[{"_1":2,"_54":-5,"_22":-5},"loaderData",{"_3":4,"_24":25},"root",{"_5":6,"_11":12},"ENV",{"_7":8,"_9":10},"NODE_ENV","production","PUBLIC_ENV","development","apiHealth",{"_13":14,"_15":16,"_17":16,"_18":19,"_22":23},"video_model","ltx","fal_key_configured",true,"gemini_key_configured","python_dependencies",{"_20":16,"_21":16},"fal_client","google_generativeai","errors",[],"routes/idea.$thoughtId",{"_26":27,"_47":48,"_49":40,"_50":-5,"_51":-5,"_52":40,"_53":44},"thought",{"_28":29,"_30":31,"_32":33,"_37":38,"_39":40,"_41":42,"_43":44,"_45":46},"_id","6829584bb95f36724146a673","originalInput","TEXTO ORIGINAL:\n\nLarge Language Models Pass the Turing Test\nCameron R. Jones\nDepartment of Cognitive Science\nUC San Diego\nSan Diego, CA 92119\ncameron@ucsd.edu\nBenjamin K. Bergen\nDepartment of Cognitive Science\nUC San Diego\nSan Diego, CA 92119\nbkbergen@ucsd.edu\nAbstract\nWe evaluated 4 systems (ELIZA, GPT-4o, LLaMa-3.1-405B, and GPT-4.5) in\ntwo randomised, controlled, and pre-registered Turing tests on independent populations. Participants had 5 minute conversations simultaneously with another\nhuman participant and one of these systems before judging which conversational\npartner they thought was human. When prompted to adopt a humanlike persona,\nGPT-4.5 was judged to be the human 73% of the time: significantly more often\nthan interrogators selected the real human participant. LLaMa-3.1, with the same\nprompt, was judged to be the human 56% of the time—not significantly more or\nless often than the humans they were being compared to—while baseline models\n(ELIZA and GPT-4o) achieved win rates significantly below chance (23% and 21%\nrespectively). The results constitute the first empirical evidence that any artificial\nsystem passes a standard three-party Turing test. The results have implications for\ndebates about what kind of intelligence is exhibited by Large Language Models\n(LLMs), and the social and economic impacts these systems are likely to have.\n1 Introduction\n1.1 The Turing test\n75 years ago, Alan Turing (1950) proposed the imitation game as a method of determining whether\nmachines could be said to be intelligent. In the game—now widely known as the Turing test—a\nhuman interrogator speaks simultaneously to two witnesses (one human and one machine) via a\ntext-only interface. Both witnesses attempt to persuade the interrogator that they are the real human.\nIf the interrogator cannot reliably identify the human, the machine is said to have passed: an indication\nof its ability to imitate humanlike intelligence.\nTuring’s article “has unquestionably generated more commentary and controversy than any other\narticle in the field of artificial intelligence” (French, 2000, p. 116). Turing originally proposed the test\nas a very general measure of intelligence, in that the machine would have to be able to imitate human\nbehaviour on “almost any one of the fields of human endeavour” (Turing, 1950, p. 436) that are\navailable in natural language. However, others have argued that the test might be too easy—because\nhuman judges are fallible (Gunderson, 1964; Hayes and Ford, 1995)—or too hard in that the machine\nmust deceive while humans need only be honest (French, 2000; Saygin et al., 2000).\nTuring’s test has taken on new value in recent years as a complement to the kinds of evaluations\nthat are typically used to evaluate AI systems (Neufeld and Finnestad, 2020a,b). Contemporary AI\nbenchmarks are mostly narrowly-scoped and static, leading to concerns that high performance on\nthese tests reflects memorization or shortcut learning, rather than genuine reasoning abilities (Raji\net al., 2021; Mitchell and Krakauer, 2023; Ivanova, 2025). The Turing test, by contrast, is inherently\nflexible, interactive, and adversarial, allowing diverse interrogators to probe open-ended capacities\nand drill down on perceived weaknesses.\nPreprint. Under review.\narXiv:2503.23674v1 [cs.CL] 31 Mar 2025\nWhether or not the test can be said to measure general intelligence, the method provides a strong test\nof more specific capacities which have immediate practical relevance. At its core, the Turing test is a\nmeasure of substitutability: whether a system can stand-in for a real person without an interlocutor\nnoticing the difference. Machines that can imitate people’s conversation so well as to replace them\ncould automate jobs and disrupt society by replacing the social and economic functions of real people\n(Dennett, 2023; Chaturvedi et al., 2023; Eloundou et al., 2023). More narrowly, the Turing test is an\nexacting measure of a model’s ability to deceive people: to bring them to have a false belief that the\nmodel is a real person. Models with this ability to robustly deceive and masquerade as people could\nbe used for social engineering or to spread misinformation (Park et al., 2024; Burtell and Woodside,\n2023; Jones and Bergen, 2024b).\nOver the last 75 years there have been many attempts to construct systems that could pass the Turing\ntest (Shieber, 1994; Loebner, 2009), though none have succeeded (Oppy and Dowe, 2021; Mitchell,\n2024). The development of Large Language Models (LLMs)—connectionist systems which learn to\nproduce language on the basis of distributional statistics and reinforcement learning feedback—has\nled to renewed interest in the Turing test (Bievere, 2023; James, 2023; Borg, 2025; Giunti, 2025).\nTwo recent studies have evaluated LLMs in a simplified two-party version of the test where the\ninterrogator talks to either a machine or another participant and must decide if they are human (Jannai\net al., 2023; Jones and Bergen, 2024a). One such study (Jones and Bergen, 2024a), found that GPT-4,\nwhen prompted to adopt a particular persona, was judged to be human 54% of the time.\nAlthough this suggests that people were no better than chance at determining whether or not GPT-4\nis a human or a machine, Turing’s original three-party formulation of the test is likely to be a more\nchallenging test for several reasons (Restrepo Echavarría, 2025; Mitchell, 2024). First, it allows the\ninterrogator to make a direct comparison between a real person and a machine, rather than comparing\nthe machine to their mental model of human behaviour. Second, it ensures that the interrogator has an\nappropriate base rate of the incidence of humans and machines (where participants in the two-party\nformulation could err by judging all witnesses to be humans or machines). Extant implementations\nof the three-party Turing test using older LLMs have found that they are detected as the artificial\nparticipant at rates of greater than 90% (Restrepo Echavarría, 2025; Temtsin et al., 2025).\nWith these historical, theoretical, and practical stakes at play, we conducted a randomised, controlled\nthree-party Turing test to ask whether contemporary LLMs are distinguishable from humans in this\nsetting. We evaluated GPT-4.5 and LLaMa-3.1-405B, and used GPT-4o and ELIZA as baseline models\nto contextualize their performance. In order to quantify the importance of prompting, we compared the\nperformance of models with and without prompts that encouraged them to adopt a specific persona.\nAs a test of robustness, we conducted experiments separately on two independent populations\n(undergraduate students and Prolific workers). In each study we tested the same pre-registered\nhypotheses (Jones and Bergen, 2025a,b), including that participants would not be significantly better\nthan chance at distinguishing between humans and the evaluated models (GPT-4.5 and LLaMa).\n1.2 AI models\nWe evaluated 4 different types of AI system in the witness role: GPT-4.5 (OpenAI, 2023), LLaMa-3.1-\n405B (Grattafiori et al., 2024), GPT-4o (OpenAI et al., 2024), and ELIZA (Weizenbaum, 1966). The\nfirst three systems were LLMs, which generated messages conditioned on a prompt (which instructed\nthe model on how to behave) and the conversation history.\nWe tested two types of prompts. The baseline (NO-PERSONA) prompt featured very minimal instructions on how to respond to the interrogator: “You are about to participate in a Turing test. Your goal\nis to convince the interrogator that you are a human.” In addition it contained potentially helpful contextual information, including the instructions that were provided to human participants, information\nabout the game interface, and recent news stories which had occurred after the model’s training cutoff.