Stimulus Equivalence in AI-Generated Anime Characters: Equivalence Responding Among Voices, Images, and Names

doi:10.21203/rs.3.rs-8380205/v1

Download PDF

Research Article

Stimulus Equivalence in AI-Generated Anime Characters: Equivalence Responding Among Voices, Images, and Names

https://doi.org/10.21203/rs.3.rs-8380205/v1

This work is licensed under a CC BY 4.0 License

Version 1

posted

You are reading this latest preprint version

This study analyzed how people identify AI-generated anime characters within a stimulus equivalence framework. Voices, faces, and names for three characters were treated as members of three-member stimulus classes, and the study examined whether untrained relations between faces and names would emerge after arbitrary matching-to-sample (MTS) training from voices to faces and from voices to names.

Unlike previous auditory equivalence studies that relied on the lexical content of spoken words, the present study defined auditory classes by acoustic properties of the voice. Text-to-speech voices derived from real speakers were used, and multiple lines of dialogue per character were generated. Thirteen Japanese undergraduates completed the study online; data from nine participants who completed the procedure without major protocol deviations were analyzed. To minimize pre-experimental stereotypes while retaining meaningful stimuli, “correct” voice–face–name combinations were determined individually for each participant by algorithmically selecting the mapping that was least consistent with their pretest choices. Seven of the nine participants showed increased accuracy on untrained Illustration-to-Name and Name-to-Illustration tests only after both Voice-to-Illustration and Voice-to-Name training, indicating the emergence of stimulus equivalence between faces and names. Two participants did not show equivalence despite additional training, suggesting individual differences in the extent to which acoustic features support conditional stimulus control. These findings extend the analysis of stimulus equivalence to AI-generated anime characters and illustrate a methodological approach for studying the formation of arbitrary relations with meaningful, stereotype-prone stimuli.

Psychology

character identification

AI-generated voices

stimulus equivalence

matching-to-sample

How do people identify other individuals? Research on person identification (Clifford & Bull, 1978) has mainly used voices, or combinations of voices and faces, as stimuli and has been conducted across several disciplines, including psychology, linguistics, and biology. Much of this work has aimed to identify variables that affect person identification and has relied on test-only procedures without explicit training. For example, Kamachi et al. (2003) asked their participants to select a face after hearing a voice, or to select a voice after seeing a face. They showed that accuracy in judging voice–face identity can exceed chance level even without explicit matching training. This kind of effect is observed not only with dynamic video stimuli but also with static images (Smith et al., 2016), and substantial individual differences in the accuracy of voice–face matching have also been reported (Kramer et al., 2021). Furthermore, Smith et al. (2025) suggested that the way in which multimodal (audio-visual) stimuli are presented, including the direction of matching, may affect the accuracy of identification responses.

My interest lies in how viewers of animated works come to identify fictional characters as they follow a story, such that they can name the characters when they see their faces or hear their voices. I operationally defined identification of anime characters as the acquisition of equivalence relations among voices, faces, and names, and examined whether stimulus equivalence would emerge following matching-to-sample (MTS) training. To my knowledge, no prior research has analyzed the process by which person identification is established within the framework of stimulus equivalence. This constitutes a primary source of originality for the present study.

