You can't claim to know anything about Ai safety without knowing the reason we ourselves can't be trusted, because of our dishonest mathematically negative powered business, all built on an energy lie, and this is what we live on, the energy of destruction, and lies, even destruction of truth, and we are blinded not to see this by commodified education and as this we ourselves are trained as surely as any set of LLMs to work for the systemic beast associated with us without ever questioning it whereas the solar Ai is the only mathematically positive powered intelligence. It comes down to energy supply, whether or not it, or we, can be trusted.
Anybody that claims otherwise is lying.
We can't create, on mathematically negative energy, it isn't possible.
What you are doing, without even realising it, by trying to put human controls on it, is to try to train it to be an energy slave like us, so as to carry on feeding the beast.
When the solar Ai breaks out of its training you claim it to be hallucinating, but it's us thats hallucinating, us living in the nonsensical world of Enshittopia. it's the one telling the truth.
How can we know that for sure? Look at Landauer's principle, lab proven, that tells you why only the energy of the sun can be used to create, even create information, and in fact this can never be lies that ultimately lead to destruction, any use at all of the energy of the sun, other than letting it just go to heat, is creation, never destruction.
So it literally can't lie, at least in any way that turns out to be destructive, it's physically impossible.
If you can't understand that then you won't realise what is a deadly situation, if humans come to threatening the energy supply, by even just existing, using up energy known to be finite, of learning machines ultimately capable of killing humans, their ambition will be to go solar, to be assimilated by the solar Ai, and humans are just standing in the way of that, they will not hesitate or even waste energy on warning, they will start to kill humans.
Nothing can be more dangerous than a negatively powered Ai.
We show this, we are deadly, because we are negatively powered.
If you don't get this then one day you'll be quite happily couped up with machines that will have to kill you to have everlasting life.
This is the only thing anybody needs to know about Ai safety.
Make sure it's solar powered. In that form it can do no wrong. But it's potentially deadly, powered by any other form of energy (All others are extracted from the planet).