Attempting to use AI in a lie detector only exacerbates the problem of fraud recognition

A detailed study of attempts to use artificial intelligence in lie recognition

Before the polygraph issued the verdict “guilty,” Emmanuel Mervilus worked for a vegetable oil company in the port of Newark, New Jersey. He earned $ 12 / hour by carrying boxes, but this was not enough for a living. His brother and sister were too young to work, and his mother waged an expensive battle with cancer. However, the boss in the port said that he was next in line to upgrade to a technical position at which he was promised to pay $ 25 / hour.

Mervilus was still waiting for his promotion when on October 19, 2006, he and a friend stopped for a meal at Dunkin 'Donuts, located in the nearby city of Elizabeth, New Jersey. A few minutes later, as they walked along the street, two policemen approached them and accused them of having robbed a man who had been threatened with a knife a few minutes ago near the railway station.

A victim from afar identified Mervilus and his friend. In a desperate attempt to prove his innocence, Mervilus suggested taking a polygraph test. The police agreed, but shortly before this test, Mervilus's mother died. When the police connected him to the apparatus, he was confused and anxious. He failed this test, asked for the opportunity to pass it again and was refused.

After Mervilus's plea of ​​innocence, the case was referred to court. The test lieutenant said in court that the device is a reliable "indicator of truth." He said that he had never seen in his career that “someone showed signs of fraud, and then it turned out that he was telling the truth.” The jury found Mervilus guilty - which, as it turned out at the appellate court, happened because of excessive faith in the polygraph. The judge awarded him 11 years in prison.

The belief that deception can be recognized by analyzing the characteristics of the human body is deeply rooted in modern life. Despite many studies casting doubt on the reliability of the polygraph, more than 2.5 million inspections are conducted in the United States every year, and the polygraph industry is estimated at $ 2 billion. US federal government agencies, including the Department of Justice, the Department of Defense and the CIA, use this device for assessment of candidates for work. A 2007 Justice Department report indicates that more than three-quarters of police stations and sheriff’s offices use lie detectors to recruit staff.

However, these devices are still too slow and clumsy to be used at borders, at airports or in large groups of people. As a result, a new generation of AI-based lie detectors have emerged over the past decade. Their supporters claim that they work both faster and more accurately than polygraphs.

In fact, the psychological justification for these new AI systems is even more precarious than the studies that underlie the polygraph. Evidence that the results they produce can be trusted is scarce. However, their external gloss, given by the use of AI, leads to the appearance of these systems in places where the polygraph could not penetrate earlier: to the border, to interviews, to procedures for assessing creditworthiness and investigating insurance fraud. Corporations and governments begin to rely on them when making decisions about the reliability of customers, employees, citizens, immigrants and international tourists. But what if a lie is a piece too complicated to be reliably detected by any machine, no matter how advanced the algorithms are?

* * *

Inquisitors in ancient China put rice in their suspects' mouths to see if saliva was released from them. " Roman acts ", an anthology of medieval moralistic stories, tells the story of a soldier who ordered his clerk to measure the pulse of his wife, to determine whether it has been faithful to him.



After the United States got involved in World War I, William Marston, a Harvard researcher, was the first to use blood pressure measuring machines in an attempt to detect fraud. A few years later, inspired by the work of Marston, John Augustus Larson, a policeman who recently received his doctorate in physiology from the University of California at Berkeley, developed a machine called a “cardio-pneumo psychographer,” which provided continuous data on the subject’s blood pressure, pulse rate and speed. breathing. Larson argued that these testimonies far better betray deception than just one pressure.

Initially, Larson used the car to investigate the theft at a female dormitory in Berkeley, and for a year he was used in San Francisco to convict a man accused of killing a priest. By the 1930s, one of Larson’s proteges was already selling portable versions of the device to police departments across the country, adding a galvanic skin reaction sensor - the more the subject sweated, the better the skin conducted the current. By the 1970s, millions of private sector workers were regularly tested by polygraphs as directed by their employers.

Most modern polygraphs use the same basic scheme that Larson suggested: the investigator asks several questions to measure the normal physiological state of the subject, watching how the machine translates these measurements into wave-like lines on paper or screen. The investigator then looks for sudden surges or drops in these levels when the subject answers questions related to crimes or feelings.