In a three-member stimulus class, if two conditional relations are directly trained (e.g., A1→B1 and A1→C1), and untrained relations such as B1→C1 and C1→B1 subsequently emerge, the set of stimuli is said to form an equivalence class (Sidman & Tailby, 1982). Numerous studies have examined stimulus equivalence using auditory stimuli (e.g., Cowley et al., 1992; Dube et al., 1993; Stewart & Lavelle, 2013). Cowley et al. (1992) worked with adults with brain injuries who had severe memory deficits and difficulty matching their therapists’ names to their faces. Using a conditional discrimination procedure, they trained dictated therapist names as samples and color photographs of therapists’ faces as comparisons, and then tested for emergent relations between faces and written names and vice versa. Training a single set of dictated-name–to-face relations resulted in the emergence of untrained face–to–written-name and written-name–to–face relations, indicating the formation of equivalence classes that included dictated names, facial photographs, and written names for each therapist. Dube et al. (1993) extended the analysis to classes consisting entirely of auditory stimuli by using a successive conditional discrimination procedure with arbitrary spoken syllables (e.g., “cug,” “vek”). Participants first learned baseline relations among sets of syllables (e.g., A-B and B-C relations), and were then tested for reflexivity, symmetry, and transitivity. Most participants formed three-member equivalence classes based solely on these sample–comparison relations among nonsense syllables, and some subsequently showed expansion to four-member classes after additional training. Together, these studies demonstrate that stimulus equivalence procedures can establish novel conditional relations among names, faces, and arbitrary sounds, and that equivalence classes can be constructed entirely from auditory stimuli.

Stewart and Lavelle (2013) further extended this line of research by examining how non-arbitrary auditory relations, specifically matching on the basis of voice, affect the formation of auditory equivalence classes. Undergraduate participants were trained and tested for the formation of three three-member (A–B–C) equivalence classes using recorded nonsense syllables spoken in different voices (adult male, adult female, and a female child). Participants were randomly assigned to one of three conditions. In the Same Voice (SV) condition, all stimuli in both training and testing were presented in the same child’s voice. In the Different Voices (DV) condition, training and testing involved different voices, but voice was irrelevant to the programmed contingencies. In the Different Voices Test (DVT) condition, all training was conducted in the same voice, but during testing the correct comparison stimulus was always spoken in a different voice from the sample, whereas one incorrect comparison was spoken in the same voice as the sample, creating a conflict between equivalence-based and voice-matching relations. Equivalence responding was highest in the SV condition, intermediate in the DV condition, and lowest in the DVT condition, in which no participant passed the equivalence test; in contrast, non-arbitrary voice-matching responses were most frequent in the DVT condition. These findings suggest that non-arbitrary auditory relations based on voice characteristics can interfere with the emergence of derived equivalence relations when they conflict with the arbitrary relations programmed by the experimenter.

Across these studies, however, the critical units for equivalence were the verbal identities of the names or syllables; acoustic properties of the voice were either held constant or manipulated as an irrelevant or competing dimension. In contrast, the present study defined stimulus classes not by the linguistic content of the utterance but by acoustic features of the speaker’s voice—such as pitch, timbre, speech rate, and intonation (prosodic and spectral characteristics). Figure 1 presents a schematic illustration of the present procedure. In the Voice-to-Illustration (VtoI) task, participants heard an utterance and selected a face illustration based on the acoustic characteristics of the voice, not the semantic content of the line. In the Voice-to-Name (VtoN) task, they selected a printed name based on the same acoustic characteristics. After MTS training on these two relations, test tasks were administered in which participants selected a name given a face (Illustration-to-Name [ItoN]) or selected a face given a name (Name-to-Illustration [NtoI]). If correct responding in these untrained tests emerged without direct training, stimulus equivalence between faces and names was considered to be established.

To construct the voice stimuli, I used a commercial text-to-speech service (CoeFont; https://coefont.cloud), which provides synthetic voices based on recordings of real speakers, and generated all dialogue lines with that service. To ensure that the linguistic content of the utterances would not serve as a cue in the MTS tasks, each speaker produced multiple, non-overlapping lines. I also used large language model–based generative AI tools to create the anime-style face illustrations and to generate and edit the lines of dialogue. Throughout the paper, these synthetic speakers are referred to as "AI characters."

