top of page

What are the faults addressed in ISO/PAS 8800 standard? How is it different from ISO 26262?

Updated: Mar 3



Welcome back to all the safety engineers who want to explore the new ISO/PAS 8800:2024 standard along with us.

 

In our previous blog, we touched upon the scope of ISO/PAS 8800 standard and provided a "big picture" view of AI Safety life cycle as it is described in ISO/PAS 8800.

 

In this blog, we will try to understand the type of "faults" addressed in ISO/PAS 8800 standard and how is it different from ISO 26262 standard. Let us try to understand this by doing a comparison with ISO 26262 standard along with some examples.

 

As we know, ISO 26262 standard addresses hazards due to systematic and/or random hardware faults within safety-related E/E systems of the vehicle. ISO/PAS 8800 standard addresses hazards due to similar faults like ISO 26262 (systematic and/or random hardware) but additionally covers another fault i.e., output insufficiencies within safety-related AI systems of the vehicle.


"Performance insufficiency", "Functional insufficiency" and "Output insufficiency” are new terms which we do not have in ISO 26262.


  • Performance insufficiency means an AI element does not perform well to meet the desired requirements, expectations etc.


  • Functional insufficiency means insufficiency in the specification of AI element or performance insufficiency.


  • Output insufficiency means wrong/insufficient output from an AI element used within a vehicle, resulting in a hazardous behaviour or inability to prevent/detect & mitigate a reasonably foreseeable indirect misuse (RFIM) or both.


Below table summarizes comparison between ISO 26262 and ISO/PAS 8800 standard with respect to goal, their applicability and the hazards that are addressed.


Topic

ISO 26262:2018

ISO/PAS 8800:2024

Goal

Absence of unreasonable risk due to hazards caused by malfunctioning behaviour of E/E systems

Absence of unreasonable risk due to AI errors caused by systematic faults, random hardware faults and output insufficiencies

Applicability

Applies to safety-related systems that includes one/more E/E systems and are installed in series production road vehicles

Applies to safety-related systems that includes one/more E/E systems that uses AI technology and are installed in series production road vehicles

Hazards  

Addresses hazards caused by malfunctions of E/E safety-related systems

Addresses the hazardous behaviour due to output insufficiencies, systematic errors and random hardware errors of AI elements used within E/E safety-related systems


Below table summarizes some examples of output insufficiency, performance insufficiency, systematic fault and random hardware fault.

Faults

Examples

Output Insufficiency

An image recognition system using AI model may fail to identify and classify an animal/may classify an object incorrectly.

Performance Insufficiency

AI model not being trained well with sufficient dataset to predict/prevent collision of ego vehicle with Kangaroo

Systematic fault

AI model producing biased result due to systematic fault of using same data to train and evaluate the AI model

Random hardware fault

Bit flip in the safety critical memory due to ionizing radiation.


2 Comments

Rated 0 out of 5 stars.
No ratings yet

Add a rating
Guest
Feb 28

How is the performance insufficiency different from SOTIF?

Like
Srisha Ramesh
Mar 03
Replying to

Hi,

Thank you for your query.

Please find my feedback below:

The Performance insufficiency concept being referred in the ISO/PAS 8800 (Safety standard for AI systems) is borrowed from ISO 21448 (SOTIF standard). The difference is in the system responsible for the performance insufficiency.

In the ISO/PAS 8800 standard the performance insufficiency is due to AI system element not meeting its desired requirement/expectation, whereas, in the ISO 21448 (SOTIF standard), it is due to limitation/shortcoming of non-AI based system in meeting its specified function of a system.

Example for AI system element based performance insufficiency - AI model not being trained well with sufficient dataset to predict/prevent collision of ego vehicle with Kangaroo.

Like
bottom of page