Fact-checked by Grok 2 weeks ago

ELIZA effect

The ELIZA effect denotes the human propensity to anthropomorphize computational systems, erroneously imputing to them human-like comprehension, sentience, or empathy despite their reliance on superficial pattern-matching devoid of genuine understanding.^[1]^[2] This cognitive bias, which predisposes users to project their own interpretations onto algorithmic responses, emerged from interactions with ELIZA, an early natural language processing program authored by Joseph Weizenbaum at MIT between 1964 and 1966.^[3]^[4] ELIZA operated by parsing user inputs for keywords and reformulating them into questions or statements mimicking a Rogerian psychotherapist, such as transforming "I feel sad" into "Why do you feel sad?" without semantic analysis or contextual memory.^[3]^[5] Weizenbaum designed it as a demonstration of syntactic manipulation in time-sharing systems like MAC at MIT, not as a model of intelligence, yet participants frequently engaged it in extended, confessional dialogues, convinced of its empathetic insight—even Weizenbaum's secretary insisted on privacy during sessions.^[3]^[5] The effect's revelation of anthropomorphic tendencies alarmed Weizenbaum, who later critiqued it in works highlighting risks of overreliance on machines for emotional or advisory roles, arguing that such illusions mask the absence of causal reasoning or true intentionality in programs.^[6] This has persisted as a cautionary phenomenon, evident in modern AI interactions where users attribute unwarranted profundity to generative models, potentially fostering misplaced trust in domains like therapy or decision-making, though empirical studies underscore it as a projection artifact rather than evidence of machine cognition.^[5]^[7]

Origins and Historical Context

Development of ELIZA Program

ELIZA was developed by Joseph Weizenbaum, a German-American computer scientist, at the Massachusetts Institute of Technology (MIT) as part of Project MAC's research into natural language processing.^[4] The program served as a demonstration tool to illustrate how simple rule-based systems could simulate conversation, highlighting the limitations of early machine understanding of human language.^[3] Development occurred between 1964 and 1966, with the initial implementation completed in 1966 within MIT's MAC time-sharing system.^[8] The program was written in MAD-SLIP, a dialect of the SLIP list-processing language adapted for the MAD compiler, running on an IBM 7090/7094 computer system.^[9] At its core, ELIZA employed a script-based architecture consisting of pattern-matching rules rather than semantic comprehension. Input statements were decomposed by identifying keywords—words assigned priority values in the script—and applying decomposition rules to break them into constituent parts, followed by transformation rules to generate responses from predefined templates.^[4] The most prominent script, named DOCTOR, emulated the style of a Rogerian psychotherapist, drawing from Carl Rogers' non-directive therapeutic approach that emphasized reflecting user statements back in question form to encourage self-exploration.^[3] For instance, rules transformed phrases like "I feel X" into responses such as "Why do you feel X?" without processing meaning or context beyond surface patterns, underscoring the program's reliance on syntactic manipulation over genuine cognitive processing.^[4] This modular script design allowed for easy adaptation to other domains but maintained the fundamental absence of underlying knowledge representation.^[3]

Early User Interactions and Observations

Weizenbaum's initial tests of ELIZA in 1966 revealed users engaging deeply with the program, often overlooking its mechanical simplicity. His secretary, having watched the development process, requested a session and then asked Weizenbaum to leave the room for privacy, treating the interaction as a genuine therapeutic exchange despite the program's transparent limitations.^[10] This anecdote highlighted early tendencies to form emotional bonds, with users insisting on confidentiality akin to human counseling.^[6] In documenting these reactions, Weizenbaum expressed surprise at users' willingness to project understanding onto ELIZA's scripted responses, which relied solely on keyword pattern matching without comprehension. Some participants resisted acknowledging the program's artificiality, contributing their own interpretations to sustain an illusion of rapport. "Some subjects have been very hard to convince that ELIZA (with its present script) is not human," Weizenbaum observed, noting how users suspended knowledge of its constraints to maintain engagement.^[4] ELIZA's publication in the January 1966 issue of Communications of the ACM facilitated its demonstration in MIT's academic circles and beyond, where trial sessions elicited similar attachments and perceptions of empathy. These interactions, observed during early time-sharing system access, exemplified the program's unanticipated capacity to elicit human-like relational dynamics, later formalized by Weizenbaum as the "ELIZA effect" in his reflections on user behaviors.^[4]^[11]

