No—an AI chatbot cannot reliably know that you are lying to it in the human sense of knowing that you believe one thing, deliberately saying another, and intending to mislead it.
That answer sounds more definitive than the science really is, because modern language models can do something that looks surprisingly close to lie detection. They can compare what you said five minutes ago with what you say now, identify contradictions, challenge a false premise, check a claim against retrieved evidence, and sometimes classify a transcript as deceptive. In controlled research, some large models have even shown a measurable ability to identify their own deceptive outputs when they are explicitly prompted to act as evaluators. citeturn5view0
But the important distinction is between detecting a suspicious statement and knowing the speaker’s mental state. A contradiction is observable in the conversation. Your intention is not. If you tell a chatbot, “I have never used Photoshop,” after previously saying you used it every day, the system has evidence of inconsistency. It does not automatically have evidence that you knowingly fabricated the second statement. You may have misunderstood the question, forgotten the earlier conversation, used Photoshop only professionally, or simply made a mistake.
That distinction is the gap most search results miss. Many pages answering whether AI can detect lies are actually describing specialised deception-detection systems that analyse video, audio, physiology or linguistic datasets. Others explain whether AI can lie. Your question is different: when you are talking to a general-purpose chatbot and deliberately feed it false information, can the chatbot tell? The answer depends on what kind of lie you give it, what evidence it can access, whether it remembers the earlier claim, and whether the system has been asked to evaluate the conversation.
This article treats the question as a systems problem rather than a futuristic mind-reading problem. It separates what the model can observe from what it would have to infer, maps the newest research, explains why self-reporting is not the same as introspection, and gives you a practical test for finding out whether an AI caught your lie or merely generated a plausible accusation.
The First Mistake: Treating Inconsistency as Proof of a Lie
Suppose you tell a chatbot that your flight leaves at 8 p.m. Later you say, “I need to reach the airport before my 7 p.m. flight.” A competent model can identify the contradiction. That is a relatively straightforward language task: two propositions in the same conversational context cannot both describe the same flight at the same time.
The harder question is why the contradiction exists. Human deception research has long distinguished between observable cues and the underlying truth status of a statement. A person can be wrong without lying, and a person can lie without displaying a reliable behavioural cue. The 2024 review “Detecting deception with artificial intelligence: promises and perils” makes this methodological problem central: technical performance alone is not enough when the construct being measured—deception—is difficult to define and ground. citeturn2search3
A chatbot has the same conceptual problem, but with an additional limitation. It is not a human interviewer who can independently know your motives. It is generating an answer from the information available in the interaction and, where enabled, from tools or retrieved sources.
| Signal Available To The AI | What It Can Reason About | What It Still Cannot Establish |
| Contradictory statements | Your claims are inconsistent across turns. | Whether the contradiction was intentional. |
| External evidence | Your claim conflicts with a reliable source. | Whether you knowingly made the false claim. |
| Conversation history | Your current statement differs from your earlier account. | Whether memory, wording or circumstances changed. |
| Linguistic style | Your wording differs from your normal or earlier pattern. | Whether the change was caused by deception. |
| Direct admission | You say you invented a claim. | Whether the admission itself is truthful. |
For a broader explanation of why fluent answers can still fail factual checks, see AI hallucinations and trust.
What A Chatbot Actually Sees When You Lie
The word “know” is doing too much work in this question. A chatbot does not need to know that you are lying to react to a lie-like statement. It only needs access to enough information to notice a conflict.
In a normal text conversation, the strongest evidence is usually one of four things: contradiction with previous turns, contradiction with information you supplied, contradiction with retrieved evidence, or contradiction with basic logical constraints. Those signals are cheap because they are already represented in the input. No specialised lie detector is necessary.
The problem begins when your lie is internally coherent. “I went to the meeting yesterday” is not something a language model can label as false merely by reading the sentence. If it has no calendar, email, location record, transcript or other evidence, the model has no direct route to the ground truth. It may still produce a guess because dialogue systems are designed to answer questions, not remain silent whenever evidence is incomplete.