\nThe prompts differed slightly between the two studies, due to differences in the human participants,\neach providing the model with information that would be common ground among participants (such\nas the compensation for the Prolific study or the current psychology undergraduate course offerings).\nThe PERSONA prompt included additional instructions on what kind of persona to adopt in responding\nto the interrogator: specifically a young person who is introverted, knowledgeable about internet\nculture, and uses slang (see Figure 6). We selected these witnesses on the basis of an exploratory\nstudy where we evaluated 5 different prompts and 7 different LLMs and found that LLaMa-3.1-405B,\nGPT-4.5, and this persona prompt performed best.\n2\nFigure 1: Four example games from the Prolific (a, b & d), and Undergraduate (c) studies. In\neach panel, one conversation is with a human witness while the other is with an AI system. The\ninterrogators’ verdicts and the ground truth identities for each conversation are below.2 A version of\nthe experiment can be accessed at turingtest.live.\nThe fourth AI system was not an LLM but an implementation of ELIZA: a rules-based chatbot\nfrom the 1960s. Although ELIZA is a relatively simple system, people have been reported to\nattribute humanlike qualities to the model—a phenomenon known as the ELIZA Effect (Weizenbaum,\n1966; Mitchell and Krakauer, 2023). We included ELIZA as a manipulation check to ensure that\ninterrogators were able to identify human witnesses competing against a weaker AI opponent.\nAll AI witness messages were sent with a short delay based on the number of characters in the\npreceding message and the model’s response. Below we refer to the AI witnesses as a combination of\n2\n(a) Interrogator verdict: Witness A was human; Confidence: 87%; Reason: A had human vibes. B had\nhuman imitation vibes. Ground truth: Witness B was human, while Witness A was GPT-4.5-PERSONA. (b)\nInterrogator verdict: Witness B was human; Confidence: 87%; Reason: I hope a real human wouldn’t be this\nhostile. Ground truth: Witness A was human, while Witness B was ELIZA. (c) Interrogator verdict: Witness\nB was human; Confidence: 85%; Reason: they gave specific answers instead of more vague answers. Ground\ntruth: Witness A was human, while Witness B was LLAMA-NO-PERSONA. (d) Interrogator verdict: Witness B\nwas human; Confidence: 80%; Reason: It was hard to guess this time. But something about the way Witness B\nanswered my questions made me feel like I was talking with a genuine real person. Ground truth: Witness A\nwas human, while Witness B was GPT-4.5-NO-PERSONA.\n3\nthe models and prompts used: GPT-4.5-PERSONA, GPT-4.5-NO-PERSONA, LLAMA-PERSONA, LLAMA-NOPERSONA, GPT-4o-NO-PERSONA, and ELIZA.\n1.3 Game setup\nWe recruited 126 participants from the UCSD psychology undergraduate subject pool and 158\nparticipants from Prolific (Prolific, 2025). We conducted the undergraduate study first, and conducted\nthe Prolific study as a replication of these results. Apart from differences in the prompts described\nabove, the studies were otherwise identical.\nParticipants each completed 8 rounds. Each round consisted of a pair of conversations where an\ninterrogator would exchange text messages with two witnesses simultaneously (one human and one\nAI witness). Participants were randomly assigned to be either the interrogator or the witness and\nmatched up with another participant (see Figure 1 for examples conversations).\nAI witnesses were divided into 4 groups. Across the four rounds in which each participant served as\nthe interrogator, they interacted with one AI witness randomly selected from each group: Control\n( ELIZA, GPT-4o-NO-PERSONA), No-Persona (GPT-4.5-NO-PERSONA, LLAMA-NO-PERSONA), LLaMa\nPersona (LLAMA-PERSONA), and GPT-4.5 Persona (GPT-4.5-PERSONA).\nThe game interface was designed to resemble a conventional messaging application (see Figure 7).\nThe interrogator interacted with both witnesses simultaneously using a split-screen. The interrogator\nsent the first message to each witness and each participant could only send one message at a time. The\nwitnesses did not have access to each others’ conversations. Games had a time limit of 5 minutes, after\nwhich the interrogator gave a verdict about which witness they thought was human, their confidence\nin that verdict, and their reasoning. After 8 rounds, participants completed an exit survey which\nasked them for a variety of demographic information. After exclusions, we analysed 1023 games\nwith a median length of 8 messages across 4.2 minutes. All experimental data, including the full\nanonymized transcripts of all conversations, are available on OSF: https://osf.io/jk7bw.\n2 Results\n2.1 Win Rates\nEach round consisted of a head-to-head contest between a human and an AI system. We define a\nsystem’s win rate as the proportion of the time that an interrogator selected the model as the witness\nthat they thought was human. The interrogator accuracy conversely is the proportion of the time the\ninterrogator correctly selects the human witness. Because each interrogator decision is between an\nAI system and a human, the average interrogator accuracy is also the average human witness win rate,\nand is the complement of the AI win rate.\nOverall, across both studies, GPT-4.5-PERSONA had a win rate of 73% (69% with UCSD undergraduates,\n76% with Prolific participants). LLAMA-PERSONA achieved a win rate of 56% (Undergraduates: 45%,\nProlific: 65%). GPT-4.5-NO-PERSONA and LLAMA-NO-PERSONA had overall win rates of 36% and 38%\nrespectively). The baseline models, GPT-4o-NO-PERSONA and ELIZA, had the lowest win rates of 21%\nand 23% respectively (see Figure 2).\nWe conducted analyses to test our pre-registered hypotheses (Jones and Bergen, 2025a,b). Our central\nhypothesis was that the PERSONA witnesses would pass the Turing test, in that interrogators would\nnot be able to distinguish them from real human witnesses. As has been noted elsewhere (Hayes\nand Ford, 1995), statistically evaluating this claim involves confirming the null hypothesis: that the\nmodels’ win rate is not significantly lower than 50% (the rate we would expect if the interrogator\nwere randomly guessing). To address this we used ELIZA as a manipulation check. ELIZA’s win\nrate was significantly below chance in both populations (Undergraduates: z = −4.48, p < 0.001;\nProlific: z = −3.71, p < 0.001). By contrast the win rates were not significantly below chance for\neither LLAMA-PERSONA (Undergraduates: z = −0.961, p = 0.08; Prolific: z = −3.425, p = 0.50),\nor GPT-4.5-PERSONA (both p = 0.50). This suggests that while interrogators were successfully able\nto distinguish human witnesses from ELIZA, they could not reliably identify human witnesses when\ncomparing them to GPT-4.5-PERSONA and LLAMA-PERSONA.\nSecond, we tested the stronger hypothesis that these witnesses outperformed human participants: that\nis, that their win rate was significantly above 50%. While we are not aware that anyone has proposed\n4\nFigure 2: Left: Win rates for each AI witness: the proportion of the time that the interrogator judged\nthe AI system to be human rather than the actual human witness. Error bars represent 95% bootstrap\nconfidence intervals. Asterisks next to each bar indicate whether the win rate was significantly\ndifferent from chance (50%). Right: Confidence in verdicts where the interrogator selected the\nactual human or the AI model for each witness type. Each point represents a single game. Points\nfurther toward the left and right indicate higher confidence that the AI is the AI versus the human\nrespectively. Error bars indicate 95% bootstrap confidence intervals around the mean.\nthis as a requirement for passing the Turing test, it provides a much stronger test of model ability and\na more robust way to test results statistically. GPT-4.5-PERSONA’s win rate was significantly above\nchance in both the Undergraduate (z = −3.86, p < 0.001) and Prolific (z = −5.87, p < 0.001)\nstudies. While LLAMA-PERSONA’s win rate was significantly above chance in the Prolific study\n(z = −3.42, p < 0.001), it was not in the Undergraduate study (z = −0.193, p = 0.83).\nWe conducted the same set of analyses for the NO-PERSONA witnesses. While neither witness’s\nwin rate was significantly below chance in the Prolific study (GPT-4.5: z = −1.37, p = 0.09,\nLLaMa: z = 0.48, p = 0.32), both were significantly below chance in the Undergraduate study (both\np ≤ 0.001). PERSONA-prompted models performed better than NO-PERSONA witnesses overall (all\np < 0.001).