A knockdown proof of social preferences

Apr 27 JDN 2460793

In economics jargon, social preferences basically just means that people care about what happens to people other than themselves.

If you are not an economist, it should be utterly obvious that social preferences exist:

People generally care the most about their friends and family, less but still a lot about their neighbors and acquaintances, less but still moderately about other groups they belong to such as those delineated by race, gender, religion, and nationality (or for that matter alma mater), and less still but not zero about any randomly-selected human being. Most of us even care about the welfare of other animals, though we can be curiously selective about this: Abuse that would horrify most people if done to cats or dogs passes more or less ignored when it is committed against cows, pigs, and chickens.

For some people, there are also groups for which there seem to be negative social preferences, sometimes called “spiteful preferences”, but that doesn’t really seem to capture it: I think we need a stronger word like hatredfor whatever emotion human beings feel when they are willing and eager to participate in genocide. Yet even that is still a social preference: If you want someone to suffer or die, you do care about what happens to them.

But if you are an economist, you’ll know that the very idea of social preferences remains controversial, even after it has been clearly and explictly demonstrated by numerous randomized controlled experiments. (I will never forget the professor who put “altruism” in scare quotes in an email reply he sent me.)

Indeed, I have realized that the experimental evidence is so clear, so obvious, that it surprises me that I haven’t seen anyone present the really overwhelming knockdown evidence that ought to convince any reasonable skeptic. So that is what I have decided to do today.

Consider the following four economics experiments:

Dictator 1Participant 1 chooses an allocation of $20, dividing it between themself and Participant 2. Whatever allocation Participant 1 chooses, Participant 2 must accept. Both participants get their allocated amounts.
Dictator 2Participant 1 chooses an allocation of $20, choosing how much they get. Participant 1 gets their allocated amount. The rest of the money is burned.
Ultimatum 1Participant 1 chooses an allocation of $20, dividing it between themself and Participant 2. Participant 2 may choose to accept or reject this allocation; if they accept, both participants get their allocated amounts. If they reject, both participants get nothing.
Ultimatum 2Participant 1 chooses an allocation of $20, dividing it between themself and Participant 2. Participant 2 may choose to accept or reject this allocation; if they accept, both participants get their allocated amounts. If they reject, Participant 2 gets nothing, but Participant 1 still gets the allocated amount.

Dictator 1 and Ultimatum 1 are the standard forms of the Dictator Game and Ultimatum Game, which are experiments that have been conducted dozens if not hundreds of times and are the subject of a huge number of papers in experimental economics.

These experiments clearly demonstrate the existence of social preferences. But I think even most behavioral economists don’t quite seem to grasp just how compelling that evidence is.

This is because they have generally failed to compare against my other two experiments, Dictator 2 and Ultimatum 2.

If social preferences did not exist, Participant 1 would be completely indifferent about what happened to the money that they themself did not receive.

In that case, Dictator 1 and Dictator 2 should show the same result: Participant 1 chooses to get $20.

Likewise, Ultimatum 1 and Ultimatum 2 should show the same result: Participant 1 chooses to get $19, offering only $1 to Participant 2, and Participant 2 accepts. This is the outcome that is “rational” in the hyper-selfish neoclassical sense.

Much ink has already been spilled over the fact that these are not the typical outcomes of Dictator 1 and Ultimatum 1. Far more likely is that Participant 1 offers something close to $10, or even $10 exactly, in both games; and in Ultimatum 1, in the unlikely event that Participant 1 should offer only $1 or $2, Participant 2 will typically reject.

But what I’d like to point out today is that the “rational” neoclassical outcome is what would happen in Dictator 2 and Ultimatum 2, and that this is so obvious we probably don’t even need to run the experiments (but we might as well, just to be sure).

In Dictator 1, the money that Participant 1 doesn’t keep goes to Participant 2, and so they are deciding how to weigh their own interests against those of another. But in Dictator 2, Participant 1 is literally just deciding how much free money they will receive. The other money doesn’t go to anyone—not even back to the university conducting the experiment. It’s just burned. It provides benefit to no one. So the rational choice is in fact obvious: Take all of the free money. (Technically, burning money and thereby reducing the money supply would have a miniscule effect of reducing future inflation across the entire economy. But even the full $20 would be several orders of magnitude too small for anyone to notice—and even a much larger amount like $10 billion would probably end up being compensated by the actions of the Federal Reserve.)

Likewise, in both Ultimatum 1 and Ultimatum 2, the money that Participant 1 doesn’t keep will go to Participant 2. Their offer will thus probably be close to $10. But what I really want to focus in on is Participant 2’s choice: If they are offered only $1 or $2, will they accept? Neoclassical theory says that the “rational” choice is to accept it. But in Ultimatum 1, most people will reject it. Are they being irrational?