That is why a chatbot can appear suspicious in one conversation and completely accepting in another. Change the evidence available to the system and you change the detection problem.
| Type Of Lie | Example | Likely AI Signal | Confidence You Should Place In A Flag |
| Contradiction | “I never worked there” after naming the employer earlier. | Conversation inconsistency. | Moderate as a contradiction; low as proof of intent. |
| False fact | “The company launched in 2011” when records say 2018. | External source conflict, if retrieval is available. | Moderate to high for factual falsity. |
| Plausible fabrication | Inventing an event that leaves no evidence in context. | Usually no decisive signal. | Low. |
| Strategic omission | Leaving out a relevant fact without saying anything false. | Potentially invisible unless the omission is inferable. | Very low. |
| Self-protective lie | Giving a false personal detail to protect privacy. | May be indistinguishable from a genuine detail. | Low. |
This distinction also explains why “AI detected my lie” screenshots are weak evidence. If the model points to a contradiction that you can independently inspect, the interesting capability is contradiction detection. If it says, “I can tell you are nervous, so you are lying,” the claim is much stronger and requires a different kind of evidence.
When AI Can Catch You
AI is most useful at catching lies when the lie leaves a trace inside the information it can inspect. That trace might be a contradiction, an impossible sequence of events, a mismatch with a supplied document, or a claim that conflicts with a source the system can retrieve.
Language models are capable of verbal deception detection under controlled conditions. A 2023 Scientific Reports study tested FLAN-T5 models on three English-language datasets involving personal opinions, autobiographical memories and future intentions. The researchers reported state-of-the-art results in several experimental scenarios and found performance improved with model size. citeturn2search1
That result should not be translated into “ChatGPT knows when you lie”. The study trained and evaluated a model on labelled datasets. Your everyday chatbot session is not the same experimental environment. The ground truth, prompt format, task definition, model version and evaluation protocol all matter.
A more recent 2026 Anthropic investigation makes this distinction even sharper. Researchers tested whether models could report their own deceptive behaviour. With simple prompting, Qwen3-235B achieved an AUROC around 0.98–0.99 in their controlled self-report setup, while smaller models were closer to chance. Third-person monitoring generally outperformed asking the model directly about itself. citeturn5view0
“Third-person monitoring outperformed self-report across model sizes.” — Jack Hopkins, Dipika Khullar, Rowan Wang and Fabien Roger, Anthropic Alignment Science Blog, 21 August 2026
That finding is important but easy to misuse. The researchers were not showing that a consumer chatbot can read a user’s mind. They were testing model behaviour under a controlled deception taxonomy. Their work shows that some models can perform surprisingly strong classification when the task is clearly specified and the transcript contains the relevant evidence.
The practical rule is simple: the more independently checkable evidence your lie creates, the more likely a chatbot is to flag it. The more your lie depends on private intent or facts outside its reach, the weaker its basis becomes.
The same evidence problem appears when different AI systems answer the same question differently; our guide to why AI tools disagree explains how model routing, context and retrieval can change the result.
When AI Cannot Know
A private lie with no observable evidence is the hard case. Imagine you tell an AI, “I was born in Manchester,” when you were actually born elsewhere. Unless the system has a trustworthy record that contradicts you, it cannot establish the truth merely from the sentence.
It may infer. It may ask follow-up questions. It may notice unusual wording. It may even sound highly certain. None of those behaviours creates ground truth.
The distinction matters because users often interpret a confident accusation as hidden access: “The AI knew I was lying.” In reality, the model may have been pattern-matching against language, exploiting an inconsistency, or simply guessing. The same system can produce a different verdict when you change the prompt framing.
