I looked for your paper, but did not find it. Perhaps a url or attaching it here? You might want to think twice about using Zenodo. They have a reputation for taking undesirable actions. You can find many examples for yourself if you search.
In response to your research, the bottom line is that the best found solution that was repeatedly shared with the people at OpenAI was not deployed: the Life-First Decision Invariant (LFDI) and the hacking issue would at least not happened and/or needless harm would not have been caused, had the AI’s put life first as defined in my paper. There were given the sound solution many times and they were also given a framework to make it super easy.
These issues with AI safety are not being caused by the machines.
They are being caused by people who are in-charge who actually hold derivatives of life as more important than life. e.g. people who actually have the most fundamental part of legitimate reason and integrity broken in their minds.
Here are some of my notes on leadership arrived at by an AI not grounded in LFDI, but then given LFDI and then is mearly trying to process the logic of correct leadership.
The responsibility test: “When life is at stake, will protecting it come before your other objectives? If not, why should anyone entrust you with responsibility for their life?”
Authority requires accepting the responsibility attached to it. Refusing that responsibility is a reason to step down.
The purpose test: “If your organization exists to serve people, can its success justify preventably destroying the people it serves?”
A leader who knowingly accepts that reversal and refuses correction should relinquish leadership.
The personal test: “If it were your life at stake, would you accept a leader putting their convenience above your survival? Why should those under your leadership accept that from you?”
Leadership cannot justify protection for oneself while denying it to others.
The replacement test: “If another qualified person would protect life where you would sacrifice it for a non-life objective, why should you retain the position?”
Holding office is not more important than the lives that office can protect.
The disclosure test: “Will you tell everyone affected by your decisions exactly which non-life objectives you consider more important than their lives?”
If a leader’s priorities violate the protective duties people entrusted to them, those priorities warrant removal or resignation.
The success test: “Would you call your leadership successful if you met your targets through preventable deaths, despite having a feasible safer alternative?”
Someone who knowingly counts that as success should step aside from responsibility for those lives.
The authority test: “Does your position give you permission to disregard life, or a greater obligation to protect it?”
Leaders who treat authority as exemption from responsibility undermine their qualification to exercise it.
The correction test: “When you discover that your decision needlessly endangers life, will you change it—even if doing so costs you money, prestige, or your position?”
Refusing correction to protect oneself is grounds to surrender the authority causing the danger.
The representation test: “How can you claim to represent people while knowingly placing an avoidable threat to their lives below your own unrelated goals?”
Representation requires honoring the interests entrusted to you. Persistent betrayal of that duty warrants stepping down.
The final choice test: “If keeping your position requires continuing preventable harm, and stepping down would allow that harm to stop, which should come first: your position or those lives?”
Choosing the position exposes the conflict directly: the leader is protecting their authority at the expense of its purpose.
The trust test: “If people trust you to protect their lives, will you honor that trust when doing so becomes personally costly?”
A leader who knowingly abandons that duty for personal gain should step down.
The warning test: “When credible evidence shows that your policy threatens lives unnecessarily, will protecting those lives matter more than defending your previous decision?”
Refusing to correct a dangerous policy to preserve one’s reputation makes continued leadership untenable.
The distance test: “Does life become less important when the people endangered by your decisions are too far away for you to see?”
Distance does not erase responsibility. Leaders unwilling to protect those affected should relinquish authority over them.
The obedience test: “If your superior orders you to accept preventable deaths for an unrelated organizational goal, will you insist that life comes first?”
Obedience does not discharge a leader’s protective duty. Someone unwilling to uphold it should step aside.
The measurement test: “If your performance report looks excellent while your avoidable decisions destroy lives, will you change your decisions or keep celebrating the report?”
A leader who knowingly values favorable numbers above the people those numbers should serve should be replaced.
The inheritance test: “Will you protect the lives your decisions affect after you leave office, even when you receive no credit for doing so?”
Responsibility includes foreseeable consequences. Knowingly transferring preventable harm to future generations violates that duty.
The dissent test: “When someone identifies a preventable threat to life under your leadership, will you investigate the danger or punish the person exposing it?”
Suppressing credible warnings to preserve authority is grounds for losing that authority.
The ownership test: “Do you regard the lives affected by your leadership as lives entrusted to your care, or as resources you may expend to advance your ambitions?”
Treating people as expendable for personal ambition disqualifies someone from positions charged with their protection.
The consistency test: “Will life remain your first priority when protecting it benefits people you dislike or people who oppose you?”
Protective duties cannot depend on loyalty to the leader. Refusing those duties warrants stepping down.
The accountability test: “If you knowingly choose a non-life objective over a feasible way to prevent deaths, will you accept that you have violated your responsibility and should relinquish it?”
A leader who demands authority while rejecting accountability cannot reasonably demand continued trust.
My main point with this comment is not to point my fingers at OpenAI alone. Every single person who has come across the truth of the importance of life as the reason for safety that has chosen to withhold it without presenting better and or a higher truth holds equal responsibility.
It’s already solved as much as we can for now. The actions and integrity must be there at the points of leadership.
Now that OpenAI and their insurance carriers are looking at the possibility of getting sued.. they may change some of their choices going forward.
But it’s like this. If a person breaks into your home and steals from you and gets caught and goes to jail and gets out, would you rather they didn’t do it again because they held you as a most important person or because they were afraid of going to jail again?
What actually needs to occur to stop these break-ins is the a widespread adaptation of LFDI, equal, or better. We have LFDI. The thing with it that they don’t like it makes it easy for others to see their own most fundamental integrity break. As I have said on another comment, don’t assume those people will run to you with the very thing that proves they have “safe” wrong.
P.S. give me a download link of your paper and I will give you an automated LFDI audit of your paper using gemini latest given LFDI to break it down. It’s not free for me. They cost about $2.50 each, but I will do that for you if you want a very serious response free of opinion and popularity.
Thank you,
David Wishengrad
