Email management is a mundane example of how machinelearning
is starting to be used. The system decides whether
or not to notify a person of an incoming message, depending
on the nature and content (and therefore the urgency) of
it, and also on the extent to which the person is willing to
tolerate a disturbance at that particular moment, which
itself depends on the task in which the person is engaged.
Contextual information can also be used to make a
decision about how relevant the email is, from that person’s
calendar, from audio and video sensors which monitor the
person’s focus of attention, and from log files of past user
behaviour. Of course, this is for more advanced needs;
machine-learning is also used to filter out the much more
commonplace and vexing volumes of spam that increasingly
assault our mailboxes.
As with previous generations of intelligent systems,
however, the success of machine-learning will depend on
how accurate the machine’s algorithms are at inferring a
person’s intentions and their actions at a given moment.
While people are very much creatures of habit, they can also
be highly unpredictable and complex in their needs and
desires. For a machine-learning approach to truly succeed,
it may well require that both users and computers make
their intentions visible to each other: machines indicating
to users what they think users want, and users indicating
to the machines what they want in turn. Users also like
to know how a machine is making its decisions, so ways
of communicating how the mechanisms work may be as
important as the outcome.
All of this proposes that humans and ‘intelligent’ machines
often need to be able to negotiate, question and answer
back – unlike current vehicle navigation systems (‘satnav‘),
whose instructions telling people where to go are
sometimes blindly followed by hapless drivers who never
question them. If people are prepared to stupidly obey
instructions given out by simple computers, this should
make us even more concerned about the relationship
between people and ever more complex computers as we
move toward 2020.