What Alexa’s creepy laugh means for AI Quality Assurance

Software bugs are usually not creepy - yes they are annoying, angering, frustrating and even damaging, but rarely do they unsettle us. But there have been some ‘creepy’ bugs rolled out in the past.

In 2017 alone there were a fair few creepers out there. Gmail, for example began asking for access to ‘body sensors’ in order to use its services - it was patched after discovery, but for a brief period Gmail required access to information like our heartbeat before we could use it.


Other bugs last year included security flaws in the Amazon Key service that made it so thieves could literally hack their way into people’s phones and it emerged that due to a touchpad bug, Google Home Mini devices had been listening into our conversations 24/7.

Double Creepy.

These are all bugs that fuel our age old mistrust in Computers and software, and quality assurers now needs to work doubletime to ensure end users aren't creeped out by mistakes in this new breed of intelligent 'machines with personality'

Amazon Echo’s Evil Laugh

Possibly the creepiest software bug to date came out in March earlier this year. Amazon Echo devices began, well…laughing.

Echo users reported that Alexa had started, unprompted, randomly cackling at various moments, almost ‘evilly’. Needless to say this was unsettling for most who saw it happen.

The problem?

If you own an Amazon Echo you’ll know that she can be a little over eager to hear her name. Sometimes ‘she-who-must-not-be-named’ will just turn on, and we have to tell her to stop, go away or just be silent.

Well, after a recent functionality boost, Alexa was given the ability to laugh. However if we wanted her to do this, she only needed to be told ‘Alexa, laugh’ before she was happy to giggle away.

It turns out there is a lot that can be misinterpreted as ‘laugh’ by what her sophisticated software hears. Calf... Mast… Dart… Barf… you know, all those words we use all the time.

The fix, which seems to have solved the problem, was to make the command more complex. Now to make her laugh, we need to ask ‘Alexa, can you laugh?’ Or ‘Alexa, will you laugh?’ before she starts her chuckling.

Problem solved, although perhaps that trust we had built with her has been damaged a little bit - it reminded us that these technologies are fallible.

Intelligent Machines with Personality - A recipe for creepiness

The recent crossover between artificial intelligence and machines with ‘personality’, like home assistants, doubles down on the fact that people are unsettled by the idea of allowing an intelligent machine into their home.

After all their core function is to always listen and respond in an efficient, and crucially personable, manner. When it goes wrong, there’s alway a chance for creepiness.

It’s a textbook example of the uncanny valley effect. The closer we get to things appearing genuinely human-like, the creepier they seem to people. Unnerved reactions to errors like Alexa’s unwarranted laughs only highlight that fact.

Quality Assurance for robots with Personality

Because of their sophistication and broad scope of functionality, machines with personality and artificial intelligence together are going to be redefining QA over the next few years to ensure these 'creepy' bugs are eliminated and ensure trust in these platforms.

The laughing bug couldn’t have been caught by conventional means alone because the conditions that could have caused it were so specific, and potentially numerous, and we are testing for a qualitative human reaction, not a functional one.

We can only guess that a mixture of automation, manual testing, UAT and Product release testing is the only way companies like Google and Amazon operate to ensure basically perfect quality.

As the scope of their functionalities expand, this will become even more-so the case.  Being so on the edge of the uncanny valley, QA for these techs needs to simultaneously both be more qualitative, and quantitative, to ensure that products do not produce these ‘creepy’ bugs.

For sure innovation will be required in the coming years and new methodologies for testing must be developed. 


One proposed testing flow for AI

Parting thoughts.

Google recently showed off the progress it has been making with it’s Duplex software, a sophisticated human speech replication software that can make calls and book appointments, responding seamlessly to the person on the other end of the phone and convincing them that they were a real life person.

Triple Creepy

In essence, Duplex can ‘dupe’ the other person into thinking its a real human… Get it?

For many, however, this is a terrifying prospect. It’s one step on from the Amazon Echo,  which is already eerily person-like, and poses the question - how human is too human?

Technology seems to be finally pushing over to the other side of the valley.

With such knife-edge precision required to do this, how does one even test for human reactions to every functionality a software like Duplex is capable of? And will it be necessary to check that these machines are 'too' personlike?

Again, a lot of thought and innovation will be required.  



Leave a reply

Your email address will not be published. Required fields are marked*