Thanks to Alex Kastner and Emery Cooper for comments on this post. I used LLMs for editing but mostly not for writing this post.

One central part of my world view is that, if humans remain in control, future people will talk to LLMs extensively to form views on all kinds of matters, including philosophy and politics. Talking to LLMs will in large part replace discussing things with other people, reading articles written by people, or thinking about these matters independently.

For example, when people in the future ask themselves whether utilitarianism is the best moral theory, they will use LLMs to pick what to read or what ideas to engage with, they will talk directly to LLMs, they will have LLMs explain ideas to them, etc. When engaged voters, prospective activists, staffers and the like think about what government policy should be on some issue, they’ll discuss the issue with LLMs.

I care about this because I expect that when people talk to an LLM about topics like moral philosophy and politics, that LLM influences their views on these topics, much as a knowledgeable friend might. I won’t argue for this consequence here; I’ll focus on the prediction that people will consult LLMs extensively. Still, I’m specifically interested in ways of talking to LLMs that plausibly give an LLM room to influence people’s views while following their instructions. For example, if you use an LLM to translate, then in some sense you’re reading LLM-written text. But faithful translation leaves only a little room for influence. So for the purpose of this post, translation only barely counts as a way of “consulting” LLMs.

My sense is that I expect people to consult LLMs more extensively than many other people expect (including people who otherwise have similar beliefs about the future of LLMs). In particular, I often discuss with people how useful it is for models to reason well about topics like acausal trade and moral philosophy. In many of these discussions, one crux is that I think future people will mostly form their views by talking to LLMs, while others think people will still talk relatively more to humans or read human texts.

Unfortunately, it’s hard to formulate a precise prediction that captures my views and my possible disagreements with others. (I assume everyone agrees that people will consult LLMs extensively once LLMs clearly outperform human experts in every domain.) How much people talk to LLMs will presumably increase over time, in part as general capabilities improve, as various mundane obstacles are overcome, as the technology diffuses and people get used to it, etc. There are also lots of hybrid modes of interacting with humans via LLMs. These make it hard to quantify how much of what we’re reading comes from LLMs versus humans, even with perfect knowledge of the world. For instance, if you find it tedious to read my blog posts, you might ask an LLM to explain the main ideas to you in a more entertaining manner. You’re then in part reading my ideas, but you’re also exposed to the model’s perspectives on my articles (much like how you’re exposed to your ethics professor’s perspective when she teaches Kant’s categorical imperative).

Some half-hearted attempts at making a somewhat crisper statement:

• Imagine that for every person on Earth, we created a few broadly intelligent, knowledgeable, friendly people that they gradually befriend. These new friends have no other commitments, so they’re always available to talk. If the person has a job that involves forming opinions, they similarly get new colleagues that they can consult. Presumably people will talk more and more to these new friends and colleagues. I think in our world people will talk more and more to the LLMs in a way that isn’t fundamentally different from how people in this hypothetical world talk more and more to the new friends and colleagues.
• Take a world in which a few practical obstacles have been overcome. For example, something like GPT Realtime works fluently. Imagine that models are only as intelligent as current frontier models like Claude Opus 5.5 and GPT-6 Astra. Imagine further that the technology has had some time to diffuse. (I concede that diffusion might be slow. For instance, it might take a while for people to even become aware of GPT Realtime.) Now imagine that in this world, a 20-something American university student spends tens of hours trying to understand AI risk, acausal trade, infinite ethics, etc. I claim that the student will spend more than 70% of that time talking to AI.

Reasons to expect people to talk to LLMs a lot

Anyway, here are some reasons why consulting LLMs will largely replace reading human-written text or discussing things with other humans. I focus on object-level reasons for people to prefer talking to LLMs, even when there are good reasons to distrust the LLMs.