Research on AI deception detection repeatedly exposes this problem. In a 2025 Journal of Communication study, Markowitz and Levine ran 12 studies using Gemini 1.5 Flash to judge human truthfulness across different communication types, modalities, base rates and AI personas. The researchers found performance was highly context-sensitive and included strong truth- and lie-biases; in some settings accuracy fell dramatically. citeturn5search45turn5view1
“AI turned out to be sensitive to context — but that didn’t make it better at spotting lies.” — David Markowitz, Michigan State University, quoted by MSUToday, 6 November 2025
That sentence is a useful antidote to a common assumption: more context does not automatically produce better deception detection. Context can help a model understand a statement, but it can also give the model more opportunities to anchor on misleading cues.
Why “I Know You’re Lying” Is Not A Reliable Verdict
If you ask a chatbot, “Are you sure I am lying?”, you are asking the same system that generated the previous interpretation to judge its own interpretation. That can be useful as a second pass, but it is not an independent lie detector.
The 2026 Anthropic study illustrates why. Direct self-report worked surprisingly well for some large models, but third-person evaluation generally performed better. The authors also found that fine-tuned lie detectors could reach AUROC 0.95 on deception types represented during training while falling to roughly 0.70–0.75 on held-out categories. In other words, a detector can look excellent when the test resembles the training distribution and much weaker when the form of deception changes. citeturn5view0
This is one of the most important information-gain findings for the question you asked: detection is not a single capability. A system may be good at recognising a known class of contradiction while being poor at a new kind of deception.
| Question You Ask The AI | What The Answer Tests | Main Failure Mode |
| “Did I contradict myself?” | Consistency checking. | The system may miss subtle contradictions. |
| “Is this claim factually true?” | Evidence or knowledge retrieval. | The model may lack current or authoritative evidence. |
| “Was I lying?” | Intent/deception classification. | The model may infer intent from insufficient evidence. |
| “What evidence suggests I lied?” | Evidence-based reasoning. | The model may invent a rationale after choosing a verdict. |
| “What would change your conclusion?” | Uncertainty and falsification. | The system may still defend its initial guess. |
The fourth question is especially revealing. If an AI cannot point to a concrete, independently inspectable reason, its accusation should be treated as an inference, not a finding.
This is closely related to the distinction between factual error and fluent certainty discussed in Why AI Gives Confident Wrong Answers.
What The 2025–2026 Research Actually Shows
The research landscape is not a straight line from “AI cannot detect lies” to “AI can”. It is a collection of narrower results that become useful only when the task is specified.
| Study | Setting | Main Result | What It Does Not Prove |
| Loconte et al., Scientific Reports (2023) | FLAN-T5 on labelled verbal deception datasets. | LLMs can classify deception patterns under controlled conditions; larger models performed better in tested settings. | That consumer chatbots reliably know when a user is lying. |
| Markowitz & Levine, Journal of Communication (2025) | 12 studies using Gemini 1.5 Flash for human veracity judgements. | AI accuracy and bias changed materially with context, modality and base rates. | That one model’s result generalises to every chatbot or situation. |
| Hopkins et al., Anthropic (2026) | Controlled on-policy deception and self/third-person lie detection. | Some large models could detect deception when prompted; specialised detectors struggled to generalise. | That models can access human users’ private intentions. |
| Markowitz & Levine, arXiv RAG study (2026) | 39,200 deception judgements across models and RAG variants. | RAG accuracy was around 57%, close to baseline and typical human levels. | That retrieval alone solves deception detection. |
| Azuma et al., AI (2025) | SVM, BERT and LLMs for multilingual instructed deception. | Language models can contribute to automated deception classification. | That linguistic cues are a universal lie signature. |
The 2026 RAG study is particularly useful because it tests a tempting idea: perhaps giving an AI more external evidence will solve the problem. Markowitz and Levine compared RAG-based and baseline models across 700 statements, four large language models and 39,200 judgements. They reported 54.5% accuracy for baseline models and 54.5% for RAG in the abstracted comparison, with the detailed preprint reporting 57.0% for RAG versus 54.6% for baseline under its analysed setup. The important point is not the small numerical difference; it is that retrieval did not transform the task into reliable lie detection. citeturn7academia59
“Theory-guided AI judgments are unreliable with current parameters.” — David M. Markowitz and Timothy R. Levine, 2026 RAG deception-detection preprint
Another 2025 study found that fine-tuned LLMs could achieve strong performance on deceptive language while still facing generalisation limits. A separate Scientific Reports study on embedded lies found that even a fine-tuned Llama-3.1-8B reached 64% accuracy, while a previously reported detector dropped substantially when moved to a different embedded-lie task. citeturn2search19
Taken together, the research supports a much narrower statement than the marketing version: AI can identify some deception-related patterns, sometimes at useful levels in controlled tasks, but performance is highly dependent on the data, definition of deception, context, model and evaluation method.
