---
title: "Ask AI to wash your car and it suggests you leave it in your garage..."
description: "5 source-backed AI claims from Ask AI to wash your car and it suggests you leave it in your garage..., with key statistics, context, and the original AI…"
url: "https://kbanc.com/claims-library/ask-ai-to-wash-your-car-and-it-suggests-you-leave-it-in-your-garage"
source: "https://aiadopters.club/p/ask-ai-to-wash-your-car-and-it-suggests"
date: "2026-06-17"
topics: ["strategy", "implementation", "tools"]
generated: "2026-08-31"
---

# Ask AI to wash your car and it suggests you leave it in your garage...

By Kamil Banc | June 17, 2026

## Claims

1. **Car Wash Test Failure** (source summary): Only five of 53 AI models correctly answered that driving to a car wash 50 metres away.
2. **Role Prompting Fails** (source summary): Role definitions alone failed to improve scores, moving the model's pass rate from zero to zero.
3. **Structure Beats Role-Play** (source summary): A structure forcing the model to name the situation and task first raised pass rate to 85 percent.
4. **Interview-First Prompting** (Kamil's interpretation): The interview prompt makes the model ask clarifying questions one at a time before answering.
5. **Pattern-Match Blind Spot** (Kamil's interpretation): Models pattern-matched short distance to walking, never realizing the car itself must reach the car wash.

## Evidence

### Quote
> "You can outsource your thinking, but you cannot outsource your understanding." - Kamil Banc

### Key Statistics
- **5 of 53**: Only five of 53 tested AI models correctly answered the car wash question more than once in ten tries.
- **85 percent**: Pass rate achieved when a structure forced the model to name the situation and task before answering, with no new information.
- **Zero to zero**: Improvement in pass rate from adding role definitions alone, showing role-play prompting did not fix the blind spot.

## Context
The article's methodology centers on a simple benchmark question posed to 53 AI models, revealing that pattern-matching on 'short distance equals walk' caused near-universal failure because models never registered that the car itself needed to reach the car wash. Ablation testing isolated the mechanism: role definitions added nothing, while a structure requiring the model to state the situation and task before answering lifted pass rates to 85 percent without supplying new facts. For practitioners, the actionable technique is an 'interview me first' prompt that makes the model ask clarifying questions one at a time, surfacing tacit context the user knew but never wrote down. The author recommends testing this against a one-line prompt on a real task, then saving the clarified context for reuse in future sessions.

## Source
- Original: [Ask AI to wash your car and it suggests you leave it in your garage...](https://aiadopters.club/p/ask-ai-to-wash-your-car-and-it-suggests)
- Cite: kbanc.com/claims-library/ask-ai-to-wash-your-car-and-it-suggests-you-leave-it-in-your-garage

## Primary Evidence
- [full test and scores here](https://opper.ai/blog/car-wash-test) (opper.ai; supports claims 1, 2, 3, 5)
- [full breakdown is here](https://arxiv.org/html/2602.21814v1) (arxiv.org; supports claims 2, 3)