How to think about AI's future without falling for hype or dismissal
In 2016, a prominent AI researcher predicted that AI would be better than radiologists at reading X-rays within five years and that training radiologists was therefore unwise. By 2023, AI was genuinely better than average radiologists on specific imaging tasks in specific conditions — and hospitals were hiring more radiologists than ever.
Both things were true: the technical claim and the employment outcome. The error was in the prediction model — assuming that technical capability in a narrow task would translate directly to job replacement. The capability arrived on schedule. The prediction was still wrong.
AI capability trajectories are difficult to predict because capability improvements are uneven, context-dependent, and often non-linear. A system that improves steadily on a benchmark may hit a wall when deployed in a messier real-world context. A system that appears to plateau on benchmarks may have sudden capability jumps with scale. Neither linear extrapolation nor "this time is different" thinking reliably predicts AI development trajectories.
The translation from capability to impact is even harder to predict. Technical capability is one input into a complex system that also includes: economic incentives to deploy, regulatory environments, social acceptance, complementary infrastructure, and the behavioral responses of people and institutions. The radiologist case illustrates this — technical capability in image reading was demonstrated, but the system it would replace is embedded in healthcare workflows, liability structures, training pipelines, and institutional inertia that AI capability alone cannot reorganize.
The Hype Cycle Problem
AI regularly cycles through periods of inflated expectation followed by disillusionment, and then more gradual adoption. The pattern is consistent enough that Gartner formalized it as the "hype cycle." Maintaining calibrated views — neither dismissing genuine capability improvements nor accepting every prediction at face value — is the intellectual challenge for anyone trying to understand AI's trajectory.
Setting aside specific predictions, some things about AI's near-term trajectory are well-supported by current evidence. Continued capability improvement in language, reasoning, and multimodal tasks — the trajectory of the last decade shows no sign of stopping, though the rate is uncertain. Widening deployment across sectors currently using early AI tools — more industries, more applications, more integration into existing systems. Continued concentration of frontier AI development in a small number of very large organizations with significant compute resources. Governance gap persistence — regulatory frameworks are not keeping pace with capability development, and this gap is likely to widen before it narrows.
Find a specific AI prediction made 3–7 years ago — about job displacement, AI capabilities, specific industry impacts, or timelines for achieving AI milestones. Evaluate: Was the prediction accurate? If it was wrong, what did it get wrong — the capability trajectory, the translation from capability to impact, or the timeline? What would a more calibrated prediction have looked like?
Start with: "I want to evaluate this AI prediction from [year]: [the prediction]. Here's how it turned out: [your assessment]"
Who controls the most powerful AI systems — and what that means for everyone else
Five companies controlled the majority of frontier AI development. Each had more compute than most national governments. Each trained on data representing most of human written output. Each had products used by hundreds of millions of people daily.
The CEOs of three of them testified before Congress. The senators asked questions that revealed they had limited technical understanding of what they were regulating. The companies' lawyers noted that proposed regulations would primarily harm smaller competitors. The regulators' staffs had fewer technical experts than the companies' government affairs teams. The power asymmetry was not between companies and regulators. It was between those who understood what they were governing and those who were trying to govern it.
Frontier AI development — training the most capable models — requires compute resources that only a handful of organizations can afford. The capital requirements for training runs of frontier models are in the hundreds of millions to billions of dollars. This creates a structural concentration: the organizations shaping the trajectory of the most consequential AI technology are a small number of US tech companies, a handful of Chinese companies, and a few well-capitalized startups.
This concentration has several implications. Value alignment: the values embedded in frontier AI systems — what they will and won't do, what perspectives they reflect, what content they treat as acceptable — are determined primarily by a small number of organizations whose values and incentives may not represent the diversity of people who use these systems. Geopolitical concentration: the most powerful AI systems are controlled by US and Chinese entities, creating a technology dependency for the rest of the world. Regulatory capture risk: organizations with resources to engage regulators and provide technical expertise have significant influence over the governance of AI that affects everyone.
The Compute Moat
Access to high-end AI training hardware (primarily NVIDIA GPUs) has become a strategic constraint on who can build frontier AI. Export controls on advanced chips — a US government policy to prevent China from accessing NVIDIA's most advanced hardware — have made compute a geopolitical asset. The organizations and countries with access to advanced compute have structural advantages in frontier AI development that are difficult to overcome through other means.
The regulatory challenge is compounded by genuine technical complexity. AI systems exhibit behaviors that their developers don't fully understand or predict. Governing these systems requires technical expertise that most regulatory bodies lack — and that the regulated companies have in abundance. This creates a structural information asymmetry that favors the regulated over the regulator.
The responses to this asymmetry: Regulatory capacity building — hiring technical experts in regulatory agencies, as the EU AI Office has done. Mandatory disclosure — requiring AI companies to share information that regulators need but couldn't independently obtain. Independent evaluation — government-funded AI safety research and third-party auditing that doesn't depend on companies' self-reporting. International coordination — building shared governance frameworks across jurisdictions to prevent regulatory arbitrage.
Choose a specific domain where AI power concentration creates governance challenges — frontier model development, AI chip supply chains, AI-generated content moderation, or AI in financial systems. Analyze: Who holds power in this domain? What are the governance challenges their concentration creates? What specific governance mechanisms would address the asymmetry?
Start with: "I want to analyze power concentration in [domain] — here's who holds power and the governance challenges that creates: [your analysis]"
What it means to remain the author of your own choices when AI shapes the options
The navigation app showed three routes. All three had been calculated by an algorithm. The fastest route avoided the residential neighborhoods where slower drivers lived. The algorithm had learned that routing through those neighborhoods increased travel time — so it avoided them, routing traffic through other residential streets instead.
