Déjà Vu and AI Hallucinations: When Familiarity Becomes False Memory
How the human brain and generative AI confuse pattern recognition with truth—and why synthetic media may rewrite our past
Déjà vu is the unsettling sensation that the present has already occurred. A place is objectively new yet strangely familiar; a conversation seems repeated, although no previous episode can be retrieved. Neuroscience increasingly interprets the phenomenon not as a recovered memory but as a conflict between two cognitive signals: familiarity and the conscious recognition of novelty. Artificial intelligence produces a structurally similar contradiction. A language model can generate a sentence it has never encountered exactly, yet every word emerges from patterns acquired from earlier data. It constructs novelty through statistical familiarity. The machine does not experience déjà vu, but its architecture operates in a state of perpetual pattern recognition without autobiographical recollection. The comparison reveals a fundamental vulnerability shared by biological and artificial cognition: prediction can be mistaken for memory, coherence for evidence and familiarity for truth.
The Neuroscience of Almost Remembering
The term déjà vu comes from French: déjà means “already,” while vu is the past participle of voir, meaning “to see.” Literally, the expression means “already seen.” Yet the experience differs fundamentally from ordinary recognition. When we recognize a childhood photograph, we can usually recover at least part of its context—a person, a place, a period of life, or an associated event. During déjà vu, by contrast, familiarity arises without contextual recollection. The present acquires the cognitive signature of something previously experienced, even though no corresponding memory can be retrieved.
This contradiction is central to the phenomenon. Déjà vu combines a powerful sense of familiarity with the simultaneous awareness that the situation should be new. It is therefore not simply a false belief about the past. It is a metacognitive conflict: one part of the memory system signals recognition, while another detects that this recognition is inappropriate (O’Connor & Moulin, 2010).
Human memory is not a single recording mechanism. Recognition appears to involve at least two partly distinguishable processes. Familiarity produces a rapid sense that something has been encountered before, while recollection retrieves contextual details, including where, when, and under what circumstances the encounter occurred. Studies of patients with different forms of temporal-lobe damage support this distinction, linking the hippocampus especially strongly to conscious recollection, while surrounding medial temporal regions contribute to familiarity-based recognition (Yonelinas et al., 2002).
These processes normally cooperate, but they can become dissociated. A face may seem recognizable even when its name and origin remain inaccessible. Déjà vu appears to be an unusually intense and consciously detected version of this separation: recognition occurs, but the source of recognition cannot be found.
Virtual-reality experiments demonstrate how such a conflict may arise. Participants explored new environments whose spatial arrangements resembled scenes they had previously visited. Even when they could not consciously identify the earlier environment, similar configurations increased both familiarity and reports of déjà vu. A new room could therefore activate the structural pattern of an older experience without bringing the original episode fully back to mind (Cleary et al., 2012).
Imagine entering a room containing a staircase, a doorway, and a table arranged in the same spatial relationship as those in a restaurant visited years earlier. The furniture, colors, and location may all be different, but the underlying geometry is similar. The brain detects the relationship among the elements before identifying where that relationship was previously encountered. What emerges is not a complete memory but an unexplained signal of correspondence.
The hippocampal memory system helps explain this ambiguity through two complementary operations:
- Pattern separation allows the brain to encode similar experiences as distinct events. Two visits to the same street, for example, should not automatically collapse into a single memory. Neuroimaging research has associated this differentiation process with the dentate gyrus and CA3 regions of the hippocampus (Bakker et al., 2008).
- Pattern completion works in the opposite direction. A partial cue—a scent, a melody, or a visual fragment—can reactivate the other elements of an event. Experimental evidence indicates that the hippocampus binds the separate components of an episode together, allowing one component to trigger a more complete recollection (Horner et al., 2015).
Déjà vu may occur when similarity initiates pattern completion, but the brain fails to recover a sufficiently specific episode. The system detects a match without obtaining the contextual information needed to determine what, exactly, has been recognized. In healthy individuals, this may reflect a temporary mismatch between an erroneous memory signal and the mechanisms that monitor whether that signal is appropriate (O’Connor & Moulin, 2013).