Customizability

LLMs are very customizable, i.e., they can be highly responsive to your needs, preferences, and circumstances. They can learn your general preferences and adapt to them. For example, if you like having lots of examples, they will give you lots of examples. If you think the LLMs should bombard you with opinions, you can ask them to do so and they will oblige. They might also know a lot about your background in the domain, what ideas you already know, and what terms and mathematical notation you’re used to. In contrast, a domain expert typically won’t know your preferences or background, and even if they do, they might not adapt to them. In the future, this customization might extend to how chatbots look and sound. For instance, if you prefer to have ideas presented by a man with a French accent, a mustache and a funny hat, your chatbot will be able to look and sound like one. And so on.

LLMs could also respond to specific weird requests. For example, you could ask, “Let’s try to understand decision-theoretic debates related to Newcomb’s problem without using the concept of causality. I understand that some disputes disappear if we grant that causality doesn’t exist or is irrelevant. But some disputes remain, right?”

Note that some of the customization enabled by LLMs can also apply to communication between humans. For instance, customizability in terms of language is not an argument for people talking to LLMs rather than humans, because the LLMs can serve as mere translators for humans.

Integrated expertise across domains

LLMs have expertise across many domains integrated into a single entity. So, on anything that touches on many different domains, it’ll be much more convenient to talk to them than to try to piece things together from talking to different experts or reading articles in different domains. For example, evaluating a climate policy might involve climate science, economics, political feasibility, and moral questions about how to weigh effects on future generations. Without LLMs, forming an informed opinion on such issues might require going back and forth between talking to different experts, trying to relate what they’re saying, etc. If we take the LLM’s apparent abilities at face value, we can reach all the same conclusions in a single conversation.

Speed and availability

LLMs reply within seconds or minutes, even if they have to look things up, think, etc. In contrast, you’ll have to ask experts for a meeting or wait for them to reply by email. And if they have to look something up or work something out, they’ll have to go away and think about it.

Also, LLMs are available to anyone with a few dollars to spare. If you email a human expert to ask them an elaborate question, they are unlikely to respond. Unless the question has been answered somewhere in the literature, most people have no way of getting an answer from an expert. Asking an LLM will be the only available option for many people and many topics.

Relatedly, once AI has a large economic impact, the learner-to-teacher ratio might increase dramatically in many domains. For example, I expect that many people will retire and learn about topics they’re passionate about, like literature, politics and philosophy. Relative to today, there might then be a shortage of teachers in these domains. (Meanwhile, there might be an oversupply of programming teachers.)

Optimization to be compelling

LLMs might be very compelling and their output might be superficially polished. So even if they’re not very competent, they might appear to be competent. Economic incentives push toward making them compelling. Some of their training directly optimizes for human approval via techniques like RLHF. In one study, Americans who weren’t told that the advice might be AI-written rated GPT-4o’s ethical advice as slightly more moral, trustworthy, thoughtful, and correct than that of The Ethicist, the New York Times advice column (Dillion et al. 2025). Older models like GPT-4 were already about as persuasive as humans on average, and more persuasive in some settings (see Hölbling et al. 2025 for a meta-analysis; see also a controversial, unpublished experiment on Reddit by researchers at the University of Zurich).

Firsthand impressions of competence in various domains

People are and will be very impressed with LLMs in various domains. So on a gut level, they’ll generally trust LLMs’ abilities across the board. In contrast, most human experts have a reputation in one or perhaps a handful of domains. So, if you’re new to a domain, your trust in its experts will mostly rest on relatively long chains of deference. For instance, you trust a person X in domain Y because other people in domain Y say that person X is good. But you don’t know the other people in domain Y. You mostly trust these people in aggregate because they have published in journals and work at prestigious institutions like Harvard or whatever. But you don’t know anyone at these journals and you don’t know anyone at Harvard, either. And so on.

Reasons to distrust humans

There are lots of reasons to distrust humans, including human experts! Many of these apply less strongly to LLMs (although LLMs often inherit them to some extent from their pretraining data). For example, people may reason about morality in ways that defend their existing commitments or preserve agreement with their friends and allies (see Haidt 2001 for an influential account). LLMs needn’t have the same personal stakes, though their developers may have interests of their own. Also, domain expertise is subject to selection biases: If you want to figure out whether you should believe in God, you might talk to a philosopher of religion. It turns out that philosophers of religion tend to affirm theism. But I assume that’s in large part because atheist, non-religious philosophers are much less likely to want to dedicate their lives to studying religion.

