The problem: a 32-year-old ECU
According to a report by IT之家, citing Tom's Hardware, an AI enthusiast named Mohammad Zuraili managed to get a 1994 Jeep Cherokee, now 32 years old, to pass California emissions tests. The vehicle repeatedly failed the tests for exceeding nitrogen oxide limits at certain speeds: initially the 40 km/h test was not passed; then, after an adjustment to the oxygen ratio in the mixture, the 80 km/h test was not passed either. Each new test cost nearly $100, according to the source.
The vehicle's ECU offered very limited functions, and according to the account, the user operated almost without precise data, proceeding by trial and error. The turning point came from the combination of language models and low-cost hardware: an Arduino Nano board costing about $3, with a total of electronic components under $15.
How the solution works
The first technical obstacle was communication between the ECU and a modern computer. Claude reportedly identified the protocol used by the Cherokee, the Chrysler SCI serial bus, with a speed of about 7,800 baud and a signal polarity opposite to that of a standard serial port. The Arduino Nano board, connected on a breadboard with a few wires, three of which were inserted into the vehicle's 6-pin diagnostic connector, served to invert the signal and make the data readable.
The first connection was successful, but initially the system recorded only the coolant temperature and battery voltage, plotting their curves. Later, according to the account, the system made two assessment errors, indicating a mixture that was too rich with 22% excess fuel, because data from other models had been used as a reference. Zuraili then had Claude collect the engine parameters and error codes during the 40 and 80 km/h tests; the model was later also credited with a role in monitoring parameters such as fuel pressure.
The idea of directly rewriting the control unit’s firmware ran into a hardware obstacle: that version would predate by a few months the variant that supports firmware rewriting. The solution therefore went through Codex, which wrote custom code onto a nearby EEPROM chip and ran it directly. In the words of the enthusiast as reported by the source, Claude reportedly handled most of the diagnosis and Codex almost all of the memory management, while wiring and driving were left to him.
What remains to be verified
The project was published on GitHub under the name xj-scanner and the GPLv3 open source license, so it is freely usable. The report has exemplary value because it shows the use of generic AI tools to diagnose and modify a closed, dated electronic system, with limited hardware expense and without resorting to a mechanic or the dealership’s proprietary diagnostic equipment.
However, some elements remain to be verified: the account is based on the video and the author’s own narrative, picked up by Tom's Hardware and relayed by IT之家, and not on independent verification of the results. Nor is it indicated which emissions tests were passed or by what margins, nor is it clear how reproducible the modifications are on other units. Finally, the case should not be confused with a homologated intervention: it is a documented experiment, not a certified procedure.
