Lie Detection Research: Can AI Beat Human Judgment?

Level: C1

Printable: Student handout (PDF) | Teacher’s key (PDF)

Reading

Judging the veracity of a statement matters in courtrooms, border interviews, workplaces, and online. For centuries, people have relied on intuition and interview technique when assessing another person’s credibility. More recently, software that analyzes language has been offered as an alternative. Comparing the two is harder than it sounds, because most studies define “better” as accuracy on staged tasks in which some participants were told to lie.

Human judgment usually begins with behavior. In a study carried out in 75 countries, the most widespread stereotype was that liars avert their gaze. Yet a meta-analysis of 158 possible cues by Bella DePaulo and colleagues found that many behaviors showed no discernible link, or only a weak one, to deception. The cues that did emerge were modest: liars were somewhat less forthcoming and more tense. Accuracy is correspondingly low. In a 2006 analysis of 206 documents and 24,483 judges, Charles Bond and DePaulo found an average of 54 percent correct. Judges identified 61 percent of truths but only 47 percent of lies, which suggests that people are susceptible to a pervasive truth bias: the tendency to assume honesty. Judges were also more accurate with audible lies than with visible ones.

Better human methods exist, but they demand discipline. Interviewers can ask unexpected questions or make the task harder, which raises the cognitive load on liars. In nine studies published in 2023, Bruno Verschuere and colleagues found that people judging statements by intuition performed at chance, whereas people told to rely on one heuristic, the level of detail, especially detail that investigators can corroborate, could discriminate lies from truths. The polygraph measures physiology rather than judgment. A 2003 review by the US National Research Council concluded that specific-incident tests separate liars from truth-tellers far better than chance but far from perfectly, and it could not confirm that they survive a countermeasure, such as a trained attempt to beat the test.

Automated approaches mostly analyze language, because words can be counted and compared. A meta-analysis of 44 studies by Valerie Hauch and colleagues found that software could detect several patterns: relative to truth-tellers, liars used fewer sensory and perceptual words, expressed more negative emotion, and distanced themselves more from events. Machine-learning classifiers are trained on thousands of statements, some honest and some written by participants asked to fabricate an account. Recent work fine-tunes large language models in the same way. These systems do not detect lies as such. They learn statistical differences between the honest and invented texts they have seen.

In head-to-head tests, algorithms often outperform people. Bennett Kleinberg and Verschuere collected 1,640 true and false statements about people’s intentions, and a model trained on them reached 69 percent accuracy. When human judges could overrule the model, accuracy fell to chance, so in that study human input did impair the machine’s performance. The authors suggest a plausible explanation: truth bias. A team led by Alicia von Schenk built a classifier that scored 66.86 percent, against 46.47 percent for human judges, in an experiment with 2,040 participants. Only 31.76 percent asked for its advice, but those who did tended to follow its accusations. In Italy, Riccardo Loconte and colleagues fine-tuned a language model that reached roughly 71 to 83 percent on the kind of statement it was trained on.

The results come with an important caveat. In the language-model study, accuracy fell to between about 49 and 56 percent on a type of statement the model had not seen, so its skill did not generalize. Most datasets contain low-stakes, invented stories, unlike real lies, which are often embedded in true accounts. Base rates matter too: when few people lie, even an accurate test produces many false accusations. For a hypothetical screening of 10,000 employees including 10 spies, the National Research Council calculated that a polygraph set to catch eight spies would fail about 1,600 innocent people. A 2025 preprint by Emilio Barkett and colleagues found that several advanced language models judged truthful statements well but deceptive ones poorly. The EU AI Act classes polygraph-type tools in migration and border control as high-risk, with the obligations applying from December 2027 under a 2026 amendment, and it has banned emotion inference in workplaces and schools since February 2025.

So which is better? In the laboratory comparisons reviewed here, algorithms beat intuitive human judgment on narrow, text-based tasks, scoring roughly 67 to 83 percent on familiar kinds of statement where human judges scored near or below chance. Whether they do so with real suspects, real stakes, and unfamiliar liars has not been shown in these studies, and humans with a simple rule can do better than chance. The evidence therefore supports a limited claim: machines are better in specific settings, and no method has been shown to be reliable enough to justify accusing someone on its own.

Vocabulary

Each word has a frequency score, a rank, and a tag. The score shows how common the word is in modern English: higher means more common, and each full point means roughly ten times more common. The rank is the word’s position on a list of about 321,000 English word forms, from most to least common. The tag shows where you are most likely to meet the word: Current, Formal, Literary, or Dated. The paragraph number shows where the word first appears in the reading. The labels are information, not instructions. Deciding which words are worth your time is up to you. How to read the scores, ranks, and tags.