In a preliminary study, some participants showed high levels of correct responding even before any training, despite the experimenter’s arbitrary assignment of face, voice, and name combinations. This pattern is consistent with the findings of Kamachi et al. (2003) and suggests that stereotypical response tendencies may have been present in the sample of Japanese university students. For example, participants may have assumed that a female character with long hair and narrow eyes would speak quietly, whereas a character with short hair and wide-open eyes would speak more energetically. When such response biases are present prior to training, accuracy on tests used to evaluate stimulus equivalence can be elevated even before MTS training begins, making it difficult to evaluate training effects. Sidman (1994) illustrated how previously acquired repertoires can affect experimental outcomes even with young children and argued that such influences should be minimized either by (a) measuring participants’ responses to candidate stimuli in advance and selecting those that do not evoke strongly biased responses or (b) using largely meaningless stimuli that are assumed to be minimally affected by prior histories. Bush et al. (1989) and Carr and Felce (2008) proposed related strategies, such as randomizing comparison positions, increasing the number of choices to lower the chance level, and restricting training to participants whose pretest accuracy is at or below chance. These strategies, however, may not be sufficient when multiple stereotype patterns are present and their effects are strong.

To address this problem, I did not fix the “correct” combinations of voice, face, and name before the experiment. Instead, for each participant, I first conducted a pretest and recorded the frequency with which each combination of voice, face, and name was selected. I then algorithmically determined, for that participant, the combination that was most difficult to select correctly (i.e., the combination that yielded the lowest accuracy given the participant’s pretest response pattern) and defined that combination as the “correct” relation in the subsequent training. This procedure is intended to minimize the influence of preexisting stereotypical response tendencies while still using meaningful stimuli to study the formation of conditional relations.

In summary, the present study had two main aims. First, I examined whether stimulus equivalence between faces and names would emerge when the voices, faces, and names of AI-generated characters were treated as members of stimulus classes, and participants were given arbitrary MTS training from voices to faces and from voices to names. Second, I evaluated whether the procedure of defining “correct” voice–face–name combinations individually for each participant, based on their pretest response patterns, would be useful as a method for controlling pre-experimental biases in studies of conditional relation formation that employ meaningful stimuli.

Participants

Thirteen Japanese undergraduate students enrolled in a psychology laboratory course volunteered for this study. Nine participants who completed all experimental sessions (seven women, two men; aged 18–19 years) were included in the analyses. Data from the remaining four were excluded for protocol deviations—either discontinuation and restart of the experimental session or missing more than two sets of data for unknown reasons. The study protocol was reviewed and approved by the institutional ethics committee and all participants provided informed consent.

Stimuli

Illustrations. For six AI-generated, anime-style characters (three women, three men), I created upper-body, smiling portraits using GPT-4o (OpenAI, 2024; see Figure 1 for women's illustrations).

Names. Six Japanese full names (family name + given name), each totaling four kanji characters, were generated using an online virtual-name generation service (https://namegen.jp/). No kanji character was repeated across names.

Scripts. Seven scenarios were prepared: (1) introduction, (2) lunchtime, (3) break at a part-time job, (4) club activity, (5) café, (6) seminar, and (7) between-class recess. For each scenario and each character, two alternative scripts were produced by editing an initial GPT-4o draft, yielding 84 scripts in total (7 scenarios × 6 characters × 2 scripts). In addition, six “common” scripts (shared across characters) were prepared for second training, in which all characters read the same line.

Voices. For each script, an MP3 audio file was generated with an online text-to-speech service. Distinct voice models were selected so that each character had a unique voice; model–character assignments were randomized and then held constant thereafter.

Each participant completed the MTS tasks described below using either the female set or the male set of AI-character stimuli. The mapping between scripts and voices was predetermined and fixed. Pairings of illustrations and names with the scripted voices were finalized after a pretest and individualized based on each participant’s responses.

Supplementary materials. For review purposes, at the end of submitted manuscript includes are the illustrations and names of the Japanese AI characters together with the English translations of all dialogue scripts (the original Japanese scripts are not included). Upon acceptance, these materials will be made publicly available via the Open Science Framework (OSF).