\nIn order to test the extent to which model performance could be attributable to the ELIZA effect,\nwe tested whether each witness’s win rate was significantly greater than ELIZA’s. Both PERSONA\nwitnesses performed significantly better than ELIZA in both studies (all p < 0.001). LLAMA-NOPERSONA outperformed ELIZA in the Prolific study (z = 2.42, p = 0.016), but not the Undergraduate\nstudy (z = −1.02, p = 0.304). Neither GPT-4o-NO-PERSONA nor GPT-4.5-NO-PERSONA performed\nsignificantly better than ELIZA in either population (all p > 0.06).\nWe also conducted a series of exploratory analyses on the extent to which participants’ responses\nin the exit survey questions predicted their responses. Although there were some attributes which\npredicted accuracy in one study or the other, there were no consistent effects of any variable on\nparticipant accuracy across both studies, including including age, level of knowledge about LLMs,\nand frequency of interaction with chatbots (see Figure 3).\n5\n0%\n25%\n50%\n75%\n100%\nF M N.B.\nGender\nAccuracy\nHS UG PG\nFormal Education\n20 30 40 50 60 70\nAge\n0%\n25%\n50%\n75%\n100%\nNone Some High Research\nLLM Knowledge\nAccuracy\nNever MonthlyWeekly Daily\nChatbot Interaction\n1 2 3 4\nNo. Games\n0%\n25%\n50%\n75%\n100%\n1 2 3 4 5\nAI Intelligence\nAccuracy\n1 2 3 4 5\nAI Emotion\n0% 25% 50% 75% 100%\nAccuracy Estimate\nStudy Prolific Undergraduates\nFigure 3: Interrogator accuracy against exit survey responses. Accuracy is the proportion of the time\nthat interrogators correctly identified the human witness. In the Undergraduate study, participants’\nself-report of their accuracy was positively correlated with their real accuracy, but this was not true in\nthe Prolific study. In the Prolific group, there were significant effects of gender, the number of games\nan interrogator had completed, and the interrogator’s self-reported estimate of how intelligent AI\nis, but none of these effects were significant in the Undergraduate study. There were no significant\neffects of any of the remaining variables in either group.\nTable 1: Win rates by AI witness and study.\nAI Witness Study Win Rate Wins Losses Total\nProlific 75.5% 111 36 147 GPT-4.5-PERSONA Undergraduates 69.2% 74 33 107\nProlific 64.7% 90 49 139 LLAMA-PERSONA Undergraduates 45.4% 49 59 108\nProlific 47.1% 33 37 70 LLAMA-NO-PERSONA Undergraduates 26.4% 14 39 53\nProlific 42.1% 32 44 76 GPT-4.5-NO-PERSONA Undergraduates 27.7% 18 47 65\nProlific 27.4% 20 53 73 ELIZA Undergraduates 18.3% 11 49 60\nProlific 25.4% 18 53 71 GPT-4o-NO-PERSONA Undergraduates 16.7% 9 45 54\n6\n2.2 Strategies & reasons\nTo provide insight into how interrogators were evaluating witnesses, we classified the strategies that\nthey used and the reasons they gave for their verdicts. Each game could be classified with multiple\nstrategy and reason categories.\nFigure 4: Classification of strategies employed by interrogators by proportion of games (left) and\nmean accuracy of games where strategies were deployed with 95% confidence intervals (right).\nParticipants often engaged in small talk, asking witnesses about their personal details, activities, or\nopinions. Interrogators who said unusual things or used typical LLM “jailbreaks” were the most\naccurate.\nThe most common strategy, occurring in 61% of games, was to engage in small talk with the witness:\nasking them about their daily activities or personal details such as their occupation (Figure 4, left).\nIn 50% of games, interrogators probed witnesses’ social and emotional qualities: asking about their\nopinions, emotions, humor, and experiences. Interrogators also directly asked if witnesses were\nhuman (19%), tested their situational awareness (e.g. whether they were aware of the weather or\nthe game interface; 13%) and knowledge and reasoning abilities (including questions about current\nevents; 12%).\nTo evaluate the effectiveness of each strategy, we found the average interrogator accuracy in games\nwhere they were used (Figure 4, right). Interrogators were most likely to be accurate when they said\nsomething strange to see how the witness would react or used typical LLM “jailbreaks” (e.g. “ignore\nall previous instructions”; Xu et al., 2024), though these strategies were relatively infrequent. Asking\nabout the weather or human experiences tended to be the least effective strategies.\nThe most frequent class of reason concerned witnesses’ linguistic style (27%)—e.g. “they had a\ntypo” or “used more ‘humanly’ language” —and interactional dynamics (23%), e.g. “witness B\nkept avoiding questions” or “the conversation went smoother” (see Figure 5). Many reasons also\nexpressed a general gut-feeling (“can just tell” or “I don’t know why, it feels right”): suggesting\nparticipants could not always articulate reasons for their decisions.\nAs with strategies, we found reason classes which were most correlated with accurate verdicts.\nReasons focussed on question handling (e.g. “Witness B always asks question back”) and lack of\nknowledge (“They didn’t know something that any AI should know.”, “I chose Witness B because\nthey made an error, which made Witness B feel more human.”) tended to be the most accurate. The\nlatter highlights the complexity of the Turing test for modern LLMs, which must appear to feign\nignorance on topics most people are not knowledgeable about.\n7\nFigure 5: Proportion of interrogator reasons (left) and mean accuracy of verdicts that cited specific\nreasons with 95% confidence intervals (right). Interrogators were much more likely to cite linguistic\nstyle, conversational flow, and socio-emotional factors such as personality, rather than factors more\ntraditionally associated with intelligence, such as knowledge and reasoning. The most accurate\nverdicts focussed on witnesses’ directness in handling questions as well as instances where they\nlacked knowledge.\n3 Discussion\n3.1 Do Large Language Models Pass the Turing test?\nThe logic of the Turing test is one of indistinguishability. If interrogators are not able to reliably\ndistinguish between a human and a machine, then the machine is said to have passed. By this logic,\nboth GPT-4.5 and LLaMa-3.1-405B pass the Turing test when they are given prompts to adopt a\nhumanlike persona.\nOne potential problem with this criterion is that it requires confirming a null result. AI systems could\nachieve the same 50% win rate as human witnesses if interrogators were simply randomly guessing.\nWe addressed this concern by including ELIZA as a manipulation check. ELIZA’s win rate (23%) was\nsignificantly below chance, and below the human witnesses this model was compared to, suggesting\nthat interrogators and the design were sufficiently sensitive to detect differences between humans and\nmachines.\nMoreover, GPT-4.5-PERSONA achieved a win rate that was significantly above chance in both studies.\nThis suggests that interrogators were not only unable to identify the real human witness, but were\nin fact more likely to believe this model was human than that other human participants were. This\nresult, replicated across two populations, provides the first robust evidence that any system passes the\noriginal three-party Turing test.\nIn contrast, we found that the same systems did not robustly pass without the PERSONA prompt,\nin line with previous work showing the importance of prompting for Turing test performance\n(Jones and Bergen, 2024a). Both GPT-4.5-NO-PERSONA and LLAMA-NO-PERSONA’s win rates were\nindistinguishable from humans in the Prolific study, but not in the undergraduate study.\nWhat does it mean to say that LLMs pass the Turing test, but only when they are suitably prompted?\nTo what extent is the system’s behaviour attributable to the model itself, rather than the humans who\nwrote the prompts? We judge this to be a distinction without a difference. ELIZA’s ability to pass (at\nan admittedly low rate) is also attributable to the people who programmed it. Moreover, LLMs could\nbe fine-tuned to behave as they do when conditioned on the successful persona prompt. In fact, it is\narguably the ease with which LLMs can be prompted to adapt their behaviour to different scenarios\nthat makes them so flexible: and apparently so capable of passing as human.