I watched an MIT professor, Joshua Tenenbaum, recently discussing AI technology in this sort of application. He said what is helping AI be something we can start to trust is when it exhibits a "common-sense core," being able to see not only patterns in data, but to then extrapolate what is likely to happen. If we see a picture of a plate on the edge of a table and a dog eyeing it, we can predict what is likely to happen to that plate just from a static photograph of these various objects. AI can also anticipate such things, especially what Tenenbaum calls intuitive physics. It's fascinating.
That level of intuition is difficult to achieve. I have seen some some research where computers will look at a series of pictures and attempt to extrapolate the next frames. They celebrate the few instances where the computer does a good job, but in order for 'trust' to exist in high risk situations, the accuracy must be order of magnitudes higher. Not sure with the current level of tools if we can achieve that level of fidelity. It may take a significant breakthrough. Ironically, we can't determine if it is possible unless we continue to invest in research. Explore the unknown.