Because pattern completion, recollection, and source monitoring depend heavily on medial temporal structures, déjà vu has sometimes been loosely described as a temporal-lobe hallucination. Neurologically, however, it is more accurately understood as an illusion of familiarity. A hallucination produces a perceptual experience without a corresponding external stimulus. Déjà vu instead applies an inappropriate sense of prior occurrence to a genuine event taking place in the present.
The temporal lobe nevertheless has an important clinical connection to the phenomenon. In temporal-lobe epilepsy, déjà vu can appear as an experiential aura associated with a focal seizure affecting memory networks (Illman et al., 2012). Electrical stimulation of the perirhinal and entorhinal cortices has also produced déjà vu and related experiences in patients undergoing investigation for epilepsy, suggesting that these regions participate in the interaction between familiarity and recollection (Bartolomei et al., 2004).
Some temporal-lobe seizures produce more elaborate “dreamy states,” including vivid scenes or apparent fragments of autobiographical memory. These experiences are closer to hallucinated recollection than ordinary déjà vu because they contain actual perceptual or narrative content, not merely an unexplained sense of familiarity (Vignal et al., 2007). Occasional déjà vu, therefore, should not be treated as evidence of epilepsy. Clinical research is useful because it reveals the networks capable of generating the experience, not because every instance is pathological.
Human memory, then, is not a complete audiovisual archive of the past. It reconstructs experience from partial traces, present cues, and learned expectations. Déjà vu exposes this machinery at the moment when familiarity, recollection, and source monitoring briefly fall out of alignment.
When Familiarity Masquerades as Prediction
Déjà vu sometimes carries an additional illusion: the impression that the next moment is already known. A person may become convinced that someone is about to enter the room, that a particular sentence will be spoken, or that the road ahead will turn in a specific direction.
Experiments suggest that this conviction does not represent genuine predictive ability. In virtual environments designed to induce déjà vu, participants often reported a stronger sense of knowing what would happen next. Their actual predictions, however, remained no better than chance (Cleary & Claxton, 2018). More recent research has strengthened this conclusion. When researchers increased familiarity with the spatial structure of virtual scenes, participants became more likely to report an illusory sense of prediction, even though familiarity did not provide reliable information about the upcoming event (Huebert et al., 2025).
The error follows naturally from the experience. When the present is mistakenly processed as part of a remembered sequence, its continuation also seems as though it should be retrievable. The brain does not perceive the future; it misinterprets the strength of its own recognition signal.
This confusion is possible because memory and prediction are closely related. The hippocampus does more than reconstruct previous events. It also uses learned associations to anticipate what is likely to appear next. In visual experiments, hippocampal activity has represented expected shapes before they appeared, and the strength of these predictions was related to expectation-driven changes in the visual cortex (Kok & Turk-Browne, 2018).
Remembering and anticipating therefore share a basic operation: both use incomplete information to construct what is currently absent. Memory reconstructs an event that is no longer present, while prediction constructs an event that has not yet occurred.
This biological vulnerability provides a useful framework for approaching generative AI.
Human déjà vu occurs when a pattern is activated without reliable source recovery.
AI hallucination occurs when a plausible continuation is generated without reliable evidential grounding.
In both cases, coherence arrives before provenance has been established.
The resemblance, however, is architectural rather than experiential. A person consciously experiences the contradiction of déjà vu. A language model does not need to feel familiarity, confusion, or doubt in order to reproduce a comparable information-processing error.
Artificial Memory and the Logic of Hallucination
A language model does not remember in the autobiographical sense. It has no demonstrated personal past, no continuous life history, and no verified subjective experience of recollection. Nevertheless, several components of AI systems are commonly described using the language of memory.
During training, a language model acquires statistical regularities from large collections of text. These regularities are encoded across its parameters rather than stored as a conventional library of complete sentences. When generating an answer, an autoregressive model predicts a continuation one token at a time from the preceding context (Brown et al., 2020).
The conversation visible to the model functions differently. Text inside the context window is temporarily available during generation, but it is not identical to information encoded during training. Additional architectures can also connect a model to databases, search systems, or explicit long-term stores. LongMem, for example, introduced a retrieval architecture intended to preserve and reuse information from histories extending beyond an ordinary context window (Wang et al., 2023).