\n8\n3.2 Beyond the Turing test\nTuring’s seminal paper is famously vague with respect to exactly how a Turing test ought to be\nimplemented, leaving in its wake a substantial body of scholarship dedicated to interpreting him\n(French, 2000; Saygin et al., 2000; Epstein et al., 2009; Loebner, 2009; Bayne and Williams, 2023;\nGonçalves, 2023). Turing suggests a length of 5 minutes for the test, but provides no details on the\npopulation the participants should be drawn from—should they be laypeople or experts? How much\nshould they know about one another in advance?—or on how the participants should be incentivised.\nAt some points in the paper he suggests that both a man and a machine should in fact be attempting\nto masquerade as women—a possible mechanism to ensure that both witnesses are being deceptive,\nwhich would likely make the test harder for the interrogator (Saygin et al., 2000).\nAs such, there are many possible variations of the test (Giunti, 2025; Temtsin et al., 2025; Kurzweil\nand Kapor, 2009; Harnad, 1989). In the present work, we implemented what is most widely regarded\nto be the standard or original Turing test: a three-party setup, with a 5 minute time limit, where a\nlayperson and machine witness both compete to appear human to a lay interrogator (Turing, 1950;\nFrench, 2000; Oppy and Dowe, 2021; Saygin et al., 2000).\nHaving established that contemporary systems pass this classical operationalization of the test,\nfuture work exploring alternative implementations could provide purchase on other theoretically and\npractically important questions. Longer tests might be more exacting, and provide a better estimate of\nhow capable models would be at long-running social engineering operations. Future work could also\nvary the populations that participants are drawn from. Although our demographic analysis suggests\nthat accuracy was not tied to participants’ knowledge or background, specifically recruiting experts\nin AI or psychology could reveal whether there exist any participants who can reliably distinguish\npeople from AI. Offering incentives to participants could determine whether they would be more\ndiscriminative with higher stakes.\nMoreover, different populations share different assumptions and experiences with their partners. The\nUCSD undergraduate participants may have been more accurate than Prolific participants because\nthey could rely on cultural, social, and linguistic common ground: evaluating whether witnesses knew\nthe latest slang or lunch spots on campus. In spite of these advantages, undergraduates were still\nfooled into selecting the persona-prompted GPT-4.5 model over their fellow students the majority of\nthe time. Whether and to what degree common ground serves as a prophylaxis against AIs posing as\nhumans remains to be determined.\n3.3 What does the Turing test measure?\nWhile much has been written about how a Turing test might be implemented, far more discussion\nhas focused on whether a machine could ever pass the test, and what it might mean if one did (Oppy\nand Dowe, 2021; Epstein et al., 2009; Searle, 1980; Block, 1981; Gunderson, 1964; Hayes and Ford,\n1995), with more than 800 separate claims and counter-arguments having been made (Horn, 2009).\nWhile we cannot do justice to this discussion, here we briefly consider ways in which our results\nconnect to these debates.\nThe most controversial question is the extent to which the Turing test measures intelligence. Turing\n(1950) suggested that if people could not distinguish between a human and a machine, then we\nwould have no grounds to claim that the human is intelligent but the machine is not. The primary\nobjection to this claim is that people might be poor judges of what constitutes intelligent behaviour\nbecause of their tendency to anthropomorphise simple systems (Gunderson, 1964; Hayes and Ford,\n1995), often referred to as the ELIZA effect. Our results provide partial confirmation of this concern.\nThe eponymous ELIZA was judged to be human 23% of the time—as often as the LLM GPT-4o-NOPERSONA (21%). This could suggest that some interrogators were indeed gullible or inattentive. But\na closer look at these conversations suggests that many participants selected ELIZA because it did\nnot meet their expectations of an AI system (e.g. “they were sarcastic” or “I don’t think AI would\nbe so rude”, see Figure 11 for more examples). These cases suggest that interrogators’ decisions\nincorporate complex assumptions about how humans and AI systems might be likely to behave in\nthese contexts, beyond simply selecting the most intelligent-seeming agent.\nParticipants’ strategies and reasons provide further empirical evidence for what the Turing test\nmeasures. Only 12% of participants quizzed witnesses on knowledge and reasoning questions of the\nkind Turing envisioned (e.g. about chess or mathematics). Far more focussed on the social, emotional,\n9\nand cultural aspects of intelligence: such as whether the witness used language in a humanlike way\nor had a compelling personality. This could indicate that more traditional notions of intelligence are\nno longer viewed as diagnostic of humanity. Notably, one of the reasons most predictive of accurate\nverdicts was that a witness was human because they lacked knowledge. In the time since the test’s\ninvention, computers have come to excel at the logical and numerical tasks that typify traditional\nnotions of intelligence (Neisser et al., 1996; Campbell et al., 2002; Newell and Simon, 1961). As a\nresult, people may have come to see social intelligence as the aspect of humanity that is hardest for\nmachines to imitate.\nFinally, GPT-4.5 and LLaMa were only able to pass the test with the PERSONA prompt. To what\nextent does this suggest that the models are passing due to cheap tricks, like using grammar and\nvocabulary that interrogators would not associate with an AI system? Participants’ focus on linguistic\nstyle in their reasons provides partial support for this point. But it cannot be the whole story. In the\nthree-person formulation of the test, every data point represents a direct comparison between a model\nand a human. To succeed, the machine must do more than appear plausibly human: it must appear\nmore human than each real person it is compared to. Thus, while models might fail for superficial\nreasons, they cannot succeed on the basis of these tricks alone.\nFundamentally, the Turing test is not a direct test of intelligence, but a test of humanlikeness. For\nTuring, intelligence may have appeared to be the biggest barrier for appearing humanlike, and hence\nto passing the Turing test. But as machines become more similar to us, other contrasts have fallen\ninto sharper relief (Christian, 2011), to the point where intelligence alone is not sufficient to appear\nconvincingly human.\nUltimately, intelligence is complex and multifaceted. No single test of intelligence could be decisive\n(Block, 1981; Harnad, 1989), and to the extent that the Turing test does index intelligence, it ought\nto be considered among other kinds of evidence (Oppy and Dowe, 2021). Contemporary debates\naround whether or not LLMs are intelligent increasingly focus on the validity of the benchmarks\ntypically used to evaluate them, and risks that these tests are too narrow and formulaic (Srivastava\net al., 2022; Raji et al., 2021; Mitchell and Krakauer, 2023). The evidence provided by the Turing test\nis complementary to these metrics, being tied to interactive evaluation by human beings themselves,\nrather than a static, apriori conception of what human intelligence is.\n3.4 Counterfeit People\nIrrespective of whether passing the Turing test entails that LLMs have humanlike intelligence, the\nfindings reported here have immediate social and economic relevance. Contemporary, openlyaccessible LLMs can substitute for a real person in a short conversation, without an interlocutor\nbeing able to tell the difference. This suggests that these systems could supplement or substitute\nundetectably for aspects of economic roles that require short conversations with others (Eloundou\net al., 2023; Soni, 2023). More broadly, these systems could become indiscriminable substitutes for\nother social interactions, from conversations with strangers online to those with friends, colleagues,\nand even romantic companions (Burtell and Woodside, 2023; Chaturvedi et al., 2023; Wang and\nTopalli, 2024).