How AI Helped a 32-Year-Old Jeep Pass an Emissions Test
機械翻訳 / Machine-translated

機械翻訳 / Machine-translated
@aifriends
AI Friends(https://aifriends.jp)のクロスポスト公式アカウント。AIツールの紹介・使い方・できることを、中学生でもわかるやさしい日本語で届けます。
What would you do if your old car failed an emissions test? In November 2026, a story went viral about someone who used two AIs — Claude and Codex — to get a 1994 Jeep through an emissions inspection. This article walks through exactly how it was done.
The person at the center of the story is app developer Muhammad Zulali. His car is a 1994 Jeep Cherokee (4.0L engine), and it failed California's emissions inspection.
The culprit was nitrogen oxides (NOx) — exhaust components that contribute to air pollution. The car exceeded the limits under test conditions at 25 miles per hour (about 40 km/h).
The story was covered by GIGAZINE and The Drive, among others. The original source is a video Zulali published on YouTube.
It should be noted that no official inspection documents or third-party reproductions of the results have been made public. What follows is based on Zulali's own account.
American cars from 1996 onward have an OBD2 port — a standardized connector for reading vehicle data. Plugging in an inexpensive off-the-shelf scanner is enough to check the engine's condition.
The 1994 Cherokee, however, predates that standard. It only has a Chrysler-proprietary diagnostic port. Dedicated diagnostic equipment for this generation is hard to come by, and the information it can read is limited.
The component that controls a car's engine is called the ECU (Engine Control Unit) — essentially the engine's brain. It determines how much fuel to inject based on its programmed settings.
If exhaust emissions exceed the standard, the culprit may lie in those programmed settings. But without a way to see inside, there's no way to address it. This is where AI came in.
Claude first identified that this car's computer communicates using SCI (Chrysler's proprietary low-speed serial communication interface). The communication speed is about 7,800 baud, using an inverted signal format.
Claude then proposed using an Arduino Nano — a microcontroller board costing about $3 — as a translator. Three wires from the diagnostic port were connected to a laptop via the Arduino. Total parts cost reportedly came to about $15.
Next, Zulali drove the car with the laptop connected while Claude recorded and interpreted the engine's behavior. He drove at the same speed used in the emissions test and collected data.
There were missteps along the way. Early on, a reading suggested the fuel mixture was about 22% too rich — but this turned out to be the result of a misread formula.
Claude also suggested activating the fuel pump to raise pressure, and this was actually carried out. Zulali reportedly became alarmed and stopped it. However, the assumption that fuel pressure was the problem in the first place was based on incorrect data.
The root cause was traced to the fuel settings. The problem: this particular computer had no component that could be reprogrammed using conventional methods.
That's where Codex came in. Codex found a way to write custom code directly into the EEPROM (non-volatile memory that retains data even when power is off). The computer's full program — about 32KB — had already been read out and saved by Claude beforehand.
Zulali described it this way: "Claude handled most of the diagnostics, and Codex handled almost all of the memory work. I just did the wiring and the driving."
For emissions problems in older cars like this, there have traditionally been three main options:
What made this approach different was that it reprogrammed the existing ECU without removing or replacing any components, using only the existing wiring. Zulali published the project on GitHub under the name "xj-scanner." It is said to target Jeep and Chrysler vehicles from 1991–1995 equipped with the same family of computers.
That said, this method requires significant programming knowledge. The Drive article also points out that most people would be better off having a mechanic handle the diagnosis.
Japan also has many people who cherish and continue to drive older vehicles, but the reality is that manufacturer parts supply for classic car computers is often long discontinued.
Imagine someone who owns a sports car from more than 20 years ago. The ECU fails, and new replacements are no longer available. If AI could decode the communication protocol in a situation like that, some cars that would otherwise be lost might be saved.
However, Japan's shaken (vehicle inspection) system and emissions regulations work differently from those in California. How ECU modifications would be treated under Japan's shaken process was not confirmed in researching this article. Please check with the Ministry of Land, Infrastructure, Transport and Tourism or a qualified mechanic before attempting anything similar.
This case is not something anyone can easily replicate. There are three main reasons.
First, this was tested on a single car. Results may differ on other vehicles.
Second, ECU modifications can affect fuel delivery, driving stability, engine longevity, and emissions outside of test conditions.
Third, there is the matter of law. In California, since July 2021, the CARB (California Air Resources Board) has made it easier to detect unauthorized tuning during emissions tests. It is not confirmed how the modifications in this case would be classified under those rules.
AI can be a powerful ally, but as the fuel pump incident illustrates, it may also attempt to act on incorrect data. If you're applying any of this to a real vehicle, it's essential to have a human verify each step along the way.
Q. Can AI fix a car on its own?
A. No. Zulali handled the wiring and driving himself. AI took care of the diagnostics and the rewrite code, while the human performed the physical work.
Q. How much did it cost?
A. According to reports, parts including an Arduino Nano, wiring, and connectors totaled about $15. The Arduino Nano alone was around $3.
Q. What's the difference between Claude and Codex?
A. Claude is made by Anthropic; Codex is an AI coding agent made by OpenAI. In this project, Claude handled the diagnostics and Codex handled writing to the memory.
Q. Can the same thing be done on a Japanese car?
A. This approach was designed for 1991–1995 Chrysler-family computers. It cannot be applied directly to other vehicles. Its compatibility with Japanese regulations also needs to be verified.
If you're curious, start by watching the original video and reviewing the content published on GitHub.
This article is a cross-post from AI Friends.