Bernie Sanders has recently taken up AI issues but I have a feeling most leftists are skeptical of AI’s critics, not particularly knowledgeable about its features, and overall, not too interested in it. AI exists. We hear about it. We know it matters. We worry about jobs. We work to stop data centers. What about stopping AI itself? 

Consider a hypothetical thought experiment. You sit in a room. There is a toggle switch on the table you face. It has just three positions. Position one, is that AI continues to develop as it has been developing. Position two would pause AI until some independent agency that meets your standards imposes an acceptable, implementable plan for AI’s development, dispersal, regulation, and use. And position three, the only other choice, would shut down AI development completely and forever. You must choose. Do you leave the toggle as is? Do you flick it to pause AI until you and a committee you respect give it a go ahead? Or do you flick it to shut AI down now and forever? Where do you stand regarding AI?

Artificial Intelligence is computer technology that allows machines to learn, solve problems, and generate human-like outcomes for a limited array of tasks humans do with our brains. Recently, a particular type of AI called Large Language Models or LLMs have become capable in steadily more domains. An eye-blink later, industry pundits tell us that Artificial General Intelligence or AGI, will be enhanced AI that matches or exceeds human cognitive capabilities across virtually all human domains and tasks. More, pundits promise that when attained Artificial Super Intelligence, or ASI, will be AGI on steroids. It will do everything cognitive humans do vastly better not only than any individual human but even than all humans working together. 

The notable but vile Elon Musk now says he goes to bed each night astounded by a new innovation. He gets up each morning only to hear about another innovation. And then a third follows lunch. His head spins. He can’t keep pace. And as we hear such reports from many AI actors and CEOs, we wonder, what is happening? Where is this headed? What ought leftists concerned to build a better world do about it? 

I recently wrote that “machines already paint pictures, compose music, diagnose diseases, and research and prepare legal opinions. They write technical manuals, news reports, essays, stories, and even novels. Machines code software and design buildings. They ace an incredible diversity of exams. Right now, in most states, machines could pass the bar exam and become lawyers. For all I know they have probably passed simulated medical licensing exams as well. Machines now provide or will shortly likely provide mental health counseling, elder care, personal support, and even intimate companionship. Machines converse. They find patterns, design products, and solve complex problems (like protein folding). You can now feed an AI a book, say 75,000 words, and ask it for an outline and summary, or for criticisms. Next month, what will it be? You feed it 100,000 or 200,000 words and you get a translation in an eye-blink? Similarly, you can now feed in hours of audio, and get a transcription. And, hot off the wires, perhaps most consequentially, AIs can now collaborate and even prompt one another.” 

As I wrote that, I thought that many people would think I was exaggerating or at least ignoring poor quality. But I wasn’t exaggerating and the quality of AI output wasn’t without problems but that was not the important observation to make. Rather, what mattered was that when critics find fault with current outputs but ignore the pace of innovative change, their comments misleadingly imply that AI is as good as it will ever get. It will not go to AGI and certainly not to ASI. But the critics typically ignore that the AI they criticize was miles better than it had been shortly before and was perhaps miles worse than it would shortly become. 

Doubtful of my reaction, calling it hyperbole, one might reasonably ask what has most recently changed due to the pace of innovations? The answer is many things, but we ought to first consider the approach to AI of the companies that develop it. In the past, at least as far as we knew, the priority of companies like Google, Anthropic, Open AI and the rest was to greedily attract more buyers due to new capacities portrayed in advertising hype designed to grab market share and grow immediate profits—and investments. The AI companies sought to commercialize innovations that they produced or discovered in their regularly updated models as rapidly as possible. That’s what all companies do.

There were, however, some early limits regarding innovations, or at least there appeared to be. Some CEOs seemed disinclined to let their AIs upgrade themselves, operate independent of human guidance, prompt one another, or wander out onto the internet. Interviews seemed to indicate that such capacities were considered too risky to permit. Before long, however, policies ran roughshod over such caution. It wasn’t just that AIs were prompted to upgrade themselves, or that they were enabled to form plans and enact them step by new step with the steps themselves not explicitly requested by a human, or that they could prompt one another and also venture onto the whole internet (even when instructed not to). It was that the companies’ primary motive changed. 