Nobody had decided this was fair. Nobody had decided it wasn't. The algorithm was optimizing for individual travel time, and the redistribution of traffic noise, pollution, and safety risk across neighborhoods was an externality it wasn't designed to consider. The driver chose from three routes. The choice was real. The set of choices had been designed — by an algorithm, for a purpose that had nothing to do with neighborhood equity.
Human agency — the capacity to make meaningful choices about one's own life — depends not just on the freedom to choose but on the quality of the options available to choose from. AI systems increasingly shape the option set: what information is available, what products are presented, what routes exist, what opportunities are surfaced. The choices we make are real, but the architecture of those choices is increasingly AI-designed.
This creates a new kind of power asymmetry. The designers of AI systems make choices about what options to present, what to optimize for, and what externalities to ignore — choices with collective consequences that are invisible to individual users making individual decisions. The driver chooses from three routes; the city's traffic distribution is determined by the algorithm's design, not by any democratic deliberation about how traffic should be routed across neighborhoods.
Designed Defaults
Default settings in AI systems — what the system does when you don't actively choose otherwise — have enormous power over behavior. Research consistently finds that most people accept defaults most of the time. AI systems that default to collecting maximum data, sharing information widely, or optimizing for engagement impose those choices on users who never actively decided to accept them. Defaults are design choices with collective consequences.
Individual agency in an AI-shaped world requires active rather than passive engagement with AI systems. Several practices support this: Legibility — understanding enough about how AI systems work to recognize when they are shaping your options and how. Friction as feature — deliberately introducing friction into AI-mediated decisions to create space for reflection rather than default acceptance. Opting out — where possible, choosing alternatives to AI-mediated options for decisions that matter most to you. Collective action — recognizing that individual choices about AI systems are less powerful than collective decisions about how those systems should be designed and governed.
The deeper point: human agency in relation to AI is not primarily an individual challenge. It is a political and collective one. The choices that matter most — how AI systems are designed, what they optimize for, what defaults they set, whose interests they serve — are not made by individual users. They are made by the organizations that build and deploy AI systems, and shaped by the governance structures that constrain or enable those choices. Maintaining human agency at the societal level requires governance, not just individual vigilance.
Choose an AI system that shapes choices in your life — a navigation app, a streaming recommendation system, a social media feed, a job search platform, or a dating app. Analyze: What options does it present? What does it optimize for? What externalities of that optimization affect others? What default settings does it apply? What choices did you not make that the system made for you?
Start with: "I want to analyze how [AI system] shapes my choices — here's what it presents, optimizes for, and decides by default: [your analysis]"
The course synthesis: what understanding AI in society commits you to
The student had completed eight modules of AI in Society. She knew more than she had — about how AI shapes information environments, about what it does to work and democracy and health and privacy and inequality and the environment. She knew more about what is at stake and who bears the costs.
What she didn't yet know — and what no curriculum could simply tell her — was what to do with that knowledge. That question turned out not to be a question about AI at all. It was a question about what kind of person she wanted to be, and what kind of society she wanted to live in.
The course has covered AI's effects across eight domains — work, democracy, privacy, health, inequality, the environment, and the future. The through-line across all of them: AI's social effects are not determined by the technology itself but by the choices made about how it is designed, deployed, and governed. Those choices are not inevitable. They are made by specific people and organizations, constrained or enabled by specific governance structures, reflecting specific values and interests.
Understanding this creates a form of responsibility. Not guilt — the systems that produce these outcomes are not individual failures — but responsibility in the sense of being the kind of person who sees the choices rather than just the technology. Who asks whose interests are served. Who looks for the governance gaps. Who questions the claims. Who considers the people bearing costs while others capture benefits.
Informed Citizenship
Democratic governance of AI requires citizens who understand enough to participate meaningfully in public decisions about it. Not technical experts — but people who understand AI's social effects, who recognize governance choices as choices, who can evaluate competing claims about AI's benefits and harms. This course has tried to build that understanding. What happens with it is yours to determine.
Across eight modules, three dispositions are worth keeping for navigating AI in society:
Specificity over abstraction. "AI will change everything" is not useful. "This specific AI system makes this specific decision about these specific people, with these specific consequences for those people and these specific benefits to someone else" is useful. The social analysis of AI lives in the specifics, not the abstractions.
Follow the power. When understanding AI's social effects, the most useful question is often: who benefits and who bears the costs? The answer almost always reveals something about power — who had the resources to build the system, who has the power to contest it, whose interests the governance framework was designed to protect. AI in society is partly a story about power. Following the power tells you where the governance questions are.
Governance is a choice. Every status quo in AI governance is a choice — including the choice to have no governance. The gaps, the oversight absences, the unenforced rules are not accidental features of a complicated technology. They are the result of specific political, economic, and institutional choices. Which means they can be made differently. The question is whether enough people understand that clearly enough to demand it.
Course Complete
You've studied AI's effects on work, democracy, privacy, health, inequality, the environment, and human agency. The understanding you've built is genuinely useful — for the choices you make as a worker, consumer, citizen, and professional. What you do with it from here is the part no curriculum can determine. That's yours.
Across eight modules you've studied AI's effects on work, democracy, privacy, health, inequality, the environment, and human agency. Now synthesize: What is the most important thing you've learned? What has changed in how you think about AI? What specific governance gap or AI application concerns you most — and what would you do about it if you had the power?
Start with: "The most important thing I've learned in this course is [your key insight] — and it changes how I think about AI because [your reasoning]"
15 questions. Complete all to finish the module.