Journalists check the veracity of a claim before publishing it.
Judging the veracity of a statement is hard.

My intuition told me to take the earlier train.
People have relied on intuition when judging honesty.

The film relied on a stereotype about teenagers.
The most widespread stereotype was that liars avert their gaze.

Tired drivers are susceptible to mistakes.
People are susceptible to a pervasive truth bias.

Corruption was pervasive in the city government.
Truth bias is a pervasive finding in deception research.

A simple heuristic, such as choosing the cheapest option, saves time.
People told to rely on one heuristic did better than chance.

Two independent witnesses corroborated her story.
Investigators can corroborate details such as receipts.

Trained tasters can discriminate between very similar wines.
People could discriminate lies from truths.

A trained attempt to beat the test is a countermeasure.
The company introduced a countermeasure against fraud.

He was fired for trying to fabricate evidence.
Participants were asked to fabricate an account.

The new model can outperform the old one on every test.
In head-to-head tests, algorithms often outperform people.

Lack of sleep can impair your judgment.
Human input did impair the machine’s performance.

The excuse sounded plausible, so nobody questioned it.
The authors suggest a plausible explanation: truth bias.

The study is interesting, but one caveat is its small sample.
The results come with an important caveat.

It is risky to generalize from a sample of ten people.
The model’s skill did not generalize to unseen statements.

Activity Page

Open the activity page in a new tab

Speed Challenge

Answer 12 questions against the clock. Every wrong answer adds 5 seconds, and the lowest score wins. Open the speed challenge in a new tab.

Shadow Reading and Typing: Reading Faces

This short text uses many of the words from this lesson. Use it for shadow reading (reading aloud along with a model) and for typing practice.

For years, I believed I could judge the veracity of what my students told me. When someone looked away while explaining a missing assignment, my intuition said they were lying. I now realize I was relying on a stereotype, and a weak one. Research suggests that avoiding eye contact has little link to deception.

What unsettled me more was learning how susceptible we all are to truth bias. It seems to be pervasive: people tend to assume honesty, so they miss many lies. Some studies suggest that a simple heuristic works better than gut feeling. Instead of watching faces, you look at the level of detail, especially details you can corroborate.

I was curious about machines, too. In laboratory tests, some software can discriminate lies from truths better than human judges. These systems learn from statements that people were asked to fabricate. In several studies, algorithms outperform people, and in one, letting humans overrule the model seemed to impair its accuracy. The authors offered truth bias as a plausible explanation.

Still, there is a major caveat. One model’s skill did not generalize to kinds of statements it had not seen, and real lies are messier than lab stories. Nobody has confirmed that polygraphs survive a trained countermeasure either. So I have changed my approach. I ask more questions, check the details, and try not to accuse anyone based on a feeling alone.

How to practice

Read the text silently once and check any words you are unsure of.
Read it aloud slowly, one sentence at a time. On the typing page, the Listen button reads each line with your device’s voice, so you can listen first and then repeat.
Shadow: read aloud again at a natural speed, trying to keep a steady rhythm without stopping.
Record yourself, listen back, and note one sentence to practice again.
Type it on the typing page below and check your speed and accuracy.

Open the typing practice in a new tab

Comprehension Questions

Why does the reading say that comparing human and AI lie detection is harder than it sounds?
Summarize what the research says about behavioral cues to deception and about human accuracy. How does the truth bias affect the results?
What advantages do the methods in the third paragraph offer over intuition, and what limits does the reading note for the polygraph?
How do the automated systems in the reading work, and why does the reading say they do not detect lies as such?
Compare the accuracy figures reported for humans and for algorithms in the fifth paragraph. What do the results suggest about combining human judgment with a machine?
Why does the reading treat the algorithms’ higher accuracy with caution? Give at least three reasons from the sixth paragraph.
Explain why base rates matter when an accurate test is used to identify liars in a population where few people lie.
What do the EU AI Act provisions mentioned in the reading suggest about how regulators view lie-detection technology?
Evaluate the writer’s final answer to the question of which is better. Which parts of the conclusion are firmly supported, and which are not?

IELTS-Style Questions

True, False, or Not Given: “Trained police officers detect lies more accurately than ordinary people.” Explain your answer using evidence from the reading.
Yes, No, or Not Given: “The writer believes that algorithms have been shown to detect lies better than humans in all settings.” Explain your answer using evidence from the reading.
Author’s purpose: Why does the writer report both the algorithms’ high accuracy and their failure on unseen statements? What effect does this have on the reader’s conclusion?