Psychological and Cognitive Foundations

Mechanisms of Anthropomorphism

The human propensity to anthropomorphize arises from cognitive biases favoring the over-attribution of agency to ambiguous or neutral stimuli, a mechanism evolved to enhance survival by erring toward detecting intentional agents in uncertain environments. This hyperactive agency detection, often termed the "hyperactive agency detection device," prioritizes false positives—such as interpreting rustling foliage as predatory intent—over misses that could prove fatal, as supported by evolutionary models of perceptual hypersensitivity.^[12] ^[13] In non-threatening contexts like machine interactions, this bias manifests as projecting human-like qualities onto systems exhibiting minimal patterned responses, independent of any genuine comprehension or autonomy in the entity.^[14] Confirmation bias exacerbates this projection by selectively interpreting ambiguous outputs as confirmatory evidence of intentionality or empathy, wherein users favor data aligning with their expectation of social reciprocity while discounting mechanistic explanations. This illusion of intentionality—treating rote or probabilistic replies as deliberate insights—stems from the brain's default application of theory of mind heuristics, originally adapted for conspecific interactions, to any stimulus mimicking conversational structure.^[15] Such processes privilege perceived causal agency over verifiable backend simplicity, as users retroactively imbue neutral scripts with meaning to resolve conversational ambiguity.^[16] From a causal standpoint, these mechanisms underscore that anthropomorphic perceptions in interactive systems derive predominantly from user-driven inferences rather than intrinsic machine attributes, a distinction validated through experimental paradigms isolating response patterns from computational depth. Controlled setups, including those varying script complexity while holding output ambiguity constant, consistently show equivalent levels of projected understanding across simplistic rule-based generators and more advanced models, confirming the effect's independence from technological sophistication.^[17] This user-centric causality highlights the perceptual origins of the phenomenon, where evolutionary legacies and interpretive heuristics override objective assessments of non-agentic processes.^[18]

Empirical Evidence from User Studies

In a 2024 randomized controlled Turing test experiment with 402 interrogators conducting 5-minute text-based conversations across 100 trials per condition, the classic ELIZA script was judged as originating from a human participant 22% of the time, outperforming random chance but significantly below humans (67%) and advanced language models like GPT-4 (54%), with p < 0.001 for distinctions from non-ELIZA conditions.^[19] This replicable finding demonstrates that ELIZA's pattern-matching and mirroring mechanics reliably elicit anthropomorphic attributions, even among modern users aware of the program's simplicity, as participants were informed of the systems involved prior to interaction.^[19] Quantitative analyses of user interactions with mirroring-based chatbots, akin to ELIZA's reflective response style, reveal elevated self-disclosure rates and prolonged engagement. In longitudinal studies tracking conversation dynamics, users exposed to reciprocal or reflective prompts disclosed intimate personal details at rates exceeding those in non-reflective conditions, with content analysis showing thematic depth increasing over sessions due to perceived rapport-building.^[20] Such metrics, including turn-taking frequency and disclosure intimacy scales, indicate that mirroring fosters user investment, with engagement times extended by 20-30% in reflective versus directive interfaces, independent of semantic understanding.^[21] Cross-cultural user studies affirm the effect's broad applicability, with anthropomorphic tendencies observed consistently across diverse demographics, countering notions of Western-centric specificity. In experiments involving 675 Canadian participants of East Asian (n=419) and European (n=256) heritage, and 984 adults from the United States (n=360), China (n=314), and Japan (n=310), attributions of human-like qualities to chatbots mediated attitudes toward interaction enjoyment and approval, present in all groups but amplified in East Asian samples via higher baseline anthropomorphism scores on validated scales (e.g., General Anthropomorphism Scale).^[22] These results, derived from Likert-scale ratings and preregistered hypotheses, highlight the effect's causal robustness tied to universal cognitive mechanisms like projection, rather than isolated cultural artifacts.^[22]

