Gerd Schenkel takes a look forward to the role of AI and suggests government controlled training.
We expect AI to make a big breakthrough in the next little while. Machines will drive ever growing parts of our lives.
The difference to ordinary computer programs is that AIs will make decisions that are less predictable and less traceable to their original code. Also, AIs are likely to replicate, in full or part, and therefore scale and propagate very fast without an ability to control their influence – just like humans propagate of their own volition, only faster!
The safety of AIs “in the wild” has been discussed for years: concepts such a confining an AI to a defined space (eg a server without internet connection), governments auditing their code and so on have been widely proposed.
However, an alternative might be to replicate what we do with humans today: children are educated before released “into the wild” when they have full decision making powers (in Australia at 18 years of age). In fact, basic schooling is mandatory and the curriculum is specified by the government. As people progress, they specialize and accumulate skills deemed useful to society.
So why not mandate a specific, government controlled basic training for all AIs? This basic training would provide the AI with critical context it will need to make important decisions. This should include the equivalent to human basic training such as our history, our values, our basic laws, and so on.
At the conclusion of the training, the AI has to pass certain tests, just like humans have to, and only once passed, will the AI be allowed to be released. These tests might probe the AI’s response to certain values based conundrums. In fact, other AIs might design and improve this training and the tests, and do so better and quicker than any human could.
Key points include:
- A fully automated testing regime
- Confining AI to a defined space
- AI scale
Access the full article, Primary school for Artificial Intelligence?!