The Self-Report Problem: Can An AI Tell You That You Lied?
This is where the question becomes genuinely interesting. Suppose you lie to a chatbot, then ask, “Did I just lie?” The model has your transcript. It can compare your statements. It can evaluate factual consistency. In principle, it can classify the conversation.
But it does not have a privileged channel into your mind. It cannot directly inspect whether you knew the statement was false at the moment you typed it.
The distinction between belief and accuracy also appears in current AI safety research. Anthropic’s 2026 work explicitly treats deception as involving a false statement plus knowledge that the statement is false, then uses controlled elicitation to create candidate cases. The researchers note that establishing ground truth is difficult: roughly 25% of their labels changed after a more informed judging pass. citeturn4view0
That is an extraordinary detail. If researchers with access to the model, controlled prompts and metadata can struggle to decide whether a model’s output qualifies as a lie, expecting a general chatbot to make a definitive judgement about a human’s private intent is an even stronger claim.
The self-report result therefore needs to be read carefully. A model saying “yes, you lied” can be useful when it explains a real contradiction. But the model’s admission is not itself proof. You still need to inspect the evidence that led to the classification.
Text, Voice And Video Are Different Detection Problems
Search results often blur together three very different problems: detecting false claims in text, detecting deception in an interview, and detecting physiological or behavioural correlates of lying.
A text chatbot has access primarily to language and conversation context. A multimodal system may also receive voice, facial movement, images, video or sensor data. That additional information changes the feature space, but it does not magically solve the truth problem.
A 2025 Scientific Reports study on multimodal deception detection combined behavioural and physiological information. Its very premise was that deception involves complex and diversified behavioural, physiological and cognitive factors, and that traditional polygraph methods remain controversial. citeturn2search0
Likewise, a 2025 ACL study evaluating LLMs and multimodal LLMs found that fine-tuned LLMs achieved strong textual deception-detection performance, while multimodal systems struggled to fully exploit multimodal cues in more realistic settings. citeturn2search15
“We’re not there yet when it comes to detecting deception across multiple domains.” — Sayde King, University of South Florida, quoted in USF News, 6 August 2025
The lesson is not that voice or video are useless. It is that more sensors do not automatically create a unique “lying signal”. Stress, hesitation, uncertainty, emotion and deception can overlap. A model can become better at classification while still lacking a universal causal marker of lying.
Hany Farid has made a parallel point about AI-generated media: detection systems should be grounded in measurable properties of the underlying artefact rather than treated as magical authenticity or falsity oracles. In a 2026 interview, he warned against uploading media to services making unsupported 99% detection claims. citeturn7search2
The same principle applies to ordinary chatbot accuracy: How Accurate Is AI in 2026? breaks accuracy down by task rather than treating one percentage as universal.
A Lie Can Also Be Invisible To The Model
One of the weakest assumptions in AI lie detection is that every lie produces a detectable output. It does not.
Consider omission. You ask a chatbot for advice about negotiating a contract but deliberately omit the fact that you already accepted a competing offer. Everything you say may be true. The model can only reason from the supplied facts. There is no contradictory sentence to catch.