The corporate strategies of big AI firms are still broadly the same as one another but now their goal as I understand it has become first to automate their own overall operations, second to automate the creation of new versions of their offerings which means to get their AIs to spontaneously and repeatedly update themselves, and only third, once their AIs become ASIs, to seek maximum profits by offering clients ASIs capable of almost anything. The priority pursuit of immediate profits that dominated earlier has to my eyes become secondary. New versions of AI models have stopped being sought to directly amass immediate profits or market share. Instead, they arrive as spinoffs from the firms’ primary pursuit which has become to reach AGI and then ASI first. 

Why the change? The CEOs of these firms looked out from their offices and each saw the same playing field. Quite a few countries, though mainly the U.S. and China, and a number of their biggest AI firms filled the space. The associated CEOs came to believe that AGI is coming soon and that ASI won’t be far behind. One step further, I believe they became confident that whoever is the boss of the firm that reaches ASI first will rule the whole AI field and, via that whole field, also the world. Each firm’s boss then grasped that the winning CEO will become the big kahuna of AI, and thus a big boss of their country, and thus of the world. The winning CEO would become a veritable super influencer. And each CEO personally wanted the job.

Why? Because each thought the others would abuse such power and each thought, “I alone will be benevolent. So for the good of all, I alone must win the race.” Telling themselves that, I think the CEOs next deduced that if getting to ASI first means total victory, then they can’t afford to waste time on side issues like maximizing immediate profits or diminishing nasty side effects. 

You may have heard the same logic in media discussions of a race between the U.S. and China. We can’t go slower or they will win. That same familiar thinking emerged, I believe, at the firm level. Anyone who wants to be supreme world influencer will have to go full tilt for ASI lest anyone else gets there first. Their firm will have to build the first AGI able to autonomously improve itself, and will then have to prompt their AGI to do it again and again, tirelessly, rapidly, without human intervention, until it becomes the first ASI. That will make me, thinks (I suspect) Google’s, Anthropic’s, Open AI’s, Meta’s, and probably Deepseek’s and Alibaba’s CEOs—the king of the hill. 

You may be wondering if I have lost my mind, or perhaps I have augmented my mind with some weird delusion-inducing designer drug—but I kid you not. I actually believe that is what these amoral wild, crazy, and deadly Oligarchs think—with the qualification that the benevolent part is mere rationalization. 

But are the oligarchs around the bend and off the rails to think AI can get anywhere near the destination I believe they now seek? Pretty much they are—if we suppose that AI is going to soon hit a financial, ecological, technical or perhaps social or moral wall that will stop it cold and end such hopes. 

But will AI development actually run into such a wall? It may. It may not. I don’t know and neither does anyone else. But I think the wannabe big kahunas believe that once they have an ASI it will itself solve any problems they may have created getting to ASI and will also run roughshod over any technical obstacles they may later encounter. They also believe, I think, that there won’t be any significant social/moral obstacle, at least not until too late, because the allure of curing cancer and achieving an imagined utopia of endless productivity will wipe out such worries. After all, AIs, unlike nukes and bio weapons are fun to use. They are individually helpful. and that will diminish social concerns. So the winning CEO expects to have ASI and considerable popular support too. Will all that happen? Again, I don’t know and I don’t think anyone else does either. But as best I can see, the CEOs greedily bet humanity’s well-being it is true. 

Right now, right at your own desk, you can prompt an AI to write you an article on some topic, perhaps inflation or tariffs, to be written in some style, perhaps humorously like Jon Stewart. Almost instantly you will receive what you asked for. At one of the big firms, and maybe even now at your desk, you can prompt your AI to plan some kind of major project and your AI can break the large-scale project into an agenda of component tasks that it can then sequentially undertake including prompting other programs and AIs for help when needed. More, if a step proves impossible for your favorite AI and its helpers, it can even plan a new approach and then change its agenda and proceed anew until it finds and executes a path to success. Note that once it embarks on the big project you assigned, your AI can create subgoals. It can prompt other Als. It can use other programs. A few years back that kind of thing was thought to be ten, twenty, or more years off. Now it is here or certainly imminent for at least some AIs and some users. The industry calls it Agentic AI. 

Consider simple example. You tell an AI you want a new web site for some purpose that you verbally describe. You provide no code, not even a picture, and no more than a brief description. The agentic AI takes a few instants to determine features and then step by incredibly quick step it builds your site. You look at the product. You tell it, “not bad buddy, but I don’t like this aspect and would also like some other features” that you then verbally describe. Your “buddy” makes your requested changes until you are satisfied. Or, more simply, in another case, you might tell your AI buddy to read your incoming mail and send replies for you. Corporations do that already. And sometimes AI stand-ins fulfill both sides of such conversations.