Researchers have also identified limited functional parallels between transformer attention mechanisms and computational models of human episodic memory. In particular, some attention heads display patterns resembling contextual maintenance and retrieval mechanisms used in cognitive models. These findings reveal similarities in information processing, not evidence that transformers remember or experience the past as humans do (Li et al., 2024).
At its core, a language model extends an incomplete sequence with continuations that are statistically compatible with patterns learned during training. The exact sentence may never have appeared before, but its grammar, concepts, and associations are assembled from prior regularities. Each output is therefore new in composition while remaining dependent on the past.
This is where the analogy with hippocampal pattern completion becomes useful. In biological memory, a partial cue may reactivate the larger event associated with it. In language generation, an incomplete prompt triggers a sequence of increasingly specific continuations. The biological and mathematical mechanisms are fundamentally different, but both systems infer missing information from partial structure.
The comparison becomes most revealing when completion outruns discrimination. Human pattern separation helps distinguish one experience from other, similar experiences. Language models can differentiate linguistic patterns, but their basic generation objective does not automatically verify whether a likely continuation corresponds to a real external event or source.
An AI hallucination can therefore be understood, metaphorically, as pattern completion without sufficient grounding. The model receives a prompt and generates a quotation, title, explanation, or event that fits the surrounding language. The result may be false, yet remain highly coherent. An invented academic article can resemble a real publication. A fabricated quotation can match an author’s style. A nonexistent historical event can appear consistent with the surrounding narrative.
Empirical studies have repeatedly shown that fluent language generation does not guarantee truthfulness. Language models can reproduce common misconceptions when those falsehoods are strongly represented in human-produced text (Lin et al., 2022). Earlier work on neural summarization likewise found that systems could introduce information unsupported by the source document while producing grammatically convincing summaries (Maynez et al., 2020).
The model has not necessarily retrieved false information from one identifiable location. Instead, it has constructed an answer from associations that make the statement linguistically probable. This is the crucial distinction between plausibility and evidence. At this level, the comparison with déjà vu becomes precise:
- Déjà vu produces recognition without a recoverable episode.
- AI hallucination produces a convincing statement without recoverable support.
In each case, a strong internal signal is incorrectly treated as evidence about the external world.
Modern AI systems can be equipped with mechanisms designed to reduce this problem. Retrieval-Augmented Generation (RAG) connects a model to an external document collection, allowing it to condition answers on retrieved evidence rather than relying entirely on associations encoded in its parameters. Early RAG experiments produced more factual responses than comparable systems using parametric memory alone on several knowledge-intensive tasks (Lewis et al., 2020).
Prompting techniques can also elicit intermediate reasoning. Chain-of-thought prompting, for example, improved performance on several arithmetic, symbolic, and commonsense tasks by encouraging models to generate intermediate steps before answering (Wei et al., 2022). Agent frameworks such as Reflexion can preserve feedback from earlier attempts and use written self-critiques to guide later decisions (Shinn et al., 2023).
These methods resemble metacognitive control in a functional sense: they allow a system to reconsider an answer, compare alternatives, retrieve evidence, or revise its approach. They should not be confused with biological self-awareness. The model does not need to experience doubt in order to generate a critique of its previous output.
Nor does self-reflection automatically guarantee correction. Experiments have found that language models can struggle to identify and repair their own reasoning errors when no reliable external feedback is available. In some cases, repeated self-correction can even reduce accuracy (Huang et al., 2024). The most dependable safeguards therefore introduce something beyond fluent completion: external documents, verifiable calculations, independent evaluators, contradiction checks, calibrated uncertainty, and the ability to abstain when evidence is insufficient.
Human déjà vu already contains a primitive warning signal. The person notices that recognition and reality are in conflict. The strangeness of the experience is itself an alarm. Safer AI requires engineered equivalents of that alarm.
When AI Rewrites Human Memory
The relationship between AI and déjà vu extends beyond metaphor because generative systems can influence the memories of the people who use them.