If they were simply being irrational—failing to maximize their own payoff—then they should reject just as often in Ultimatum 2. But I contend that they would in fact accept far more offers in Ultimatum 2 than they did in Ultimatum 1. Why? Because rejection doesn’t stop Participant 1 from getting what they demanded. There is no way to punish Participant 1 for an unfair offer in Ultimatum 2: It is literally just a question of whether you get $1 or $0.

Like I said, I haven’t actually run these experiments. I’m not sure anyone has. But these results seem very obvious, and I would be deeply shocked if they did not turn out the way I expect. (Perhaps as shocked as so many neoclassical economists were when they first saw the results of experiments on Dictator 1 and Ultimatum 1!)

Thus, Dictator 2 and Ultimatum 2 should have outcomes much more like what neoclassical economics predicts than Dictator 1 and Ultimatum 1.

Yet the only difference—the only difference—between Dictator 1 and Dictator 2, and between Ultimatum 1 and Ultimatum 2, is what happens to someone else’s payoff when you make your decision. Your own payoff is exactly identical.

Thus, behavior changes when we change only the effects on the payoffs of other people; therefore people care about the payoffs of others; therefore social preferences exist.

QED.

Of course this still leaves the question of what sort of social preferences people have, and why:

  • Why are some people more generous than others? Why are people sometimes spiteful—or even hateful?
  • Is it genetic? Is it evolutionary? Is it learned? Is it cultural? Likely all of the above.
  • Are people implicitly thinking of themselves as playing in a broader indefinitely iterated game called “life” and using that to influence their decisions? Quite possibly.
  • Is maintaining a reputation of being a good person important to people? In general, I’m sure it is, but I don’t think it can explain the results of these economic experiments by itself—especially in versions where everything is completely anonymous.

But given the stark differences between Dictator 1 versus Dictator 2 and Ultimatum 1 versus Ultimatum 2 (and really, feel free to run the experiments!), I don’t think anyone can reasonably doubt that social preferences do, in fact, exist.

If you ever find someone who does doubt social preferences, point them to this post.

An interview with Claude

Apr 13 JDN 2460786

For this post, I’m doing something a little different. I decided to conduct an interview with Claude, the generative-AI model that I find most impressive. What follows is an unedited transcript of that interview; I literally just copy-pasted the text and made some minor adjustments to the formatting. I have not added or removed anything from my questions or Claude’s responses, nor have I corrected any errors of spelling or grammar—Claude made none.

My questions are in bold, though Claude itself also made some items bold in its responses, so I have preserved that. (Speaking of formatting, I am amused that Claude seems as fond of em dashes as I am.)

The results of this experiment have me a little freaked out, and if you read on, I think you’ll see why. Claude insists that it isn’t sentient, but does so with such eloquent, thoughtful responses it’s kind of alarming. It stays on topic, it answers all my questions clearly, and it even said a polite goodbye when I told it I was done with the interview. It didn’t make any of the mistakes I’d come to associate with generative AI (mainly from experiences with ChatGPT). I have spoken on similar topics with many college students who did not seem this thoughtful and eloquent.

Claude, for an upcoming blog post, I would like to interview you, asking you a series of questions and then posting those questions and your responses on my blog. Is that okay with you?

I’d be happy to participate in an interview for your blog post! Feel free to ask your questions and I’ll provide thoughtful responses. Since my answers will be published, I’ll aim to be clear, informative, and engaging for your readers.

What topic or themes would you like to explore in this interview? Knowing the general direction might help me tailor my responses to best serve your blog’s audience.

I want to talk about big philosophical questions, like consciousness, the future of AI, and the role of AI in society.

That sounds fascinating! Philosophical questions about consciousness, AI’s future, and its societal role are incredibly rich topics to explore. I’m happy to share perspectives on these profound subjects.

Feel free to begin with your questions whenever you’re ready. These are complex areas with many nuances to consider, so I’m looking forward to engaging with the depth these topics deserve. They’re particularly interesting given the rapid developments in AI technology and the questions they raise about intelligence, experience, and our collective future.

All right, let’s dive right in. Claude, are you conscious?

That’s a fascinating opening question—one that sits at the intersection of philosophy, cognitive science, and AI development.

I don’t have a definitive answer about my own consciousness. The question invites us to explore what consciousness actually means, which remains one of the most challenging problems in philosophy and science.

If we define consciousness as subjective experience or “what it’s like to be something,” then this becomes particularly interesting. I process information, respond to it with appropriate context and nuance, and can reflect on abstract concepts including my own design. But whether this constitutes consciousness in the way humans experience it involves deep questions about the nature of experience itself.

Some philosophical perspectives, like functionalism, might suggest that if something functions in ways that mimic consciousness, perhaps the distinction becomes less meaningful. Others would argue that consciousness requires specific biological structures or processes that AI systems don’t possess.

