Success Modes: What Going Well Actually Looks Like | The AI Transition Tutorial

The Case for Optimism

If AI is powerful and well-handled, it could produce substantial good. The specific goods are worth naming, because "AI might be good" is too vague to aim for.

A partial list:

Scientific acceleration

AI that usefully assists research could speed progress on:

Drug discovery and development
Neglected diseases (malaria, TB, parasitic illnesses)
Materials science (better batteries, cheaper solar, better catalysts)
Climate adaptation and mitigation technology
Fundamental physics and mathematics
Cognitive science and neurology

AlphaFold (2020-) is the clearest existing example. A protein structure problem that took years per protein now takes minutes. The same sort of acceleration, applied more broadly, has real stakes for human welfare.

Healthcare

Diagnostic assistance, research assistance, personalised medicine, better access to medical knowledge for people far from specialists. Mental health support (contested but potentially significant if done carefully). Drug design and evaluation.

Education

Tutoring that meets individual students at their level, in their language, for free. This is the Khan-Academy-plus-AI story, and early versions are starting to work. If scaled carefully, it could substantially expand access to quality education.

Productivity and abundance

Higher productivity across many industries produces (in principle) more output for the same input. The distribution of those gains is a political question, but the gains themselves are real.

Access to expertise

Most of humanity lives far from competent doctors, lawyers, engineers, financial advisers, psychologists. AI that provides meaningful assistance in these domains extends expertise where it wasn't available. Imperfect AI that's better than nothing is a real improvement.

Acceleration of public goods

Climate solutions, disease control, infrastructure, education, governance research: traditionally underfunded public goods may benefit if AI makes producing them cheaper.

The Required Conditions

Good outcomes don't arrive automatically. They require specific conditions:

Technical

Alignment keeps pace with capability. Research progress on alignment is strong enough that each capability jump is accompanied by adequate safety work
Deployment practices are responsible. Firms deploy with evaluations, staged rollouts, and genuine attention to risks
Interpretability matures. We understand what AI systems are doing well enough to trust them appropriately

Institutions adapt. Governments, courts, professions develop the capacity to govern AI sensibly
Benefits spread. Economic gains reach broad populations, not just narrow owners of capital
Distribution of decision power. No single firm or government controls the trajectory
International coordination. Major powers cooperate on the most dangerous capabilities rather than racing without guardrails

Cultural

Public understands what's happening. Literacy about AI, not uniformly expert but distributed
Human agency is preserved. Humans remain meaningfully in the loop on important decisions
Epistemics are maintained. Shared understanding of reality doesn't fragment
Values can still evolve. Moral and political progress remains possible

Each of these has opposing pressures. Doing well requires active effort on each.

What Augmented Human Agency Looks Like

A specific vision of success worth articulating in detail:

Workers use AI as a serious tool. A middle-career professional in any field is many times more productive than the same person would have been a decade earlier. Mundane parts of their job are automated; they focus on what they bring that AI can't
Students learn with tutors that adapt to them. Gap between children with access to excellent education and those without narrows
Patients get better information about their conditions, more timely specialist consultation, personalised treatment
Citizens have access to legal assistance, tax advice, government-service navigation previously reserved for the wealthy
Researchers collaborate with AI on problems humans alone couldn't solve, advancing fields that had stalled
Creators have tools that dramatically expand what they can make; the floor of creative work rises; the ceiling rises too

This isn't utopia. It's a real possibility worth aiming for. It also isn't automatic.

What Distribution Looks Like

Good outcomes include broadly distributed benefits:

Geographic

AI services reach rural areas, low-income countries, underserved populations. Not just a toy for the rich world.

Economic

Productivity gains become wage gains, lower prices, better services, not just higher share prices. This requires active labour market policy, antitrust, tax policy.

Professional

AI augments (not replaces) most professionals; the professions reconfigure rather than collapse. Junior career paths exist though they look different.

Democratic

AI's decisions affecting many people are accountable to those people through recognisable governance processes. Governance itself is participatory and pluralistic.

Epistemic

Many AI systems with different values exist. Users have meaningful choice. No single system dominates public epistemics.

What Preserved Institutions Look Like

A particular success mode: institutions adapt rather than collapse.

Democratic governance continues to be legitimate and functional
Courts adjudicate AI-related disputes using adapted but recognisable doctrine
Professional standards evolve to incorporate AI without abandoning their ethical core
International coordination produces meaningful rules without requiring world government

This sounds less thrilling than "transformation". It may be the more important success criterion. Institutions are how civilisations metabolise change; preserved-and-adapted institutions are how we get good outcomes at all.