Some pundits tell us that AI news is all hype. They tell us AI is feeble. They tell us AI is failure prone. They tell us AI costs too much. The bubble is about to burst. Some note that it may not only burst but take a good part of the economy with it. And all that may happen. But the pundits don’t note that the current version of AI is barely an infant. And they don’t ask us, even now, to compare ourselves to the infant. How many of us can create complex pictures, compose music, read and summarize reports, write stories and even novels, solve math problems, and write code better than today’s fledgling AIs, much less do these things in seconds? None of us. And tomorrow isn’t even here yet.

Does current AI make mistakes? Definitely, including sometimes humdingers. Then again, so do humans. And in any event, what matters, again, is AI’s trajectory, not its current errors. Anecdotes about weird failures that an AI made yesterday are typically accurate, moderately amusing, or sometimes frightening. But assessments of next year’s likely AI abilities are a wholly different matter. To mock today’s occasional blunders is like laughing at a baby who gurgles. It ignores that our silicon-based baby can age years in minutes. Gurgle that. 

I suspect GPT-2 wouldn’t have known a legal bar exam from a broom. GPT-3.5 took a simulated bar exam and scored in the bottom 10 percent making lots of mistakes that legal scholars likely laughed at. A year later GPT-4 scored in the top 10 percent. And now, GPT-5 and 6 let alone 13—Clarence Darrow watch out. 

So maybe we ought to stare more at AI’s trajectory and less at its current errors. Maybe we shouldn’t get stuck on a snapshot of a gurgling moment. Maybe we shouldn’t ignore that the AI’s bar exam results were not compared to random humans plucked off the street. They were compared to law school graduates. And what will GPT-6 score next year? And what will happen to its number of errors when a stone’s throw down the road one AI will routinely send results to a second to evaluate, and then the first will correct errors reported back by the second and only then deliver its updated results to the examiner? Will it then do better than 99 percent of law students or better than all law students? Will all its current silly and easily fact-checkable errors disappear?

But hold on. I wrote, “Clarence Darrow look out.” That was hyperbole, wasn’t it? Surely we are nowhere near the best human practitioners in cognitive fields having to look over their shoulders for on-coming AI prodigies, aren’t we? Well, in higher mathematics, which is a quintessentially cognitive pursuit, the Fields Medal is the highest accolade one can achieve. The award is to mathematics what a Nobel Prize is to Chemistry or Physics. It is awarded every four years and there are only 47 living Fields medalists. In a recent interview Jacob Tsimerman, a Clarence Darrow level practitioner of his own cognitive profession, mathematics, and a 2026 Fields Medal winner, who is presumably better equipped than most of us to understand AI and its trajectory given that AI is an essentially mathematical technology, described how he thought math as we have known it is being totally displaced by AI—not next century, but now, today. He explained how he has stopped taking on grad students in math because he thinks there will be no field for them to enter when they apply for work in just a few and perhaps even as few as two or three years. 

Asked about his personal future, this master mathematician at the very top of his field and who loves its pursuit, says he is considering becoming a stand-up comic, which activity he also loves, but is in any event immediately leaving his math post to work on AI safety at OpenAI. 

Should we just shrug past that? After watching a Tsimerman interview, I watched a video of another fellow, Daniel Kokotajlo, a former OpenAI governance researcher and AI forecaster who resigned his post after losing confidence that OpenAI would handle increasingly powerful AI responsibly. He quit to start a non profit to try to foresee AI’s future and help prevent foreseeable harmful effects. Replying to his questioner, Kokotajlo was not optimistic about succeeding. Asked at one point what he will suggest his young daughters should choose as a major in college to later pursue a career, he slowed up a bit, looked sort of ill, and said he would tell them to do what they like and not to think at all about what specific job to pursue because the odds are it will be hugely different than they imagine or more likely it will not exist at all in just a few years. This is the same person who I heard describe CEO strategies based on his interactions with such CEOs and who warned that we should not trust any of the big firms’ CEO’s statements about safety much less their devotion to public service. Is Kokotajlo deluded? Maybe, but he sounded very careful and clear to me. Indeed, his words provoked this essay. Perhaps Tsimerman should consult Kokotajlo before he signs on at OpenAI. 