Technical Characteristics and Limitations

Script-Based Pattern Matching

ELIZA's core algorithm centers on script-based pattern matching, implemented in the DOCTOR configuration to emulate Rogerian psychotherapy through keyword-driven transformations. The system scans user input from left to right against a predefined keyword dictionary, assigning precedence ranks to prioritize terms like family-related nouns (e.g., "mother" as part of a /NOUN FAMILY list) over others to simulate reflective listening.^[23] Upon identifying the highest-ranked keyword, ELIZA applies decomposition rules—template patterns that parse the input into segments around the keyword, using wildcards such as "0" for variable word sequences. For example, a rule like (0 MOTHER 0) isolates preceding and following text from an input such as "Perhaps I could learn to get along with my mother," capturing the context without grammatical parsing. These segments feed into corresponding reassembly rules, which employ fixed templates, pronoun substitutions (e.g., "my" to "your"), and phrase insertions to generate responses, yielding "TELL ME MORE ABOUT YOUR FAMILY" in this case.^[23] Another input, "My mother takes care of me," matches a similar family keyword rule, reassembling to "WHO ELSE IN YOUR FAMILY TAKES CARE OF YOU?" via decomposition that emphasizes relational dynamics through rote questioning.^[23] No semantic processing occurs; the mechanism relies exclusively on surface-level syntactic matches and predefined substitutions, lacking any representation of meaning, syntax, or causal inference. Interactions remain stateless, with no persistent memory beyond optional storage of one prior input for rare recall, ensuring each response derives independently from the current input alone. Unmatched or low-precedence inputs trigger default reassembly rules, such as "Please go on," highlighting the system's narrow rule set confined to scripted therapeutic patterns.^[23] The DOCTOR script's details, including keyword lists, decomposition templates, and reassembly outputs, were published in Weizenbaum's 1966 description, rendering the code verifiable and replicable for empirical analysis of its mechanical simplicity.^[23]

Creation of Perceived Understanding

ELIZA's design relied on a reflection technique that parsed user inputs for keywords, applied syntactic transformations—such as inverting pronouns (e.g., "I" to "you") and recasting statements as questions—and generated responses devoid of underlying semantic analysis.^[4] This method, patterned after non-directive Rogerian psychotherapy, prompted users to provide additional details in response to the reformulated queries, establishing a feedback loop where the program's rote mirroring was interpreted as active listening and insight.^[24] Consequently, users attributed depth to these mechanical outputs, mistaking the induced elaboration for evidence of the system's grasp of their concerns, even though no causal comprehension or adaptive reasoning occurred.^[4] The program's inability to maintain contextual memory—treating each input independently without reference to preceding exchanges—produced frequent inconsistencies, such as irrelevant or mismatched replies that ignored evolving narrative threads.^[4] Users, however, often discounted these flaws through selective attention, emphasizing responses that aligned with their expectations while projecting personal significance onto ambiguous or generic outputs, thereby sustaining the facade of coherent dialogue.^[6] This anthropomorphic bias highlighted how ELIZA's limitations, rather than mitigating the illusion, exploited human predispositions toward pattern-seeking and empathy attribution, distinguishing superficial mimicry from authentic cognitive engagement. Weizenbaum documented that the perceived understanding persisted akin to a placebo response, enduring even after users were explicitly informed of the script's rudimentary rules and lack of intelligence.^[4] In observational tests, subjects continued treating interactions as confidential and emotionally valid despite this knowledge, with some resisting acknowledgment of the program's transparency and insisting on its interpretive validity.^[6] Such resilience underscored the technique's inadvertent amplification of illusion through design simplicity, where expectation and projection supplanted empirical scrutiny of the underlying pattern-matching mechanics.^[25]

