[SystemSafety] AI and safety

Rob Alexander rob.alexander at york.ac.uk
Mon Nov 12 11:09:00 CET 2018


The DeepMind safety blog has taken to calling this the distinction
between the "revealed" and "ideal" specifications ---
https://medium.com/@deepmindsafetyresearch/building-safe-artificial-intelligence-52f5f75058f1

(This is, of course, not a problem restricted to machine learning, or
even to software. Machine learning just provides new ways to achieve
such a discrepancy)


Rob
On Mon, 12 Nov 2018 at 08:58, Matthew Squair <mattsquair at gmail.com> wrote:
>
> For the short/medium term AI challenges these issues could be characterised as ‘The Monkey’s Paw’ problem that is you get what you specify but it’s sure not what you want. Unfortunately with AI we can’t just wish away the results…
>
> I think it was Nancy Leveson who pointed out in Safeware that safety should always be evaluated against what is desired rather than what is specified, because the specification can always be wrong.
>
> Regards,
>
> On 12 Nov 2018, at 12:22 pm, Dariusz Walter <dariusz at walterki.com> wrote:
>
> Based on the brief descriptions in the database, all the AI solutions seem to play within the set rules that are explicitly defined. None of the AI solutions seemed to have "failed" any of the set rules. They definitely seem to be reliable. In fact, on review of the issues in the database, I find the AI solutions ingenious.
>
>
>
> If anything, in my mind they identify the holes/gaps/assumptions present in the explicit set of rules defined for the task, including
> 1. the specifications/behaviour of the environments that these AI systems are grown to work in
> 2. completeness/correctness of the rules/cost functions that these AI systems are supposed to meet
>
>
>
> Imagine if some of the cost criteria/requirements that these AI systems are grown for were safety requirements. It would be interesting to see what ingenious solutions would be identified, and if they would in fact be safe to a human interpretation.
>
> E.g. If harm to humans as interpreted through a thermal sensor means the thermal readings do not drop or go above a certain level, then say, chopping the humans head off and putting them on the right setting in a slow cooker may be a interpreted as a perfectly safe solution...
>
> It almost seems like the bar for defining the simulation environment would need to be raised to the same level as for defining the safety requirements in order to even begin a claim for AI safety. In the case where the AI is interacting with the real world, translation of the safety requirements into those that can be obtained through the AI's sensors needs to be closely considered.
>
> In either case, I look forward to the translation of the current legalese definition of "safe" into an unambiguous set of rules/requirements for AI consumption.
>
> Dariusz
>
> On Sat, Nov 10, 2018 at 10:32 PM Olwen Morgan <olwen at phaedsys.com> wrote:
>>
>>
>> On 10/11/2018 04:50, Peter Bernard Ladkin wrote:
>>
>> <snip>
>>
>> There is no connection made from these features which consitute "AI
>> safety" to harm caused to or the environment, and damage to things,
>> avoidance of which is the usual definition of safety.
>>
>> <snip>
>>
>>
>> With due respect, Peter, this seems to me to be missing the wood for the
>> trees. The only way we'll ever address the problems associated with
>> using AI in critical systems is to build experience of what can go
>> wrong. AFAI can see (maybe wrongly - it's not my field) with current
>> knowledge, we would be hard pressed even to classify different types of
>> AI cock-up. Until we can do that, we won't be able to devise effective
>> systemic countermeasures.
>>
>> Olwen
>>
>> _______________________________________________
>> The System Safety Mailing List
>> systemsafety at TechFak.Uni-Bielefeld.DE
>> Manage your subscription: https://lists.techfak.uni-bielefeld.de/mailman/listinfo/systemsafety
>
> _______________________________________________
> The System Safety Mailing List
> systemsafety at TechFak.Uni-Bielefeld.DE
> Manage your subscription: https://lists.techfak.uni-bielefeld.de/mailman/listinfo/systemsafety
>
>
> _______________________________________________
> The System Safety Mailing List
> systemsafety at TechFak.Uni-Bielefeld.DE
> Manage your subscription: https://lists.techfak.uni-bielefeld.de/mailman/listinfo/systemsafety



-- 
Dr Rob Alexander
Lecturer in High Integrity Systems Engineering
Department of Computer Science
The University of York, Deramore Lane, York, YO10 5GH, UK
Tel: 01904 325474  http://www.cs.york.ac.uk/~rda/

If we can disclaim it, we do --- http://www.york.ac.uk/docs/disclaimer/email.htm


More information about the systemsafety mailing list