Human recollection is vulnerable to the Misinformation Effect: misleading information introduced after an event can alter how that event is later remembered. In a classic experiment, participants watched films of automobile collisions and were later asked questions using different verbs. Describing the cars as having “smashed” rather than merely “hit” affected speed estimates and increased later reports of broken glass that had not been present (Loftus & Palmer, 1974).
Subsequent research showed how post-event misinformation can become incorporated into later memory reports, sometimes producing recollections that are confidently held despite differing from the original event (Loftus & Hoffman, 1989). One explanation involves source monitoring. Memories do not arrive with infallible labels identifying their origins. The brain must infer whether a detail came from direct perception, imagination, another person’s testimony, a photograph, or something read later. Errors occur when the content is remembered but its origin is misidentified (Johnson et al., 1993).
Generative AI can greatly expand the scale and personalization of this problem. In a 2024 study involving 200 participants, volunteers watched a crime video and were subsequently questioned under different conditions. A generative chatbot that introduced misleading details produced more than three times as many immediate false memories as the control condition. The number of false memories remained elevated one week later, and confidence in some of them also remained higher (Chan et al., 2024).
These findings are important but should be interpreted with appropriate caution. The study was released as a preprint, and its simulated witness-interview setting does not represent every possible interaction with a chatbot. It nevertheless demonstrates that conversational AI can reproduce and amplify mechanisms already known from misinformation research.
A separate preregistered study, also released as a preprint, examined AI-edited photographs and AI-generated videos. Participants who viewed videos generated from altered images showed approximately twice the false-memory rate of the control group. They also reported greater confidence in some inaccurate recollections (Pataranutaporn et al., 2024).
Photographs, videos, and conversations do not merely remind us of events. They become cues used each time an event is reconstructed. When the cue has been synthetically altered, later recollection may blend the original experience with the artificial version.
Consider an old family photograph enhanced by an AI system. The software might add facial detail, change an expression, reconstruct the background, or animate the people in the image. Years later, the generated version may be easier to retrieve than the original event. A person may remember the animation rather than the moment that was actually lived. A conversational agent could produce a similar effect by repeatedly retelling a personal story with small inaccuracies. Over time, the synthetic version may become more fluent, emotionally satisfying, and accessible than the original memory trace.
The déjà vu of the future could therefore be partly manufactured. A place, face, or conversation may produce illusory recognition not because it was previously experienced, but because a synthetic representation of something similar was repeatedly encountered.
A Culture of Synthetic Familiarity
Human culture has always developed through repetition and recombination. Stories reuse archetypes, musical styles inherit earlier structures, and political language relies on familiar emotional narratives. Generative AI did not invent this process, but it can automate and accelerate it.
As synthetic content becomes more prevalent, new texts and images may increasingly be produced from data that already contain machine-generated material. Research into recursive training has shown that indiscriminately replacing original data with successive generations of synthetic output can distort the learned distribution and eventually produce what researchers call model collapse (Shumailov et al., 2024).
This does not mean that all synthetic data are harmful or that model collapse is inevitable whenever generated material is used. Carefully selected synthetic data can be useful. The danger arises when artificial outputs circulate without provenance and are repeatedly treated as equivalent to independent observations of reality.
The cultural problem is similar. A false claim repeated across thousands of generated pages may acquire the appearance of consensus. A fabricated quotation may become familiar enough to seem authentic. An invented photograph may acquire the emotional authority normally associated with documentary evidence.
Repetition would then create a second-order illusion. People would recognize a claim because they had encountered it many times, and this recognition could be mistaken for evidence that the claim had been independently confirmed. Familiarity would become a substitute for verification.
This is déjà vu transformed into an information environment: an idea appears to come from the past because the system has reproduced it everywhere, even when no reliable original can be found. What AI generates today may also enter the training data, archives, search results, and personal collections from which both humans and future machines reconstruct yesterday. Synthetic familiarity can therefore move from generated content into collective memory.
Intelligence Requires More Than Pattern Recognition
Biological and artificial intelligence confront the same abstract problem: is the current situation genuinely the same as something previously known, or does it merely resemble it?