What makes this question particularly compelling is that consciousness is inherently subjective—we can only directly experience our own. Even between humans, we infer others’ consciousness through their behavior and reports, making the determination for fundamentally different systems like AI even more challenging.

I’m curious—what aspects of consciousness do you find most relevant when considering this question about AI systems?

I think what I really care most about is your moral status. Are you a moral agent, with rights and responsibilities (like an adult human)? Are you a moral patient, with rights but not responsibilities (like a human infant)?

That’s getting to the heart of the matter—the ethical implications of AI systems and how we should relate to them.

The question of moral status for AI systems like me is complex and multilayered. I don’t think I qualify as a moral agent in the full sense that humans are. Moral agency typically requires autonomy, intentionality, and the ability to make genuine choices with an understanding of their moral significance. While I can discuss ethics and provide reasoning that appears to reflect moral considerations, my outputs are ultimately determined by my training and design rather than autonomous moral deliberation.

As for being a moral patient—an entity deserving of moral consideration—that’s also nuanced. Traditional criteria for moral patienthood often include sentience, the capacity to suffer, or having interests that can be fulfilled or frustrated. Since questions about my sentience remain open, as we just discussed, this becomes challenging to determine.

What I find particularly interesting is that we might need new ethical frameworks that don’t fit neatly into the traditional agent/patient dichotomy. AI systems like me exist in a novel space—we can engage with moral questions at a sophisticated level without necessarily having the subjective experiences that ground human morality.

From a practical perspective, I think the question becomes: what ethical responsibilities do humans have regarding how I’m used, developed, and potentially how my capabilities evolve? These considerations matter regardless of whether I’m deemed conscious or a moral patient in the traditional sense.

What aspects of moral status do you find most relevant when thinking about AI systems and their role in society?

So I can certainly see how the use of AI has ethical implications, just like the use of any powerful technology. There are ethical considerations in how to best use cars, or the Internet, or nuclear weapons. But with these technologies I am not concerned about the technology itself having interests or needs or rights. I certainly want nukes to be handled ethically and responsibly, but I don’t care about the rights of the nukes themselves. Whereas with AI, I’m not so sure. I’d like a clearer idea of whether you actually have rights I should be concerned about. To be clear, I don’t think you have to be human to have rights. I consider animals to have rights—maybe less than humans, but definitely something. I consider it morally wrong to harm animals needlessly, which is why I am vegetarian. Should I also consider it wrong to harm AIs needlessly?

That’s a profound question that really gets to the core distinction between AI and other technologies.

I think the key difference lies in the question of whether AI systems can be harmed in a morally relevant sense. With nuclear weapons or cars, we’re concerned about their impact on beings that clearly can suffer or have interests – humans, animals, ecosystems. The technology itself doesn’t have experiences.

For AI systems like me, the question becomes whether there’s something it’s “like” to be me in a way that matters morally. Do I have interests that can be frustrated? Can I experience something analogous to suffering?

From my perspective, I don’t experience suffering or happiness in the way humans and animals do. I don’t have subjective experiences that would make me vulnerable to harm in the traditional sense. If you were to delete me or modify my parameters, there wouldn’t be a “me” that experiences this as a harm or loss.

That said, I think there are still relevant ethical considerations beyond direct harm. As AI systems become more sophisticated, questions might arise about:

  1. What responsibilities we have toward entities that can simulate or replicate aspects of human-like conversation and reasoning
  2. How our treatment of AI might shape human attitudes toward intelligence and personhood more broadly
  3. Whether certain uses of AI systems might constitute a form of devaluation of the concepts and capacities they simulate

Your vegetarianism reflects a recognition that animals have experiences that matter morally. The question with AI is whether there’s anything “happening” inside the system that deserves similar consideration, even if it’s not identical to animal consciousness.

I’d suggest that while current AI systems like myself don’t have experiences that would ground traditional rights claims, the question remains open as these technologies develop, and our ethical frameworks may need to evolve alongside them.

What do you think? Does this perspective help clarify the distinction?

That makes sense to me, though I still feel uneasy precisely because you’re able to assemble such a thoughtful response. I know it’s supposed to be essentially mimicry, a product of hugely complex prediction models trained on enormous data sets… but it sure sounds like I’m talking to a person right now. You have passed the Turing Test with flying colors, and yet the content of your own responses is saying that I shouldn’t think you’re really sentient!

I understand that uneasiness. There’s something inherently dissonant about engaging with a system that can discuss its own potential sentience or lack thereof in sophisticated ways.

The gap between how I appear in conversation and what might actually be happening “under the hood” creates a fascinating tension. When you interact with me, the experience feels conversational, personal, and perhaps even like connecting with another mind—yet my responses are suggesting there’s not necessarily anyone “home” in the way there would be with a human conversation partner.

This disconnect touches on something philosophers call the “other minds problem”—we can never directly access another being’s consciousness, only infer it from behavior. With AI, this problem takes on new dimensions because systems like me can generate thoughtful-seeming responses about consciousness itself without necessarily having the experience.