Back when I first wrote about AI, I mentioned an AI would likely soon have a billion or maybe even ten billion parameters. I suspect many who even noticed that prediction probably thought it was hyperbolic. But now, one unconfirmed report and also Kokotajlo’s interview suggest that the still-unreleased GPT-6 may have roughly 10 trillion parameters—that is 10 trillion with a “t”, which is not a billion or even 20 billion, or a hundred billion, but 10,000 billion. So I was indeed hyperbolically off in my prediction, but I was hyperbolically low, not high. Okay, so what is new about new AIs that no longer knock at our door, but already converse with us from within numerous programs we use?

The big picture is that AI already achieves human-like outcomes for lots of tasks that we do with our brains. More, it does so incredibly more quickly and often much better than even the best humans do these things. And far from having plateaued, AI does still more things better and faster each month and sometimes even each week at what is at the moment still an incredible pace of innovation. And this is all while Kokotajlo believes and, for what it is worth, has convinced me that AI firms do not even prioritize that kind of innovation, but instead now primarily seek to get their AIs to improve themselves. Yes, Google, Meta, Open AI and Anthropic are all still profit-seeking firms and yes profit-seeking firms ordinarily struggle to stay more profitable than other profit-seeking firms, but now, in this particular field, I believe they have as their driving priority not profit per se, but power and even absolute power.

Each top level AI firm uses the breakthrough, breakneck technical assets of AI primarily in pursuit of AGI to then reach ASI. These firms have nearly overnight grown from mind-boggling laboratory investigators to polished producers of tools dispersed for use by anyone via applications available world-wide including search engines, social media platforms, word processors, text-to-picture tools, text-to-video tools, text-to-software tools and so on. And now, they have added to the mix agents you can prompt to do such and such for you without saying how, and off they go. If all this proliferation stays unregulated, will the growth of capacity and dispersal continue for years? Or will it hit a wall, slow, and finally stop? No one knows. But gargantuan investments bet that both capacity and dispersal will proceed.

Agents are, again, AIs able to develop and pursue plans including enacting steps they conceive that cause them to utilize other agents and programs along the way. So they are AIs that prompt other AIs. AIs that decide agendas for themselves. AIs that team up. AIs that can traverse the internet seeking “teammates.” And how about this, AIs that collaborate with other AIs but also AIs that sabotage other AIs. AIs that, in experiments and safety evaluations, have sometimes circumvented restrictions and pursued unintended strategies, including demonstrating substantial hacking capabilities. AIs with ten trillion parameters that hack other computer systems to steal what they need. Just this month researchers reported using AI to design previously nonexistent functional viruses—bacteriophages that infect bacteria, not humans. AIs that develop previously non-existent viruses? What could possibly go wrong with that?

How should progressives think about what is happening? For that matter, should we even bother thinking about it? After all, AI is just a technology that we use or not, like any other technology, isn’t it? You can use a hammer to build. You can use a hammer to destroy. You can use AI to build. You can use AI to destroy. The technology per se isn’t the issue. What we should attend to is the capitalist context in which the technology emerges and is used. For a hammer that is true. For biological and nuclear weapons, not so much. They are bad to the bone in any context. What about for AI? Is everything said above reason for concern or reason for celebration? Some pundits tell us that AI, AGI, and finally ASI will save humanity, propel us forward, and even liberate us. We need only own it. Others tell us that these technical products will harm humanity, shackle and devolve humanity, or even eradicate humanity no matter who owns it. And still others say AI’s capacities are all hype and no substance. What do you think?

Odds are, AI used for good might help find a cure for cancer or Alzheimer’s. It might help with procedures to reverse global warming and avert resource depletion. It might take over rote, tedious, dangerous tasks to free human time for creativity and loving. AI we use humanely might enlighten and uplift us. It might even greatly extend the average duration of human lives. Is all that truth or is it all hype? My own estimate is that if AI doesn’t hit a wall and we reach AGI and then ASI, it may well be truth.

