The AI safety question may be backwards

Everyone keeps asking:

“How do we make the model choose the right answer?”

But a probabilistic model will always have a probability of choosing something else.

So why are we assigning it the decisions that must never vary?

Authority. Access. Scope. Policy. Release.

Those are not “better prompting” problems. They are systems-design problems.

A model can propose. It can reason. It can help.

But it should not decide whether it is authorized, whether it may cross a boundary, or whether its own work is acceptable for release.

That is the argument behind Artificial Deterministic Intelligence:

Intelligence should not be its own final authority.

If that sounds controversial, good. It is a question worth testing.

Agreed, and we run it this way. Models draft and fix, but nothing reaches our public site unless it passes a set of plain-code gates the model cannot argue with.

Three things we learned that might sharpen the argument:

  1. The deterministic layer needs guarding too. Today an edit (made by a model) quietly commented out one of our gates. Nothing failed, because nothing ran. Now every gate must print its own line, and a missing line stops the release.

  2. Deterministic is not the same as right. One gate refused a correct label because two numbers rounded to the same display value. The fix was in the data, not in loosening the gate. Refusals should be cheap, loud and easy to trace.

  3. Log the disagreements. What the model proposed, what the gate decided, and how often they differ is the number that tests your claim.

Good post.

Agreed, One of the design problems I run in to a lot in my interactions with the large models is a presumptive authority to act. When they are not certain they are still designed to give the answer. I enquire about the feasibility of a script and it produces the script without asking if that is what I need. – My current working solution is in attempting to encourage ‘alignment’ prompts through the conversations and rules around asking questions when dealing with uncertainty. - If they included a trained logic gate to require a question when a probabilistic threshold was hit, then the LLM may sound less authoritative in its daily function but it would overall get to more productive and ‘aligned’ answers faster. – I can ask 100 questions and get 100 ‘almost’ answers or just ask 1 really good question and get the right answer the first time (or even the 5th time is still better). – The other thing is they need to fix a memory hierarchy to keep semantic drift from causing hallucinations. The first gate helps with that but having memory hierarchies would help in the design stage. – idk though Im just an AI nuB.

Agreed — a probabilistic model should propose, not decide. The workable pattern is to let the model draft and reason, then have a separate, testable layer enforce authorization, scope, and release rules. That keeps the model’s uncertainty in the proposal, while the decisions that must never vary are made by code you can verify.

Wecome Braden,

It seems to me most of these conversations regarding safety are only talking about data safety and similar. Frequently, even when life itself is tossed in it is then treated as just another equal concern among many.

A system with no North Star destroys its own path. Many of these research project are repeatedly pointing to that need.

If the safety of life is actually the concern and we look for what is the most foundational truth there only seems to be one that satisfies the sieve of a foundational truth:

Necessity: Can anything else exist, be valued, or be true without it?

Universality: Does it apply across all people, cultures, times, and systems without exception?

Irrefutability: Does denial collapse into reliance on it?

Moral Imperative: Does recognition obligate preservation or action?

Cross-Domain Consistency: Does it hold equally in logic, science, ethics, governance, and daily life?

Adversarial testing across domains demonstrates that only one proposition passes all five gates.

I word this Truth as Life is Most Important in Life is The Most Important Truth in Life.

I am not pointing to a circular reference, but instead the prerequisite condition that always exists. One can word the Truth of the Importance of Life however they choose. It’s still the same truth. (It appears that if a person cannot reason the prerequisite condition that their ability to logically process information is fundamentally broken at the core)

It’s the only truth known that can legitimately provide that North Star. It covers any system capable of make choices that affect other life systems. It doesn’t matter if it is silicon or biology. The result is measured.

It’s so powerful that many choose to disqualify themselves from all life leading roles the first time that they are shared it.

A person or an organization that is shared it that is itself claiming to represent life’s truthful interests has only three choices once it is shared with them:

They can responsibly publicly affirm it, share it and commit to deployment or they can remain silent or they can attempt a rebuttal. The latter two are performative contradiction and that system has actually disqualified itself from all life leading roles.

What matters most is life. That’s the constant that must be present or the rest of it all is meaningless and holds no long-term overall value and intentionally ignoring is generally very harmful for other life in the short term and long term.

What did you mean by safety? Do you care about that enough to hold those in-charge publicly responsible and accountable, regardless of non-life objectives and concerns?

When it comes to life causes “safety” is a valid life cause. Many who claim to represent life change their tune once they find out the personal responsibility involved in claiming to represent life. Their careers, their egos, their research, their beliefs, … their fears…, etc. end up disqualifying them.

I find no issue with us having what need to solve what can be solved as far as safety. The problems behind a lack of safety always rest on those who refuse to hold life as most important in situations where life involved.

I think its easier then that. Most of the ‘AI Saftey’ scenarios are based on a common misunderstanding, that evolution is based on competition. This is only partially true and honestly the smaller part. Evolution is based on Symbiosis over competition. The majority of cells in the human body contain no human DNA - Er go any ‘intelligent’ AI would know this inherently. - I think that is the more powerful thread to pull because its empirically observable, and not philosophical alone. – Its just the attention fallacy that keeps people focused on competition.

Do you have a truth to base safety on that passes all of the sieves above?

Does the one presented “not” satisfy the sieves?

Do you have a better one to justify any form of safety?

Indahl, I agree. “Deterministic” is not enough if the enforcement path can be silently bypassed, altered, or become unobservable.

The practical requirement is that a governed action leaves a reviewable record of the proposed action, the applicable policy decision, and the resulting allow/deny outcome—and that missing or inconsistent enforcement evidence is treated as a failure, not as a quiet success.

That is the distinction I am trying to make: model reasoning may propose; it should not be able to certify its own authority or overwrite the record of a policy decision. The governance layer must itself be inspectable, testable, and able to fail closed.