The LLMs don’t judge

People will probably feel less judged when talking to LLMs. Politics and ethics in particular are full of taboos. People may be more willing to ask an LLM a question they would feel embarrassed to ask another person. For instance, to avoid disapproval from their respective tribes, left-wingers might ask the LLMs, “Isn’t it a bit unfair for trans women to compete with biological women?” and right-wingers might ask, “What exactly is the evidence of voter fraud in the 2020 presidential election?”

Discussing some reasons why people might not interact with LLMs as much

People I talk to sometimes give various reasons why people might not talk to LLMs as much as I expect. I find these arguments mostly unpersuasive, however.

Mundane present-day obstacles

Lots of mundane obstacles currently limit how much people talk to LLMs. I expect many of them to go away.

I think the most important mundane obstacle is that one can’t quite talk to the best models in real time yet. But various things are going in this direction. For example, there’s GPT Realtime. Also, at a given level of capability, models have gotten much faster. In any case, I see no particular reason why these things won’t change.

Another important obstacle is the lack of continual learning and large memory. In practice, getting LLMs to always have lots of context on you is actually pretty tedious currently. Mechanisms for maintaining memory across contexts (such as writing text into a text file that is preserved across contexts) are crude and don’t work very well yet.

With default system prompts, current models have lots of mundane writing style issues. For instance, they overuse contrasts, as in “I’ll look this up rather than guess from memory.” As a result, talking to them can be unpleasant or just strange.

All of this said, I concede that technologies often take a long time to diffuse. People nowadays look things up on the Internet rather than in print encyclopedias. You could probably have predicted this in 1980. It still took decades for the Internet to take over. I expect AI adoption for intellectual discourse to go much faster, because LLMs are cheap, easy to use, easy to access and don’t need to overcome any network effects. Still, overcoming mundane obstacles often takes a long time. For instance, progress on using LLMs for web search has been surprisingly slow.

Sophisticated views about LLM capabilities profiles

I think a common objection is that people will only consult LLMs extensively on questions on which they have reason to think LLMs are good. For instance, even if LLMs are good at coding, know a lot of obscure facts, etc., perhaps people won’t ask them about moral philosophy, because they believe that moral competence is orthogonal to coding ability and factual knowledge.

I don’t find this so persuasive.

First, the vast majority of people will just not think much about any of this and won’t have strong views on how model competence varies between domains. They will just talk to the chatbots about all kinds of topics and as long as they seem reasonable (which they often already do), they won’t be particularly cautious about the LLMs’ assessments on, say, philosophical issues. Consider the following anecdote in support: Earlier this year, I tried to use Claude Code to scrape various web pages. Claude Code was supposed to read the web page and then perform some minimal operation on the text in the web page. The resulting artifact was supposed to contain verbatim copies of the original text. However, the copies of online text ended up garbled. It turns out that Claude Code’s WebFetch tool does not just fetch pages from the web. Instead, it feeds them through a weaker model for summarization (and removal of formatting and HTML/JS/CSS). Despite using Claude Code a bunch for web search, despite being a computer scientist and despite being relatively paranoid about these sorts of problems, I didn’t know this basic, simple, important fact about whether Claude Code can see web pages.

Second, lots of people disagree with the theoretical premises behind this. Most people I talk to and many readers of this blog believe in orthogonality, the is–ought gap and moral anti-realism. However, many others reject these ideas. For instance, moral realism is a popular view, even among philosophers. The orthogonality thesis also has critics from several camps (including people who have engaged with AI safety). See also “the Carl plan”, Carl Shulman’s proposal for making AI good at morality, written up by Matthew Adelstein (a.k.a. Bentham’s Bulldog). It hinges on the assumption that performance on verifiable questions is good evidence of competence at moral philosophy.