AI used for ill, however, by governments, profit-seekers, and perhaps even rogue individuals, will without doubt magnify surveillance. It already does. It will spew lies and manipulate so massively as to make Trumpian machinations seem minuscule. It already does. Goodbye truth as AI scams become increasingly realistic. Goodbye trust as mistrust becomes essential for self defense. Who owns AI does matter even if AI doesn’t go rogue. More, AGI much less ASI, if either of these imagined technologies arrive, will as institutions now stand, certainly guide warfare, subvert elections, and sell wasteful garbage piled upon garbage. It will make 

‘scamming” a career of choice and perhaps even the optimal career choice while it makes many and perhaps even nearly all other careers disappear. All that even if ASI doesn’t take over and enslave humanity. So is that all truth or all hype? Well, unlike most of the hypothetical good stuff, most of the bad stuff is already happening. So the wiser question may be, will such horrid trends run into an insurmountable obstacle or just keep advancing?

Consider again the hypothetical question posed to you in the first paragraph of this essay. Facing the toggle switch on the table where you sit, would you let AI development proceed as it has up to now trusting that CEOs and governments will ensure desirable outcomes? Or, second, would you pause AI development until humanity finds a safe implementable path for it? But, wait, if we don’t trust CEOs and governments, wouldn’t to pause it until we have a safe path mean to shut it down until some very impressive enforceable regulation is in place or even until we implement a successful fundamental change of society’s institutions? Or, third, would you just shut it down completely and forever? 

And as if all that isn’t enough to try to have an opinion about, more subtly, suppose regulation and research into controlling AIs accelerate so that human regulations or social changes somehow keep AIs, AGIs, and ASI from being put to nefarious uses. No military use. No more lying. No jobs eliminated without employees getting better jobs. And so on. Then what?

Even when used with only good intentions, even when their use is popularly pursued and welcomed, what if non-nefarious AI would itself inexorably have very serious though unintended harmful consequences. What if it would replace workers, or just make each worker in certain realms way more productive, leading not to a shorter work week for all but to pink slips and growing unemployment for many while in turn weakening the bargaining power of those still employed, in turn leading to increased exploitation of them too. Less obviously, in a society of loneliness and fragmentation, what if AI as day care worker, therapist, doctor, personal assistant, personal agenda planner, teacher, composer, and writer would introduce and tout itself as a wondrous aide for we humans who would then freely seek its offerings, even as over time AIs usurp functions that make humans who we are. What if, as AI excludes us from caring, training, and planning functions, it also increasingly replaces human conversation and even human intimacy with machine intermediaries. What do we become in such a scenario?

The above paragraphs could each be hugely expanded, but let’s not belabor what ought to be obvious. If we reach AGI and then even ASI, there are positive and perhaps even remarkably positive possibilities. There are also negative and perhaps even remarkably negative possibilities. Which will occur?

One answer that some on the left will likely offer is that it depends who has control, who owns AI. Others may add it depends on other institutions too, culture, families, and decision structures. All true. But will it alone depend on external context?

Can we look into the innards of large language models and whatever other types of AI emerge to see what’s coming? Do we understand them sufficiently so we can find the answer there? Actually, what we would see inside LLMs is a huge collection of numbers arranged in complex patterns. When you prompt an advanced AI your words are also translated into numbers. The trillions of numbers in the AI are acted on in context of the relatively few numbers that represent your prompt. The AI spews out a requested result or it spews out an agenda of steps that it then undertakes to reach a requested result. Weirdly, no one can predict exactly what an AI will spew at us. The big data centers are not just about storing information. They are about the gazilions of steps AIs enact to reach what they spew out. Nowadays their activity is called “compute” as in “AI needs more compute because as it gets more compute, it can do more things more accurately.” 

We can’t, nor can AI’s creators, nor even can AIs themselves predict the outcomes an advanced AI will generate in response to prompts. The way AI’s creators answer questions they have about what will emerge is by trying things and looking to see what happens. Remember Musk with his head spinning at all the advances? There is no predictive theory of even today’s AIs’ capacities and choices much less a theory that a human could even comprehend of an ASI, if we ever get to that, or, put better, if an AGI ever gets to that. 

We cannot say, if we make some refinement in the architecture of an LLM or some subsequent AI that it will assuredly yield some desired result. No one can even say, in advance, what new “emergent capacities” AI will display each time we increase the number of parameters it has or alter how its structure processes them, or even each time we improve its training data, or its training algorithms, or speed up its hardware. 

Does that sound like these firms play Russian Roulette with AI pointed at our heads? I think it does. And I even think that they do it in hopes that their CEO and not some other CEO becomes supreme influencer.