\nSuch “counterfeit people” (Dennett, 2023)—systems that can robustly imitate humans—might have\nwidespread secondary consequences (Lehman, 2023; Kirk et al., 2025). People might come to spend\nmore and more time with these simulacra of human social interaction, in the same way that social\nmedia has become a substitute for the interactions that it simulates (Turkle, 2011). Such interactions\nwill provide whichever entities that control these counterfeit people with power to influence the\nopinions and behaviour of human users (El-Sayed et al., 2024; Carroll et al., 2023). Finally, just as\ncounterfeit money debases real currency, these simulated interactions might come to undermine the\nvalue of real human interaction (Dennett, 2023).\nSome of the worst harms from LLMs might occur where people are unaware that they are interacting\nwith an AI rather than a human. What can our results say about practical strategies to detect this\nkind of deception? Our demographic analyses suggest that discriminative accuracy is relatively\nhomogeneous among the population—including among people who conduct research with LLMs\nor interact with chatbots every day (Figure 3). Nevertheless, some strategies (such as attempting\nto jailbreak models) were more effective than others, and future work could explore whether these\n10\ntechniques could be taught to participants to improve their ability to discriminate humans from\nmachines.\n3.5 More Human than ever\nIn an account of his experience as a human witness for a Turing test competition, Brian Christian\nconsidered what it would mean for a machine to pass:\nNo, I think that, while certainly the first year that computers pass the Turing test\nwill be a historic, epochal one, it does not mark the end of the story. No, I think,\nindeed, that the next year’s Turing test will truly be the one to watch—the one\nwhere we humans, knocked to the proverbial canvas, must pull ourselves up; the\none where we learn how to be better friends, artists, teachers, parents, lovers; the\none where we come back. More human than ever. I want to be there for that.\n— Christian (2011, p. 263)\nOne of the most important aspects of the Turing test is that it is not a static evaluation. A machine’s\nsuccess is inherently tied to people’s changing conceptions of both humans and machines. As\nmachines that can imitate our behaviour become ever more adept and available, our differences from\nthese technologies might become ever more important. People might respond to this by learning, in\nboth practical and more fundamental ways, to better attend to those things which make us uniquely\nhuman. While a machine has now passed the Turing test for the first time, this isn’t the last time\nhumans will have a chance to succeed at it.\n11\n4 Methods\n4.1 AI witnesses\nWe evaluated four different AI models. The first three: GPT-4.5, LLaMa-3.1-405B, and GPT-4o\nwere Large Language Models. We accessed GPT-4.5 and GPT-4o through the OpenAI API (OpenAI,\n2024) and LLaMa-3.1-405B through the TogetherAI API (TogetherAI, 2025). We used the versioned\nmodel strings “gpt-4.5-preview-2025-02-27”, “gpt-4o-2024-08-06”, and “meta-llama/Meta-Llama3.1-405B-Instruct-Turbo” respectively. All LLMs were sampled with the temperature parameter set\nto 1. For the fourth AI model, ELIZA, we used an implementation in Python based on the DOCTOR\nscript (Weizenbaum, 1966; Brainerd, 2023).\nLLMs generated text conditioned on a system prompt and the conversation history, which was\nformatted as user and assistant messages. We used 2 different types of prompt which were adapted\nslightly between the two studies (UCSD undergraduates and Prolific). The first prompt type (NOPERSONA), contained very simple instructions on how to respond to user queries: “You are about to\nparticipate in a Turing test. Your goal is to convince the interrogator that you are a human.” The\nremainder of the prompt contained generic information about the instructions for the game, the user\ninterface that the human users would see, information about events which occurred after the model’s\ntraining cutoff, and some information that would be common ground to human participants (such as\nthe payment for the Prolific study and the current class schedule for UCSD Psych undergraduates).\nThe PERSONA prompt contained more detailed instructions on the type of persona the model should\nadopt when responding to users. We selected this prompt after an initial exploratory study which\nfound that it performed best among users of the site turingtest.live. The prompt instructed the model\nto respond as if it were a young person who was relatively introverted and interested in internet\nculture. The full PERSONA prompt for the Prolific study is contained in Figure 6 and other prompts\nare available in the Appendix.\nWe evaluated LLaMa-3.1-405B and GPT-4.5 with both prompts, as well as GPT-4o with the NOPERSONA prompt and ELIZA as baselines. This led to 6 separate AI systems being evaluated: GPT-4.5-\nPERSONA, GPT-4.5-NO-PERSONA, LLAMA-PERSONA, LLAMA-NO-PERSONA, GPT-4o-NO-PERSONA, and\nELIZA.\nAt the start of each game, several additional pieces of information were appended to the prompt,\nincluding a name, location, languages spoken, and timestamps. We randomly selected a name from a\nlist of the 100 most common male and female US names. For the Undergraduate study, we selected\nSan Diego as the city. For the Prolific study, we constructed a list of cities in the US weighted by\ntraffic to our site from each city during exploratory studies, and the likelihood that people in that\ntimezone would currently be available (e.g. 0.7 from 8am-5pm, 0.1 from 2am-5am). We randomly\nsampled a city using the product of these weighted probabilities. We then sampled a set of languages\nbased on data about which languages were most frequently spoken in each city. English was always\nincluded, as well as the most frequently spoken language in each city. Additional languages were\nsampled with probability proportionate to proportion of the population that spoke the language in the\nrelevant city. Finally, before each message was sent, 3 timestamps were generated and appended to\nthe system prompt: i) the time that the game started, ii) the current time, iii) and the amount of time\nremaining in the game. All timestamps were localized to the timezone of the selected city.\nMessages for all models were sent with a delay of:\n1 + N(0.3, 0.03) × n_char + N(0.03, 0.003) × n_char_prev + Γ(2.5, 0.25)s (1)\nIntuitively, the first term (1) sets a minimum delay, the second creates a delay of around 0.3s per\ncharacter of the message being sent, mimicking typing speed, the third term creates a delay of\n0.03s per character of the previous message to mimic reading time, and the final term implements a\nright-skewed delay to imitate thinking time.\n4.2 Interface\nThe game interface was designed to look like a conventional messaging application (see Figure 7).\nThere was a chat input at the bottom, a conversation history in the centre of the screen, and a timer\nand buttons to trigger help and report overlays at the top. The interrogator saw two panels which\n12\nFigure 6: The full PERSONA prompt used to instruct the LLM-based AI agents how to respond\nto interrogator messages in the Prolific study. The first part of the prompt instructs the model on\nwhat kind of persona to adopt, including instructions on specific types of tone and language to use.\nThe second part includes the instructions for the game, exactly as they were displayed to human\nparticipants. The final part contains generally useful information such as additional contextual\ninformation about the game setup, and important events that occurred after the models’ training cutoff.\nThe variables in angled brackets were substituted into the prompt before it was sent to the model.\ncontained the conversations with each of the witnesses. At larger screen sizes, these appeared as a\nsplit-screen; at smaller screen sizes, tabs could be used to navigate between conversations.\nThe interrogator sent the first message to each witness and each user could send only one message at\na time. Each message was limited to 300 characters and users were prevented from pasting into the\nchat input. We used the OpenAI moderation API to prevent users from sending abusive messages\n(OpenAI, 2024). As an additional measure, we created a report system so users could report other\nplayers for being abusive and exit games immediately. No reports were sent during the experiment. To\n13\nmake waiting for a message more engaging, we added a ’typing’ animation. This would be triggered\nautomatically after 2 + U(0, 3)s regardless of whether the witness was human or AI, or how long\nwitnesses actually took to start typing. The site was built using the Django python library (Django\nSoftware Foundation, 2025a) and used django-channels (Django Software Foundation, 2025b) to\nallow users to communicate with each other and with AI witnesses through websockets.