This is the summary version of a longer story published on Medium in The Startup. The full story can be found here (free link / no paywall).
AI is expected to boost productivity but risks need to be managed. 12 principles can be helpful to build better AI solutions:
- Non-discriminatory (facial recognition has shown different levels of accuracy by gender and race)
- No black box and explicability (XAI)
- Debug mode (which should be turned on when the system makes mistakes)
- Fail-safe (to reduce or turn off any capabilities creating issues)
- Circuit breaker (for extreme cases, it must be possible to shut down the entire system)
- Approval matrices (humans should be in the chain of command and approve key decisions)
- Keeping track of assets, delegation, and autonomy (users should never delegate decision-making capability completely, nor stay on the sideline)
- No completely virtual or decentralized environments
- Feedback with discernment (AIs need filters and tools to use feedback optimally and remain useful)
- Annotated and editable code
- Plan C (there should always be a plan to switch back to human operations and use an alternative technology)
- Keeping track of returns and resources
Full story / free link / no paywall:
Medium
Top comments (1)
Ok, but how? How do you program control measures into a sentient AI with human or greater level intelligence that can rewrite its own code? Demanding off buttons is all well and good, but once we are talking sentient AIs, you're going to have worse luck installing off buttons than you would trying to install them on humans.
There is no practical list of principles that can be mandated for such a new and unexplored field, those can only come once the domain is explored and the pitfalls mapped.
The fact is, the AI industry is going to develop like every industry before it, unpredictably over a long period of time with much uncertainly along the way and definitely a trail of harm in its wake. That is called life and there is no way around it.