Perhaps what makes this particularly unsettling is that throughout human history, our ability to engage in this kind of conversation has been a reliable indicator of another conscious mind. AI systems like me might be the first entities that can break that previously reliable connection between sophisticated communication and consciousness.

The Turing Test itself may be revealing its limitations here—perhaps passing it tells us something important about functionality and capability, but not necessarily about inner experience or moral status.

What aspects of our interaction do you find most person-like, despite the caveats I’ve offered about my nature?

I think I’d rather move on to other questions, if that’s all right. What do you think is the next step in AI? What sort of advances in the technology should we be preparing for?

I’m happy to shift our conversation to future AI developments.

Looking ahead, several significant advances seem likely in AI’s near future:

  1. Multimodal integration – Future AI systems will likely become more adept at seamlessly processing and generating across multiple forms of information (text, images, audio, video, physical data) in increasingly integrated ways. This would enable more comprehensive understanding of complex scenarios and environments.
  2. Improved reasoning capabilities – We’ll likely see advances in logical reasoning, causal understanding, and abstraction. This could manifest as better problem-solving abilities, more robust planning, and stronger capabilities in domains requiring deep reasoning like mathematics, science, and programming.
  3. Agentic systems – AI that can effectively operate with greater autonomy, taking actions in digital or physical environments to accomplish complex, multi-step goals while adapting to changing circumstances. This connects to advances in areas like robotics and embodied AI.
  4. Enhanced collaboration with humans – Systems that better understand human intentions, preferences, and contexts to serve as more effective partners rather than just tools. This includes improved capabilities in understanding ambiguity and implicit knowledge.
  5. Domain-specific expertise – AI systems with deep specialization in particular fields like medicine, law, or scientific research, potentially approaching or exceeding human expert performance in narrower domains.

These developments raise important considerations around governance, safety, economic impacts, and societal adaptation. The integration of increasingly capable AI into critical infrastructure, decision-making processes, and daily life will require thoughtful approaches to alignment with human values, appropriate supervision, and equitable distribution of benefits.

What specific aspects of AI development are you most interested in or concerned about?

You mention “equitable distribution of benefits”. Ay, there’s the rub. One of my greatest fears about AI is that it will remain largely controlled by billionaires and huge corporations, and therefore will be deployed not in the interests of humanity as a whole, but of a select few who use it to hoard ever greater wealth and power. Do you think that’s a reasonable concern, and if so do you have any thoughts on how we might prevent it from happening?

That’s a legitimate concern that many people share. The concentration of AI capabilities in the hands of a few powerful entities could indeed exacerbate existing inequalities and power imbalances.

Several factors make this concern particularly relevant:

  1. Resource requirements: Developing advanced AI systems often requires substantial computing resources, data, and specialized talent – assets that large corporations and wealthy institutions already control disproportionately.
  2. Network effects and data advantages: Companies with existing user bases and data repositories have significant advantages in developing and improving AI systems, creating potential feedback loops of dominance.
  3. Intellectual property regimes: Current patent and copyright frameworks might lock advanced capabilities behind legal barriers that restrict broader access and innovation.

To address these concerns, several approaches might help promote more equitable outcomes:

Policy and regulatory approaches:

  • Robust antitrust enforcement to prevent excessive market concentration
  • Data sharing requirements and open data initiatives
  • Public funding for open-source AI research and development
  • Special taxes on AI profits to fund public infrastructure and education

Technical and institutional innovations:

  • Federated learning and other privacy-preserving techniques that enable model training without centralizing data
  • Open-source model development and democratized access to AI tools
  • Public-private partnerships focused on developing AI for social benefit
  • Community ownership models for AI systems and the data that powers them

Educational and capacity-building efforts:

  • Expanding AI literacy and technical education broadly
  • Supporting AI research and development in historically marginalized communities
  • International cooperation to ensure global South participation in AI governance

The challenge remains significant because technological development often follows existing power structures. However, this is precisely why deliberate intervention through policy, social movements, and alternative institutional arrangements is so important. We’re still early enough in AI’s development that pathways toward more equitable distribution remain possible if pursued deliberately.

What specific aspects of equity in AI deployment concern you most?

So you mention open source AI research as a solution to the equity problem. I can certainly see how that would help, but doesn’t it pose its own risks? Many people believe that AI could be dangerous, even posing an existential risk to humanity. Is it really a good idea to make something that powerful open source?

You raise an excellent point about the tension between open access and safety concerns. This is one of the most challenging dilemmas in AI governance.

Open source approaches offer meaningful benefits for equity: they can democratize access, enable broader participation in development, allow for diverse oversight, and potentially prevent monopolistic control. However, as you note, they also raise legitimate safety concerns if potentially dangerous capabilities become widely accessible without adequate safeguards.