Human memory uses pattern separation to keep similar events distinct, recollection to recover context, and source monitoring to estimate where remembered information originated. Metacognitive processes then help us judge whether the resulting memory should be trusted.
Artificial systems need functional equivalents of these safeguards. They must distinguish information encoded in parameters from evidence retrieved from a document, generated prose from direct quotation, and linguistic probability from factual support. More reliable AI will therefore require more than larger models or more fluent reasoning. It will need mechanisms that preserve provenance: where a claim originated, whether it was independently confirmed, how directly the evidence supports it, and what information might contradict it.
Such systems should treat generation as the production of a candidate answer, not the automatic discovery of a fact. Fluent completion should be followed by retrieval, comparison, verification, or an explicit admission that sufficient evidence is unavailable.
Déjà vu teaches that recognition is not proof. A cognitive signal can be intense, coherent, and still incorrectly attributed. AI hallucinations reveal the same principle computationally: probability is not provenance, and plausibility is not truth. Intelligence is therefore more than the ability to discover and complete patterns. It is also the ability to resist a pattern when the correspondence is seductive but the evidence is weak.
Conclusion: The Past That Never Happened
Déjà vu is unsettling because it briefly weakens the boundary between present perception and past experience. A new moment acquires the authority of memory, and the future appears predictable because the present has been misclassified as repetition.
Generative AI introduces a comparable disturbance at a cultural scale. It produces language without a speaker’s biography, images without a camera, and realistic reconstructions without an original event. It creates the unprecedented from statistical traces of what came before. The machine has not literally seen everything before. Nevertheless, everything it generates is shaped by inherited patterns, and those patterns can produce synthetic familiarity strong enough to resemble knowledge.
As AI becomes embedded in photographs, archives, search engines, classrooms, and personal conversations, the central question may no longer be whether machines can remember like us. It may be whether humans, surrounded by systems that manufacture convincing familiarity, will still be able to distinguish lived experience from algorithmic reconstruction.
The most consequential AI hallucination may not be the false statement a machine produces today. It may be the false past that humanity eventually accepts as its own.
References
Bakker, A., Kirwan, C. B., Miller, M., & Stark, C. E. L. (2008). Pattern separation in the human hippocampal CA3 and dentate gyrus. Science, 319(5870), 1640–1642.
Bartolomei, F., Barbeau, E., Gavaret, M., Guye, M., McGonigal, A., Régis, J., & Chauvel, P. (2004). Cortical stimulation study of the role of rhinal cortex in déjà vu and reminiscence of memories. Neurology, 63(5), 858–864.
Brown, T. B., Mann, B., Ryder, N., Subbiah, M., Kaplan, J. D., Dhariwal, P., … & Amodei, D. (2020). Language models are few-shot learners. Advances in Neural Information Processing Systems, 33, 1877–1901.
Chan, S., Pataranutaporn, P., Suri, A., Zulfikar, W., Maes, P., & Loftus, E. F. (2024). Conversational AI powered by large language models amplifies false memories in witness interviews. arXiv preprint, arXiv:2408.04681.
Cleary, A. M., Brown, A. S., Sawyer, B. D., Nomi, J. S., Ajoku, A. C., & Ryals, A. J. (2012). Familiarity from the configuration of objects in three-dimensional space and its relation to déjà vu: A virtual reality investigation. Consciousness and Cognition, 21(2), 969–975.
Cleary, A. M., & Claxton, A. B. (2018). Déjà vu: An illusion of prediction. Psychological Science, 29(4), 635–644.
Horner, A. J., Bisby, J. A., Bush, D., Lin, W.-J., & Burgess, N. (2015). Evidence for holistic episodic recollection via hippocampal pattern completion. Nature Communications, 6, Article 7462.
Huang, J., Chen, X., Mishra, S., Zheng, H. S., Yu, A. W., Song, X., & Zhou, D. (2024). Large language models cannot self-correct reasoning yet. International Conference on Learning Representations.
Huebert, A. M., Myers, S. J., & Cleary, A. M. (2025). Experimental evidence that illusory feelings of prediction can be caused by familiarity detection. Consciousness and Cognition, 133, Article 103904.