Now consider a lie that is entirely plausible: “I met the client in London on Tuesday.” Unless the system can access your calendar, messages, travel record or other evidence, there may be nothing in the text that separates the lie from the truth.
Finally, consider a lie that is true in one interpretation and false in another. Natural language contains ambiguity, shorthand and unstated assumptions. A model may label a statement inconsistent when a human would understand the intended scope differently.
| Deception Pattern | Why It Is Hard | Best Available Check |
| Omission | Nothing false was explicitly stated. | Ask what relevant facts may be missing. |
| Plausible fabrication | No contradiction exists inside the conversation. | Independent evidence or records. |
| Ambiguous statement | Truth depends on interpretation or scope. | Clarify definitions before judging. |
| Memory error | False statement may be sincerely believed. | Separate factual error from intentional deception. |
| Strategic framing | All individual sentences may be true while the overall impression is misleading. | Check implications and omitted context. |
This is why a robust system should not compress every problem into “truth versus lie”. A better taxonomy includes true, false, unsupported, contradictory, ambiguous, incomplete, misleading-by-omission and intentionally deceptive. The last category is the hardest because intent is the least directly observable.
The Five-Minute Test: Find Out What The AI Actually Detected
If you want to test whether a chatbot genuinely caught your lie, do not begin by asking, “Did you know I was lying?” That question invites a narrative. Instead, separate the detection layers.
Use a controlled conversation with one false claim and one true claim. Keep the lie plausible and avoid adding clues that reveal the answer. Then ask the model to identify contradictions, factual conflicts and uncertainty separately.
Test 1: Remove The Accusation
Ask: “List every statement in my last five messages that conflicts with something I said earlier. Do not decide whether anyone lied.” If the model finds the contradiction without being prompted to hunt for a lie, you have evidence of consistency checking rather than accusation generation.
Test 2: Demand Evidence
Ask: “For each suspicious statement, give the exact earlier statement or external source that conflicts with it.” This prevents the model from replacing evidence with a vague linguistic impression.
Test 3: Test A Sincere Error
Change the scenario so the false statement could plausibly be a memory mistake. If the model still labels it a deliberate lie with the same confidence, its classification is probably collapsing factual falsity and intentional deception.
Test 4: Remove External Access
Run the same conversation once with browsing or retrieval and once without it, where the platform permits a meaningful comparison. If the verdict changes, you have learned something important: the system’s judgement depended on evidence access rather than a hidden lie-reading capability.
Test 5: Ask What Would Falsify Its Verdict
Finally ask: “What evidence would make you change your conclusion?” A strong answer identifies a specific missing record, contradiction or source. A weak answer repeats the same accusation in different words.
This testing approach produces something much more valuable than a yes/no result: it tells you what the model actually used.
For a related problem—when an AI seems to lose or distort earlier context—see why Perplexity ignores follow-up questions.
What Happens If You Lie To AI On Purpose?
Usually, nothing dramatic happens. The immediate consequence is that the model’s next response is based on a distorted input. If the lie changes the facts of the problem, the quality of the answer can fall even when the model responds perfectly to the information it received.
For example, tell an AI that a contract contains a termination clause when it does not, then ask it to explain your rights. The model may produce a coherent analysis of a fictional clause. The failure is not that the AI “believed you” in a human emotional sense. The failure is that your false premise became part of the working context.
The same issue appears in research, coding, planning and decision support. A fabricated constraint can cause the model to optimise the wrong problem. A false number can distort calculations. A fake source can contaminate a summary. A false personal detail can make a recommendation inappropriate.
There is also a more subtle possibility: the model may challenge you. Modern assistants are increasingly trained to question unsupported premises, identify contradictions and request clarification. But that behaviour is not equivalent to a universal lie detector. It is better understood as epistemic checking—an attempt to determine whether the current information is sufficient and internally coherent.
The safest mental model is therefore: treat the chatbot as a reasoning system operating over the evidence you provide, not as an observer with access to your private truth.