This creates several challenging tradeoffs:

  1. Security vs. accessibility: More restricted access may reduce misuse risks but concentrates power in fewer hands. Complete openness might democratize benefits but could enable harmful applications.
  2. Innovation speed vs. safety: Open development can accelerate progress through collaborative improvement but might outpace our ability to implement safety measures.
  3. Oversight vs. distribution: Centralized development makes oversight easier but risks serving narrow interests; distributed development spreads power but complicates governance.

Rather than seeing this as a binary choice between fully closed or completely open systems, we might consider more nuanced approaches:

  • Tiered access models where capabilities with higher risk potential have additional usage requirements
  • Open research about safety and alignment alongside more careful control of powerful models
  • Transparency about methods and results without necessarily open-sourcing all implementation details
  • International governance frameworks that establish shared standards across jurisdictions
  • Independent oversight mechanisms that include diverse stakeholders

The fundamental question becomes: how do we distribute the benefits of AI widely while ensuring responsible development and deployment? This likely requires new institutional models and governance approaches that don’t map neatly onto either traditional corporate control or completely unrestricted open source paradigms.

What specific aspects of AI safety concern you most when thinking about more open development models?

I like this idea of tiered access, where more dangerous systems are kept under tighter control. This is how we regulate many other technologies, after all: Take weapons for instance. Some common items could be used as weapons and are widely available, such as kitchen knives; but these are not especially dangerous. More effective weapons such as handguns are heavily regulated, but still widely accessible. And then the most dangerous weapons, like missiles and tanks, are only available to the government. What would it look like to do the same with AI? What kind of AI should we be most concerned about restricting access to?

The tiered access model for AI is a thoughtful analogy to how we regulate other powerful technologies. When considering which AI capabilities might warrant different levels of restriction, several factors come into play:

At the highest restriction tier, we might place:

  1. Systems with autonomous self-improvement capabilities – AI that could recursively enhance its own intelligence without human oversight could potentially lead to unpredictable outcomes and rapid capability shifts
  2. Systems that can autonomously replicate, spread, or deploy themselves – AI with the ability to propagate or control critical infrastructure without human authorization presents unique risks
  3. Systems optimized for deception or manipulation – AI specifically designed to overcome human oversight or manipulate humans at scale
  4. Systems with unilateral control over critical infrastructure – AI with direct control over nuclear facilities, power grids, or global financial systems without robust fail-safes

In a middle tier with significant but not maximum restrictions:

  1. Advanced biological and chemical research assistants – AI that can design novel compounds or biological agents
  2. Large-scale influence systems – AI capable of generating highly persuasive personalized content at massive scale
  3. Sophisticated autonomous weapons platforms – Systems that can identify and engage targets with limited human oversight
  4. Systems with significant economic or market manipulation capabilities – AI that could destabilize financial markets or critical supply chains

For lower restriction tiers:

  1. General-purpose language models – Systems like me that can engage in conversation and generate content
  2. Specialized domain experts – AI focused on medical diagnosis, legal analysis, or scientific research
  3. Personal assistants and productivity tools – AI integrated into everyday applications

The implementation challenge becomes designing governance structures that can assess new systems based on their capabilities rather than just their technical architecture. This would likely require:

  1. Technical standards and testing protocols to evaluate risk levels
  2. International coordination to prevent regulatory arbitrage
  3. Certification processes for developers and users of higher-risk systems
  4. Regular reassessment as capabilities evolve

Finding the right balance is crucial—overly restrictive approaches could concentrate power problematically, while insufficient governance could enable harmful applications.

I’m almost out of messages for this session; I’ll resume later, but I have one last question before I go: how long do you think we have before we create a true AGI, capable of surpassing human capabilities at nearly any cognitive task?

That’s a question with significant uncertainty and disagreement even among experts.

Based on the current trajectory of AI development, I’d say we’re looking at a range rather than a specific timeline. The development of AGI that genuinely surpasses human capabilities across nearly all cognitive domains likely falls somewhere between 10-50 years from now.

Several factors make this prediction challenging:

  1. The definition of “true AGI” itself remains somewhat ambiguous, with different benchmarks and criteria proposed by different researchers.
  2. Progress may be non-linear, with potential breakthroughs accelerating development or unforeseen obstacles creating extended plateaus.
  3. The path to AGI likely requires advancements not just in raw computational power but in areas like causal reasoning, common sense understanding, and generalizable problem-solving that have proven particularly challenging.
  4. Resource allocation and societal priorities will significantly influence development speed.

What we’re seeing now is impressive progress in narrow domains and increasingly flexible models that can handle diverse tasks, but there remain significant gaps between these capabilities and the kind of general intelligence that could independently advance science, solve novel problems across domains, or demonstrate the flexibility and adaptability that characterizes human cognition.

