Speaking Engagements & Private Workshops - Get Dean Bubley to present or chair your event

Need an experienced, provocative & influential telecoms keynote speaker, moderator/chair or workshop facilitator?
To see recent presentations, and discuss Dean Bubley's appearance at a specific event, click here

Wednesday, September 23, 2015

Post-VW thought for the day: more software will detect tests and reviewers

A very quick post today. 

Everyone is watching the VW scandal evolve - the company appears to have programmed some of its cars' engine management systems to work out when they were being tested for emissions. And then put the vehicles in a special mode to "game" the tests, before switching back to "normal" (more-polluting) modes afterwards.

I suspect that's the tip of the iceberg. And that the iceberg is going to grow.

I'm not an AI or machine-learning specialist. But I bet one of the easiest things for a computer to "learn" is when it's being subjected to a test of some sort. Tests look different to normal use - formulaic and often repeated. Could be a legal or regulatory test, could be a product reviewer, could be an internal acceptance-test, could be a benchmarking exercise - and so on.

Once the system knows it's being tested, it can then react to try to optimise itself to give a different set of test results.

I once lightheartedly suggested (in 2009) that we'd see broadband networks learn to give maximum speed/QoS to speed-test services. They might also be programmed to "look" very neutral when they see something from voip.regulator.com, but react differently on other occasions. Devices or cloud services might recognise the names or IP addresses (or cookies) of prominent journalists or product-reviewers. Self-driving cars will probably be very good at spotting traffic cops.

I'd be willing to bet that a lot of this type of "gaming" of tests and reviews already happens. VW won't have been the first to think about it.

In other words, testing software will be like giving exams to humans. We'll need invigilators of some sort, to make sure they don't cheat. And testers and product reviewers are going to have to be a lot more subtle, and change/obfuscate their work.

That's going to be a non-trivial problem if we want comparable data. And I haven't even thought about how you invigilate a Turing test to ensure an AI isn't "playing dumb".

2 comments:

Hannah said...

This is a brilliant article forecasting the day AI outsmarts its maker. For the developers who haven't thought of it already, I'm sure the VW scandal has generated more than a few ideas about "gaming" technology tests.

Sonam said...
This comment has been removed by a blog administrator.