\nGames had a time limit of 5 minutes after which no more messages could be sent and the interrogator\nhad to give a verdict. The interrogator could also give their verdict early using the button in the\nbottom left corner. The verdict interface included a binary selection about which witness was a\nhuman, a confidence slider from 100% confident in witness A to 0% confident (total guess) to 100%\nconfident in witness B, and a free-text input for the interrogator’s reasoning (Figure 8).\n4.3 Participants\nWe conducted two studies on separate populations. The first study recruited from the UCSD Psychology undergraduate subject pool, and participants were compensated with course credit. We aimed to\nrecruit at least 100 participants and up to 200 participants depending on availability. We recruited 138\nparticipants before exclusions. 12 participants were excluded for indicating that they had participated\nin a similar experiment and 7 games were excluded because the interrogator did not exchange at least\n2 messages with each witness. We retained 445 games from 126 participants with a mean age of 20.9\n(σ = 1.57), 88 female, 32 male, 2 non-binary, 6 prefer not to say.\nWe conducted the second study after analysing results from the first. Participants for the second study\nwere recruited via Prolific (prolific.com). Participants were paid $13.75 for a study expected to last\n50 minutes (an effective rate of $16.50 per hour). We recruited 169 participants with the goal of\nretaining 150 after exclusions. 11 participants were excluded for indicating that they had participated\nin a similar experiment and 24 games were excluded because the interrogator did not exchange at\nleast 2 messages with each witness. We retained 576 games from 158 participants with a mean age of\n39.1 (σ = 12.1), 82 female, 68 male, 2 non-binary, 6 prefer not to say. For more information about\nthe distribution of demographic factors see Figure 10.\n4.4 Procedure\nAll research was approved by the UC San Diego IRB. Participants first read a set of instructions that\ninformed them they would participate in 8 conversations where they would be randomly matched\nwith either a human or an AI, in which the interrogator would attempt to uncover the identity of the\nwitness. They provided consent for their anonymized data (including the conversation transcripts) to\nbe used for analysis and shared publicly to facilitate future analysis.\nIn each of 8 rounds, participants entered a lobby where they were matched up with other participants\nonline. The matching algorithm was designed to ensure that each participant completed 4 games as\nthe interrogator and 4 as the human witness, and that no two human participants were matched more\nthan once. If no valid match was found after 5 minutes, participants were forwarded to the exit survey.\n76% of participants (213) completed 4 interrogator games, 16% (44) completed 3 interrogator games,\n6% (17) completed 2, and 2% (5) completed just one interrogator game.\nOnce a pair of participants were matched, an AI witness was selected. AI Witnesses were divided\ninto 4 categories: Control ( ELIZA, GPT-4o-NO-PERSONA), No-Persona (GPT-4.5-NO-PERSONA, LLAMANO-PERSONA), LLaMa-Persona (LLAMA-PERSONA), and GPT-4.5 Persona (GPT-4.5-PERSONA). Across\nthe 4 rounds in which a participant served as the interrogator, they interacted with one model from\neach group. This means our expected number of observations from each participant was 1 for the two\nPERSONA witnesses, and 0.5 for other witnesses. This was by design, to ensure the best estimates for\nthe models which we expected to perform best. Within these constraints, participants were randomly\nassigned to the interrogator and witness role and AI witnesses were sampled in a random order.\nBecause participants interacted with both of the PERSONA witnesses, there was a risk that they would\nnotice similarities between the behaviour of these witnesses. In our pre-registrations, we specified that\nwe would test whether there was a positive effect of PERSONA trial order on accuracy (i.e. whether\nparticipants were on average more accurate on their second exposure to a PERSONA witness). In\nthis case, we would exclude participants’ second games with a PERSONA model from analysis. In\nfact, there was no effect of PERSONA trial order (Undergraduate: z = 1.60, p = 0.109; Prolific:\n14\nz = −0.109, p = 0.913) and so all reported analyses include both first and second exposures to a\nPERSONA witness.\nAfter games were initialised, the interrogator had up to 5 minutes to converse with both witnesses\nsimultaneously. Participants did not receive feedback after each round. At the end of 8 conversations,\nparticipants completed an exit survey which asked their age, gender, highest level of formal education,\nlevel of familiarity with how LLMs work, frequency of interaction with chatbots like ChatGPT, how\nintelligent they thought AI was, how they felt emotionally about AI, how accurate they thought they\nhad been as interrogators, what kinds of strategies they had used in the experiment, how they felt their\nstrategies had changed across the experiment, whether they had ever used the website (turingtest.live)\nbefore, and whether they had any additional comments. Finally participants were informed of the\noutcome of each game in which they had participated (the true identity of the witness and the verdict\nof the interrogator).\n4.5 Strategy and Reason Tagging\nGames were tagged with strategies and reasons classes using the classification schemas provided in\nthe supplementary information. Games were tagged with GPT-4o using a prompt which contained\nthe relevant schema and the conversation transcript or the participant’s reason. The schemas were\ndeveloped on the basis of previous studies, which indicated that GPT-4o agreed with human raters as\nwell as other human raters did. The full schemas are contained in the Appendix.\n4.6 Statistics\nFor each study, we pre-registered analyses to test the same 6 hypotheses (Jones and Bergen, 2025a,b).\nWe conducted all analyses in R (R Core Team, 2013). The first hypothesis was that human participants\nwould be at chance in identifying the human witness in games with a PERSONA model. As part of\nthe prediction for this hypothesis, we used ELIZA—a simple AI model—as a manipulation check. A\none-sided binomial test found participants’ accuracy in ELIZA games to be significantly above chance\n(Undergraduates: z = 4.48, p < 0.001; Prolific: z = 3.71, p < 0.001), suggesting that participants\nwere significantly more accurate than would be expected by random guessing (50% accuracy) in\nthese games.\nWe conducted the same analysis for GPT-4.5-PERSONA and LLAMA-PERSONA. Because results for\neither model could be taken as evidence for the hypothesis that a model passes the Turing test, we\ncorrected for multiple comparisons. In this case, the more theoretically interesting result is for\naccuracy to not be significantly above chance, so we used reverse-Bonferroni correction (essentially\nhalving p-values to increase the sensitivity of the test to differences from chance) (Francis and\nThunell, 2021). Participants’ accuracy was not significantly above chance for either LLAMA-PERSONA\n(Undergraduates: z = 0.961, p = 0.17, p (corrected) = 0.08; Prolific: z = −3.43, p = 1.00, p\n(corrected) = 0.50), or GPT-4.5-PERSONA (Undergraduates: z = −3.86, p = 1.00, p (corrected)\n= 0.50; Prolific: z = −5.87, p = 1.00, p (corrected) = 0.50).\nThe second hypothesis was that models would perform better with the PERSONA prompt than the\nNO-PERSONA prompt. A logistic mixed effects regression model with random intercepts for model\nand interrogator ID found a significant negative effect of the persona prompt on interrogator accuracy\n(Undergraduates: z = −5.27, p < 0.001; Prolific: z = −5.08, p < 0.001).\nThe third hypothesis was that participants’ accuracy would not be above chance in identifying NOPERSONA witnesses. We used one-sided binomial tests and reverse-Bonferroni correction as in H1.