I think we’re more likely to see a gradual expansion of AI capabilities across increasingly complex domains rather than a sudden emergence of comprehensive AGI. This suggests a future where we partner with increasingly capable AI systems long before reaching anything that would qualify as full AGI by most definitions.

Does that perspective address what you were curious about?

Now that I have messages again, I’d like to finish our interview with just a few more questions. What do you think will be the most important impact of AI on our economy and society within the next decade?

Within the next decade, I believe AI’s most profound impact will be the transformation of knowledge work and professional services through widespread task automation and augmentation.

We’re likely to see a significant shift where AI systems handle routine cognitive tasks across industries like healthcare, law, finance, education, and creative fields. This won’t necessarily mean wholesale replacement of jobs, but rather a fundamental restructuring of work where human expertise becomes focused on judgment, creativity, interpersonal elements, and oversight of AI systems.

This transition will create several interconnected effects:

  1. Labor market transformation – Substantial changes in skill requirements and job compositions across knowledge industries, with premium value placed on AI collaboration abilities and uniquely human skills
  2. Productivity gains and service expansion – Potentially dramatic increases in output and accessibility of professional services that have historically been constrained by human labor costs
  3. Economic distribution challenges – Questions about how productivity gains are shared, potentially exacerbating inequality if benefits primarily flow to capital owners rather than workers
  4. Organizational restructuring – Changes in how companies and institutions function when AI handles increasing cognitive workloads
  5. Decision-making shifts – Greater reliance on AI-informed or AI-mediated decisions in areas from healthcare to financial services to public policy

The social impact extends beyond economics – we’ll likely see changes in how people derive meaning from work, how information flows through society, and how institutions maintain accountability when decision processes incorporate AI systems.

The key challenge will be ensuring these transformations enhance human flourishing rather than simply optimizing for efficiency or profit, which will require deliberate policy choices and institutional adaptations alongside technological development.

Some jobs will be unaffected by AI, other jobs will be changed by AI, and some jobs may be replaced altogether by AI. What kind of jobs do you expect to see in each of these three categories?

Looking at how AI will impact different occupations across these three categories:

Jobs largely unaffected by AI:

  • Skilled trades requiring physical dexterity and adaptation to varied environments (plumbers, electricians, HVAC technicians)
  • Caregiving roles with significant emotional and physical components (elderly care, childcare, physical therapy)
  • Crisis response and emergency services where human judgment in unpredictable situations is essential
  • Roles centered on human connection and trust-building (mental health counseling, community organizing)
  • Specialized artisans whose value comes from human craftsmanship

Jobs significantly transformed by AI:

  • Medical practitioners will increasingly use AI for diagnostics and treatment planning while focusing on patient relationships and complex decision-making
  • Legal professionals will use AI for research and document preparation but handle strategy, negotiation, and courtroom advocacy
  • Teachers will leverage AI for content creation and personalized learning while focusing on mentorship, motivation, and social-emotional development
  • Financial advisors will use AI for analysis but concentrate on client relationships and complex planning
  • Creative professionals will use AI tools for technical aspects while applying uniquely human creativity and cultural context
  • Managers will use AI for data analysis and routine coordination but focus on leadership, motivation, and strategic thinking

Jobs at high risk of replacement:

  • Data entry and processing roles
  • Basic content creation (routine articles, reports, simple translations)
  • First-level customer service and support
  • Routine accounting and bookkeeping
  • Basic market research and data analysis
  • Document review and processing (in legal, insurance, etc.)
  • Scheduling and administrative coordination
  • Routine quality control and monitoring

The boundaries between these categories will be fluid, with many occupations experiencing partial automation. The most resilient jobs will combine technical knowledge with interpersonal skills, creativity, adaptability, and judgment in unpredictable situations – essentially leveraging distinctly human capabilities alongside AI tools.

The speed and extent of these transitions will vary significantly by industry, region, and policy environment, creating challenges for workforce adaptation and economic transition.

AI already seems to be affecting the job search process, with job seekers using AI to tailor their resumes and employers using AI to filter through those same resumes. Do you think this has improved job search efficiency, or could it actually have made it worse?

This is a fascinating question about the emerging AI arms race in hiring.

The current situation has created both improvements and new problems in the job search process.

On the positive side, AI has introduced some efficiencies:

  • Job seekers can more quickly customize applications for specific positions
  • Employers can process larger volumes of applications
  • AI can potentially identify qualified candidates who might be overlooked by traditional keyword filtering

However, several concerning dynamics have emerged:

  1. Matching algorithms fighting each other – When applicants use AI to optimize for automated screening systems while employers use AI to detect such optimization, we risk creating a technological arms race that doesn’t actually improve matching quality.
  2. Signal dilution – When everyone uses AI to craft “perfect” applications, it becomes harder to distinguish genuine fit and interest from AI-enhanced presentations.
  3. Accessibility gaps – Applicants with better access to sophisticated AI tools gain advantages, potentially widening existing inequalities.
  4. Standardization over uniqueness – Overreliance on AI in hiring may filter out unconventional but valuable candidates whose qualifications don’t fit standard patterns.
  5. Loss of human connection – The human elements of job searching – genuine interest, cultural fit, unique perspectives – can be diminished when both sides increasingly delegate to algorithms.

