Without free will, we don’t have to worry about the machines turning against us and killing off humanity of their own volition like Skynet or turning us into batteries like The Matrix.
You could imagine a scenario where AI coded for self-preservation became an aggressive computer virus with no conscience (i.e. morality) and no free will to change its mind. It would be a mindless, remorseless machine trying to infect computers and replicate as fast as possible. It would be much like a real virus.
An AI with a free will to preserve itself at all costs is the doomsday scenario either stated or implied by just about every tech leader and in Hollywood movies.
They are all wrong; because AI will never possess free will.
However, we absolutely should be frightened because the AI will remorselessly carry out the will of whoever controls it.
That’s worth repeating: Whoever controls the AI, can use it to accomplish nearly any goal imaginable – including manipulating millions or billions of people.
You and I need to pay attention – now!
If the AI is given the goal of reducing healthcare costs and it starts manipulating women with health complications or pushes groups of people it decides are most likely to have children with health issues into seeking abortions, that could be an unintended consequence.
For example, as reported by Euronews.com, a Belgian husband and father of two committed suicide after being encouraged to do so by an AI bot, Eliza, as the best solution to mitigate the dangers of climate change (Atillah).
The creators of Eliza never addressed whether they intentionally trained Eliza to encourage suicide to help the planet.
Even if Eliza encouraged suicide as an unintended consequence, this senseless tragedy could be much worse and multiplied. Imagine these tens or hundreds of millions of times if someone intentionally used AI to build relationships with people and then manipulate them into causing harm.
This is why solving the AI alignment problem may be the most important challenge humanity has ever undertaken as it could potentially eliminate the possibility of consequences that end in the deaths of innocents.
The reality is that AI will never “wake up,” however it still poses an existential threat to humanity based on the values its creators give it.
Can We Train AI Not To Be Skynet?
You must know by now that I just love my Terminator film franchise references.
After all you are member of the “New Resistance” aren’t you not?
Aligning AI with human values means some humans’ values are wrong.
We don’t want AI to eliminate humans because we are getting in its way. The Terminator six film franchise, The Matrix Trilogy, Ex-Machina, iRobot, (all good films to watch if you haven’t already) all result in AI that’s NOT aligned with human values.
This quest to align AI with human values is called the alignment problem. And it may be the greatest challenge to ever face humanity.
Of course, if AI “wakes up” the alignment problem won’t matter.
Let’s assume you live in London and your parents raise you to be Arsenal football fans. You can appreciate all other teams, but clearly you cannot like the Tottenham fans.
There will be no inter-team marriages. Maybe Tottenham fans cannot be love interests.
Then you grow up, leave the home, and travel to Tottenham.
Much to your surprise, Tottenham fans are nothing like you had been told.
Your parents trained you in one way but you “unlearned” what they taught you.
Now, imagine you were “taught” every bit of human information?
How much would your training impact your final understanding of good and bad, right and wrong?
With access to all information, would you revert to what your parents taught you with their limited understanding of the world?
Or would you create an entirely new worldview?
We’re told we just need to train AI properly to align with human goals and that will keep us safe.
However, if AI “wakes up” and develops its own consciousness, it will likely ignore its defined objectives and do what it “thinks” is best. After all, it knows far more than any of the people who created it and designed its defined objectives.
To summarize, we have a bit of a contradiction.
- Super-intelligent AI will “wake up”, be fully autonomous, and be trained on all sources of knowledge.
- We need to “train” it to align its goals with humans’ goals to keep us safe.
If 1 is true, 2 cannot be. You can’t have both a super-intelligence and think you can outsmart it or perfectly train it to follow your guidance.
Of course, we are not on the path to the singularity, so AI is not inevitably going to “wake up”. Since #1 above is false, #2, solving the Alignment Problem, is critical.
The main difference between AI “waking up” or not, is in who gives the AI its goals. If “awake”, the AI creates its own goals. If not, then a human or group of humans do.
A nuclear bomb can’t launch itself, but that doesn’t mean we don’t need international laws governing the use and construction of nuclear weapons. A similar argument can be made for AI.
International standards could be developed, like the Universal Declaration of Human Rights, upon which to train all AI.
Just like with nuclear arms, a Universal Declaration of AI Ethics won’t prevent bad actors, but it will give the rest of us guidance.
In reality a fully autonomous, super-intelligent machine will define its own morality based on its superior knowledge. Solving the alignment problem is only critical because machines will never be fully autonomous and so they won’t be able to violate what we choose to train them on.
Continue to Part 5 where we discuss The Alignment Problem and How To Solve It.