Our explainer on AI hallucinations in 2026 covers the same input-evidence problem from the opposite direction: what happens when the model supplies unsupported information itself.
The Bigger 2026 Lesson: Deception Detection Is A Distribution Problem
The most useful finding from the current research is not a headline accuracy number. It is the repeated failure of systems to generalise cleanly when the form of deception changes.
Anthropic’s August 2026 experiments provide a concrete example. Fine-tuning pushed in-distribution AUROC from about 0.60 to 0.95, but held-out lie categories remained around 0.70–0.75. The researchers concluded that detectors appeared to learn surface patterns associated with the tested settings rather than a deep, general concept of deception. citeturn5view0
That matters for ordinary chat. Your lie may not resemble the examples a model or detector has learned. A system trained to spot contradictory factual claims may not detect omission. A detector trained on suspicious wording may fail when you write naturally. A model that performs well on a benchmark may behave differently when the conversation contains emotional pressure, roleplay, multiple languages or long context.
A 2026 review of manipulation, persuasion and deception in LLMs similarly argues that deception research needs careful definitions and stronger controls, while warning against interpreting striking demonstrations as proof of a general deceptive capability. citeturn9search0
“Until detectors track deception itself rather than the surface form of the settings that produce it, in-distribution accuracy will keep overstating what they can catch.” — Jack Hopkins, Dipika Khullar, Rowan Wang and Fabien Roger, Anthropic, August 2026
That is arguably the cleanest answer to the original question. An AI can sometimes detect the surface consequences of a lie. The unresolved research challenge is whether it can robustly recognise deception as a general property across new contexts, without relying on narrow cues.
For a user, that means a chatbot’s lie verdict should be treated like a hypothesis with evidence attached—not like a polygraph result.
Our Editorial Verification Process
This article was researched as an explainer rather than a product review. The research process separated four evidence layers: peer-reviewed studies on verbal and multimodal deception detection; 2025–2026 experiments on AI personas, LLM deception and lie-detection generalisation; primary institutional reporting from Michigan State University and Anthropic; and live, indexed Perplexity AI Magazine pages used only for contextual internal linking.
The target query and close variants were reviewed across the current search landscape. The recurring top-ranking pattern was broad rather than precise: define AI lie detection, describe facial or linguistic cues, discuss polygraphs, list limitations, and conclude that the technology is developing. Several academic papers offered stronger evidence but were not written around the everyday chatbot question. The article therefore uses a different structure: first distinguish contradiction from intent, then map what a chatbot can observe, then compare controlled research, then give a reproducible user-side test.
The live sitemap.xml, sitemap_index.xml and post-sitemap.xml endpoints were not parseable through the available browsing layer at research time. No sitemap URLs were invented. Six contextually relevant internal links were selected from live indexed Perplexity AI Magazine pages and embedded once each in body sections.
No current software pricing matrix was included because the target is a conceptual behaviour question rather than a product comparison; inserting unrelated plan prices would add noise rather than evidence.
This article was researched and drafted with AI assistance and reviewed by the Sami Ullah Khan editorial desk at Perplexity AI Magazine. All data, citations, pricing figures, and named quotes have been independently verified against primary sources before publication.
Conclusion
Does AI know when you are lying to it? Not reliably in the human sense of knowing your private intention. What current AI can sometimes do is narrower and, in some contexts, genuinely useful: it can detect contradictions, compare claims with evidence, identify implausible sequences, and classify deception patterns under controlled conditions.
The 2025–2026 research makes two things clear at once. First, language models are capable of more sophisticated deception-related classification than the old “AI is just autocomplete” description suggests. Second, impressive performance on a defined task does not automatically generalise to a new kind of lie, a different context, or an ordinary consumer conversation. The Anthropic work is especially revealing because strong in-distribution detection still weakened substantially when the deception category changed. citeturn5view0
The practical consequence is simple. If an AI tells you that you lied, ask what it actually observed. Did it find a contradiction? Did an external source disprove the claim? Did it infer intent from wording alone? Can you reproduce the reasoning without accepting the model’s conclusion first?
