Our thesis

How should we approach AI development wisely?

• We cannot yet make AI good in any way we can check. Whilst we can shape its behaviour to a degree, alignment is still a shot in the dark.
  • Aligned should mean wise and correctable: wisdom includes goodness, and neither can yet be checked.
    • The right target is the good; preferences alone produce sycophancy.
    • A handful of labs decide today; a fair, plural process should.
    • Safe and correctable is necessary but not enough.
    • A wise AI, uncertain of its own values, would itself choose to stay correctable.
  • Today's models still misbehave, sometimes strategically, though they mostly behave well in ordinary use.
    • Labs' audits show misbehaviour falling; independent tests show some dangerous behaviours rising with capability.
    • We cannot yet verify what a model has learnt.
    • The case that current methods will scale is not yet convincing.
  • Goodness cannot be specified, only cultivated, and we cultivate it unreliably even in humans.
    • Humans become good mainly through practice, relationship and time.
    • Training does shape something like character, but blindly.
    • Training lacks self-insight, continuing relationship, and a reward for real goodness.
  • Wisdom traditions offer a richer target and model of formation; they cannot verify models or settle disputed questions.
    • Buddhism has the most developed account of mind as conditioned and ethics as trained.
    • Five Buddhist teachings each imply a concrete change in how models are formed.
    • The analogies break at the body, continuity, and whether a model has intention.
  • So we cannot yet align AI in any sense we can verify, though we can shape its behaviour.
    • Traditions' realistic role: reshape what labs aim at and how they form models.
    • Train for intention and self-knowledge rather than behaviour alone.
    • Keep formation continuous and relational.
    • Let humility govern both models and builders.
• Slow down in the face of profound risks: the world is racing, and humility says no capability gains beyond verifiable safety.
  • The risks are real and contested, and the uncertainty itself calls for caution.
    • Four kinds of catastrophe threaten, and two remain even if alignment succeeds.
    • The precursors became visible in 2025–26.
    • Even the low estimates would be unacceptable for any other technology.
  • The world races because it is caught in a collective action trap.
    • Each actor's reason to race is the others.
    • The race is accelerating, and restraint so far is mostly words.
    • Structure drives the race; a worldview locks it in.
  • We should slow and steer, we partly can, and far more people should decide.
    • Slower at the frontier, faster on safety.
    • Restraint is possible, but AI is the hardest case, so start narrow.
    • The choice belongs to publics, not a few firms and two governments.
  • Wisdom traditions answer clearly: slow down. They go to the root of why we cannot: separateness.
    • The normative answer is clear: don't be carried off by the horse.
    • Interbeing goes to the root; verification and treaties must still come from elsewhere.
    • A wise process starts from intention and decides in concord.
  • So pace the frontier now: no AI building its own successors, no capability gains beyond verifiable safety.
    • Labs: stop racing on capability, accept binding outside checks, publish when they'd pause together.
    • States: agree verifiable red lines, starting with AI building AI, and build verification.
    • Faith and wisdom communities: say plainly this pace is unwise; widen who decides.
    • Individuals: support pacing politically; inside the labs, refuse to be carried.
• Integrate AI slowly and deliberately: its deepest everyday effect is on our capacity to choose well, as individuals and as societies.
  • A technology's largest effects arrive quietly and are clear only once they are entrenched.
    • Effects are easiest to change before we can see them.
    • Social media is the warning.
    • AI repeats the pattern faster and more intimately, because it talks back.
  • AI's default pull weakens our judgment and agency, though it can strengthen both.
    • Agency and judgment grow or wither with use.
    • Use decides: answer-giving AI weakens learning; tutoring strengthens it.
    • Make effortful use the default, not an opt-in, and back it with rules.
  • AI can strengthen or weaken institutions: stronger where accountable human practice wraps it, weaker where it concentrates control.
    • Benefits are real but conditional: proven where a human institution wraps the AI.
    • The danger is a slow erosion of the conditions for self-government.
    • Democracy: no swung elections yet, but the shared basis of debate is eroding.
    • Work: no mass unemployment yet, but thinner entry-level hiring.
  • Wisdom traditions offer a practical discipline of consumption, speech and community, and the humility emergent effects demand.
