How good are you at catching a lie? Research suggests that despite the fact many liars expose their own deceit in the way they talk, we humans are terrible at picking up on it. Half the time we're right. Half the time we're wrong. Computers, it turns out, are much better at detecting these clues, and they're getting better all the time. Programs now exist that can digest transcribed speech and judge the veracity of the statements by looking for give-away patterns in the language.
The people who wrote these programs say that, while not perfect, they're already accurate enough to prove useful in many situations, from interviewing job candidates to catching insurance fraud.
"We wouldn't want to say that you should send someone to jail based on the results. But what we would like to show you is where to take the questioning and what to follow up on," says Joan Bachenko, a computational linguist.
Bachenko is the co-founder of Deception Discovery Technologies, a company that takes transcripts of interviews and flags them for suspicious statements.
"We were interested in deceptive language and whether there could be a way to program a computer to be able to detect language in an interview or a deposition or testimony," she says.
Over the years, data from cognitive science and criminal psychology have identified aberrations in the speech of people who are motivated to deceive. But the methods of this research have left something to be desired. Most of the work has been done in a laboratory setting, asking student participants to fabricate stories that contradict their own convictions.
While these experiments have opened new avenues of inquiry, Bachenko felt that in their premise they lacked the stress of a real-world situation. No one stood to really lose anything, and there was no saying how this would change the language the students were using.
"The work that had been done so far was coming up short," says Bachenko.
With the help of Eileen Fitzpatrick, the chairwoman of the linguistics department at Montclair State University in New Jersey, she turned to a new source of data. Together, they compiled speech samples using public records-depositions from criminal trials, police interrogations, interviews before Congress-and performed painstaking background work to fact check every statement contained in them.
"It's high stakes. They can go to jail. They can lose their retirement pension. They lose face," says Bachenko. "This is the kind of data we want."
Next, they set about building a program that could detect deceit in these transcripts, asking it to flag well-known indicators of mendacious speech. Most of the indicators come from experimental laboratory research, but they will be familiar to anyone who has watched enough CSPAN. Hedging phrases, like "to the best of my knowledge" and absolute statements like "under no circumstances," have both been associated with deception, as have negative terms and prefixes like "non" and "un."
Beyond word choice, the way that people organize their speech can reveal deceit, as well. It's been shown that liars put all the meaty details in the beginning and end of a story, as though to distract the listener.
"In the middle where the truth teller will have a lot of concrete narrative about what happened, the liar will give a very skimpy report," says Fitzpatrick.
Even the most adept liars will drop these hints when they are pressed to give their story, an observation that has been explained by the hypothesis that they are either spending too much energy juggling the consistency of their own fabrications to closely monitor their speech or that their feelings of guilt give them away.
"All of these things are considered to be leakage," says Fitzpatrick.
After being trained to look for these semantic leaks, the program Bachenko created picked up on more than 75 percent of the false statements that were made.
Her approach is particularly useful because it judges each of the facts in the transcript rather than the overall intent of the person who is speaking. The system goes through the statement line by line and rates the veracity of the language.
Once analyzed, a transcript can be used to determine much more than whether the speaker is a liar or not. It shows what parts of the story are most suspicious, functioning as a tool for further investigation.
Fitzpatrick says improving the accuracy is within reach but will require much more data. This is the most expensive part of the work, as it requires her to become a bit of a detective herself.
She is applying for funding to broaden her transcript database-what linguists call a "corpus." But the work, she says, has already reached a level of accuracy that will be useful and which far outshines human performance.
"You have to think what your baseline is. How good is 76.5 percent?" she says. "There's evidence that people perform randomly."
And while us humans continue to be our gullible selves, computers will continue to hone their powers of skepticism.
Morgen E. Peck is a contributor to IEEE Spectrum, Innovation News Daily and Scientific American.
This article originally appeared at Txchnologist.