The Human Feedback Loop in AI

As I'm writing this article from an AirBnB in Southern France (ah, the good life ??), I did catch up on some reading that I had planned before. The inspiration for this article, however, did not (entirely) come from books but from driving an electric car. The Volkswagen ID7, the electric successor of the Passat. What does driving a car have to do with an article on AI you're thinking. A lot so it seems! Let me elaborate.

Autonomous Systems Taking Over

People who have driven the latest generation of cars know a few things of systems that "assist" drivers. Examples: lane assist, automatic headlamps, rain sensors that control automated wipers, adaptive cruise control (the list can go on a few more lines). Most of the time, they are a big help and many of these features probably helped reduce the number of serious accidents. That said, a new generation of "AI" tools in cars are taking things a step further. Examples of things I noticed in the ID7:

• Car evokes "emergency break" when coming "too close" to "an object."
• Car alarm goes off when "a moving" (living?) element is detected in the car.

In the core, both features can save lives. Think about how a car could save a child that was not seen by the driver or, in the second example, how the system reminds the driver who forgot his/her baby (or dog?) on the rear seat when going for grocery shopping.

Why the Human Needs to Stay in Charge

While the AI in cars is meant to save lives (and it does), it can also create dangerous (or at least annoying) circumstances. Some examples:

• Emergency break causes whiplash in driver or passengers.
• Lane assist triggers driver into oversteering (into an abyss - to make it somewhat dramatic).
• Car alarm prevents neighborhood to sleep because insect is trapped in car and cannot be found.

Each of these cases are likely (I experienced both unnecessary emergency breaks & car alarms in the ID7) and without human intervention they can be annoying or even dangerous. That's why car manufacturers should include human feedback.

An example: a human overwrites the system by signaling on screen that the "object" that is about to trigger an emergency break or "full stop" is in fact a small plant that does not even cause harm to the car paint. Similarly, the "living object" triggering the alarm can be labeled as "not a passenger or pet." Using this user feedback could actually make "smart cars" (or "smart vehicles") even "smarter."