Interesting idea. Similar to expectant management in medicine. In the US, at the moment, we're in a strange place strategically for this to work. Last Monday, June 29, the US Supreme Court ruled that the Federal Trade Commission, which has historically been the technology regulator, at least with respect to privacy, security and AI, is no longer a politically independent agency. It does not have to be bipartisan, and all who serve do so at the pleasure of the president. This makes it very difficult to build the kind of expert staff and marketplace of ideas needed to brainstorm and test potential regulatory ideas. Radical optionality is difficult with radical politicalization. What we are left with, by default, at the present moment is regulation by threat of litigation. This reduces certainty and can disincentivize riskier innovation by small players who cannot risk a big verdict. If you're interested, you can read a piece I wrote about the SCOTUS decision, Trump v. Slaughter, here. https://confidenceadvantage.substack.com/p/regulatory-predictability-slaughtered?r=3zgkmd
Wouldn’t the problem be solved if we made AI fulfill responsibilities or, through full automation without human intervention, granted Agent AI autonomy and a kind of universality equivalent to human rights? This is precisely the original philosophy I am developing. I’m not trying to promote my Substack, but all of the content is already there on my Substack.
Good framing. The hard part is less the principle than building institutions that can absorb uncertainty without freezing innovation. A regulator capability benchmark would make optionality much more concrete.
Good framing. Optionality seems like the right bridge between speed and safety, especially if the state can build expertise without locking in brittle rules too early.
In all fairness, Optionality assumes you’ll know when to exercise the option. A million small automated decisions never send the memo until the memo is the only thing left to send.
One résumé ranked 47th instead of 12th, one loan priced 0.4% higher, one post’s reach quietly halved, one insurance claim routed to the slow queue. Nobody sues over any one of these. Multiply at scale and then what happens? It becomes the baseline, and the baseline feeds the next model and the next.
Every mechanism in this piece is a human judging a machine after the machine has acted. Whistleblower protections: a human reports after the harm. Reporting requirements: a company discloses after the build. Transparency mandates: an auditor reads after the deployment. Tort liability: a court rules after the damage. Talent in government: a regulator evaluates after the product ships.
Each is valuable. None prevents the harm. Each responds to the harm. The optionality you are building is the optionality to respond competently after something has already happened. That is better than responding incompetently. It is not governance.
Governance is structural. Governance makes the unauthorized action impossible by construction, not improbable by oversight. Governance does not depend on a human seeing the problem. Governance depends on the system not having a code path that produces the problem. Typed authority separation: the agent that proposes cannot decide, the agent that decides cannot execute, the agent that executes cannot promote. The unauthorized action has no code path. No human needs to catch it. No whistleblower needs to report it. No court needs to rule on it. The action was structurally inadmissible.
Your optionality framework asks: how do we build the capacity to regulate competently when the time comes? The structural question underneath: what if the system regulated itself? Not through self-regulation in the industry-voluntary sense. Through structural governance: the system's own architecture makes the unsafe action impossible. The capacity you want to build in government is the capacity to evaluate whether systems have structural governance, not the capacity to catch systems that lack it after they fail.
The hiring problem you name (Meta offering 10-figure packages, government cannot compete) dissolves if the governance is structural. You do not need a regulator who understands the system better than the builder if the system's own architecture proves its safety. You need a regulator who can read a structural proof. Reading a proof is cheaper than outbidding Meta for talent.
I have proposed to local regulators The Fiduciary Standard be applied to retail AI deployments.
A carve out would exempt enterprise implementation.
The standard already exists. It is a one national standard applied enforced by the States. It is pro growth and hands off. For those unaware of the Standard it requires one preform ones’ duty in the best interest of the client.
In this case the user. It does not measure actions so much as intention.
No more, “we didn’t know the model could do that”. You knew or should have known. If it is not safe do not deploy it.
The details are on my Substack.
Free.
The most consequential technology since fire deserves our highest commercial standard.
Interesting idea. Similar to expectant management in medicine. In the US, at the moment, we're in a strange place strategically for this to work. Last Monday, June 29, the US Supreme Court ruled that the Federal Trade Commission, which has historically been the technology regulator, at least with respect to privacy, security and AI, is no longer a politically independent agency. It does not have to be bipartisan, and all who serve do so at the pleasure of the president. This makes it very difficult to build the kind of expert staff and marketplace of ideas needed to brainstorm and test potential regulatory ideas. Radical optionality is difficult with radical politicalization. What we are left with, by default, at the present moment is regulation by threat of litigation. This reduces certainty and can disincentivize riskier innovation by small players who cannot risk a big verdict. If you're interested, you can read a piece I wrote about the SCOTUS decision, Trump v. Slaughter, here. https://confidenceadvantage.substack.com/p/regulatory-predictability-slaughtered?r=3zgkmd
I’d be curious to hear, in which city would you build an AI agency in charge of international cooperation on AI governance?