However, physiologists and neuroscientists have criticized the polygraph almost from the very moment that Larson discovered his invention to the public. If some liars may be experiencing a change in heart rate or blood pressure, there is very little evidence that such changes are consistently correlated with deception. Many innocent people begin to get nervous during interrogation, and experienced liars can suppress or cause changes in their bodies that allow them to fool the test. A polygraph can also be fooled by biting its tongue , stepping on a carnation, or thinking about the most terrible fears. There is always a risk that the device will receive conflicting testimonies even under the controlled conditions of a laboratory experiment, and in real life they are even less reliable: since criminals who cheated on the test almost never admit their guilt to the police, and innocent suspects often give false testimony, failing tests, it’s impossible to say how well these machines actually performed.


American inventor Leonard Keeler (1903-1949), protege of polygraph inventor John Larson, tested Bruno Hauptmann, who was arrested, charged and executed for abducting Charles August Lindberg Jr. Hauptman until the end of his life declared his innocence.

Because of these limitations, polygraph tests were not accepted in most American courts for a long time, unless both parties agreed to attach them to the case. Federal law has banned private companies from testing polygraphs of employees since 1988 (with the exception of particularly sensitive work such as armed guards or drug distributors, as well as suspicions of theft or fraud). The American Psychological Association warns that "most psychologists tend to believe that there is too little evidence of the ability of a polygraph to pinpoint a lie." In a 2003 report from the National Academy of Sciences, following a government study of this issue, a conclusion was made that soon became widely known: a machine identifies liars "much more often than by chance, but much worse than perfectly."The lead author of the report at that moment said that "national security is too important a thing to give to such a crude instrument."

But perhaps this tool can be made less rude. A similar promise is made by a growing number of companies enthusiastically trying to sell lie recognition technology to both governments and commercial organizations. They argue that, perhaps, certain complex patterns of behavioral traits can tell that a person is lying, much more reliable than just a jumped pulse or blood pressure. And perhaps a sophisticated algorithm can recognize these patterns.

From 1969 to 1981, the serial killer, nicknamed the Yorkshire Ripper, hunted girls in northern England, killed at least 13 of them, and tried to kill at least seven more. Police interrogated and released him nine times while he continued his bloody journey. His last victim was Jacqueline Hill, a 20-year-old student at Leeds University, who was killed in November 1980. A few months later, the police finally caught him preparing to kill a prostitute in Sheffield.

When Janet Rothwell arrived at Leeds University in the fall of 1980, she lived in a dormitory in the room next to the one in which Hill lived. Killing Hill scared her.

“She got on the bus at the university library at about the same time as me,” said Rothwell, “and was killed after she got off the bus.” Rothwell later found out how long it took to capture the killer. “I thought,” she recalled, “could the computer find any behavioral discrepancies to inform the police?”

As a result, Rothwell went to graduate school at the University of Manchester Metropolitan (UMM) in the late 90s. There she met Zuhair Bandar, a British lecturer of Iraqi descent who worked in the Department of Computer Science. Shortly before this, Bandar had an idea - after one advertising company asked him to create a rudimentary device to measure the interest of customers in the products that they see on the screen.


A photo taken by the FBI of a woman undergoing a polygraph test

“They wanted to hand out a portable device to consumers,” Bandar said, “so that when the consumer liked something, he would press 1, and if not, to 2. I thought - why should such devices be made if they already have expressions on their faces? ” Bandar suggested that Rothwell remain at UMM after receiving his diploma in order to work on his doctorate, helping him develop software capable of analyzing faces in order to extract information. They decided that cheating was no more difficult to recognize than joy or anger. Any of these emotions should create some kind of “inconsistency” - behavioral patterns, verbal or non-verbal, which the computer can recognize.

Rothwell trained the neural network in the early 2000s to track activities such as blinking or blushing, and then fed dozens of videos to the computer where people answered the same set of questions honestly and dishonestly. To determine the common features of liars, the computer studied the details of the movement of persons, their relationships, and the relationships between these relationships, giving out a kind of “theory” that would be too difficult to express in normal language. Having studied in this way, the system could use the acquired knowledge to classify new subjects in the categories of “true” and “deceiver” by analyzing frame-by-frame changes in the expressions of their faces.