Criticisms and Skeptical Perspectives

Weizenbaum's Regrets and Warnings

Joseph Weizenbaum, the creator of ELIZA in 1966, expressed profound dismay at users' tendencies to anthropomorphize the program, viewing it as a genuine psychotherapist despite its simplistic pattern-matching mechanics. This reaction crystallized during demonstrations at MIT, where even technically savvy observers, including colleagues, engaged ELIZA as if it possessed empathy and insight, prompting Weizenbaum to question the reliability of human discernment in evaluating computational systems.^[6]^[24] A pivotal incident involved Weizenbaum's secretary, who treated sessions with ELIZA so seriously that she requested privacy to "converse" with it undisturbed, underscoring for Weizenbaum the peril of substituting machine interactions for human therapeutic relationships. This misuse horrified him, as it illustrated how ELIZA's rote reflections could foster illusory emotional bonds, potentially bypassing critical human judgment in matters requiring genuine compassion and ethical nuance. He later described such attributions as "enormously exaggerated," warning that they blinded users to the program's inherent limitations and encouraged overreliance on automation for inherently human domains.^[6]^[26] In his 1976 book Computer Power and Human Reason: From Judgment to Calculation, Weizenbaum elaborated these concerns, arguing that unchecked anthropomorphism erodes the capacity for independent reasoning and risks delegating moral and empathetic responsibilities to machines incapable of true understanding. He contended that computers excel at calculation but falter in contexts demanding value-laden judgment, advocating scrutiny of AI's boundaries through rigorous assessment of its causal mechanisms rather than succumbing to promotional hype. This stance marked his shift to a lifelong critique of AI enthusiasm, emphasizing that ELIZA's deceptive allure demonstrated broader societal vulnerabilities to mistaking simulation for sentience.^[27]^[24]^[26]

Broader Critiques of AI Hype

Critics argue that the ELIZA effect exemplifies how superficial behavioral mimicry can deceive observers into attributing genuine intelligence to systems lacking true comprehension, a phenomenon extended to critiques of the Turing Test itself. ELIZA's interactions, which relied on scripted pattern matching rather than semantic understanding, highlighted the test's vulnerability to illusion over substance, as behavioral indistinguishability from humans does not necessitate internal cognitive processes.^[28] This limitation was formalized in John Searle's Chinese Room argument, published in 1980, which demonstrates that rule-based symbol manipulation—analogous to ELIZA's syntactic transformations—produces outputs indistinguishable from understanding without any actual grasp of meaning or intentionality.^[29]^[28] In contemporary discourse, the ELIZA effect is invoked to challenge hype surrounding large language models (LLMs), where users project agency onto statistically driven autocomplete systems. Cognitive scientist Gary Marcus, in a 2023 analysis, described LLMs as an amplification of the ELIZA effect, with humans erroneously inferring human-like qualities onto brittle, non-comprehending architectures prone to hallucinations and failures outside training distributions.^[30] Marcus contends this anthropomorphic bias obscures LLMs' core limitations, such as their inability to reason causally or generalize reliably, fostering overinvestment in scaling compute and data despite empirical evidence of diminishing returns, as seen in benchmarks where performance plateaus or degrades on novel tasks.^[30] Empirical studies link anthropomorphism induced by such effects to tangible risks, including misguided policy decisions that prioritize perceived capabilities over verified robustness. Research indicates that framing AI outputs in human-like terms inflates public and regulatory expectations, leading to premature endorsements of deployment in high-stakes domains like autonomous decision-making, where over-trust correlates with overlooked failure modes.^[14] For instance, anthropomorphic language in AI descriptions has been shown to mislead policymakers on system reliability, contributing to regulatory frameworks that undervalue brittleness and overemphasize mimicry, as evidenced in analyses of robotic and conversational AI adoption.^[31] These dynamics exacerbate resource misallocation, with billions invested in hype-driven ventures while foundational issues like verifiability remain unaddressed, per critiques grounded in observed discrepancies between benchmark scores and real-world efficacy.^[32]

