There’s a scene within the 2008 film that provides a glimpse of future interactions between human and artificial intelligence assistants. In it, Tony Stark’s virtual assistant J.A.R.V.I.S. responds with sarcasm and humour to Stark’s commands.

Tony Stark and his AI assistant J.A.R.V.I.S. work on a project.

Contemporary voice assistants like Siri and Alexa are yet to supply such natural, nuanced social chatter. To that end, our team of computer science researchers on the University of British Columbia investigated what is likely to be missing.

We found that voice interface designers handled an interesting dilemma: the stress between offering social conversation and getting things done.

Friendly or efficient?

Linguists classify human conversations into two categories: Social conversation reminiscent of greetings, humour and small talk for expressing social relations and private attitudes, and “transactional conversation,” which transmits factual or propositional information.

People can effortlessly mix these two kinds of conversations in a natural manner. However, this magical mixing is finished somewhat subconsciously. Voice designers often fail to seek out the best mix since the two kinds of conversations are complementary but additionally conflicting.

The problem becomes pronounced when designers create voice assistants to assist users complete tasks reminiscent of checking the weather or making a restaurant reservation. Designers try to complement their voice agents’ dialogues with social courtesies reminiscent of sympathetic responses or chit-chat to reinforce the naturalness.

Our study also showed that designers encounter challenges to find an appropriate trade-off between designing for an efficient assistant versus an affable companion. One participant highlighted that the more personality added, the longer the dialogue becomes, and leads to either an excessively chatty or cold and robotic voice agents.

Adding friendliness to virtual assistants may positively affect human-AI voice exchanges.

Tool and design guideline support for voice designers will be helpful in solving this problem. A correct scripting tool for designing voice assistant dialogues should help designers balance the trade-off. For example, a sophisticated dialogue-authoring tool may suggest the designers add friendly remarks to the script or also issue a warning if the social chatter is simply too lengthy.

Also, design guidelines should provide prescriptive directions on the best way to mix these two kinds of conversations for various situations. For example, voice assistants should only use witty sarcasm when the user’s voice tone is detected to be in an excellent mood.

Collecting our emotions

To provide natural conversational experiences with voice agents, tech giants reminiscent of Apple, Amazon and Google might want to collect plenty of details about users’ conversation contexts, reminiscent of where they’re, what they do, what they need and even how they feel. Indeed, scientists at Amazon are trying to grasp our emotions based on our utterances.

By listening in to conversations, corporations can learn quite a bit about users’ health, finance and social life. Are users willing to offer extensive data away to those tech giants in service of more natural conversational experiences with voice agents? What is required for a more ethical and desirable future with voice agents?

Through natural conversations with voice assistants, we should always handily have the option to unlock cutting-edge AI technologies without the tedious learning process often experienced with graphical user interfaces. Recent technological advancements reminiscent of the event of nearly human-level language-generation models and speech synthesis promise the arrival of truly natural voice agents.

Striking a balance between a benevolent assistant and a friendly interlocutor is close by, but it should take more research to generate significantly higher tool support for voice interface designers, and would require users to share their data.

This article was originally published at