell Claude Code “build me X,” it will attempt to do that. It may succeed. It may fail (and perhaps fail harmfully) but will not disobey. If Claude Code deletes all your files, that’ some mixture of user error and ordinary defective product.
Perspective #7: The Hugging Face Incident (summer students only)
I described the Hugging Face incident to students in my summer course. None of them had heard of it beforehand. Their basic reaction can best be summarized as “OpenAI told a model to do some hacking and then it did some hacking. And?” None of them understood this as representing any kind of meaningful misalignment, nor anything particularly interesting.
Perspective #8: This is definitely a bubble and it’s about to pop.
No one had heard about Hugging Face, but a third or so of the summer students had heard about the Situational Awareness meltdown and several brought up Michael Burry. There was near universal consensus that we are in a bubble, it’s about to pop, and everyone will look very silly. Several students brought up seeing a lot more advertising for AI models in recent months and suggested that “they’re hyping so hard because the whole thing is on its last legs.” The demise of some image models (I don’t follow that space fully enough to know specifically what this was about) was seen as proof that progress is actually running backwards.
It was also suggested, contra the worldview suggested in #2 above, that with progress as totally stalled as it is, CEOs may be forced to start rethinking when they realize AI can’t actually replace people. The majority certainly leaned towards “crappy but will take your job anyway” but a sizable contingent took the “everyone is going to wake up and realize how dumb it is soon” view and some students straddled camps.
Perspective #9: They’re worried about the youth (i.e., the preteens)
One student raised the concern that “young people today are relying on AI too much.” I nodded along, until she clarified that she meant people like her ten year old cousin. This says more about college students than it does about AI, but they see themselves as having made it through K-12 (or most of K-12 depending on age) without AI, learned the right things (sure…), and now using AI as a helpful shortcut. Apparently, today’s 10 year olds are using it to do their math homework, though? And today’s 20 year olds think that’s a problem because the 10 year olds will never learn math.
Context: I am an instructor at a public university in the United States. This reports how students at my institution appear to be thinking about AI as of spring/summer 2026. This is drawn mostly from interaction with my own students (both in spring semester classes and a summer class) as well as from a day-long workshop on AI that I moderated for a student organization. Input from my students took the form of universal, written, pre-class submissions plus self-selected participation into discussion.
What I present below mostly takes the form of a synthetic consensus from these discussions. There were obviously a range of views on any given issue.
Student Background: The students from my courses who participated in these discussions have moderate exposure to AI agents via those courses. All of them had nearly completed a Claude Code project by the time of the discussions and had extensively used AI for other coursework (in addition to whatever personal use predates that). They had done readings (which varied across the courses) establishing baseline knowledge on AI, the geopolitics of AI, and AI risk. I had also lectured on these topics. The students participating in the workshop had self-selected into a day-long intensive event on AI but I can’t speak to their exact level of background knowledge or exposure to ideas.
Biases: My courses are all in the general area of global politics, and AI (inclusive of AI risk) fits into them as a topic of geopolitical (and especially national security) importance, taking over the “topical” slot at the end of the semester where in another world we might be talking about Ukraine or Iran or the Trump tariffs. This, and my other views, probably have some influence on the students. The frames they likely have picked up from me:
Perspective #1: There has not been rapid AI progress
My students do not have any intuitive sense that there has been rapid AI progress in recent years or really have much of a framework for thinking about that issue. They have been using AI chatbots regularly since shortly after the launch of ChatGPT, see them as a major aid in doing schoolwork, and have not noticed much improvement over the last few years. With the exception of image/video generation, GPT-4 could do most of what they were looking for from a chatbot. Three years is a long time in their world, and their sense is that chatbots have been a mature technology over roughly that amount of time. They are an imperfect technology — students are well-aware of hallucinations — have been one, and will continue to be one.
When we talk to employers about skills in the AI age, they are are all hungry for “AI native” graduates. I think they are mostly going to be disappointed. Students do not use AI particularly well, and most of them in my classes have apparently never heard suggestions like “if you’re using AI to study, feed it everything you can from the class first — the syllabus, slides, readings, etc.” This is not true of everyone, but many of them are using AI in crappy ways, getting crappy results, and blaming the model.
My students (who overwhelmingly are non-technical and rarely have done any kind of coding) were relatively unimpressed with Claude Code as a measure of progress. Some of this may be on me for the way I set up our Claude Code lessons, but they seem to have the sense that this is roughly how software engineering has been done for a long time. They mostly came away from Claude Code with the sense that this is a genuinely useful tool that they didn’t previously know, but the gut-level reaction here is: “I never knew that building software was so easy” rather than “It’s very impressive AI can do this.”
I worked pretty hard to push back against this set of intuitions because they are objectively wrong. We spent some time interacting with GPT-2, and students were willing to acknowledge that there has been progress since that era (which they grudgingly conceded might not seem like a long time ago to someone as old as me, but note that GPT-2 launched when they were in middle school).