AI Discussion Prompt

I just read a C1 English lesson called How Humans and AI Detect Lies. It covered the research on human cues and accuracy (Bond and DePaulo’s 54 percent, the gaze aversion stereotype, truth bias), better human methods and the polygraph, language-based AI and fine-tuned language models, head-to-head results in which algorithms scored roughly 67 to 83 percent on familiar statements while humans scored near or below chance, the limits of those results (low-stakes lab tasks, poor generalization, base rates), and the EU AI Act. I am an advanced English learner at about C1 level. Please have a balanced conversation with me about whether machines or people should be trusted to judge honesty, present more than one perspective, and do not push me toward a conclusion. Ask one question at a time, starting with something accessible. Encourage me to use words like veracity, intuition, heuristic, plausible, and caveat. Respond to my ideas first, then briefly point out one or two language errors or more precise word choices.

AI Vocabulary Prompt

I want to practice vocabulary from a lesson about how humans and AI detect lies. Please quiz me on the following words one at a time, using a mix of question types: multiple choice, true/false, and sentence completion. Give me feedback after each answer before moving on, and bring back the words I miss later. Here are the words: veracity, intuition, stereotype, susceptible, pervasive, heuristic, corroborate, discriminate, countermeasure, fabricate, outperform, impair, plausible, caveat, generalize.

Discussion Questions

How good do you think you are at telling when someone is lying? What do you look for, and how sure are you that those signs are reliable?
The most widespread stereotype about liars is that they avert their gaze, yet the research finds weak links. Why might a false belief like this stay popular?
People are more accurate with audible lies than visible ones, according to the reading. What might explain that?
The truth bias makes people assume honesty. In everyday life, is this bias a weakness or a useful social habit?
Verschuere and colleagues found that relying on the level of detail helped people. Would you trust a simple rule like this in a real situation? Why or why not?
The polygraph is still used in some places. Should governments or employers use tests that are far from perfect? Under what conditions, if any?
Algorithms in the studies scored higher than humans, but mostly on staged, low-stakes tasks. How much should these results influence real decisions?
When human judges could overrule the model, accuracy fell to chance. What does this suggest about the role of human oversight in automated decisions?
Only about a third of participants asked for the classifier’s advice, and those who did tended to follow its accusations. What might this suggest about how people use AI?
If very few people lie in a given setting, even an accurate test produces many false accusations. How should a society weigh the cost of a false accusation against the benefit of catching a liar?
Should employers or schools be allowed to use tools that infer emotion or honesty? What are the arguments for and against the EU’s approach?
Liars could learn to fool a language-based detector. How might that change how the technology is used, and who might benefit?
Is a lie detector that is right two times out of three useful? What would you need to know before using it?
Which do you trust more to judge honesty in a high-stakes case: a trained person, a machine, or both together? What evidence would change your mind?

Language Focus: Comparatives with Degree Modifiers

What it is: Comparative forms (better, more accurate, less forthcoming, fewer words) show the direction of a difference. A modifier before the comparative (far, much, considerably, somewhat, slightly) shows its size, and “than” names what is being compared.

Why writers use it: Research writing must say not only that one result is better but how much better, and compared with what. A modifier stops readers from treating a small difference as a large one, or the reverse.

Examples from the reading:

“…specific-incident tests separate liars from truth-tellers far better than chance…”
“…liars were somewhat less forthcoming and more tense.”
“Judges were also more accurate with audible lies than with visible ones.”
“…liars used fewer sensory and perceptual words, expressed more negative emotion, and distanced themselves more from events.”

Practice task: Combine into one sentence with a comparative and a modifier that shows the size of the difference: “The model scored 69 percent. The human judges scored 50 percent.”

(Example answer: “The model was considerably more accurate than the human judges.”)

Recommended Resources

Accuracy of Deception Judgments (Bond and DePaulo, Personality and Social Psychology Review, 2006). The abstract of the meta-analysis behind the 54 percent figure, including the truth and lie accuracy rates.
The Use-the-Best Heuristic Facilitates Deception Detection (Verschuere et al., Nature Human Behaviour, 2023). Nine studies on whether people can do better when they rely on one cue. Compare its optimism with the earlier meta-analysis.
Lie Detection Algorithms Disrupt the Social Dynamics of Accusation Behavior (Köbis et al., iScience, 2024). The abstract of the experiment on how people use a lie-detection algorithm. The accuracy figures appear in the working paper version.
Verbal Lie Detection Using Large Language Models (Loconte et al., Scientific Reports, 2023). An open-access paper on fine-tuned language models. Read the results and the limitations section together.
The Polygraph and Lie Detection: Summary of Scientific Evidence (National Research Council, 2003). The committee’s summary of what the science does and does not support, and why polygraph estimates may overstate real-world accuracy.