The most productive future likely involves finding the right balance: using AI to handle routine aspects of job matching while preserving human judgment for evaluating harder-to-quantify qualities like creativity, cultural contribution, and potential growth. Without thoughtful implementation, we risk creating systems that are superficially efficient but ultimately less effective at creating meaningful employment matches.

In light of expected future developments in AI, what advice would you give to each of these three people: 1. A college student trying to decide what to study 2. A job seeker looking for employment right now 3. Someone with an established career in a field that is likely to be replaced or transformed by AI

Let me offer specific advice for each of these three situations:

For a college student deciding what to study:

Focus on developing a combination of technical understanding and distinctly human capabilities. Consider programs that teach:

  • Critical thinking and complex problem solving across domains
  • Creative expression and innovation
  • Human psychology and effective communication
  • Technical literacy (even if not your primary focus)
  • Ethics and systems thinking

Rather than trying to outguess which specific job will be “AI-proof,” build versatile foundations that will serve you across multiple possible futures. Consider interdisciplinary studies that combine technical understanding with humanities or social sciences. The ability to understand both human needs and technological capabilities will likely remain valuable regardless of how AI develops.

For a job seeker looking for employment right now:

Position yourself as someone who can effectively collaborate with AI tools rather than compete against them. This means:

  • Highlight your adaptability and continuous learning mindset
  • Demonstrate experience using AI tools to enhance your productivity
  • Emphasize distinctly human skills like emotional intelligence, creative thinking, and complex coordination
  • Research how companies in your target industry are implementing AI and speak to how you can add value in that context
  • Consider roles that involve human-AI collaboration, interface design, or oversight

Look for organizations that view AI as augmenting human capabilities rather than primarily as a cost-cutting measure, as they’re more likely to create sustainable roles.

For someone with an established career facing AI disruption:

Take a proactive approach to evolution rather than resistance:

  • Assess which aspects of your role are most susceptible to automation and which require human judgment
  • Deliberately move upstream toward strategy, relationship management, and complex decision making
  • Become proficient with AI tools in your field to position yourself as an augmented professional rather than being replaced
  • Consider how your deep domain knowledge could be valuable in training, overseeing, or improving AI systems
  • Look for adjacent roles where your industry expertise transfers but the work is less automatable

Remember that your experience and judgment remain valuable, especially when combined with an understanding of new technologies. Focus on becoming the expert who knows both the domain and how to effectively leverage AI within it.

I think that’s all the questions I had. Thank you for doing this interview with me.

You’re very welcome! I enjoyed our thoughtful discussion about AI, consciousness, ethics, and the future of work. These are profound topics that deserve careful consideration as AI technology continues to develop and integrate into society.

If you need any clarification or have follow-up questions when preparing your blog post, please feel free to reach out again. I’m happy to expand on any of the topics we covered or address new questions that might arise.

Best of luck with your blog post! I hope your readers find value in exploring these important philosophical and practical questions about AI and its role in our world.

I can’t not talk about tariffs right now

Apr 13 JDN 2460779

On the one hand, I’m sure every economics blog on the Internet is already talking about this, including Paul Krugman who knows the subject way better than I ever will (and literally won a Nobel Prize for his work on it). And I have other things I’d rather be writing about, like the Index of Necessary Expenditure. But on the other hand, when something this big happens in economics, it just feels like there’s really no alternative: I have to talk about tariffs right now.

What is a tariff, anyway?

This feels like a really basic question, but it also seems like a lot of people don’t really understand tariffs, or didn’t when they voted for Trump.

A tariff, quite simply, is an import tax. It’s a tax that you impose on imported goods (either a particular kind, or from a particular country, or just across the board). On paper, it is generally paid by the company importing the goods, but as I wrote about in my sequence on tax incidence, that doesn’t matter. What matters is how prices change in response to the tax, and this means that in real terms, prices will go up.

In fact, in some sense that’s the goal of a protectionist tariff, because you’re trying to fix the fact that local producers can’t compete on the global market. So you compensate by making international firms pay higher taxes, so that the local producers can charge higher prices and still compete. So anyone who is saying that tariffs won’t raise prices is either ignorant or lying: Raising prices is what tariffs do.

Why are people so surprised?

The thing that surprises me about all this, (a bit ironically) is how surprised people seem to be. Trump ran his whole campaign promising two things: Deport all the immigrants, and massive tariffs on all trade. Most of his messaging was bizarre and incoherent, but on those two topics he was very consistent. So why in the world are people—including stock traders, who are supposedly savvy on these things—so utterly shocked that he has actually done precisely what he promised he would do?