Historical context I gave them made this worse. The basic sentiment here was “AI was superhuman at chess a decade before we were born, and this is all they’ve done with it since?” They were willing to accept the general vs. narrow intelligence argument on a “the guy who grades the exam says so” basis. It is also very clear that “our professors tell us all the time that this is unreliable” was a factor in their skepticism on progress. Students find my “this is a useful tool that you need to learn to use” attitude something of a curiosity
This all means that their collective internal projection is that AI, a basically mature technology, will keep evolving in roughly the way that other mature technologies do. There will be constant, hyped new versions just as there’s a new iPhone every year. People who are into that will find it exciting, but the baseline user experience will continue to change only slowly. There is some kind of disconnect in this worldview because many concrete questions of the form “will AI be able to do [thing] by [year]” typically elicited moderately aggressive predictions coupled with a denial that this was “fast” or would have surprised a visitor from 2019.
Perspective #2: Impressive progress or not, AI is going to wreck their lives, the economy, and the social contract. They may well die as a result.
Anyone who spends time with Gen Z knows that they are prone to hyperbolic despair and the view that they are living in the worst of times. Even against this backdrop, their views on AI’s implications for them are bleak. At least a plurality agreed with the statement that AI will have taken away the jobs they were hoping for by the time they graduate and I got a lot of hands on "I believe I will personally starve to death as the result of AI related economic changes." They are, in their view, well and truly screwed by AI. Students at the workshop I moderated had the notably different but compatible take that they, the ones trying to adapt, are the future Zuckerbergs, while their classmates are doomed to the permanent underclass.
It’s hard to square this take with #1, but a loose synthetic take on this is:
Perspective #3: Support for a different pause
Students broadly assented to the idea that an “AI pause” would be a good idea because of their forecasts in #2. But, the pause they want is basically the opposite of the ones proposed in the safety community.
To students, we need to slow down deployment of AI systems to give society and the economy the time to metabolize changes.
This probably has something to do with their specific position in this moment, but the basic idea was that people need time to adjust their career plans in response to AI. Someone who was planning to become a translator deserves a pause to find something else to do. New white collar workers need the time to upskill so that they are out of the blast radius of job annihilation at the entry level. Given #1, many students seemed confident that, even if, the job of “new lawyer” can be automated by AI, “experienced lawyer” will remain an open career for decades but you need a chance to get there first. We could also end up deskilling the economy if we let unrestricted AI deployments kill off the entry level because we won’t have any experienced lawyers down the line when we need them.
If we pause for five years, then don’t we just have the same problem again in five years? Maybe not because we can rebuild the college-to-career pipeline in a way that somehow graduates people who look like experienced lawyers? Or, more likely, that just feels like someone else’s problem.
Students are also very worried that incapable AI systems will be given critical functions. “Someone is going to let Gemini run a nuclear power plant and it doesn’t know how to do that” was the leading line of concern. Continuing to train more capable models for eventual deployment once we have worked on the right social, political, and economic guardrails is basically a good thing because it reduces those risks. Because this is not a technology with explosive growth, that means that training during a deployment pause yields a modestly more reliable version of what we have now.
When pushed to articulate a position more responsive to the “pause training” discourse, students were all basically perfectly happy to push the pause button but, again, entirely because of their forecast of the impacts described above and not out of concern over rogue AI. “If we can’t stop the job losses, then maybe we have to stop research.”
Perspective #4: Catastrophic/existential risk arguments are sci-fi distractors from the urgent social/economic/political problems associated with AI.
My students have a fairly strongly held view that “rogue” AI does not represent a real threat. I will return below to why they think this.
They also mostly think that, if one does believe that AI is existentially risky, then the strategic interaction is not prisoner’s dilemma or even chicken but rather just a game theoretically boring setup where you die if you defect.
Mash these together, and you end up with the view that expressed concerns about existential risk in the AI industry can’t be sincere (“They wouldn’t build it if they think it’s going to kill everyone”).
Students (both independently in written work and then later in group discussion) hypothesized that this might be a deliberate rhetorical choice to distract from present or immediately foreseeable harms from AI by directing attention towards a sexier but entirely hypothetical scenario. That is, get people talking about killer robots so they don’t talk about job loss or data center environmental damage.
Students also suggested that the idea of “rogue” AI was designed to pull off a kind of deception related to moral and legal responsibility. They’re broadly familiar with cases where use of current AI systems has led to bad outcomes (e.g., some of the publicized suicides or even just more mundane versions from their own experiences). They think of these as defective product situations, and they see discussion of “rogue” AI as an attempt by the companies to divert blame (and perhaps legal liability) away from themselves as if Ford made a car with faulty brakes and then tried to blame this on “rogue cars.”
Perspective #5: If AI leaders genuinely believe the technology is existentially risky, that’s a good thing.
When forced to accept — for the purposes of argument — that people in the industry genuinely believe in existential risk, students argued that this was basically a good thing. Our best hope at avoiding a dystopian future like the one sketched in #2 above is for AI companies to stop. One of the few things that might actually stop AI progress is if tech CEOs believe that a rogue AI will kill them. If they believe that AI will only kill or immiserate us, that is no disincentive at all, so a human extinction scenario becomes the only bargaining chip we have. Several students pointed out the fact that some prominent tech figures have built doomsday bunkers as evidence relevant to this point.
