3
$\begingroup$

Providing that

  • more and more decisions about human life are (to be) decided by machine (like access to loans, housing, scholarship, jobs, healthcare, insurance, etc.)

  • at the same time, in many countries there are laws and codes of conduct against (negative/positive) discrimination

Does there any industrial-accepted way to examine the AI system its legal compliance?

I believed that ACM/IEEE software engineer professional code of conduct can be applied here, but also like to learn more about auditing process from examiner side as well, if there's any.

Thank you.

$\endgroup$
  • $\begingroup$ Second point on the list is exactly the reason point 1 is true. That's if business want to discriminate some people, it will use AI and all discussions are finished because oh-ah-yeah AI is objective! That's the case.... $\endgroup$ – kakaz Dec 17 '17 at 9:49
0
$\begingroup$

I'm not sure if one doesn't exist already but it is readily acknowledged that the method of operation of most machine learning algorithms is unknown, thus there should be no way of auditing AI behaviour as of now.

See: The Dark Secret at the Heart of AI for more information.

$\endgroup$
3
$\begingroup$

Understanding how an algorithm works can be realized with two possibilities. At first, with Opensource software. That is a well known method and prevents proprietary non-documented code. Opensource software alone doesn't solve the problem, as seen by @k.c. sayz 'k.c sayz' correctly. For example, if we are implementing a neural network as opensource, it remains a blackbox which is not predictable, especially for non-experts.

The shared language between legal laws and engineering capabilities is the natural language, especially English. So the engineers have to create their system in a way, that it communicates with the world in normal English. That means not, that the car speaks like K.I.T.T., it means only that the interface for controlling the car has commands like “start”, “stop” and “drive slower”. If the car detects with his neural network a pedestrian, than it should print out on the console a simple “pedestrian detected”. That seems natural, but most programs today printing out “memoryadress $3400 = -65” or “sigmoid activation function is set to sin(4-x)”.

In a potential law case it makes sense to cite status-messages of the onboard computer. Because if they are written in normal English, lawyers will understand it. How the natural language output was produced is in the scope of the engineers and is only a detail question.

$\endgroup$
  • $\begingroup$ Legal documents depend on technical "terms of art" and "legalese" (the term laypeople use for legal jargon and concepts) is distinct from natural language. $\endgroup$ – DukeZhou Mar 15 '18 at 21:17
2
$\begingroup$

I agree with Manuel Rodriguez. You can also go some way to example behaviour of a closed system through simulation. This can add a certain percentage of predictability to an algorithm, but with the understanding it isn’t a guarantee

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.