2014-06-10 12:45:21
By Chris Baraniuk
Intelligent machines that can pass for humans have long been dreamed of, but as
Chris Baraniuk argues, they re already among us.
Sometimes it s the promise of sex that fools you. Sometimes it s because they
seem wise, friendly or just funny. The bots don t really care how they trick
you their only objective is to make you think they re human. In fact, if you
use social media or spend any time online, it s quite possible you ve already
been a victim.
This week, a controversial claim was made that a chatbot passed the Turing
test at an event at the Royal Society in London. During a series of text-based
conversations, a computer program named Eugene Goostman persuaded judges it was
a 13-year-old Ukrainian boy, thus passing a benchmark for artificial
intelligence proposed years ago by the computer scientist Alan Turing.
So does this announcement mark the era of human-like AI, as has been claimed?
Not really. Turing s test stopped being important for AI research years ago,
and many scientists see the contests as flawed because they can be won with
trickery such as pretending to be a non-native English speaker.
However, what chatbots are fully capable of in everyday life is far more
interesting. We re already surrounded by bots capable of tricking us into
thinking they are real people, and they don t enter competitions. Some are
sophisticated enough to infiltrate social networks and perhaps even influence
public opinion.
There are certainly plenty of them out there. Although most people think of the
web as a place primarily frequented by humans, the reality turns out to be
quite different. A recent report found that 61.5% of internet traffic is
generated by automated programs called bots.
Honey trap
The bots most likely to fool us employ colourful trickery, explains Richard
Wallace of Pandorabots, which makes chatbots for customer service and other
uses. Wallace is the creator of a bot called Alice, which on three occasions
has won the Loebner Prize a Turing-like contest in which chatbots vie to
convince judges that they are human.
The people who are the most skilful authors of these bots are not people who
are computer programmers, they are people who work in a creative field, says
Wallace. That s really the key to creating a believable chatbot writing
responses which are believable, entertaining and engaging.
Scammers are well aware of this phenomenon. Security research firm Cloudmark
has documented the rise of a flirtatious bot called TextGirlie . After
obtaining a victim s name and telephone number from their social media profile,
TextGirlie would send the victim a personalised message asking them to continue
the conversation in an online chatroom. A few coquettish exchanges later and
the victim would be asked to click on a link to an adult dating or web cam
site.
Cloudmark estimates that as many as 15 million initial TextGirlie text messages
could have been sent to mobile phones and they confirm that the scam operated
for several months. According to Andrew Conway, a research analyst at the firm,
this is a good indication that the attack was in some measure successful.
Automated deceit
People are more likely to be fooled by a bot in a situation where they d expect
odd behaviour or broken English. Back in 1971, for example, psychiatrist
Kenneth Colby was able to convince a few fellow practitioners that they were
talking to a patient via a computer terminal. In fact, Colby had simply set up
sessions with a program that simulated the speech of a paranoid schizophrenic.
And more recently, in 2006, psychologist Robert Epstein was fooled by a
cleverly programmed computer which wore the guise of a Russian woman who said
she was falling in love with him. Lately, bots have been turning up on online
dating networks in droves, potentially ensnaring more hapless singletons in a
web of automated deceit.
Sometimes, bots can even trick the web-savvy. Birdie Jaworski knows what it
feels like. Jaworski is a seasoned contributor to Reddit and fan of the digital
currency called dogecoin, a playful alternative to Bitcoin. On the Reddit forum
for dogecoin aficionados, a user called wise_shibe emerged recently, posting
witty remarks in the style of ancient proverbs. He would reply to you with a
fortune cookie style response, remembers Jaworski. It would sound like
something Confucius might say.
These comments even started making wise_shibe money, since the forum allows
users to send small digital currency tips to each other if they like a
comment that s been made. The wise_shibe rejoinders were popular, so were
showered with tips. But things soon started to look suspicious: the account was
active at all hours and eventually started repeating itself. When wise_shibe
was unmasked as a bot, the revelation divided members of the forum. Some were
incensed, while others said they didn t mind. Jaworski was amused, but also
felt cheated. All of sudden you realise this little robot is collecting all of
these tips, she says.
Phantom tweeters
If a bot s presence and interactions appear natural enough, it seems to be the
case that we are unlikely to even question its legitimacy we simply assume
from the outset that it s human. For Fabricio Benevenuto, this phenomenon has
become the subject of serious research. Recently he and three other academics
published a paper which explains just how easy it is to infiltrate Twitter with
socialbots so long as they look and act like real Twitter users.
Benevenuto and his colleagues created 120 bot accounts, making sure each one
had a convincing profile complete with picture and attributes such as gender.
After a month, they found that almost 70% of the bots were left untouched by
Twitter s bot detection mechanisms. What s more, the bots were pre-programmed
to interact with other users and quickly attracted a healthy band of followers,
4,999 in total.
The implications of this are not trivial. If socialbots could be created in
large numbers, they can potentially be used to bias public opinion, for
example, by writing large amounts of fake messages and dishonestly improve or
damage the public perception about a topic, the paper notes.
It s a problem known as astroturfing , in which a seemingly authentic swell of
grass-root opinion is in fact manufactured by a battalion of opinionated bots.
The potential for astroturfing to influence elections has already raised
concerns, with a Reuters op-ed in January calling for a ban on candidates use
of bots in the run-up to polls.
'More sophisticated'
The ramifications of astroturfing are in fact so serious that the US Department
of Defense has jointly funded research into software which can determine
whether a Twitter account is run by a bot. The application, called BotOrNot, is
available publicly online and provides a predictive analysis based on account
activity and tweet semantics which suggest whether the account operator is
likely to be a human or a bot.
But Emilio Ferrara, a lead researcher on the project, admits that the system
may already be outdated. Trained on Twitter data which is now three years old,
it s possible that today s best bots could still evade detection.
Now bots are more sophisticated, he says. They are better at disguising
their identity and looking more like humans. Therefore the task becomes harder
and harder we don t even know the accuracy of the system in detecting the
most recent and most advanced bots out there.
And so the rise of bots only looks set to continue with or without Turing
test approval. For Fritz Kunze of Pandorabots, the hope is that people will get
better at questioning innocent-looking users who contact them online so that
they re not so easily duped. But he is also acutely aware of how hard a task
that will be in the near future.
It s going to be a big shock to most people, he says. And these bots are
going to be really, really good they re going to be good at fooling people.