Imagine a future where electronic text messaging is tracked by an intelligent algorithm that can identify truth from lies. A new study from two US researchers suggests this kind of online polygraph is entirely possible, with early experiments showing a machine learning algorithm can separate truth from lies based just on text cues over 85 percent of the time.
"I think we all have good common sense about the people we meet face to face, but how much common sense do we have with the strangers we encounter online where you can meet a lot of people very fast," says Shuyuan Ho, a researcher from Florida State University working on the new study. "This research is so important because it can provide another reference point offering more protection."
In order to analyze whether truth and lies can be discerned from simple text-based communication the researchers designed an online game that randomly assigned players the roles of either "Saint" or "Sinner". Forty subjects took part in 80 specific game sessions and a machine-learning system was trained to look for discrepancies between those truth-telling Saints and the lying Sinners.
A large variety of language cues appeared to separate those lying from those telling the truth. Liars were found to reply faster, using words such as "always" and "never" to affirm a sense of certainty. Those telling the truth, on the other hand, were found to be more speculative in their responses, using words such as "guess" and "perhaps", and taking more time to respond.
"Truthtellers tend to say 'no' a lot," says Ho. "Why? They like to provide emphasis when they explain their reasons. If you ask them, 'Is this true?' They tend to say 'no' because there is another true reason."
Micro-pauses between responses – and even between words when typing – were a fascinating cue the machine-learning system picked up on that would most likely be imperceptible to humans. Those telling the truth seemed to type slower and deliver more considered responses than those telling lies.
Ultimately, the machine-learning system demonstrated a stunning ability to identify deception and lies, with an accuracy rate of between 85 and 100 percent. This compared to the human subjects only accurately spotting lies around 50 percent of the time.
The researchers readily admit this is an early study with a small sample size, but the statistical significance of these cursory results demonstrate incredible potential for this kind of technology to effectively detect truth or lies from simple text-based communication cues. Ho is keen to develop the technology into a system that can be widely deployed, although it is yet to be seen whether the general public is eager to have a computerized polygraph attached to messaging apps like WhatsApp or Facebook Messenger.
"This basic research offers great potential to develop an online polygraph system that helps protect our online communication," says Ho. "I want to get the world's attention on this research so we can hopefully make it into a commercial product that could be attached to all kinds of online social forums."
The study was published in the journal Computers in Human Behavior.
Source: Florida State University