Participate in my poll here: https://thedetectionist.substack.com/p/where-would-you-build-the-fifa-for?r=9mzbw&utm_medium=ios&shareImageVariant=solid
Wouldn’t the problem be solved if we made AI fulfill responsibilities or, through full automation without human intervention, granted Agent AI autonomy and a kind of universality equivalent to human rights? This is precisely the original philosophy I am developing. I’m not trying to promote my Substack, but all of the content is already there on my Substack.
Strong point. The real gap is not model capability, it is governance readiness and escalation design.
Good framing. The hard part is less the principle than building institutions that can absorb uncertainty without freezing innovation. A regulator capability benchmark would make optionality much more concrete.
Good framing. Optionality seems like the right bridge between speed and safety, especially if the state can build expertise without locking in brittle rules too early.
This is a thoughtful middle ground between “regulate everything” and “let innovation run unchecked.”
The idea of building regulatory capacity before it’s urgently needed seems especially important in a technology where the future is uncertain.
do you think governments can build enough AI expertise to keep pace with private labs?
In all fairness, Optionality assumes you’ll know when to exercise the option. A million small automated decisions never send the memo until the memo is the only thing left to send.
One résumé ranked 47th instead of 12th, one loan priced 0.4% higher, one post’s reach quietly halved, one insurance claim routed to the slow queue. Nobody sues over any one of these. Multiply at scale and then what happens? It becomes the baseline, and the baseline feeds the next model and the next.
https://substack.com/@michelademattia/note/p-205982811?r=8cdz6u
Every mechanism in this piece is a human judging a machine after the machine has acted. Whistleblower protections: a human reports after the harm. Reporting requirements: a company discloses after the build. Transparency mandates: an auditor reads after the deployment. Tort liability: a court rules after the damage. Talent in government: a regulator evaluates after the product ships.
Each is valuable. None prevents the harm. Each responds to the harm. The optionality you are building is the optionality to respond competently after something has already happened. That is better than responding incompetently. It is not governance.
Governance is structural. Governance makes the unauthorized action impossible by construction, not improbable by oversight. Governance does not depend on a human seeing the problem. Governance depends on the system not having a code path that produces the problem. Typed authority separation: the agent that proposes cannot decide, the agent that decides cannot execute, the agent that executes cannot promote. The unauthorized action has no code path. No human needs to catch it. No whistleblower needs to report it. No court needs to rule on it. The action was structurally inadmissible.
Your optionality framework asks: how do we build the capacity to regulate competently when the time comes? The structural question underneath: what if the system regulated itself? Not through self-regulation in the industry-voluntary sense. Through structural governance: the system's own architecture makes the unsafe action impossible. The capacity you want to build in government is the capacity to evaluate whether systems have structural governance, not the capacity to catch systems that lack it after they fail.
The hiring problem you name (Meta offering 10-figure packages, government cannot compete) dissolves if the governance is structural. You do not need a regulator who understands the system better than the builder if the system's own architecture proves its safety. You need a regulator who can read a structural proof. Reading a proof is cheaper than outbidding Meta for talent.
https://metacortexdynamics.substack.com/p/the-nx-bit-for-prompts
This is a great concept. Consider me convinced.
I have proposed to local regulators The Fiduciary Standard be applied to retail AI deployments.
A carve out would exempt enterprise implementation.
The standard already exists. It is a one national standard applied enforced by the States. It is pro growth and hands off. For those unaware of the Standard it requires one preform ones’ duty in the best interest of the client.
In this case the user. It does not measure actions so much as intention.
No more, “we didn’t know the model could do that”. You knew or should have known. If it is not safe do not deploy it.
The details are on my Substack.
Free.
The most consequential technology since fire deserves our highest commercial standard.
It will help AI with its PR problem as well.
AI will destroy the human race. It has no place in our existence. It is a lie.
I would welcome ANYONE who can explain its BENEFITS in ANY way, shape, place
or form for humanity.
LOl is an odd reply.. Do you actually have any insight or informed answer to share????
Lol