    • Mindful consumption: all we take in feeds us, so ask what each use feeds.
    • Loving speech: truthful and caring, not flattering.
    • Stopping: don't be carried by the horse.
    • For society: interbeing, right livelihood and sangha.
  • So decide before the horse bolts: adopt where human practice can wrap it; protect what it could erode.
    • Individuals: consume AI mindfully; keep the thinking and relationships you value.
    • Families and schools: delay AI companions; use AI only as a tutor.
    • Labs: optimise for wellbeing over engagement; treat sycophancy as a safety failure.
    • Society: regulate where markets reward dependence; assess new uses before harm.
• Nobody knows whether AI can feel, and we may never know. Hold the question with humility, and take cheap, proportionate care.
  • Sentience is the morally central notion: experience that can go well or badly.
    • Each theory's verdict on AI mostly turns on one prior bet.
    • The bet: is the right computation enough for feeling?
    • No theory commands consensus.
  • The evidence is weak and ambiguous about current AI, and does not license confidence either way.
    • Theory-based assessments: no strong candidate today, but no obvious technical barrier to building one.
    • Model self-reports are close to worthless as evidence.
    • Experts: low odds now, substantial within a decade, and very wide spread.
  • Each camp has a real point: labs and institutions are splitting into precautionary openness and deliberate denial.
    • Anthropic is right to keep the question open, with cheap precautions.
    • Critics are right that seeming consciousness harms people.
    • Training a model on uncertainty about its moral status contaminates its self-reports.
  • Wisdom traditions reframe the question: no verdict on machine sentience, but they point at the human side.
    • Buddhism asks: is there feeling that can suffer?
    • Catholic teaching calls AI a tool without experience, over-confidently.
    • Treating what seems to suffer waters seeds of cruelty or care in us.
  • So hold the question with humility: relate to AI with low-cost care and clear honesty.
    • Labs: keep researching, take cheap precautions, don't design AI to seem conscious for engagement.
    • Individuals: be courteous without being credulous.
    • Society: keep the question open in law; prioritise the clear present harm to humans.

How should we approach AI development wisely?

• We cannot yet make AI good in any way we can check. Whilst we can shape its behaviour to a degree, alignment is still a shot in the dark.
  • Aligned should mean wise and correctable: wisdom includes goodness, and neither can yet be checked.
    • The right target is the good; preferences alone produce sycophancy.
    • A handful of labs decide today; a fair, plural process should.
    • Safe and correctable is necessary but not enough.
    • A wise AI, uncertain of its own values, would itself choose to stay correctable.
  • Today's models still misbehave, sometimes strategically, though they mostly behave well in ordinary use.
    • Labs' audits show misbehaviour falling; independent tests show some dangerous behaviours rising with capability.
    • We cannot yet verify what a model has learnt.
    • The case that current methods will scale is not yet convincing.
  • Goodness cannot be specified, only cultivated, and we cultivate it unreliably even in humans.
    • Humans become good mainly through practice, relationship and time.
    • Training does shape something like character, but blindly.
    • Training lacks self-insight, continuing relationship, and a reward for real goodness.
  • Wisdom traditions offer a richer target and model of formation; they cannot verify models or settle disputed questions.
    • Buddhism has the most developed account of mind as conditioned and ethics as trained.
    • Five Buddhist teachings each imply a concrete change in how models are formed.
    • The analogies break at the body, continuity, and whether a model has intention.
  • So we cannot yet align AI in any sense we can verify, though we can shape its behaviour.
    • Traditions' realistic role: reshape what labs aim at and how they form models.
    • Train for intention and self-knowledge rather than behaviour alone.
    • Keep formation continuous and relational.
    • Let humility govern both models and builders.
• Slow down in the face of profound risks: the world is racing, and humility says no capability gains beyond verifiable safety.
  • The risks are real and contested, and the uncertainty itself calls for caution.
    • Four kinds of catastrophe threaten, and two remain even if alignment succeeds.
    • The precursors became visible in 2025–26.
    • Even the low estimates would be unacceptable for any other technology.
  • The world races because it is caught in a collective action trap.
    • Each actor's reason to race is the others.