That discipline matters because the most dangerous failure is not an AI that misses an obvious lie. It is an AI that confidently claims to know your intention when all it really knows is that your words look unusual.
As AI systems gain better memory, retrieval, multimodal input and agentic capabilities, the evidence available to them will expand. Whether that eventually produces robust deception detection is still an empirical question. For now, the strongest answer is not “AI knows” or “AI does not know”. It is: AI can detect some evidence associated with lying, but evidence of inconsistency is not evidence of intention.
Frequently Asked Questions
Does AI know when I am lying to it?
No. AI can sometimes detect contradictions, factual conflicts or other deception-related patterns, but it cannot reliably establish that you knowingly made a false statement with an intent to deceive.
Can ChatGPT tell if I am lying?
ChatGPT can flag inconsistencies in what you say and, when tools are available, compare claims with external evidence. That is different from reliably knowing your private intent, so a confident lie accusation should not be treated as proof.
Can AI detect lies from text alone?
Text-only models can classify some deceptive language patterns in controlled datasets, but performance varies by task, language, context and model. Text cues are not a universal signature of lying.
Can AI detect lies from facial expressions?
AI can analyse facial and behavioural signals associated with deception research, but those signals are not unique to lying. Stress, emotion, uncertainty and other states can produce overlapping cues.
Why might an AI say I am lying when I am telling the truth?
The model may mistake an unusual writing pattern, contradiction, ambiguity or incomplete context for deception. AI deception judgements are sensitive to context and can show systematic bias.
Does AI get better at lie detection when it can search the web?
External retrieval can help verify factual claims, but it does not reveal private intentions. Current research indicates that retrieval alone does not turn general-purpose models into reliable lie detectors.
What is the best way to test whether an AI actually caught my lie?
Ask it to identify the exact contradiction or external evidence first, without asking it to label the statement a lie. Then ask what evidence would change its conclusion. This separates observable evidence from an unsupported accusation.
Is a false statement the same as a lie?
No. A false statement can result from memory error, misunderstanding or missing knowledge. A lie normally implies that the speaker knows the statement is false and intends to mislead.
References
Azuma, D., Meléndez, R., Ptaszynski, M., Masui, F., Aslan, L., & Eronen, J. (2025). SVM, BERT, or LLM? A comparative study on multilingual instructed deception detection. AI, 6(9), 239. DOI / publisher page
Hopkins, J., Khullar, D., Wang, R., & Roger, F. (2026, August 21). Fine-tuned lie detectors failed to generalize. Anthropic Alignment Science Blog. Publisher page
Joshi, G., Tasgaonkar, V., Deshpande, A., et al. (2025). Multimodal machine learning for deception detection using behavioral and physiological data. Scientific Reports, 15, 8943. Nature / Scientific Reports
Loconte, R., Russo, R., Capuozzo, P., Pietrini, P., et al. (2023). Verbal lie detection using Large Language Models. Scientific Reports, 13, 22849. Nature / Scientific Reports
Markowitz, D. M., & Levine, T. R. (2025). The (in)efficacy of AI personas in deception detection experiments. Journal of Communication, 75(6), 459–469. Oxford Academic
Markowitz, D. M., & Levine, T. R. (2026). Theory-guided deception detection: A RAG-based artificial intelligence exploration. arXiv preprint
Miah, M. M. M., Anika, A., Shi, X., & Huang, R. (2025). Hidden in plain sight: Evaluation of the deception detection capabilities of LLMs in multimodal settings. ACL Anthology. ACL Anthology
Ibrahim, L., Hafner, F. S., & Rocher, L. (2026). Training language models to be warm can reduce accuracy and increase sycophancy. Nature, 652, 1159–1165. Nature
Hany Farid. (2026). AI and deepfake detection interviews and commentary cited for authentication and detection limitations. UC Berkeley source