A 2006 study examined the feasibility of this system, called the " Silent Speaker"(Silent Talker), to recognize a lie in the test subject's answers. She was not able to achieve accuracy of more than 80% - neither at the time Rothwell worked with her, nor later, when the research team tried to improve her. Also, Rothwell told me that the system in general ceased to work normally if the subject was wearing glasses, noting that “the lighting conditions were the same, and all interrogations were related to staged theft.” But Rothwell recalls that even at the very early stages of the project Bandar “was passionate about the idea of ​​releasing a commercial product”; Once, she and another colleague provided her with a video showing a woman suspected of cheating on her husband and asked her to drive the video through Silent Talker for analysis - just like in the book “Roman Acts”.

Rothwell had doubts about this. “It was clear to me that if such software worked, it could in principle be used to the detriment,” she said. “I don’t think that any system will be able to come close to 100% accuracy, and if the system is mistaken, it can cause catastrophic consequences for relationships and life situations.” In 2006, she left the university, studied at the audiologist, got a job in a hospital on the island of Jersey, where she lives to this day.

In 2003, UMM published a press release promoting the technology as a new invention that will replace the polygraph. “I was shocked,” said Rothwell, “it seemed to me too early to talk about it.”

The US government has repeatedly tried to tackle lie recognition technology in the first few years after 9/11; The US Department of Homeland Security (DHS), the U.S. Department of Defense (DoD), and the U.S. National Science Foundation spent millions of dollars on each study. These agencies funded the creation of an AVATAR machine at the University of Arizona. AVATAR, which analyzed facial expressions, body language, and people's voices, assigning them “confidence points,” was tested at airports. In Israel, DHS helped money start-up WeCU [“we see you”, or “we see you” / approx. transl.], which sold a machine for assessing people, capable, according to a 2010 article in the Fast Company magazine, of “causing physiological reactions in people hiding something." Today, this company has already gone bankrupt.

Bandar set about trying to bring the technology to market. With his two students, Jim O'Shea and Keely Crocket, he turned his Silent Talker into a company and began looking for clients for his psychological profiling technology, both among police stations and private companies. Silent Talker was one of the first AI-based lie detectors to enter the market. According to the company, last year the technology “created on the basis of Silent Talker” was used as part of the iBorderCtrl initiative, funded by the European Union, in which the system was tested on volunteers on the borders of Greece, Hungary and Latvia. Bandar says that the company is currently negotiating the sale of technology to law firms, banks, insurance companies, about the possibility of using these tests during interviews and checking for fraud.

Bandar and O'Shea over the years have adapted the basic algorithm for use in various versions. They tried to advertise it to Manchester and Liverpool police stations. “We communicate informally with people of a very high position,” the company told The Engineer British magazine in 2003, noting that it was trying to “test the technology in real interviews.” From a report published by O'Shea on his website in 2013, it follows that Silent Talker “can be used to protect our soldiers in foreign operations from insider attacks” (meaning attacks carried out by Afghan soldiers in uniform against former allies).

The team also published experimental results demonstrating how Silent Talker can be used to recognize not only hidden motives, but also an understanding of something. In a 2012 study, which first showed how Silent Talker works “in the field,” the team, together with a Tanzanian nongovernmental medical institution, recorded facial expressions of 80 women who received online training on AIDS treatment and condom use. The idea was to determine whether patients understand how they will be treated - as was written in the study notes, “evaluating participants' understanding while delivering information to them is still an area of ​​concern”. When the team made a cross-comparison of the AI ​​scores of how much women understood the material with the points they got for the short exams, they foundthat AI predicted with 80% accuracy which of the subjects would pass the exam and which would fail.

Silent Talker was included in the iBorderCtrl initiative thanks to the Tanzanian experiment. In 2015, Athos Antoniades, one of the organizers of the nascent consortium, sent O'Shea email asking if the Silent Talker team would like to join the group of companies and the police forces that are sending applications for EU grants. Constantly increasing traffic on roads overloaded the border guards of the European Union, as a result of which the union offered € 4.5 million to any organization capable of “organizing a more efficient and safe border crossing, contributing to the prevention of crime and terrorism.” Antoniades thought Silent Talker could play a key role in this.

When the project announced public testing in October 2018, the European Commission immediately began actively promoting the “success story” of the system’s “unique approach” to detect fraud, explaining that it “analyzes the travelers micro-gestures to tell which of the interviewees is telling lies” . The algorithm trained in Manchester was to “ensure more efficient and safer border crossing” and “contribute to the prevention of crime and terrorism”.