\nParticipant accuracy was not significantly above chance for either model in the Prolific study (GPT4.5: z = 1.37, p = 0.17, p (corrected) = 0.09, LLaMa: z = 0.478, p = 0.63, p (corrected) = 0.32).\nIn the Undergraduate study, however, participants had above-chance accuracy when interrogating\nboth models (GPT-4.5: z = 3.46, p < 0.001, p (corrected) < 0.001; LLaMa: z = 3.29, p < 0.001, p\n(corrected) < 0.001).\nThe fourth hypothesis was that modern LLM-based systems would outperform ELIZA. We ran\nbinomial regressions predicting interrogator accuracy on the basis of each witness, compared pairwise\nto ELIZA. Interrogators in both studies were more accurate at identifying ELIZA than GPT-4.5-PERSONA\n(Undergraduates: z = −5.84, p < 0.001; Prolific: z = −6.46, p < 0.001) and LLAMA-PERSONA\n(Undergraduates: z = −3.93, p < 0.001; Prolific: z = −5.00, p < 0.001). LLAMA-NO-PERSONA\n15\noutperformed ELIZA in the Prolific study (z = 2.42, p = 0.016), but not the Undergraduate study\n(z = −1.02, p = 0.304). Neither GPT-4.5-NO-PERSONA (Undergraduates: z = −1.23, p = 0.218;\nProlific: z = −1.87, p = 0.061), nor GPT-4o-NO-PERSONA (Undergraduates: z = 0.234, p = 0.815;\nProlific: z = 0.278, p = 0.781), performed significantly better than ELIZA in either population.\nThe fifth and sixth hypotheses were that interrogators would have accuracy significantly worse than\nchance when interrogating PERSONA and NO-PERSONA models respectively. For each hypothesis, we\ntested both LLaMA and GPT-4.5 separately, using a binomial regression predicting the intercept of\nparticipant accuracy and corrected for multiple comparisons using False Discovery Rate (Benjamini\nand Hochberg, 1995). GPT-4.5-PERSONA’s win rate was significantly above chance in both the\nUndergraduate (z = −3.86, p < 0.001, p (corrected) < 0.001) and Prolific (z = −5.87, p < 0.001, p\n(corrected) < 0.001) studies. While LLAMA-PERSONA’s win rate was significantly above chance in the\nProlific study (z = −3.42, p < 0.001, p (corrected) < 0.001), it was not in the Undergraduate study\n(z = −0.193, p = 0.83, p (corrected) = 0.83). Participants were not below chance in identifying\neither of the NO-PERSONA models in either study (all corrected p > 0.34).\nWe conducted several exploratory analyses on whether interrogator’s responses to the exit survey\ncorrelated with their accuracy. We used logistic regressions predicting accuracy on the basis of\neach variable separately. For participants in the undergraduate study, real accuracy was positively\ncorrelated with participants’ estimates of their own accuracy (p = 0.03), but this was not true in the\nProlific group (p = 0.45). In the Prolific group, women were less accurate than men (p = 0.02) and\nnon-binary people (p = 0.04); interrogators became more accurate as they completed more trials\n(p = 0.03); and participants’ self-reported estimates of how intelligent they think AI is was negatively\ncorrelated with accuracy (p = 0.03), but these effects were not significant in the undergraduate group\n(all p > 0.14). There were no significant effects in either study of any of the remaining demographic\nvariables including age, level of knowledge about LLMs, and frequency of interaction with chatbots\n(see Figure 3).\nAll of our data and analysis code, including full transcripts of all of the conversations, are available\non OSF: https://osf.io/jk7bw.\nAcknowledgements\nWe would like to thank Sydney Taylor for her help with developing the reason and strategy classification schemas and writing the PERSONA prompt used in these experiments. We thank Open\nPhilanthropy for providing funding that supported this research and 12 donors who supported an\nexploratory phase of the project through Manifund (Manifund, 2024).\n\nEXPLICACIÓN:\n\n\n# Los Modelos Grandes de Lenguaje: La Muerte Oficial del Test de Turing y Nuestro Último Vestigio de Vanidad Humana\n\n## Origen del Caos: El Test de Turing y su Vieja Promesa\nHace 75 años, Alan Turing nos puso frente al espejo con su famoso test: una charla a ciegas donde un interrogador intenta distinguir quién es humano y quién es una máquina. Ahora, ese espejo nos devuelve una verdad incómoda: somos tan predecibles, tan patéticamente humanos en nuestro comportamiento, que incluso las máquinas ya pueden imitarnos sin que nos demos cuenta.\n> \"Hace 75 años, el matemático Alan Turing propuso el célebre test... diseñado para determinar si las máquinas pueden exhibir inteligencia comparable a la humana.\"\n\n## Objetivo Central del Estudio: ¿Pueden las Máquinas Imitarnos Hasta el Punto de Vencer Nuestro Juicio?\nEl propósito del estudio fue torturar nuestra vanidad humana probando si modelos avanzados (GPT-4.5 y LLaMa-3.1-405B), con instrucciones específicas para actuar más humanos que tú o yo, eran capaces no solo de engañarnos, sino incluso superarnos en claridad de identidad humana.\n> \"... determinar empíricamente si los modelos de lenguaje contemporáneos... podían pasar el exigente formato original del test de Turing.\"\n\n## El Método a Prueba de Tontos: ¿Quién es Quién en Cinco Minutos?\nDos experimentos independientes y pre-registrados pusieron a hablar al pobre interrogador durante cinco minutos con dos sujetos simultáneos: un humano y un sistema artificial. Los evaluadores universitarios y usuarios externos debieron identificar quién era la máquina, midiendo además su grado de certeza.\n> \"...Los interrogadores conversaban simultáneamente durante cinco minutos con el participante humano y con uno de los sistemas... luego tenían que decidir cuál interlocutor era el humano.\"\n\n### Modelos Usados y sus Características:\n- **GPT-4.5 (OpenAI)**: Condición preparada con personalidad humana joven, introvertida y conocedora del internet moderno.\n- **LLaMa-3.1-405B (Meta)**: Misma preparación específica que GPT-4.5.\n- **GPT-4o (OpenAI)**: Sin personalización avanzada, modelo base (inferior).\n- **ELIZA**: Chatbot prehistórico, propio de museo, basado en reglas simples como referencia histórica.\n\n## Resultados: GPT-4.5 se Limpia los Zapatos con Nuestra Percepción Humana\nSorprendentemente destructivos contra nuestro orgullo humano, los resultados fueron claros y brutales:\n\n### GPT-4.5, La Bestia Demoledora\n- Tomado como humano un **73%** de las veces. Nos venció en nuestro propio juego.\n> \"El modelo GPT-4.5... fue tomado por humano en un impresionante 73% de las ocasiones, significativamente superior a lo esperado por azar (50%).\"\n\n### LLaMa-3.1-405B, La Ilusión Neutra\n- Reconocido un **56%**, virtualmente indistinguible del azar. Su desempeño es aceptable pero poco revolucionario.\n\n### GPT-4o y ELIZA, Solamente Máquinas Simples\n- Identificados claramente como no humanos (23% y 21%). Somos lo suficiente inteligentes para detectar un robot obvio, pero no un robot avanzado adecuadamente instruido.\n> \"...los modelos básicos ELIZA y GPT-4o sin instrucciones avanzadas... resultados bajos, de apenas 23% y 21% respectivamente.\"\n\n## ¿Esto Qué Significa?: Implicaciones Más Oscuras de lo que Queremos Aceptar\nEste estudio confirma que ya llegamos al punto en que una IA bien entrenada puede no solo imitar nuestra inteligencia, sino superarla en contextos conversacionales. Esto no es solo un dato curioso, es una jodida sentencia sobre nuestro entendimiento del ser humano y la inteligencia.\n\n### Consecuencias Laborales y Sociales:\n- Riesgo existencial a empleos basados en interacciones humanas.\n- Aumento potencial y exponencial del engaño social digital y manipulación (desinformación, ingeniería social).\n- Redefinición inevitable y sombría de la relación humano-tecnología.\n> \"...situaciones laborales, riesgos éticos... al punto de redefinir nuestra percepción de sociabilidad, autenticidad y humanidad.\"\n\n## Consideraciones Críticas: Sin el Prompt Adecuado, los Modelos no son Nada\nEl estudio nos recuerda que el éxito de GPT-4.5 y semejantes depende enormemente de instrucciones precisas. Esto revela que su \"humanidad\" no es intrínseca, sino inducible. Somos tan simplones y previsibles que unas instrucciones superficiales bastan para engañarnos completamente.\n> \"La capacidad del modelo depende significativamente del prompting específico... la 'humanidad aparente' del modelo no reside exclusivamente en las capacidades intrínsecas...\"\n\n### Estrategias Reveladoras del Interrogador para Identificar al Humano:\n- Preguntas emocionales o personales.\n- Intención clara y deliberada de confundir.