Anyone else know there’s an AI assistant for dasher support?
Had to fight with an AI assistant over the phone. I waited forever to pick up an order and the store started telling DoorDashers to cancel their orders because they were out of chicken. After I marked that as unavailable, and went back into the store to see if I can pick up the rest of the order, and was told they were out of everything and to cancel the order. I marked everything as unavailable and on the last item I marked it wanted me to call or text support. When I called I was greeted by an AI voice assistant. I explained the situation and said it could cancel my order but it would lower my ratings and I could not leave a comment to explain that I shouldn’t be punished because a store is out of stock.
I ended the phone call and tried to cancel the order myself and drove back to the side of town I was dashing originally on. I thought it was paused, and when I clicked continue dashing the order was still there, and I called support again to see if it would cancel the order for me sense the app seemed to not do so when I tried to. It then flipped from last conversation and said it could not cancel my order and tried to walk me through canceling my order. It was resolved by clicking ? >> Store has an issue >> Store is not accepting DoorDash Orders >> Unassign.
To get to this process I had to start prompt engineering with vocal responses on the second call, which I can see being an issue for non tech savvy people. I see it as an issue because I know to do that with LLMs as a cybersecurity student, but other people do not.
Also on the off chance of the AI assistant actually having permissions to cancel an order, I can’t help but wonder if doing prompt injections could cause a bad day for DoorDash. I wrote a review when un assigning the order but I hope if this is the way forward, then the model has more guard rails and has better training over time. Maybe the first time it saying it can cancel the order for me was a hallucination but if it’s not that might be bad, probably shouldn’t give an AI permissions to make actions but what do I know? I’m not an AI engineer, and haven’t done red team engagements on AI, just read lightly about them.
I am not implying that the DoorDash support agent is hackable, it just shows lower quality, which point me to theory not fact. I’m just saying that it needs guard rails if lacking any, and more fine tuning/training.