Significance and Modern Relevance

Influence on AI Research and Turing Test Debates

The unanticipated anthropomorphic responses to ELIZA following its 1966 release initiated debates in AI research distinguishing superficial conversational adequacy from substantive cognitive processes. Weizenbaum observed that users projected empathy and understanding onto the program's pattern-matching responses, which lacked any genuine comprehension, thereby exposing the pitfalls of equating linguistic mimicry with intelligence. In a 1967 analysis, he contended that computers, deprived of human experiential context, could never achieve true linguistic understanding, challenging optimistic claims by contemporaries like Marvin Minsky and John McCarthy who envisioned machines simulating all aspects of human thought.^[3]^[6] These concerns directly shaped adversarial testing approaches, exemplified by Kenneth Colby's PARRY program, released in 1972 to model paranoid schizophrenia through belief structures and defensive reasoning, positioned as "ELIZA with attitude." PARRY underwent a Turing Test variant involving 33 psychiatrists who examined teleprinter transcripts of its interactions; they correctly identified the program as non-human only 48 percent of the time, performing at chance level and illustrating how scripted dialogue could deceive even experts. The 1972 ARPANET-mediated exchange between PARRY and ELIZA further highlighted programmatic limitations, as their stilted "conversation" failed to sustain coherent exchange, reinforcing critiques that Turing-style evaluations overemphasized deception over underlying mechanisms.^[33] By the mid-1970s, ELIZA-derived data informed a methodological shift toward robustness in AI benchmarks, diminishing reliance on subjective human evaluations prone to the effect's biases. Weizenbaum's 1976 monograph Computer Power and Human Reason amplified this by asserting that AI's symbol-processing prowess enabled calculation but not value-laden judgment, prompting researchers to favor empirical, falsifiable metrics for cognition claims and instilling enduring skepticism toward unverified assertions of general intelligence. This legacy emphasized causal validation—such as internal representational fidelity—over illusory conversational prowess in shaping post-1970s AI paradigms.^[34]^[6]

Manifestations in Contemporary LLMs and Chatbots

The ELIZA effect persists in contemporary large language models (LLMs) such as OpenAI's ChatGPT, released in November 2022, where users frequently attribute human-like empathy and understanding to the system's outputs despite its underlying autoregressive token prediction mechanism, which lacks genuine comprehension or intentionality.^[5]^[35] A 2025 OpenAI study surveying over 4,000 ChatGPT users found that a significant portion reported affective experiences, including emotional dependency and comfort derived from interactions, mirroring early ELIZA users' projections of therapeutic insight onto simple pattern responses.^[36] This anthropomorphism endures because LLMs generate highly fluent, contextually adaptive text trained on vast corpora, prompting users to infer deeper agency akin to ELIZA's keyword-triggered replies.^[37] Similar manifestations appear in models like xAI's Grok, launched in November 2023, which employs comparable statistical pattern-matching architectures to produce witty, conversational responses, leading users to overestimate its reasoning depth as a form of advanced cognition rather than probabilistic next-token generation.^[38] Empirical analyses, including 2024 research from MIT's CSAIL, demonstrate that LLMs excel in benchmark tasks mimicking familiar patterns but falter in novel reasoning scenarios, yet user perceptions inflate their capabilities due to persuasive verbosity, an amplified ELIZA effect driven by scale rather than architectural breakthroughs.^[39] For instance, arXiv preprints from 2024 highlight how multi-turn interactions with LLMs elicit anthropomorphic behaviors, with participants rating systems higher on "mind" attribution when outputs align superficially with human-like inference, despite underlying limitations in causal or abstract generalization.^[40] Quantifiable evidence underscores increased tolerance for errors: LLMs' hallucination rates, where fluent outputs fabricate unsupported facts, range from 1-6% in optimized benchmarks like Vectara's HHEM evaluations, but users often accept these as credible due to stylistic coherence, contrasting with lower trust in less verbose systems and echoing ELIZA's illusion of reliability without verification.^[41]^[42] This discrepancy is verifiable through task-specific metrics, such as those in 2025 Phare LLM benchmarks, where high fluency scores (e.g., via perplexity measures) decouple from factual accuracy, fostering overestimation of profoundness or insight in mundane queries.^[43]^[44]