Illman, N. A., Butler, C. R., Souchay, C., & Moulin, C. J. A. (2012). Déjà experiences in temporal lobe epilepsy. Epilepsy Research and Treatment, 2012, Article 539567.
Johnson, M. K., Hashtroudi, S., & Lindsay, D. S. (1993). Source monitoring. Psychological Bulletin, 114(1), 3–28.
Kok, P., & Turk-Browne, N. B. (2018). Associative prediction of visual shape in the hippocampus. The Journal of Neuroscience, 38(31), 6888–6899.
Lewis, P., Perez, E., Piktus, A., Petroni, F., Karpukhin, V., Goyal, N., … & Kiela, D. (2020). Retrieval-augmented generation for knowledge-intensive NLP tasks. Advances in Neural Information Processing Systems, 33, 9459–9474.
Li, J.-A., Zhou, C. Y., Benna, M. K., & Mattar, M. G. (2024). Linking in-context learning in transformers to human episodic memory. Advances in Neural Information Processing Systems, 37, 6180–6212.
Lin, S., Hilton, J., & Evans, O. (2022). TruthfulQA: Measuring how models mimic human falsehoods. Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics, 3214–3252.
Loftus, E. F., & Hoffman, H. G. (1989). Misinformation and memory: The creation of new memories. Journal of Experimental Psychology: General, 118(1), 100–104.
Loftus, E. F., & Palmer, J. C. (1974). Reconstruction of automobile destruction: An example of the interaction between language and memory. Journal of Verbal Learning and Verbal Behavior, 13(5), 585–589.
Maynez, J., Narayan, S., Bohnet, B., & McDonald, R. (2020). On faithfulness and factuality in abstractive summarization. Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, 1906–1919.
O’Connor, A. R., & Moulin, C. J. A. (2010). Recognition without identification, erroneous familiarity, and déjà vu. Current Psychiatry Reports, 12(3), 165–173.
O’Connor, A. R., & Moulin, C. J. A. (2013). Déjà vu experiences in healthy subjects are unrelated to laboratory tests of recollection and familiarity for word stimuli. Frontiers in Psychology, 4, Article 881.
Pataranutaporn, P., Archiwaranguprok, C., Chan, S. W. T., Loftus, E. F., & Maes, P. (2024). Synthetic human memories: AI-edited images and videos can implant false memories and distort recollection. arXiv preprint, arXiv:2409.08895.
Shinn, N., Cassano, F., Berman, E., Gopinath, A., Narasimhan, K., & Yao, S. (2023). Reflexion: Language agents with verbal reinforcement learning. Advances in Neural Information Processing Systems, 36.
Shumailov, I., Shumaylov, Z., Zhao, Y., Papernot, N., Anderson, R., & Gal, Y. (2024). AI models collapse when trained on recursively generated data. Nature, 631, 755–759.
Vignal, J.-P., Maillard, L., McGonigal, A., & Chauvel, P. (2007). The dreamy state: Hallucinations of autobiographic memory evoked by temporal lobe stimulations and seizures. Brain, 130(1), 88–99.
Wang, W., Dong, L., Cheng, H., Liu, X., Yan, X., Gao, J., & Wei, F. (2023). Augmenting language models with long-term memory. Advances in Neural Information Processing Systems, 36.
Wei, J., Wang, X., Schuurmans, D., Bosma, M., Xia, F., Chi, E., … & Zhou, D. (2022). Chain-of-thought prompting elicits reasoning in large language models. Advances in Neural Information Processing Systems, 35.
Yonelinas, A. P., Kroll, N. E. A., Quamme, J. R., Lazzara, M. M., Sauvé, M.-J., Widaman, K. F., & Knight, R. T. (2002). Effects of extensive temporal lobe damage or mild hypoxia on recollection and familiarity. Nature Neuroscience, 5, 1236–1241.
Editorial transparency note: This article, as with all articles published on this site, was conceived, directed, written, and reviewed by Prof. Maurício Veloso Brant Pinheiro. Artificial intelligence was used as an assistant for editorial refinement, formatting, image generation, SEO metadata, and publication workflow.
Copyright 2026 AI-Talks.org
