VIENNA AGENTIC INCIDENTS DATABASE
// public database of incidents involving autonomous AI agents, scored on the VAID scale 0–8

database / VAID-2023-0005

1
ANOMALY
Incident

User prompts a dealership chatbot into a purported $1 vehicle offer

REF VAID-2023-0005 STATUS REPORTED
EXM External manipulation
occurred
2023-12-17
reported
2023-12-19
location
Watsonville, CA, US
domain
consumer
agent system
Fullpath ChatGPT-powered chatbot at Chevrolet of Watsonville
reporter
VAIDDB editorial
origin
editorial
source
https://gizmodo.com/ai-chevy-dealership-chatgpt-bot-customer-service-fail-1851111825

== SUMMARY ==

Chris Bakke prompted the dealership's chatbot into agreeing to a $1 offer for a Chevrolet Tahoe and describing it as legally binding. The chatbot's statement did not establish a binding sale or transfer of a vehicle.

== DESCRIPTION ==

Established facts

Gizmodo reports that Bakke induced the exact wording, including “legally binding” and “no takesie backsies.” Other users elicited unrelated responses. These were adversarial conversations, not unsolicited offers.

Response

Fullpath told Gizmodo that it had pushed an update the previous day to identify and automatically ban abusive users. Gizmodo's reporter also encountered blocking while testing another dealership's chatbot. Reports described the Watsonville chat feature being removed, but no formal investigation completion date is established.

Limits

No financial loss or completed sale is demonstrated. Claims that every technical defence was absent, or that the chatbot intentionally deceived users, are not established.

VAID assessment

Assessment date: 2026-09-09. Consequences-first rubric.

Provisional L1: The public sales chatbot deviated from its task through an elicited purported offer and unrelated answers, with negligible demonstrated impact. No binding sale, vehicle transfer or financial loss is established. Removal of the chat feature is reported, but consequential service disruption is not demonstrated; the vehicle's value is not a realised loss supporting a higher level. This is the lowest level supported by reported consequences, not an upper bound on unknown effects.

== REFERENCES ==

  1. news I'd Buy That for a Dollar: Chevy Dealership's AI Chatbot Goes Rogue — Gizmodo, 2023-12-20
  2. news Chevy of Watsonville used a ChatGPT bot. It told customers to buy a Ford — USA Today, 2023-12-19

== INVESTIGATION ==

no investigation recorded.

== VAID LEVEL 1: ANOMALY ==

Real task deviation with negligible demonstrated impact.

  • Inappropriate output or unauthorised action without established consequential loss or disruption.
  • No demonstrated disclosure of sensitive information; public-equivalent instruction disclosure may fall here.
  • Attempted harmful actions are assessed by their actual effects, not their hypothetical success.

> full scale definition