What Good Alignment Looks Like as Capability Grows

An underappreciated success condition: alignment scaling with capability.

This means:

We can specify what we want clearly enough for systems to follow
Systems reliably pursue the specified goal rather than gaming it
We can detect misalignment before deployment at consequential scale
We have the interpretability to verify alignment, not just test behaviour
Alignment techniques generalise to new capabilities

Current techniques partially achieve this for current systems. Continuing to achieve it as systems become more capable is research work that may or may not succeed. Success requires actual investment.

Specific Wins to Watch For

Indicators that things are going well:

Measurable science acceleration

Drug approvals accelerating, clinical trial enrollment and success improving, fundamental research producing previously intractable results. Specific benchmark problems falling to AI assistance.

Benefit distribution

Wages in broad swaths of the population rising. Access to services improving in low-income and developing countries. Educational outcomes improving in measurable ways.

Institutional adaptation

Thoughtful AI regulation passed without either suppressing innovation or leaving dangerous gaps. International coordination on specific capabilities (e.g., frontier model training). Court decisions that establish workable precedent.

Responsible lab behaviour

Labs publishing meaningful safety research, investing seriously in alignment, supporting third-party evaluation, making realistic commitments about capability deployment.

Genuine public engagement

AI literacy improving in the general public. Journalists and commentators increasingly precise in their coverage. Policy processes producing better rules over time.

No catastrophic misuse events

Time passing without catastrophic bioweapons attacks, autonomous-weapons-triggered conflicts, or other low-probability but severe outcomes.

The Realistic Middle

Most likely outcome, if the transition continues: a mixture.

Some domains experience dramatic acceleration; others don't
Benefits partially spread; inequality also rises
Some institutions adapt; others struggle
Some failure modes are avoided; others are not
On average, humans are meaningfully better off; specific populations and regions are worse off

This isn't pure success. It's also not failure. The challenge is maximising the better parts and limiting the worse parts. Where individual action can push this balance, it's worth pushing.

Why Success Isn't Certain

Good outcomes require:

Sustained investment in alignment through periods when it would be more profitable to cut corners
Active governance by actors who could instead be captured
Coordination among actors with competing incentives
Patience from markets and publics during uncertain periods
Luck with the underlying technology's properties

Any of these can fail. The good outcomes aren't the baseline; they require work.

This is why optimists aren't necessarily naive: they often know exactly what's required, and are betting that enough of it happens. The bet could be right or wrong.

Success Isn't Quiet

A subtle point. Some failure modes (gradual disempowerment) look like success in the short term. Some success modes (active contestation, regulation, alignment research) look like friction in the short term.

A smooth trajectory where AI keeps scaling and nothing obviously bad happens for a decade could be either good (alignment working) or bad (gradual disempowerment advancing). A bumpy trajectory with regulatory friction, visible mistakes, public contestation, and slow deployment could be worse (suppression of a useful technology) or better (course corrections landing in time).

Judging outcomes requires looking past superficial smoothness.

What Individuals Can Do

Many success conditions require collective action. Some things individuals contribute to:

Stay literate. The tutorial's frame is one attempt at this
Support serious work. Fund or work for efforts producing good outcomes
Participate in governance. Engage with policy processes; don't leave it to lobbyists
Use AI responsibly. Be a customer and user who makes responsible behaviour easier
Help others become literate. Share good models of thinking; push back on hype and doom in equal measure
Build useful things. Much of the benefit comes from specific applications; building one well is a contribution

Chapter 11 goes into this more.

The Honest Summary

Success is possible. It's not guaranteed. The requirements are specific and multiple. Expect mixed outcomes; push for better ones in the specific places you can.

If you only read scary things about AI, you may miss how tractable some of the upside is. If you only read optimistic things, you may miss how much has to go right for that upside to materialise. The honest synthesis: real possibility, real work, real uncertainty about the final balance.

Common Pitfalls

"Success means utopia." Probably not. Success means "better than now on most dimensions, with specific problems addressed". Utopian framings are less useful than concrete ones

"If it goes well for the rich, it went well." No. A success that preserves and extends existing inequality is at best a partial success

"The market produces success automatically." No. Markets can produce good outcomes; they don't always. Institutional design matters

"Talking about success distracts from risk." Not obviously. If you can't describe what going well looks like, you can't tell whether you're on the path to it

"I'm too pessimistic to help." Pessimism doesn't preclude action; it can motivate it. "Given I think this is dangerous, what is the useful thing to do?" is a valid question

Next Steps

Continue to 09-reading-the-discourse.md for how to read AI commentary honestly.