Two years since its demise, the spectre of Microsoft's animated paperclip, Clippy , still haunts anyone hoping to develop a virtual assistant to help people get things done. Few have tried to push virtual assistants to the public since.
But Clippy's unpopularity hasn't deterred the US Defense Advanced Research Projects Agency (DARPA) from spending an estimated $150 million on its own virtual helper.
And although intended to ease the US military's bureaucratic load, an artificially intelligent helper based on the project is heading the way of consumers later this year.
Begun in 2003 the CALO, for Cognitive Assistant that Learns and Organizes, project involved over 60 universities and research organisations and is the largest ever non-classified AI project. It ends this Friday and has produced a virtual assistant that can sort, prioritise, and summarise email; automatically schedule meetings; and prepare briefing notes before them.
That focus could make the crucial difference between CALO being an annoyance like Clippy, and a genuinely useful helper.
Most software capable of learning needs large numbers of examples for something to stick – a spam filter trained on millions of emails, for example. But CALO needs to be quicker on the uptake. If it takes thousands of examples to learn how someone likes their email sorted, frustrated users will soon switch it off.
So the developers have built in tricks such as "transfer learning", which applies lessons from one domain to another. For example, if a person consistently marks emails from one person, perhaps their boss, as high-priority, CALO can use that knowledge to order their meeting schedule too.
A spin-off of DARPA's project, an app called Siri, will be coming to Apple's iPhone later this year. Siri has been designed to assist with mundane tasks, such as checking online reviews to find a good local restaurant and booking a table.
Rather than having to personally trawl through multiple websites to find a likely eatery, get the contact details and address, and make a reservation, the user can verbally instruct Siri to, say, "find me a romantic Thai restaurant in this area".
Siri uses navigates the various web services for the user, even booking restaurants and taxis through web forms where possible. The person it is helping can also book cinema tickets, and search for flights or weather forecasts without typing a word.
Another CALO spinoff is Social Kinetics, a social-network analysis package that helps people organise their contacts by criteria that can include relationship and expertise.
The system is already being used by the military to track how information flows through the ranks, identify experts, and generate a repertoire of answers to standard questions.
The consumer version focuses on healthcare, connecting people to the experts and information they need to make decisions about their health and treatment.
Bart Selman, an AI researcher at Cornell University who is not involved with the CALO projects, says that virtual assistants are not yet comparable to human help. "It's safe to say that the system does not yet perform at the level of a dedicated personal assistant," he told New Scientist.
But, he continues, in most organisations, human helpers are a luxury most people do not have. In these situations, automated, dedicated, and personalised assistants could be helpful – as long as they don't bug the hell out of their users.
This article originally appeared on New Scientist.