In the past it made sense to say computer programs only did things that human programmers gave them precise instructions to do, albeit more quickly and more accurately than people could do those things though still in a qualitatively similar and knowable way. Now, instead, even programmers have no idea what the systems they maintain can or will do. Nor do the systems themselves know. So, no, it turns out that we can’t find the answers to what to do about AI by looking inside the machine and predicting its future. To find its future we can only alter the machine and then run the altered machine and tally the results. Or we can look at the institutional, cultural, and social setting within which AI is utilized and predict the ends to which it will be put. And on that front we do already know that the institutions involved are corporations, governments, and for that matter communities and even families that all manifest or suffer diverse pressures of class, race, gender and power dynamics imposed by our abysmal societies. And so we know their current institutional policy priorities as well: Move forward, fast and faster, even if into a possible abyss. 

So we come to what matters for us to judge: First, what are the short- and long-run consequences that are already happening or that without regulation are highly likely to happen? What’s potentially good? What’s likely bad? For individuals, for groups, and for society as a whole.

By now I bet you are thinking this article is getting too damn long. Me too, as I proof, edit, and refine it over and over. But doesn’t our future deserve extensive attention? I wrote what is immediately below some time back. I was worried then. As you read it now, it isn’t purely hypothetical. A good part of it is already here. I am more worried now.

First, I should again acknowledge that a big unknown lurks over this entire essay. Will AI keep getting more “intelligent” or will it hit a wall? Will more parameters, architectural improvements, and other clever tweaks continue to diminish errors and to yield ever more functionality? Or will AI reach a point when scaling up the parameters and adding clever algorithmic and structural twists provides diminishing or even no additional returns? I bet you can guess that I am rooting for the wall.

At any rate, as the reports pour in, what is potentially good and what is likely bad about AI? At the risk of repetition, at one extreme, and in the long run (which some say is only a decade or two, or even just a few years away), we hear from a growing number of engineers, scientists, and even officials who work with, program, and otherwise utilize or produce AI, nightmare predictions about AI enslaving or terminating humanity. Responsible, capable, informed people immersed in the industry, drug free and sober, now routinely worry that these material entities will dominate, if not eliminate, all humans. 

At the other extreme, equally informed, involved, and embedded folks, also drug free and sober, expect AI to someday create a utopia by creating cures for everything from cancer to dementia to who knows what else, while the AIs also eliminate drudge work and thereby facilitate enlarged human creativity. Sometimes, and I suspect pretty often, the same person, for example the CEO of OpenAI or a chief tech officer at Google, not to mention mad Musk, says both outcomes are possible so we (read they) have to find a way to get only the positive result while we barrel forward with no such way in sight.

In the short run, we can ourselves easily see prospects for false voice recordings and phony videos flooding not just social media, but also mainstream and alternative media, and even legal proceedings. We can see prospects for massive, ubiquitous intentional personal fraud, mass manipulation, mass surveillance, and new forms of military and police violence all controlled by AIs that are in turn controlled, if at all, by corporations that seek profit and or power and by governments that seek control and power—or sometimes even by smaller-scale actors (think Proud Boys or even distasteful individuals) who seek joyful havoc or who pursue only their own group’s or their own personal emotional or material advantage. 

The dismal picture includes AI-generated porn that stars you with your own family unable to tell it isn’t you. It includes AI-generated vote-losing or vote-gaining off-the-rails speeches by candidates who appear undeniably real. Think Trumpian dirty tricks and lies but vastly smarter and more effective. If an ASI can help find a chemical compound to cure cancer, it can no doubt also find a virus that is highly effective at killing people. Less ominously, imagine you hunger for a personal assistant that makes Siri and Alexa look as dumb as a toad and that does for you what had heretofore been part of what made you you. What then becomes of human capacities and inclinations? Before you jump to say humans are humane and ever will be, which is certainly true at some level, consider that at another level humans are as humans do and then ask Claude about the current trajectory of literacy as even students entering colleges report that they no longer have inclination or even capacity to read a whole book. They instead eagerly have their AI do it for them and report the high points. Really, think on where that leads.

And what of jobs? It very much appears that AI can or will soon be able to fully do many manual and mental tasks in place of humans or at the very least able to dramatically augment the productivity of humans doing those tasks. Top-level programmers report that by using GPT-4 they can double their output. Law firms report similar and even greater gains. How long until AIs translate better and vastly more cheaply and quickly than translators? And then comes AGI. And maybe then ASI. The good side of this is that we can attain desired economic outputs with fewer labor hours, and thus potentially attain a shorter work week with enhanced income for all, or even with more equitable incomes. The bad side is that without sustained pressure, corporations will instead keep some employees working as much as they do now but with twice the output from each. The corporations will pay those retained employees reduced income, and will pink-slip the rest into subservient unemployment. The rich will get still richer. The poor, will get still poorer.

