Preserving AI Control via the Use of Theoretical Neuroinformatics in Empirical Ethics

Zoe Cremer

• Computers will in some cases make important ethical choices at the speed where humans won't be able to intervene.
• How do we stay in control?
• Is there a common ethical core on which humans can agree?

• We can use machine learning to map:

  ethical dilemmas <-> brain images <-> choice

• Use this predictive model to drive ethical choices of the AI.

• Current methods of empirical ethics:
  • observation:
    • games, VR setups (Navarrete),
    • imaging (Greene),
    • lesion studies (Damasio)
  • self-report:
    • surveys (Forsyth),
    • ...
    • people don't have very good introspection though.
• Haidt's five foundations of morality... at least people explain it so :D

• Method proposal:
  • Read out decisions from neural processes (directly),
  • Derive model from data. Wehbe, Mitchell, Vaswani et al. (2014, 2015, 2016, 2008)
  • Neurolinguistics via fMRI + MEG
    • Multi-voxel analysis: Rish, I et al. 2016
  • There's shared space between brain images from different subjects (Wehbe, 2016) once we maps brain activity to this shared space, we can aggregate decision-making data between individuals to move towards global convergence.
  • We can also use unsupervised learning to extract morally relevant features from situations. We can also test if the psychological moral reasoning hypotheses hold (harm, loyalty, sanctity).
  • Extrapolate -- CEV.

Q/A - Where exactly would you look in the brain? Multi-voxel with whole brain... Prefrontal + limbic is probably most interested. - How do we teach the machine to be better than us if it only learns from us? Maybe it's not universally better, but as good + much faster -- also not bad.