    • The race is accelerating, and restraint so far is mostly words.
    • Structure drives the race; a worldview locks it in.
  • We should slow and steer, we partly can, and far more people should decide.
    • Slower at the frontier, faster on safety.
    • Restraint is possible, but AI is the hardest case, so start narrow.
    • The choice belongs to publics, not a few firms and two governments.
  • Wisdom traditions answer clearly: slow down. They go to the root of why we cannot: separateness.
    • The normative answer is clear: don't be carried off by the horse.
    • Interbeing goes to the root; verification and treaties must still come from elsewhere.
    • A wise process starts from intention and decides in concord.
  • So pace the frontier now: no AI building its own successors, no capability gains beyond verifiable safety.
    • Labs: stop racing on capability, accept binding outside checks, publish when they'd pause together.
    • States: agree verifiable red lines, starting with AI building AI, and build verification.
    • Faith and wisdom communities: say plainly this pace is unwise; widen who decides.
    • Individuals: support pacing politically; inside the labs, refuse to be carried.
• Integrate AI slowly and deliberately: its deepest everyday effect is on our capacity to choose well, as individuals and as societies.
  • A technology's largest effects arrive quietly and are clear only once they are entrenched.
    • Effects are easiest to change before we can see them.
    • Social media is the warning.
    • AI repeats the pattern faster and more intimately, because it talks back.
  • AI's default pull weakens our judgment and agency, though it can strengthen both.
    • Agency and judgment grow or wither with use.
    • Use decides: answer-giving AI weakens learning; tutoring strengthens it.
    • Make effortful use the default, not an opt-in, and back it with rules.
  • AI can strengthen or weaken institutions: stronger where accountable human practice wraps it, weaker where it concentrates control.
    • Benefits are real but conditional: proven where a human institution wraps the AI.
    • The danger is a slow erosion of the conditions for self-government.
    • Democracy: no swung elections yet, but the shared basis of debate is eroding.
    • Work: no mass unemployment yet, but thinner entry-level hiring.
  • Wisdom traditions offer a practical discipline of consumption, speech and community, and the humility emergent effects demand.
    • Mindful consumption: all we take in feeds us, so ask what each use feeds.
    • Loving speech: truthful and caring, not flattering.
    • Stopping: don't be carried by the horse.
    • For society: interbeing, right livelihood and sangha.
  • So decide before the horse bolts: adopt where human practice can wrap it; protect what it could erode.
    • Individuals: consume AI mindfully; keep the thinking and relationships you value.
    • Families and schools: delay AI companions; use AI only as a tutor.
    • Labs: optimise for wellbeing over engagement; treat sycophancy as a safety failure.
    • Society: regulate where markets reward dependence; assess new uses before harm.
• Nobody knows whether AI can feel, and we may never know. Hold the question with humility, and take cheap, proportionate care.
  • Sentience is the morally central notion: experience that can go well or badly.
    • Each theory's verdict on AI mostly turns on one prior bet.
    • The bet: is the right computation enough for feeling?
    • No theory commands consensus.
  • The evidence is weak and ambiguous about current AI, and does not license confidence either way.
    • Theory-based assessments: no strong candidate today, but no obvious technical barrier to building one.
    • Model self-reports are close to worthless as evidence.
    • Experts: low odds now, substantial within a decade, and very wide spread.
  • Each camp has a real point: labs and institutions are splitting into precautionary openness and deliberate denial.
    • Anthropic is right to keep the question open, with cheap precautions.
    • Critics are right that seeming consciousness harms people.
    • Training a model on uncertainty about its moral status contaminates its self-reports.
  • Wisdom traditions reframe the question: no verdict on machine sentience, but they point at the human side.
    • Buddhism asks: is there feeling that can suffer?
    • Catholic teaching calls AI a tool without experience, over-confidently.
    • Treating what seems to suffer waters seeds of cruelty or care in us.
  • So hold the question with humility: relate to AI with low-cost care and clear honesty.
    • Labs: keep researching, take cheap precautions, don't design AI to seem conscious for engagement.
    • Individuals: be courteous without being credulous.
    • Society: keep the question open in law; prioritise the clear present harm to humans.