Technology 03 · Neuromorphic AI
AI draft
Nobody on the team has reviewed this page yet. An AI model wrote it as an outline for the team to react to, so treat every line and every rating as unchecked.
Would AI built like a brain think like us?
Neuromorphic, or brain-inspired, AI borrows parts of how brains work, such as spiking neurons or the circuits behind empathy, without copying any particular brain. The hope is that AI built this way would learn values closer to ours, though it might mostly just make AI more capable sooner.
Where we land
Full scorecardFive of the six ratings are Unclear for now. In each case the reason is that we haven't finished looking.
- Alignment case
- Intelligence CurseThe risk that increasingly capable AI reduces the economic and political value of human work and participation, and so weakens human agency. case
- Human welfare risk
- Artificial-mind welfare and rights risk
- Governance risk
- Safeguard adequacy
No milestones mapped yet
How to read the marks
- From a sourceQuoted word for word from a cited source.
- Our readThe team's own interpretation.
- AI-suggestedSuggested by an AI model, set in a dashed outline.
- UnclearWe don't know yet, and the tag gives the reason.
At a glance
AI-suggested and not yet reviewed, unless marked
- What it claims to solve
Alignment. AI that learns the way brains do might form concepts and values closer to ours, and might be easier for people to understand and predict.
- How it would work
Neuromorphic systems copy features of brains, such as spiking neurons, learning rules that work locally, or brain-like wiring.
- Strongest case
Human brains are the only systems we know of that develop general intelligence and human values together.
- Strongest objection
A system with brain-like parts could still end up with values very different from ours if its training, body and incentives differ, and it might look more trustworthy than it is.
- Artificial minds
Could a brain-like AI be conscious or able to suffer, and would brain-like hardware make that more likely than brain-like software?
- Main welfare risks
Two possible harms, neither yet checked against evidence: AI built on human social instincts might be better at reading and persuading people, and cheaper, more efficient AI could speed up the replacement of human work.
How it would work
AI-suggested and not yet reviewed, unless marked
What it claims to solve
Alignment. AI that learns the way brains do might form concepts and values closer to ours, and might be easier for people to understand and predict.
The mechanism
Neuromorphic systems copy features of brains, such as spiking neurons, learning rules that work locally, or brain-like wiring. Some run on ordinary computers, others on chips designed to work like neurons.
One version of the safety argument goes further: work out how specific brain systems give rise to social instincts like empathy and fairness, then build those into AI on purpose.
Strongest case and objections
AI-suggested and not yet reviewed, unless marked
For
The strongest case
- 01
Human brains are the only systems we know of that develop general intelligence and human values together. Understanding how they do it could be more dependable than hoping values emerge from training data.
- 02
Borrowing a few mechanisms is much easier than scanning and emulating a whole brain, so this approach could matter on a realistic timeline.
- 03
Shared structure could make AI easier to inspect, since neuroscience already has decades of tools and theories for studying brain-like systems.
Against
The strongest objections
- 01
A system with brain-like parts could still end up with values very different from ours if its training, body and incentives differ, and it might look more trustworthy than it is.
- 02
The insights that make AI more brain-like could also make it more capable and efficient before anyone knows how to align it, so this work may speed up capabilities more than safety.
- 03
Neuroscience doesn't yet know how brains produce values. The circuits the safety case depends on are less well understood than, say, vision or movement.
Artificial minds
AI-suggested and not yet reviewed, unless marked
Open questions
Could a brain-like AI be conscious or able to suffer, and would brain-like hardware make that more likely than brain-like software?
If we deliberately build in social emotions like empathy, do we also build in the capacity to be hurt?
How could anyone tell from the outside whether such a system has welfare that matters, given that it might report feelings whether or not it has them?
Welfare and governance risks
The team hasn't assessed welfare and governance risks in detail yet. The current risk ratings are in the full scorecard.
Milestones
The team hasn't mapped any milestones for this technology yet.
Full assessment
The full scorecard with the basis for each rating.
Show the full assessmentHide the full assessment
AI-suggested and not yet reviewed, unless marked
Full scorecard
The ratings use the scales in the team’s draft scorecard. “Not claimed” means the proposal doesn’t claim that outcome, so there is no case to grade. We never rate a risk Low just because nobody has studied it, and the cutoffs for the risk ratings aren’t set yet. How we rate
| Dimension | Rating | Confidence | Basis |
|---|---|---|---|
| Alignment caseCase · Alignment | UnclearSearch incomplete | Low | The motivation is clear, since brains are the only known systems that develop general intelligence and human values together, but we haven't yet looked at the evidence on whether brain-like design leads to more human-like values in practice. |
| Intelligence Curse caseCase · Intelligence Curse | Not claimed | Low | Neuromorphic AI is pitched as a route to alignment and to cheaper, more efficient computing. No proposal found so far says it would help with the Intelligence Curse. |
| Human welfare riskRisk | UnclearSearch incomplete | Low | Two possible harms, neither yet checked against evidence: AI built on human social instincts might be better at reading and persuading people, and cheaper, more efficient AI could speed up the replacement of human work. |
| Artificial-mind welfare and rights riskRisk | UnclearSearch incomplete | Low | It's an open question whether a brain-like AI could be conscious or able to suffer, and whether building in emotions like empathy on purpose would also build in the capacity to be hurt. The search for evidence on either hasn't started. |
| Governance riskRisk | UnclearSearch incomplete | Low | We haven't reviewed the governance work that exists, but so far brain-inspired AI seems to get little specific attention in AI policy. Dual use is also a concern: methods that make AI cheaper or more capable could arrive well before methods that make it safer. |
| Safeguard adequacyCase | UnclearSearch incomplete | Low | The search so far found no safeguards proposed specifically for brain-inspired AI, and it hasn't covered whether general AI safety rules would apply. |
Grading key