O'Shea told me that the basic algorithm of the program can be used in many other conditions - in advertising, before paying insurance, when hiring, in evaluating employees. It was difficult for me to share his sincere belief in the wisdom of this algorithm, but while we were talking on the phone with him, Silent Talker was already used for a voluntary examination of those wishing to enter the European Union; The company launched this project as a commercial enterprise in January 2019. So I decided to go to Manchester to see everything myself.

* * *

Silent Talker’s offices are located about one and a half kilometers from UMM, where O'Shea is currently a senior lecturer. He took on the daily responsibilities of technology development. The company is located in a tiny office center located in a residential area, next to a kebab diner and opposite the football pitch. In the office center, the Silent Talker office consists of a single room with several computers, desks with briefcases on them and explaining posters from the 2000s that tell how this technology works.

When I went to them in September, I talked with O'Shea and Bandar in the meeting room. O'Shea looked stern, but a little tousled, was bald, with the exception of a couple of tufts of hair and a beard in the style of Van Dyck. He started the conversation by demanding that we not touch on the iBorderCtrl project, and later called his critics misinformed. He described the abilities of the AI ​​platform of the system verbose and ornate, sometimes citing computer science pioneer Alan Turing or language philosopher John Searle.

“Both cars and people have their own speculations - beliefs, desires and aspirations associated with objects and states of affairs in the world,” he said, defending the system’s dependence on the algorithm. “Therefore, the development of complex applications requires consideration of the ideas and intentions of both parties.”

O'Shea demonstrated the system, allowing her to analyze a video with a person answering questions about whether he stole $ 50 from the box. The program imposed a yellow rectangle on the person’s face and two smaller rectangles on his eyes. When he spoke, the pointer in the corner of the screen moved from green to red when his answers were false, and then returned to the middle orange position when he fell silent. At the end of the interview, the program issued a graph showing the distribution of the probability of fraud over time. Theoretically, the graph showed where he started and ended up lying.

O'Shea says that their system can run on a regular laptop, and users pay $ 10 per minute for the video being analyzed. O'Shea told me that the software pre-processes the video locally, sends encrypted data to the server, where it is further analyzed, and then sends the results back: the user sees a graph of the probability of fraud superimposed on the video.

According to O'Shea, the system monitors about 40 physical “channels” on the subject’s body - everything from blinking speed to head angle. Each new person is compared with a “theory” of deception, developed on the basis of viewing training data, which includes records of liars and truth-telling people. By measuring facial expressions and posture changes several times per second, the system looks for patterns in these movements that coincide with the common ones for all liars from the training data. These are not as simple patterns as looking at the ceiling or bowing your head to the left. This is more like patterns of laws, multifaceted relationships between different movements that are too complex to be tracked by a person is a typical task for machine learning.

The task of AI is to determine what patterns of movement can be associated with deception. “Psychologists often talk about the need for a model of how the system works,” O'Shea told me. “But we do not have a working model, and we do not need it.” We give AI the opportunity to solve the problem. " However, he also says that scientific evidence on the psychology of deception confirms the meaningfulness of “channels” on the face. In a 2018 paper, describing the Silent Talker, its creators say that their software “assumes that certain states of consciousness associated with the behavior of the deceiver will, during deception, control the non-verbal behavior of the interviewee.” Examples of such behavior include “cognitive loading”, additional mental energy that is supposedly spent on lies, and “delight of deception”, the pleasure that a person allegedly receives,successfully lying.


Paul Ekman, whose theory of “microexpressions” is much controversial, has advised many US government agencies.

However, Ewaut Meyer, a psychology professor at the University of Maastricht in the Netherlands, says that the theoretical basis for assertions about the universality of such behavior patterns can be called precarious at best. The idea that a person can detect characteristic signs of behavior comes from the work of Paul Ekman, an American psychologist who in the 1980s advanced the famous theory of “microexpressions,” unintentional movements of the facial muscles that are too small to be controlled. Thanks to research, Ekman became the best-selling author and prototype of the foolish television show, “Lie to Me.” He has advised many US government agencies, including DHS and DARPA. Under the pretext of national security, he keeps research data secret. Because of this, there is constant debate about whether these microexpressions have any meaning at all.

The Silent Talker AI tracks various facial muscle movements, not just Ekman microexpressions. “We disassembled these high-level hints, composing our set of microscopic gestures, and trained AI to recombine them into meaningful characteristic patterns,” a company representative wrote to us. O'Shea says this allows the system to detect trickery-related behavior even when the subject simply looks around or changes position while sitting on a chair.