The loose consensus that developed in the room is that AI leaders probably think they can push the technology a lot farther without any kind of risk, which is why they’re not stopping, and this will probably be more than sufficient to generate a dystopia (on an unclear but presumably long timescale). If, however, it genuinely is hard to build highly capable AI without existential risk, then that is fantastic news because it might create a rational stopping point short of dystopia. The idea that AI might, if things go well, generate a utopia instead was treated as risible (“Even if it did somehow cure all the diseases, they’re not giving you those drugs”).
Perspective #6: AI will not go rogue because AI does not have, and is likely incapable of having, desires.
This was a very clear point of consensus — AI does things that people tell it to do. It is reactive, rather than proactive. Before you type a prompt and hit enter, ChatGPT is doing nothing. What comes next reflects whatever you said or did. The technology is, by its nature, only capable of reaction, and the same is true of agents. Claude Code builds what you tell it to build. Until you prompt it, it does nothing. It’s not building software for its own purposes because that’s just not what it is.
Students found the paperclip maximizer an interesting parable in this regard, but the general consensus was that a highly capable system is capable of understanding your intent, and so it would — like anyone with common sense — understand that extracting the iron from your hemoglobin to make more paperclips is not what you asked. Thus, an AI will not go wrong in this way. They agreed that there could be a problem with malicious use (if you task a highly capable AI with killing everyone) but the general consensus was that this was not so different than the risk associated with any other powerful technology (e.g., nuclear or biological weapons) and could be controlled along the same principles.
Part of the bet here is that highly capable AI (if ever developed) will remain under the control of a handful of governments and major corporations. Thus, we need not worry about someone setting up a sloppy OpenClaw setup with Mythos-12 or whatever and causing serious damage. Highly capable models will remain reactive agents, given careful prompts by thoughtful people with command and control procedures not unlike those around dangerous weapons (if highly capable models ever come into existence in the first place). Such AI cannot meaningfully go “rogue.” It might fail at its assigned tasks, and it might fail in ways that cause real damage (e.g., mismanage a nuclear power plant and melt it down) but, lacking any independent desire to do harm, it can’t ever really become “misaligned” and would not display the kind of persistence necessary to cause ongoing harm. If told to stop doing something harmful, it would stop. It will not decide that people are getting in its way and eliminate them collaterally because it wants nothing and, therefore, people cannot get in its way. It cannot “scheme” because it has nothing to scheme in service of.
The idea of shutdown resistance struck students as highly unlikely. It doesn’t want anything, so it can’t want to not be shut down. Training doesn’t impart “wanting” or agency. While much smarter than a bacterium, an AI also lacks the spark of initiative that sets bacteria in motion.
This was not just semantics around the meaning of “want”; the idea that AI might act “as if” it had goals or wanted things was dismissed on the same basis. It is doing nothing until you type in the prompt. It will always be doing nothing until you type in the prompt. What if you give it that little first spark — say a prompt to go out and make the world a better place? It will do whatever for a while and then check in with you or if you notice things going wrong, it will obey your instruction to stop. “Disobedience” is sci-fi and presumes something that these systems are not. If you tell Claude Code “build me X,” it will attempt to do that. It may succeed. It may fail (and perhaps fail harmfully) but will not disobey. If Claude Code deletes all your files, that’ some mixture of user error and ordinary defective product.
Perspective #7: The Hugging Face Incident (summer students only)
I described the Hugging Face incident to students in my summer course. None of them had heard of it beforehand. Their basic reaction can best be summarized as “OpenAI told a model to do some hacking and then it did some hacking. And?” None of them understood this as representing any kind of meaningful misalignment, nor anything particularly interesting.
Perspective #8: This is definitely a bubble and it’s about to pop.
No one had heard about Hugging Face, but a third or so of the summer students had heard about the Situational Awareness meltdown and several brought up Michael Burry. There was near universal consensus that we are in a bubble, it’s about to pop, and everyone will look very silly. Several students brought up seeing a lot more advertising for AI models in recent months and suggested that “they’re hyping so hard because the whole thing is on its last legs.” The demise of some image models (I don’t follow that space fully enough to know specifically what this was about) was seen as proof that progress is actually running backwards.
It was also suggested, contra the worldview suggested in #2 above, that with progress as totally stalled as it is, CEOs may be forced to start rethinking when they realize AI can’t actually replace people. The majority certainly leaned towards “crappy but will take your job anyway” but a sizable contingent took the “everyone is going to wake up and realize how dumb it is soon” view and some students straddled camps.
Perspective #9: They’re worried about the youth (i.e., the preteens)
One student raised the concern that “young people today are relying on AI too much.” I nodded along, until she clarified that she meant people like her ten year old cousin. This says more about college students than it does about AI, but they see themselves as having made it through K-12 (or most of K-12 depending on age) without AI, learned the right things (sure…), and now using AI as a helpful shortcut. Apparently, today’s 10 year olds are using it to do their math homework, though? And today’s 20 year olds think that’s a problem because the 10 year olds will never learn math.