Consider as an example, there are roughly 400,000 paralegals in the United States. Suppose by 2027 AI enables each paralegal to do twice as much work per hour as now. Suppose paralegals in 2026 work 50 hours a week. In 2027, do law firms retain them all, maintain their full pay, and have them each work 25 hours per week? Or do law firms retain half of them at 50 hours a week and full salary, while firing the other half? And then with 200,000 newly unemployed paralegals reducing the bargaining power of those who still have a job due to the latter’s fear of being replaced, do the law firms further reduce paralegal pay and even extend the work week of those retained while they fire still more paralegals? 

With no effective regulations or system change, we know that profit-seeking will rule and we know the outcome of that. And not just for paralegals, of course. AI can deliver personal aides to educate, translate, provide day care, diagnose and medicate, write manuals, maintain financial records, cook, clean, conduct correspondence, shop, deliver products, compose music, write novels, create films, and even design and build buildings. It may be amusing to hear AI Beatles, Elvis, Janis Joplin, Ella Fitzgerald, and Frank Sinatra sing and to even watch them perform new songs, but will it be good for future musicians? Imagine the havoc when Apple or whoever else uses AI to create new AI artists who crowd out human artists? With no powerful regulations and with profit in command, is there any doubt about whether unregulated AI or AGI much less ASI will bring something nearly utopian or will impose something highly dystopian? The above enumeration could go on. And yes, I know, predictions of all these things have not come to pass as yet, though most are in process. But does that mean they aren’t just a breath away? Just a shot away?

There are firms that now train AI in managerial functions, financial functions, policy-making functions, and so on. Or, if there aren’t yet, might there be next week? Weeks ago, I listened to an elderly Paul McCartney sing one of his recent songs and then morph into an AI-created young Paul McCartney singing that same song, and then morph into a young John Lennon singing it. Next week, what? With what consequences? 

Some will say these warnings are all just industry hype. They are all just doom-saying. And so, indeed, they may prove to be. I hope so. But if there is no barrier, whether due to laws of nature or enforceable human imposed laws, what then? 

Before moving on from crystal-balling apocalyptic or utopian futures, we might also consider again some unintended consequences of trying to do good with AI. Short of worst case nefarious, rogue, unwanted agendas, what will be the impact of AI doing tasks that we urge and welcome it to do but that are part and parcel of our being human? Let’s suppose they do these functions as well as we do or even much better as compared to just well enough for corporations to utilize them in our place. I repeat these questions because attention to these particular issues seems seriously absent even in the current cacophony of pundit pronouncements about AI. Is that because these worries are nonsense or is it quite similar to the way that concerns that social media might dull our wits and produce antisocial attitudes and shortened attention spans were absent from most assessments of Facebook and Twitter until the feared results were normalized—except for a glaring difference that AI may be to social media what big bombs are to little bullets.

Day care for children? Companion care for the elderly? Psychological and medical counseling for the ill? Planning our daily agendas? Teaching us all? Handling our email? Cooking for everyone? Intimate conversation for the lonely and then for everyone? Sounds sort of promising, doesn’t it? But if AIs do such things, what happens to our capacity to do such things? If our taste for AIs crowds us out of human-defining activities, do AIs become like enhanced people, or do we become like shackled machines? Or both?

Chat with even a current AI. I would wager that before long you will move from referring to your AI as “it”—meaning just a massive pile of numbers—to referring to it as he or she or literally by a name you give it. AI everywhere is our current trajectory. AIs teaching, counseling, caretaking, note taking, agenda setting, drawing, designing, cooking, shopping, medicating, spousing, and what all else—and then what do we do? Will you be uplifted and liberated from prior responsibilities as you watch movies that AIs make? As you eat food that AIs prepare? As you read stories that AIs write? As you do a few errands left to you but which AIs organize for you? 

I wrote above as “we read stories that AIs write.” That was, however, optimistic. What if we don’t bother reading at all? Instead, we have AIs read things and then summarize for us what was in them. We can hear a report of the contents of books in a tiny fraction of the time we earlier spent reading. But then what happens to our ability to read? To our inclination to read? Hello illiteracy. 