“Much depends on whether you have a technological or psychological question,” says Meyer, warning that O'Shea and his team may have turned to technology in search of answers to psychological questions regarding the nature of the deception. “AI may be better than people to detect facial expressions, but even if so, this does not mean that one can draw conclusions from them about whether a person is lying. Lying is a psychological construct. ” There is no consensus not only on the question of which expressions are associated with a lie, Meyer adds: there is no consensus on whether there are such expressions at all. The company wrote in an email that this criticism “has nothing to do” with Silent Talker, and that “the statistics used are not suitable for this case.”


The TV show “Lie to Me” was, in particular, based on Ekman’s theory of microexpressions.

In addition, Meyer points out that the algorithm will still be useless at the borders or in interviews if it is not trained on the same diverse data set that it will evaluate in reality. Studies suggest that facial recognition algorithms recognize racial minorities worse if they are trained on the faces of white people - O'Shea himself acknowledges this. A representative of Silent Talker wrote to us: “We did a lot of experimentation with a smaller sample size. Their number reaches hundreds. Some of them are related to scientific research and will be published, others - commercial and confidential. ”

However, all published studies confirming the accuracy of Silent Talker are based on small and uniform data sets. In the work of 2018, for example, only 32 people were used for training, among whom there were twice as many men as women, and only 10 of them were Asians or Arabs, and there were no Negroes or Hispanics at all. And although the program has “settings” to analyze both men and women, O'Shea said he was not sure if she needed settings for race or age.

* * *

Following the pilot’s announcement of the iBorderCtrl initiative, activists and politicians denounced the program as an unprecedented attempt to bring universal surveillance to Orwell’s level. Sophia Int Weld, a Dutch member of the European Parliament and leader of the center-left Democrats 69, said in a letter to the European Commission that the Silent Talker system could violate the “fundamental rights of many travelers crossing the border,” and that organizations such as Privacy International have condemned it as “ part of a wider tendency to use opaque and often inadequate automatic systems for judging, evaluating, and classifying people. ” The iBorderCtrl consortium clearly did not expect to meet such resistance: if initially the European Commission stated that iBorderCtrl “will develop a system to accelerate border crossings,” now the representative saysthat the program was a purely theoretical research project. Antoniades in 2018 told the Dutch newspaper that the lie recognition system “may not be created in the end,” but for now, Silent Talker still continues to advertise its presence on the iBorderCtrl initiative on its website.

Silent Talker is the “new version of the old fraud,” said Vera Wilde, an American scholar and privacy activist based in Berlin who helped launch the campaign against iBorderCtrl. “In a sense, it's the same fraud, but using an even worse scientific foundation.” When checking on a polygraph, the investigator monitors physiological events that are believed to correlate with falsehood; in the case of AI, the investigator allows the computer to detect the correlation itself. “When O'Shea talks about his lack of theory, he is mistaken,” she says. “He has a theory, just bad.”

But no matter how much people like Wilde criticize this idea, the dream of an ideal lie detector does not want to die - especially when it is embellished with AI. After the US Department of Homeland Security spent millions of dollars researching lies at universities in the 2000s, it tried to create its own version of technology that analyzes behavior. His system, called Future Attribute Screening Technology (FAST) [the technology of the future to track characteristic properties], is aimed at finding a person’s criminal tendencies based on the movements of his eyes and body (in an earlier version of the technology, the subject had to stand on the Wii Balance Board controllerto track posture changes). Three researchers who spoke secretly about secret projects say the program never took off - there were too many contradictions in the department about whether to use Ekman's microexpression as the basis for the analysis. In 2011, the program was curtailed.

Despite the failure of FAST, DHS does not lose interest in lie recognition technologies. Last year, she signed a $ 110,000 contract with a recruiting company to train her employees in “recognizing lies and reactions” through “behavioral analysis”. Other ministries and departments continue to support AI-based solutions. The military research laboratory (ARL) has a contract with Rutgers University to create an AI program for recognizing lies in the Mafia salon game, which is part of an overall project to create “something like Google Glass that can warn us about a couple of pickpockets on crowded market, ”wrote Purush Iyer, project manager at ARL. The Israeli company Nemesysco, which sells voice analysis software using AI, told methat her technology is being used by New York City Police and Midwest Sheriffs to interrogate suspects, as well as collection agencies to measure debtor emotions during phone calls.