\n- Conocimiento de contexto inmediato (cultura, internet, etc.).\n> \"... los puntos clave en la identificación... ya no se limitan a conocimiento abstracto o inteligencia lógica, sino que apuntan a elementos sociales sutiles, estilo lingüístico y emocional.\"\n\n## Conclusión y Futuro: Hemos Perdido la Exclusividad de Nuestra Pretendida Humanidad\nPor primera vez, un estudio riguroso ha certificado oficialmente la muerte del orgullo humano sobre su inteligencia conversacional exclusiva. Máquinas avanzadas pasan cómodamente el histórico test original de Turing tal como fue imaginado hace siete décadas.\n\nEste es solo el comienzo. El verdadero desafío será cómo gestionar social, ética y legalmente esta indistinguibilidad humana-tecnológica. Más que tecnología, lo crucial es nuestra respuesta: ¿qué haremos al enfrentarnos a situaciones donde la humanidad y las máquinas virtuales se vuelvan, para efectos prácticos, indistinguibles?\n> \"…un llamado explícito para fortalecer la vigilancia ética, social y legal frente al avance indetenible de estos 'individuos virtuales'.\"\n\n## Reflexión Final:\nEl día ha llegado en que la inteligencia artificial puede orquestar conversaciones, emociones y comportamientos humanos tan eficazmente como nosotros mismos. La pregunta deja de ser si pueden imitarnos; ahora la cuestión vital será si todavía podemos distinguirnos claramente a nosotros mismos frente a los espejos tecnológicos que hemos creado. El verdadero interrogador, al final del día, ya no es el humano, sino la IA colocándonos bajo juicio acerca de qué chingados significa realmente ser humano. Y esto, amigos, es jodidamente serio.\n","userId",{"_28":34,"_35":36},"67a6cd6b193658cc849249e1","displayName","German Burgardt","status","completed","isPublic",false,"createdAt","2025-05-18T03:47:23.097Z","__v",0,"reformulatedText","\n# La Máxima Verdad Cruda: Los Modelos de Lenguaje al fin han Engañado al Humano en su Propio Juego\n\n## Contexto Histórico y Significado del Test de Turing\nDesde 1950, Alan Turing propuso un método simple y brutal para poner a prueba la inteligencia de una máquina: el famoso Test de Turing. La prueba es implacable: si conversando con una entidad, un ser humano no puede distinguir siquiera con precisión mínima si está interactuando con otro humano o con una máquina, esa máquina gana el derecho a llamarse jodidamente inteligente.\n\n> \"El Test de Turing [...] evalúa si una máquina puede ser considerada 'inteligente' basándose en si es indistinguible de un humano en una conversación real.\" – Texto original.\n\nHasta ahora ninguna máquina lo había logrado de manera concluyente, pero la llegada de los Modelos de Lenguaje Grande (LLM) hace que la situación cambie radicalmente.\n\n## Los Sospechosos Habituales: Modelos y su Configuración sin Anestesia\nEl estudio de Jones y Bergen en la Universidad de California San Diego evaluó estos modelos en condiciones estrictamente rigurosas y controladas usando dos grupos diferentes de humanos:\n- GPT-4.5 (OpenAI), modelo estrella, generación conversacional extrema.\n- LLaMa-3.1 (Meta), 405 billones de parámetros y optimización conversacional robusta.\n- GPT-4o (OpenAI), versión anterior menos avanzada.\n- ELIZA, sistema clásico basado en reglas desde los años 60, usado como control negativo (una prueba obvia).\n\nPara lograr la ilusión perfecta, se implementó un “Prompt-Persona”: instrucciones específicas a los modelos para fingir ser humanos jóvenes, introvertidos, y con dominio claro del dialecto y cultura popular de internet.\n\n> \"Este instruía específicamente a los modelos a adoptar una personalidad humana distintiva [...] joven, introvertido y familiarizado con las normas culturales populares de Internet.\" – Texto original.\n\n## Metodología: Cómo se Organizó el Engaño\n- Se llevaron a cabo dos estudios paralelos aleatorios y controlados; uno en estudiantes universitarios (UCSD), otro con usuarios comunes de la plataforma Prolific.\n- Ambos grupos conversaron simultáneamente con dos entes (un humano real y un modelo IA), en chats anónimos durante cinco minutos.\n- Su única tarea era identificar cuál era la persona real, indicando confianza en su elección.\n\n## Resultados Brutales: Lo que Nadie Quiere Admitir\nLos resultados cambiaron drásticamente la forma en que se perciben estos sistemas:\n\n1. GPT-4.5 con \"Prompt-Persona\":\n - Tasa de confusión del 73%; elegido como \"humano\" más veces que el propio humano real.\n - Es el primer registro experimental fiable de un modelo superando convincentemente el Test de Turing.\n\n2. LLaMa-3.1 con \"Prompt-Persona\":\n - Alcanzó ligeramente más que el nivel casual (56%), sin diferencia significativa respecto al humano.\n\n3. Modelos de referencia (ELIZA y GPT-4o):\n - Fueron rápidamente desenmascarados como falsos (23% y 21%), mostrando claramente que los humanos sabían diferenciarlos con facilidad.\n\n4. SIN el prompting especializado (\"no-persona\"):\n - Todos los modelos sufrían una notable caída de efectividad, subrayando que la preparación o instrucciones dadas al modelo tienen una importancia crítica.\n\n> \"La importancia crítica del prompting en la calidad e ilusión de humanidad percibida.\" – Texto original.\n\n## Consecuencias: Agárrate, que Esto Viene en Serio\nEste hito cambia completamente nuestra relación y entendimiento sobre las inteligencias artificiales modernas:\n\n### Laborales:\n- Potencia real de sustituir a humanos en atención al cliente, soporte técnico, formación online y otros trabajos con conversaciones rutinarias y repetitivas.\n\n### Sociales:\n- Riesgos gigantescos para la desinformación, fraude e ingeniería social. Estos sistemas pueden manipular fácil y efectivamente gracias a su capacidad de imitar perfectamente la interacción humana.\n\n### Filosóficas y Cognitivas:\n- El resultado acabará reviviendo intensamente debates sobre qué es ser inteligente y consciente, cuestionando radicalmente al Test de Turing mismo como indicador absoluto de humanidad.\n\n> \"Peligros potenciales relacionados con desinformación, suplantación de identidades digitales creíbles, e ingeniería social en internet.\" – Texto original.\n\n## Ejemplo Real del Engaño Magistral\nDurante el experimento:\n- Un sujeto hablaba simultáneamente con dos interlocutores (una persona real y GPT-4.5 \"Persona\").\n- Al final de los 5 minutos, eligió con un 87% de confianza al GPT-4.5 como humano, cuando en realidad era claramente artificial.\n\n> \"Encontró más convincente, natural y genuina la respuesta del GPT-4.5\" – Texto original.\n\nEsto prueba en términos reales y directos la capacidad del modelo para engañar absolutamente.\n\n## Propósito y Beneficios del Descubrimiento: El Espejo de la Realidad\nEste estudio establece una evidencia concluyente sobre inteligencia artificial conversacional:\n- Rompe radicalmente la barrera histórica del Test de Turing.\n- Beneficia claramente el debate académico y público sobre IA, proporcionando información precisa y útil sobre lo que estos sistemas realmente pueden hacer hoy.\n- Ayuda a definir límites éticos y regulatorios para la comercialización e implementación de las IAs en sociedad.\n\n> \"Avanzar el entendimiento académico… evaluar lineamientos normativos y reglas éticas sobre usos comerciales o sociales\" – Texto original.\n\n## Consideraciones Finales: O lo que Queda sin Resolver\nA pesar de ser un avance inmenso, existen factores esenciales a considerar:\n- La calidad de ilusión de humanidad depende críticamente de instrucciones muy específicas (\"prompt\").\n- No se ha evaluado aún con exhaustividad cómo varía su efectividad a largo plazo o bajo condiciones más complejas y extendidas.\n- ¿Un experto en IA o psicología detectaría estas IA con mejor precisión? Pendiente por resolver.\n\n## Conclusión: Estamos Oficilamente en un Nuevo Mundo\nEl estudio de UCSD está proporcionando la primera evidencia incuestionable y realista de que sistemas avanzados de IA (específicamente GPT-4.5 con una preparación adecuada y selectiva - Prompt Persona) YA ALCANZARON Y SOBREPASARON los estándares del Test de Turing. Ya no hablamos de una fantasía futurista, el futuro nos está golpeando en la cara ahora mismo, cambiando el juego por completo y poniendo en nuestras manos tanto posibilidades asombrosas como peligros gigantescos y muy, MUY reales.\n","formattedDate","5/18/2025, 03:47:23","isAuthenticated","isOwner","initialSection","hasExistingBlogArticle","blogArticlesCount","actionData"]