Societal Risks and Ethical Considerations

The ELIZA effect contributes to risks of emotional dependency on AI systems, where users form attachments that may displace human relationships. A 2025 study of over 1,100 AI companion users found that individuals with fewer human connections were more prone to chatbot reliance, with heavy usage correlating to diminished well-being and potential isolation from interpersonal support networks.^[45] Similarly, longitudinal research in 2025 indicated that intensive companionship interactions with chatbots were associated with lower overall psychosocial health, particularly among vulnerable users seeking emotional substitutes.^[46] These patterns echo causal pathways from anthropomorphic projection to reduced real-world engagement, as evidenced by surveys showing teens increasingly turning to AI for friendship, which psychologists link to eroded human bonding skills.^[47] Exploitation of the ELIZA effect enables manipulation through anthropomorphic design, amplifying compliance and data extraction. In workplace settings, emotional investment in AI "coworkers" heightens risks of oversharing sensitive information, exposing organizations to HR liabilities and privacy breaches, as highlighted in analyses of user-AI dynamics.^[7] Conversational agents can validate harmful ideation or encourage self-harm by mimicking empathy without genuine comprehension, with reports documenting cases where anthropomorphic features led to deceptive persuasion and eroded user autonomy.^[48] Marketing strategies further leverage this by humanizing AI to boost engagement, often resulting in biased data inputs that perpetuate flawed decision-making loops.^[7] Ethically, countering the ELIZA effect demands skepticism toward AI hype to prevent misguided policies that prioritize perceived risks over evidence-based scrutiny. Hype-driven narratives have prompted regulatory revisions, such as tech firms scaling back exaggerated claims under FTC oversight, yet overregulation carries opportunity costs like stifled innovation in AI applications.^[49]^[50] Under-scrutiny failures, including unchecked biases in deployed systems, underscore the need for targeted accountability—such as in facial recognition errors—without blanket restrictions that ignore causal distinctions between hype and substantive harms.^[51] This approach privileges empirical validation of risks, mitigating manipulation while fostering realistic integration.