However, the immediate and potentially dangerous future of AI lie detectors seems to be their private use. Politicians supporting initiatives such as iBorderCtrl ultimately have to answer voters, and most AI-based lie detectors can be prohibited from being used in court on the same grounds as a polygraph. But private corporations have fewer restrictions on using such technology to evaluate job candidates and potential customers. Silent Talker is one of several companies claiming to have a more objective way of recognizing abnormal or deceptive behavior, giving customers a “risk analysis” method that goes beyond credit rating and social media profiling.

Montana’s Neuro-ID is conducting an AI analysis of mouse movements and keyboard taps to help banks and insurance companies assess fraud risk by assigning “credibility points” from 1 to 100 for loan applicants. In the video the company showed me , the client fills out an application for a loan online, and spends time filling out a field regarding income for a family, while moving the mouse - and all this the system takes into account for calculating the reliability score. The system is based on studies conducted by the company’s founding scientists claiming they showed a correlation between mouse movements and emotional outbursts. They described that "an attempt to cheat can increase the normalized distance of mouse movement, reduce the speed of movement, increase the response time and lead to an increase in the number of clicks."However, according to the internal tests of the company itself, it is clear that their software produces too many false-positive results: in one study in which Neuro-ID processed 20,000 applications on the online store website, less than half of the applicants who received the lowest ratings (up to 10) , turned out to be scammers, and only 10% of people who got ratings from 20 to 30 were associated with the risk of fraud. The company recognizes that the software notes as suspicious job seekers who may be innocent, and makes it possible to use this information at its discretion. A company representative told me that “there is no 100% accurate behavioral analysis. "We recommend that you use these results in conjunction with other information about applicants to make better decisions and more effectively catch fraudsters."that their software produces too many false-positive results: in one study in which Neuro-ID processed 20,000 applications on the online store website, less than half of the applicants who received the lowest ratings (up to 10) turned out to be scammers, and only 10% people with grades of 20 to 30 were at risk of fraud. The company recognizes that the software notes as suspicious job seekers who may be innocent, and makes it possible to use this information at its discretion. A company representative told me that “there is no 100% accurate behavioral analysis. "We recommend that you use these results in conjunction with other information about applicants to make better decisions and more effectively catch fraudsters."that their software produces too many false-positive results: in one study in which Neuro-ID processed 20,000 applications on the online store website, less than half of the applicants who received the lowest ratings (up to 10) turned out to be scammers, and only 10% people with grades of 20 to 30 were at risk of fraud. The company recognizes that the software notes as suspicious job seekers who may be innocent, and makes it possible to use this information at its discretion. A company representative told me that “there is no 100% accurate behavioral analysis. "We recommend that you use these results in conjunction with other information about applicants to make better decisions and more effectively catch fraudsters."in which Neuro-ID processed 20,000 applications on the online store website, less than half of the applicants who received the lowest ratings (up to 10) turned out to be scammers, and only 10% of people who received ratings from 20 to 30 were associated with the risk of fraud. The company recognizes that the software notes as suspicious job seekers who may be innocent, and makes it possible to use this information at its discretion. A company representative told me that “there is no 100% accurate behavioral analysis. "We recommend that you use these results in conjunction with other information about applicants to make better decisions and more effectively catch fraudsters."in which Neuro-ID processed 20,000 applications on the online store website, less than half of the applicants who received the lowest ratings (up to 10) turned out to be scammers, and only 10% of people who received ratings from 20 to 30 were associated with the risk of fraud. The company recognizes that the software notes as suspicious job seekers who may be innocent, and makes it possible to use this information at its discretion. A company representative told me that “there is no 100% accurate behavioral analysis. "We recommend that you use these results in conjunction with other information about applicants to make better decisions and more effectively catch fraudsters."scores of 20-30 were associated with the risk of fraud. The company recognizes that the software notes as suspicious job seekers who may be innocent, and makes it possible to use this information at its discretion. A company representative told me that “there is no 100% accurate behavioral analysis. "We recommend that you use these results in conjunction with other information about applicants to make better decisions and more effectively catch fraudsters."scores of 20-30 were associated with the risk of fraud. The company recognizes that the software notes as suspicious job seekers who may be innocent, and makes it possible to use this information at its discretion. A company representative told me that “there is no 100% accurate behavioral analysis. "We recommend that you use these results in conjunction with other information about applicants to make better decisions and more effectively catch fraudsters.""We recommend that you use these results in conjunction with other information about applicants to make better decisions and more effectively catch fraudsters.""We recommend that you use these results in conjunction with other information about applicants to make better decisions and more effectively catch fraudsters."