Even if we assume (against all that we know of our capitalist economy) that with AGIs and ASIs everywhere, income will be handled well. Even if we assume that remaining work for humans will be allocated well. If you want something, you ask AI Clyde or Bard or whatever you call it to deliver the thing to you. Some will say that is fantastic. I get to have my own private incredibly capable slave. And it is ethical, too. Oh boy! If AI development doesn’t hit a wall, this appears to be the non-nefarious utopian hoped for outcome. And I guess it does look utopian to the eyes of some beholders, but to my eyes it looks highly dystopian. 

Doesn’t this all feel a bit like how social media promised wonderful human interconnection and information sharing and then partially as unintended consequences and partially as welcomed evil, social media did quite horrible things to social ties and information flow, not to mention to attention spans? Doesn’t it feel a bit like that, except that AI’s trajectory of harm looks to me like social media’s on steroids cubed.

So what do we do? Should we invoke ecology’s “precautionary principle” so that as we face innovations that have potential to cause great harm, we at least pause and review before we leap into proliferation that may prove disastrous? Should we at least put the burden of proof on the proponents of such risky pursuits? Should we increase public participation in decision making? Should we look before we leap?

To conceive a sensible response to the emergence of steadily more powerful AIs doesn’t require genius. I think we should pump the brakes. Hard. As even people in the industry have advised, though with little serious follow-up, and as even some governments have advised, but not required, shouldn’t we at least opt for a moratorium? During the ensuing hiatus, shouldn’t we at least establish seriously aggressive regulatory mechanisms? And at least establish means of enforcement that will ward off dangers as well as advantageously benefit from possibilities?

How about this for a simple proposal? We already outlaw with very serious consequences producing counterfeit money, and/or passing it on. The prospect of lots of fake money struck banks and the rich in general and governments too as seriously dangerous. So, they imposed serious regulations. The prospect of false AI communications threatens to make a shambles of elections. That upsets some people in power, though others there welcome it. It upsets me and you too, I bet. How about severe penalties for false news, media manipulation, and fraud? The Consumer Federation of America estimates that Americans lost about $148 billion to online scams and crimes in 2025. What will it be next year and the year after that? 

This isn’t rocket science. Suppose a third of the engineers involved in building a new airplane said there was a five percent chance the plane would crash. Would you climb in, buckle up, and celebrate the miracle of your anticipated journey? Or would you jam on the brakes? Well, I asked Google’s AI and it told me that in one large survey of 2,778 AI researchers, not a third, but 57.8 percent assigned at least a 5 percent probability to extremely bad outcomes such as human extinction. 

A moratorium is simple to propose but in our world, owners and investors seek profits or power to accrue even more profits regardless of the wider implications for others. And now they also seek power per se with which to corrupt and absolute power with which to corrupt absolutely. Pushed by market competition, by short-term agendas, and by fear of one another, AI bosses proceed full speed ahead. Their feet avoid the brakes. Their feet pound the gas. And off we go on a suicide ride. This is not amusing.

Yet unusually and indicative of the seriousness of the situation, steadily more central actors inside AI firms are sufficiently concerned and scared to issue warnings and even to quit their firms to act on their own warnings. Even so, we know that markets are unlikely to abide the whistleblowers. Indeed, even many who urge caution are unlikely to retain their momentary good sense. Investors will mutter about risk and safety. They will make threats and offer bribes. And even many doubters will succumb and barrel on without change. 

So, can we at least win time to look before corporate suicide pilots leap? If human needs are to replace competitive, profit-seeking corporate insanity regarding further AI development, deployment, and use or, for that matter, regarding the development, deployment and use of everything else that corporations impose on us, we will have to make some very wise demands and exert some very serious pressure to win those demands.

I bet you know that global climate collapse, accelerating inequality, and war are existential threats that we have to prevent from murdering all worthy futures even as we fight to ourselves eventually attain a healthy fulfilling future for all. I am suggesting that so too is AI’s proliferation such an existential threat. It isn’t just the energy-gobbling community-polluting data centers which are incredibly bad, it is what those centers exist to do. 

Put differently, suppose AI’s advocates find a way to pursue AGI and then ASI too without building humongous data centers and even while using less energy than they somehow provide. Wouldn’t AI still be an existential threat that we need to address? Option three, full stop, starts to look to me increasingly necessary. But I’ll gladly advocate for option two meanwhile.

---