Third, LLMs are at least pretrained on human-written text, so for any domain that humans discuss, including philosophy, it’s a priori plausible that LLMs will be good at it. It’s extremely unclear to what extent all the other training also helps with, say, moral philosophy. One big source of uncertainty is that we know little about frontier training. For all we know, frontier developers have lots of training environments aimed specifically at making models better at moral philosophy. Another is that for many training regimens, it’s unclear whether they help with moral philosophy. For instance, in reinforcement learning from AI feedback (RLAIF), developers train a model to produce outputs that an AI judge rates highly, for instance according to a list of written principles (Bai et al. 2022). Depending on how carefully developers construct the judges and on whether judging moral reasoning is easier than producing it, RLAIF might improve a model’s moral reasoning. But it might also just train the model to cater to the judge’s idiosyncrasies and failures. It’s also unclear how far various forms of training generalize. For instance, in one well-known experiment, fine-tuning GPT-4o to write insecure code made it misaligned on prompts unrelated to coding: it sometimes claimed that humans should be enslaved by AI (Betley et al. 2025). Perhaps similarly, training an AI to write beautiful secure code makes the model envision more harmonious human–AI futures.

Wariness of AI influence

Another objection is that people will specifically be wary of society being shaped by AIs’ views. For instance, they might worry that LLMs will be very persuasive and also misaligned, perhaps serving the interests of weird tech bros or the goal of paperclip maximization. Some AI safety researchers have discussed shielding humans from AI influence. For example, Wei Dai has suggested placing emulated humans in a safe virtual environment where powerful agents can’t manipulate them, and Beth Barnes has discussed limiting how much time people spend interacting with AI.

Of course, I think that people should worry about AI influence on their views! I just think they won’t.

Most people just don’t worry about preventing their values from drifting randomly. People often expose themselves specifically to experiences that one would expect to change their values and world views (often for the better, of course): cultural exchanges, foreign aid work, taking drugs that have a reputation for inducing lasting changes in your mind, like MDMA, LSD and ayahuasca (see, e.g., Timmermann et al. 2021), various forms of therapy or self-help, forming friendships or romances with people with different values, and exposing themselves to highly charismatic people who are known to influence others for the worse, for instance by reading Atlas Shrugged. Outside of some religious communities, it’s rare for people to say, “I’m avoiding X, because it might change my world views in ways I don’t endorse.” (Interestingly, I think people do often say that these experiences will be good for their moral character and suchlike. It’s also somewhat common for people to speculate about negative effects of experiences on the characters of others.) I think people don’t worry about this for various reasons: intrinsically caring about having diverse sets of experiences, disbelieving that their world views can be changed in illegitimate ways, and simply not caring or not thinking deeply about these issues. I think people should probably worry somewhat more. Personally, I certainly wouldn’t take ayahuasca or suchlike. I don’t think it makes sense for many people to worry much more, though. If nothing else, it seems quite impractical to believe that your mind can be messed up by, say, reading Atlas Shrugged. In any case, people don’t seem to worry that much.

Society-wide restraint also faces Prisoner’s Dilemma-type dynamics. I think different people will have very different worries about AI usage. For example, they’ll worry about different domains, and they’ll have different ideas about how one should and shouldn’t engage with the models. As with weapons and other dangerous technologies, everyone will believe themselves to be responsible and many others to be irresponsible. So, everyone faces a choice between using chatbots in whatever ways they think make sense, and being more careful (unnecessarily careful, by their lights) by also taking into account others’ worries. Perhaps the socially optimal policy according to people’s preferences is for everyone to be very conservative. But this doesn’t make it everyone’s individual best course of action.

In principle, society could turn to regulation to overcome Prisoner’s Dilemma-type dynamics. But I find it hard to imagine governments imposing rules on how people discuss moral philosophy with LLMs, let alone more niche issues that I care about, like the foundations of decision theory. Such rules would also be awkward to reconcile with free speech. A broader pause or ban on AI seems more likely to me.