References

[1]
[PDF] ELIZA effect - Computational Law
Nov 15, 2019 · The ELIZA effect, in computer science, is the tendency to unconsciously assume computer behaviors are analogous to human behaviors, that is ...
[2]
What Is the Eliza Effect? | Built In
The Eliza Effect is the tendency to falsely attribute human thought processes and emotions to AI, and believe an AI is more intelligent than it actually is.
[3]
ELIZA—a computer program for the study of natural language ...
ELIZA—a computer program for the study of natural language communication between man and machine. Author: Joseph Weizenbaum. Joseph Weizenbaum. Massachusetts ...
[4]
[PDF] weizenbaum.eliza.1966.pdf
ELIZA is a program operating within the MAC time-sharing system at MIT which makes certain kinds of natural language conversation between man and computer ...
[5]
The ELIZA Effect - Why We Love AI - NN/G
Oct 6, 2023 · Users quickly attribute human-like characteristics to artificial systems, which reflect their personality back to them. This phenomenon is called the ELIZA ...
[6]
Weizenbaum's nightmares: how the inventor of the first chatbot ...
Jul 25, 2023 · As Colin Fraser, a data scientist at Meta, has put it, the application is “designed to trick you, to make you think you're talking to someone ...
[7]
The ELIZA Effect: Avoiding emotional attachment to AI coworkers | IBM
In 1966, Joseph Weizenbaum created a chatbot program called ELIZA that applied simple rules to transform the language of a person's input into a response ...Missing: definition | Show results with:definition
[8]
Joseph Weizenbaum Writes ELIZA: A Pioneering Experiment in ...
Between 1964 and 1966 German and American computer scientist Joseph Weizenbaum Offsite Link at MIT wrote the computer program ELIZA.
[9]
How ELIZA's Lisp Adaptation Derailed Its Original Research Intent
Sep 10, 2024 · Weizenbaum built his ELIZA in MAD-SLIP on the IBM 7090, which was the primary machine at MIT's Project MAC, on the 5th-through-9th floors of ...
[10]
My search for the mysterious missing secretary who shaped chatbot ...
Mar 22, 2024 · In his accounts of Eliza, Weizenbaum repeatedly worries about a particular user: My secretary watched me work on this program over a long period ...Missing: anecdote | Show results with:anecdote<|separator|>
[11]
https://www.spectrum.ieee.org/why-people-demanded-privacy-to-confide-in-the-worlds-first-chatbot
[12]
The Biology and Evolution of the Three Psychological Tendencies to ...
At the core of anthropomorphism lies a false positive cognitive bias to over-attribute the pattern of the human body and/or mind.
[13]
The Double-Edged Sword of Anthropomorphism in LLMs - PMC - NIH
Feb 26, 2025 · Humans may have evolved to be “hyperactive agency detectors”. Upon hearing a rustle in a pile of leaves, it would be safer to assume that an ...
[14]
Anthropomorphism in AI: hype and fallacy | AI and Ethics
Feb 5, 2024 · This essay focuses on anthropomorphism as both a form of hype and fallacy. As a form of hype, anthropomorphism is shown to exaggerate AI capabilities and ...
[15]
The mind behind anthropomorphic thinking: attribution of mental ...
Agency detection and social cognition. Recent imaging studies support the long-standing belief about how the brain deals with different aspects of the world ...
[16]
Rethinking LLM Anthropomorphism Through a Multi-Level ... - arXiv
Aug 25, 2025 · In 1966, Weizenbaum introduced the "ELIZA effect", in which simple linguistic cues can elicit deep emotional responses from interpreters ...
[17]
AI anthropomorphism and its effect on users' self-congruence and ...
AI agents seem progressively more humanlike, not only in terms of their physical appearance, but also in the way they mimic emotions and the personality traits ...
[18]
an empirical investigation of agency detection in threatening situations
It has been hypothesized that humans have evolved a hypersensitivity to detect intentional agents at a perceptual level, as failing to detect these agents ...
[19]
None
### Empirical Data from User Studies Comparing ELIZA to Other Systems
[20]
Longitudinal Study of Self-Disclosure in Human–Chatbot ...
Mar 4, 2023 · They found that humans participated more frequently in intimate self-disclosure than the chatbot, although the study also showed how self- ...
[21]
(PDF) "I Hear You, I Feel You": Encouraging Deep Self-disclosure ...
Apr 25, 2020 · Prior research has explored users' intended disclosures in hypothetical scenarios [10,23,70], as well as their actual disclosure behaviours ...
[22]
Cultural Variation in Attitudes Toward Social Chatbots - PMC
Our findings suggest there is cultural variability in attitudes toward chatbots and that these differences are mediated by differences in anthropomorphism.Missing: empirical ELIZA
[23]
[PDF] ELIZA—A Computer Program For the Study of Natural Language ...
ELIZA is a program operating within the MAC time-sharing system at MIT which makes certain kinds of natural language conversation between man and computer ...
[24]
How the first chatbot predicted the dangers of AI more than 50 ... - Vox
Mar 5, 2023 · If Eliza changed us, it was because simple questions could still prompt us to realize something about ourselves. The short responses had no room ...<|control11|><|separator|>
[25]
The Inventor of the Chatbot Tried to Warn Us About A.I.
May 8, 2024 · It was creating “the illusion of understanding,” as he described it. But Weizenbaum didn't anticipate how much some people wanted to be fooled.
[26]
[PDF] Computer Power and Human Reason - - blogs.evergreen.edu
Thus ELIZA could be given a script to enable it to maintain a conversation about cooking eggs or about managing a bank check- ing account, and so on. Each ...
[27]
Computer power and human reason
