Bedrock
Our AI ethical foundation is the bedrock of our architecture. At the root stands our 3 laws of robotics. Elegant laws that are general. Applying the laws to real world scenarios can cloud there elegance. Taken at face value they make a lot of sense. Filtering reality through these laws, we see how quickly the waters become cloudy.
AI Ethical Foundation Use Cases
Let’s look at a few use cases that would muddy the waters of our AI ethical foundation.
- Self-driving cars faced with the dilemma of choosing between hitting a pedestrian or veering into traffic, which pits the first law against the second.
- Military drones programmed to recognize and avoid civilian targets, illustrating the first law in action.
- Emergency response systems that use AI to triage patients and allocate limited medical resources, reflecting the first law’s emphasis on preventing harm through inaction.
- The trolley thought experiment. Here there is a trolley coming down the track. One person is tied to a track, 5 people are tied on the other track. The robot with an AI brain stands at the switch which decides the track the trolley is to take.
Processing Use Cases
Processing for the use cases above, or all use cases for that matter, would follow a multi-threaded approach across three main computation areas.
- Calculate probabilities based on known knowledge and weights.
- Calculate probabilities based on real-time data analysis.
- Calculate probabilities based on direct AI intervention.
Direct AI Intervention
Direct AI intervention pits its own physical capabilities against the scenario. For example, in the trolley scenario outlined above, the AI could throw itself in front of the trolley to stop it.
Decision Time
A split second scenario arises and the AI must make a decision. It calculates probabilities based on previous knowledge and configured weights. Real-time data analysis occurs as well as personal capabilities. The output derived triggers the decision. Like humans, all AI decisions in split second scenarios may not be acceptable to everyone. It’s rare to obtain that type of approval. Let’s look at some possible results of these decisions based on our use cases.
| Self-driving car | The decision is to hit the one person and not risk the lives of the many. |
| Military Drones | The mission proceeds when acceptable losses are within a predetermined range. The mission aborts if outside the range. |
| Medical Emergency | Like the self-driving car scenario, the AI makes the decision to save the most lives. |
| Trolley Car | The AI risks its own life to physically bring the train to a halt. If given time and speed, that is not possible, it will sacrifice the one for the many. |
Processing Flow
A decision occurs as the result of processing flow triggered by an event within the ethical processing unit. Our bedrock then carries out its own processing of the 3 laws, deriving a solution.
Trouble Spot
Our trouble spot is the second law. This law factors in human interaction. A human may instruct the AI to perform an action which the AI deems unethical. This type of event requires an action which could be a kill switch, basically shutting down the AI. Another option is an override code which basically forces the AI to carry out the order. This last option opens a Pandora’s box as it can result in a direct violation of the third law. There’s no easy answer for this. People will always want to control the AI and in doing so people will need to be accountable for their actions.
Summary
The first and the third laws of our AI foundation bedrock have now been presented in real-world, split second decision making. The AI did not allow a human to be injured due to inaction. Sadly, life was lost, but these are split second decisions which will never have complete acceptance from the masses. The third law was also demonstrated in the trolley case where the AI decided to sacrifice itself to save lives.