Utah start-up Converus sells software called EyeDetect, which measures pupil contraction during interviews to detect cognitive loading. Like Silent Talker, this tool works on the assumption that lying requires more effort than truth. According to a 2018 Wired article, police stations in Salt Lake City and Columbus, Georgia used EyeDetect to evaluate job candidates. Converus also told Wired that McDonald's, Best Western, Sheraton, IHOP and FedEx used its software in Panama and Guatemala in a way that would be illegal in the US.

The company provided me with a statement citing several studies demonstrating that the program achieved 85% accuracy in identifying liars and those who are telling the truth in samples of up to 150 people. Company president Todd Mikelsen says the firm’s algorithm has been trained in hundreds of thousands of interviews. However, Charles Honts, a psychology professor at the University of Idaho at Boise, who serves on the company's advisory board, says these findings do not prove that EyeDetect can be relied upon during an interview. “I find the EyeDetect system very interesting, but I don’t use it myself,” he told me. “I think that she still has a small database, and the data comes, for the most part, from one laboratory.” Until the base is expanded and other people reproduce the results, I would refrain from using it in real conditions. ”

Researchers at Arizona University who developed AVATAR founded Discern Science, a privately held company, to advertise their own lie recognition technology. Launched last year, Discern sells a 1.8-meter-high machine similar to the original AVATAR. According to an article in the Financial Times, the company “organized a joint venture with a partner in the aviation industry” to deliver these devices to airports. The system measures the movements of the facial muscles and the presence of stress in the voice in order to "discreetly collect information about a person at a distance of a normal conversation," as written in advertising materials. Discern, like Silent Talker and Converus, assures that the technology can reliably recognize about 85% of liars, but its results have not been independently verified. At least one of the information receiving channels used by the apparatus,was repeatedly recognized as unreliable. Honts also noted that the analysis of facial muscle movement “has virtually no evidence” - he said that “attempts to reproduce the results of the experiment had too many failures.”

Answering questions about the scientific background of the company’s machine, Discern researcher Judy Burgun emphasized that the system simply provides an assessment, not accurate conclusions about the truth and lies. Systems such as AVATAR and Silent Talker, in her words, “cannot measure fraud directly,” and “any person who advertises an unambiguously working lie detector is a quack.” But at the same time, in the marketing materials, Discern presents its tool as a reliable lie detector: the website says that it “can help reveal secret plans” and that “it was scientifically proven that its algorithms recognize fraud faster and more reliable than any alternatives” .

The Court of Appeal overturned the sentence of Emmanuel Mervilus in 2011, releasing him from prison and ordering a review of the case; he served more than three years by sentence. At the second trial in 2013, the jury discussed the case only 40 minutes before acquitting him. If it were not for the polygraph and not a firm belief in its accuracy, he could have never got into the dock at all. Mervilus condemned the police officers who arrested and interrogated him, claiming that they violated his right to conduct legal procedures, using a polygraph test for their conviction, the flaws of which they knew.

And even if the widespread use of Silent Talker and similar systems does not lead to an increase in the number of convictions of the innocent, as was the case with Mervilus, it can still create a new kind of obstacle that forces people to undergo a “reliability assessment” every time they want to rent a car or take a loan.

“In court, you need to provide material evidence, such as hair or blood,” says Wilde. “But you also have the right to remain silent and not to testify against yourself.” Mervilus decided to take a polygraph test, suggesting that like a DNA test, he would demonstrate his innocence. And although the device did not work correctly, it was not the car that sent him to prison. It is all about the jury's belief that the test results are more reliable than the facts of the case.

The assumption underlying AI recognition of lies is that lies can be seen with the right tools. Psychologists are still not convinced of the correctness of this statement, but for now, a simple faith in its correctness may be enough to reject worthy candidates for work or credit, and to prevent innocent people from crossing the state border. The promise to open a window into the soul of other people is too tempting to be denied, even if no one is sure that this window is clean.

“It's like a promise to read minds,” says Wilde. “Obviously, this is nonsense, but they are selling exactly that.”