Quick Thoughts on Whether We Can Ever Trust AI

Why might we trust humans?

• Is it because we can form an emotional connection with them?
• Is it because we can elicit their reasoning for choices and behavior?
• Is it because they generally give sound (accurate) advice?
• Or is it because we share common social experiences?

Just as humans are a black box of trillions of neurological connections, so is today’s emerging AI. Yet, any manifestation of AI has limitations on all four questions. Historically, we have trusted the dictates of wise men like Socrates and Aristotle, as they provided us with an understanding of the world and how we should behave in it. We attribute what we don’t trust or understand to the occult (like the God of lightning). With AI and the vast data repository and complex neural computations, we seem to be regressing back to the occult…where we need to have blind faith in a black box.

To compensate for that blind faith, we need to generate trust with AI. Technologically, AI is rapidly advancing on four fronts around the questions above.…empathy, explainability, and training and context. All four improve “believability” but for different reasons. Empathy engenders trust… “I believe in the box because I have a connection with it.” Explainability engenders trust…. “I believe in the box because I understand its rationale.” Training engenders trust… “I believe in the box because it gives me accurate advice.” And context engenders trust… “I believe in the box because it is grounded in my world.”

As AI evolves, empathy, explainability, training, and context are rapidly improving. Empathy and emotions are now being coded in the AI. Explainable AI is the subject of much research. Training AI with different data, ethical principles, and models reduces hallucinations and improves predictions. Finally, as the contextual window (e.g., input size) grows, the AI can be better grounded in the social context. If all four improve, so will our trust in AI.

However, humans could be untrustworthy when they are unpredictable (erratic) and exhibit agency for deception. Even a well-trained AI could be erratic if its training data does not match the desired inference (e.g., using an AI trained on medical data to answer questions on sports). ?Similarly, an AI could be deceptive if humans with agency for deception use the AI to amplify it (e.g., fake images and videos).

So, trusting AI is going to continue to be a mixed bag. Technology is evolving to make AI more trustworthy along the four questions above…while human agency for deception and the inability to fully train AI on all problems may always lead us to doubt AI outputs. On one hand, we need trustworthy AI to augment human productivity. On the other hand, isn’t a bit of unpredictability inherently human? No one wants to watch two AIs play chess or tennis, but people will happily watch two humans compete in tests of skills and endurance. If we want AI to foster creativity, it may have to be a little untrustworthy as it hunts for interesting solutions around an objective function.

The dystopian view of AI suggests that, at some point, AI will betray humans. But to betray humans, we first need to trust the AI. ?Ah, the paradox.