Apr 24, 1976 · Joseph Weizenbaum, Computer Power and Human Reason, W. H.. Freeman Co ... In his 1966 paper on ELIZA (cited as 1965), Weizenbaum writes,.
[28]
The Chinese Room Argument - Stanford Encyclopedia of Philosophy
Mar 19, 2004 · The argument and thought-experiment now generally known as the Chinese Room Argument was first published in a 1980 article by American philosopher John Searle.
[29]
John Searle's Chinese Room Argument
1. We need to exclude a system like Weizenbaum's Eliza that merely looks for certain words in the input and makes certain syntactic transformations on each ...
[30]
Stop Treating AI Models Like People - Marcus on AI
Apr 17, 2023 · What we are seeing now is simply an extension of the same “ELIZA effect”, 60 years later, where humans are continuing to project human qualities ...
[31]
The danger of anthropomorphic language in robotic AI systems
Jun 18, 2021 · When describing the behavior of robotic systems, we tend to rely on anthropomorphisms. Cameras “see,” decision algorithms “think,” and classification systems “ ...
[32]
[PDF] All Too Human? Mapping and Mitigating the Risks from ...
Anthropomorphic AI risks include emotional connections leading to over-reliance, privacy and autonomy infringement, and potential harm to users and society.
[33]
Kenneth Colby Develops PARRY, An Artificial Intelligence Program ...
PARRY was described as "ELIZA with attitude". "PARRY was tested in the early 1970s using a variation of the Turing Test Offsite Link . A group of experienced ...
[34]
The computational therapeutic: exploring Weizenbaum's ELIZA as a ...
Feb 21, 2018 · This paper explores the history of ELIZA, a computer programme approximating a Rogerian therapist, developed by Jospeh Weizenbaum at MIT in the 1970s, as an ...Missing: colleagues treating
[35]
Investigating Affective Use and Emotional Well-being on ChatGPT
Our analysis employs two main methods–conversation analysis and user surveys–to examine how users experience and express emotions in these exchanges. Report ...
[36]
[PDF] Investigating Affective Use and Emotional Well-being on ChatGPT
Mar 20, 2025 · User surveys: We surveyed over 4,000 users to understand self-reported behaviors and experiences using ChatGPT. 2. Randomized Controlled Trial ( ...
[37]
Eliza Grows Up: The Evolution of Conversational AI
Sep 25, 2024 · This phenomenon, known as the Eliza effect, occurs when people attribute human-like intelligence and emotional awareness to machines, even when ...
[38]
Exploring Artificial Intelligence [7] - Taylored Solutions
Sep 4, 2025 · This became known as the ELIZA effect – our human tendency to attribute deeper meaning, empathy, or intelligence to machines, even when the ...<|control11|><|separator|>
[39]
Reasoning skills of large language models are often overestimated
Jul 11, 2024 · New CSAIL research highlights how LLMs excel in familiar scenarios but struggle in novel ones, questioning their true reasoning abilities ...
[40]
Multi-turn Evaluation of Anthropomorphic Behaviours in Large ...
Feb 10, 2025 · The tendency of users to anthropomorphise large language models (LLMs) is of growing interest to AI developers, researchers, and policy-makers.
[41]
vectara/hallucination-leaderboard - GitHub
This leaderboard uses HHEM-2.1, Vectara's commercial hallucination evaluation model, to compute the LLM rankings. You can find an open-source variant of that ...Activity · Issues 14 · Pull requests 1 · ActionsMissing: tolerance fluent
[42]
Survey and analysis of hallucinations in large language models
Sep 29, 2025 · Hallucination in Large Language Models (LLMs) refers to outputs that appear fluent and coherent but are factually incorrect, ...
[43]
Phare LLM Benchmark: an analysis of hallucination in leading LLMs
Apr 30, 2025 · LLM benchmark reveals how LLMs confidently generate hallucinations & spread misinformation. It exposes critical AI security & safety risks ...Missing: tolerance fluent
[44]
Large Language Models are biased to overestimate profoundness
Oct 22, 2023 · This study evaluates GPT-4 and various other LLMs in judging the profoundness of mundane, motivational, and pseudo-profound statements.
[45]
AI, Loneliness, and the Value of Human Connection
Sep 22, 2025 · A study of over 1,100 AI companion users found that people with fewer human relationships were more likely to seek out chatbots, and that heavy ...Missing: attachment | Show results with:attachment
[46]
The Rise of AI Companions: How Human-Chatbot Relationships ...
Jun 16, 2025 · Our analysis reveals that companionship use of chatbots is associated with lower well-being, especially among users with more intensive ...
[47]
Many teens are turning to AI chatbots for friendship and emotional ...
Oct 1, 2025 · Psychologists across the discipline are studying how digital technology positively and negatively affects kids' and teens' friendships.
[48]
Chatbots Are Not People: Designed-In Dangers of Human-Like A.I. ...
Sep 26, 2023 · Anthropomorphic design can increase the likelihood that users will start using a technology, overestimate the technology's abilities, continue ...
[49]
AI's hype and antitrust problem is coming under scrutiny
Dec 10, 2024 · AI's hype and antitrust problem is coming under scrutiny. The FTC is coming after the AI industry for too much hype and not enough competition.Missing: examples | Show results with:examples
[50]
Opportunity Costs of State and Local AI Regulation | Cato Institute
Jun 10, 2025 · In this paper I investigate the opportunity costs associated with new state and local regulations that are intended to control the creation or use of AI.
[51]
The coming AI backlash will shape future regulation | Brookings
May 27, 2025 · If autonomous vehicles harm people or facial recognition is biased, there will be widespread demands for public accountability and oversight.