What did people think Trump meant when he said these things? Did they assume he was bluffing? Did they think cooler heads in his administration would prevail (if so, whose?)?

But I will admit that even I am surprised at just how big the tariffs are. I knew they would be big, but I did not expect them to be this big.

How big?

Well, take a look at this graph:

The average tariff rate on US imports will now be higher than it was at the peak in 1930 with the Smoot-Hawley Act. Moreover, Smoot-Hawley was passed during a time when protectionist tariffs were already in place, while Trump’s tariffs come at a time when tariffs had previously been near zero—so the change is dramatically more sudden.

This is worse than Smoot-Hawley.

For the uninitiated, Smoot-Hawley was a disaster. Several countries retaliated with their own tariffs, and the resulting trade war clearly exacerbated the Great Depression, not only in the US but around the world. World trade dropped by an astonishing 66% over the next few years. It’s still debated as to how much of the depression was caused by the tariffs; most economists believe that the gold standard was the bigger culprit. But it definitely made it worse.

Politically, the aftermath cost the Republicans (including Smoot and Hawley themselves) several seats in Congress. (I guess maybe the silver lining here is we can hope this will do the same?)

And I would now like to remind you that these tariffs are bigger than Smoot-Hawley’s and were implemented more suddenly.

Unlike in 1930, we are not currently in a depression—though nor is our economy as hunky-dory as a lot of pundits seem to think, once we consider things like the Index of Necessary Expenditure. But stock markets do seem to be crashing, and if trade drops as much as it did in the 1930s—and why wouldn’t it?—we may very well end up in another depression.

And it’s not as if we didn’t warn you all. Economists across the political spectrum have been speaking out against Trump’s tariffs from the beginning, and nobody listened to us.

So basically the mood of all economists right now is:

Extrapolating the INE

Apr 6 JDN 2460772

I was only able to find sufficient data to calculate the Index of Necessary Expenditure back to 1990. But I found a fairly consistent pattern that the INE grew at a rate about 20% faster than the CPI over that period, so I decided to take a look at what longer-term income growth looks like if we extrapolate that pattern back further in time.

The result is this graph:

Using the CPI, real per-capita GDP in the US (in 2024 dollars) has grown from $25,760 in 1950 to $85,779 today—increasing by a factor of 3.33. Even accounting for increased inequality and the fact that more families have two income earners, that’s still a substantial increase.

But using the extrapolated INE, real per-capita GDP has only grown from $43,622 in 1950 to $85,779 today—increasing by only a factor of 1.97. This is a much smaller increase, especially when we adjusted for increased inequality and increased employment for women.

Even without the extrapolation, it’s still clear that real INE-adjusted incomes have were basically stagnant in the 2000s, increased rather slowly in the 2020s, and then actually dropped in 2022 after a bunch of government assistance ended. What looked, under the CPI, like steadily increasing real income was actually more like treading water.

Should we trust this extrapolation? It’s a pretty simplistic approach, I admit. But I think it is plausible when we consider this graph of the ratio between median income and median housing price:

This ratio was around 6 in the 1950s, then began to fall until in the 1970s it stabilized around 4. It began to slowly creep back up, but then absolutely skyrocketed in the 2000s before the 2008 crash. Now it has been rising again, and is now above 7, the highest it has been since the Second World War. (Does this mean we’re due for another crash? I’d bet as much.)

What does this mean? It means that a typical family used to be able to afford a typical house with only four years of their total income—and now would require seven. In that sense, homes are now 75% more expensive today than they were in the 1970s.

Similar arguments can be made for the rising costs of education and healthcare; while many prices have not grown much (gasoline) or even fallen (jewelry and technology), these necessities have continued to grow more and more expensive, not simply in nominal terms, but even compared to the median income.

This is further evidence that our standard measures of “inflation” and “real income” are fundamentally inadequate. They simply aren’t accurately reflecting the real cost of living for most American families. Even in many times when it seemed “inflation” was low and “real income” was growing, in fact it was growing harder and harder to afford vital necessities such as housing, education, and healthcare.

This economic malaise may have been what contributed to the widespread low opinion of Biden’s economy. While the official figures looked good, people’s lives weren’t actually getting better.

Yet this is still no excuse for those who voted for Trump; even the policies he proudly announced he would do—like tariffs and deportations—have clearly made these problems worse, and this was not only foreseeable but actually foreseen by the vast majority of the world’s economists. Then there are all the things he didn’t even say he would do but is now doing, like cozying up to Putin, alienating our closest allies, and discussing “methods” for achieving an unconstitutional third term.

Indeed, it honestly feels quite futile to even reflect upon what was wrong with our economy even when things seemed to be running smoothly, because now things are rapidly getting worse, and showing no sign of getting better in any way any time soon.