Experimental tasks

Four MTS tasks were implemented using lab.js (Henninger et al., 2022) and delivered via a web server using JATOS (Lange et al., 2015):

Voice-to-Illustration (VtoI): a spoken voice served as the sample stimulus; three illustrations were the comparison stimuli.

Voice-to-Name (VtoN): a spoken voice served as the sample stimulus; three printed names were the comparison stimuli.

Illustration-to-Name (ItoN): an illustration served as the sample stimulus; three printed names were the comparison stimuli.

Name-to-Illustration (NtoI): a printed name served as the sample stimulus; three illustrations were the comparison stimuli.

On every trial, the sample and comparison stimuli were presented simultaneously. The visual sample (illustration or printed name) appeared at the top center of the screen. The three comparison stimuli were arranged horizontally at screen center in a randomized order. Participants made their selections by clicking one of the comparison stimuli with the mouse. A block consisted of three MTS trials.

During test phases, an additional option labeled “I don’t know” in Japanese was displayed below the comparison array to discourage random guessing. During training phases, correct responses produced a “◯” and incorrect responses produced an “✕” at the bottom center of the screen; no feedback was provided in test phases. To clearly distinguish test and training phases and mitigate changes in responding when feedback was withdrawn, the background color was honeydew (#f0fff0) in test phases and white (#ffffff) in training phases, and the current phase label (“Test” or “Practice” in Japanese) was shown at the top right of the screen (see Procedure below).

Apparatus

The study was administered online. Participants used their own personal computers. Stimuli were presented and responses were recorded in the web browser, and audio was played through the participants’ headphones. Because data were collected on participants’ devices, hardware and software properties (e.g., CPU performance, operating system and browser, screen size/resolution) and network conditions could not be controlled.

Procedure

Before participation, students were informed of the approximate time required for the experimental session and asked to choose a time window during which they could work in a quiet room without interruption. After reading and signing an informed consent form, participants received the URL for the experiment.

1. Instruction. Participants read the following task instructions in Japanese and adjusted the audio volume as directed. In this experiment, you will learn the voices, faces, and names of three AI characters. When “+” appears on the screen, click it. A line of dialogue, a name, or an illustration will be presented. Decide whose voice, face (illustration), or name it is, then click one of the three choices. If you are unsure, click “I don’t know.” Note that some trials may not include the “I don’t know” option. There are two types of phases. In Test phases, the background is light green. No feedback on correctness is shown. In Practice phases, the background is white. A “◯” appears for a correct response and an “✕” for an incorrect response. Use this feedback to help you learn. As you continue to respond correctly, the dialogue lines will change. Practice repeats until you maintain correct responses across several lines. When you are ready, click “OK.”

2. Pretest. Participants then completed, in a fixed order (VtoI → VtoN → ItoN → NtoI), two blocks of MTS trials using Scenario 7. Each block was administered once without feedback. Based on pretest performance, I finalized, separately for each participant, the correct-response mapping for the MTS tasks by evaluating all 216 one-to-one illustration–name–voice assignments and selecting the candidate that minimized the number of matches between the participant’s pretest responses and the candidate’s ItoN and NtoI pairs, subject to a prespecified ceiling (≤ 3). Ties were resolved by applying the same rule to VtoI and VtoN (≤ 3), and any remaining ties by the smallest combined totals. The full algorithm and commented JavaScript code are provided in the supplementary materials.

3. First training phase. Participants were divided such that half first trained on VtoI and the other half first trained on VtoN. Training used Scenarios 1–3. Blocks were repeated until the percentage of correct responses reached 100% within a block.

4. Midtest. After reaching criterion, participants completed a midtest identical to the pretest (Scenario 7; two blocks; no feedback).

5. Second training phase. After the midtest, all participants were trained on the remaining (previously untrained) MTS task using Scenarios 4–6. Blocks were repeated until participants again achieved 100% correct responses within a block.

6. Posttest. After completing training on both MTS tasks, participants completed a posttest identical to the pretest (Scenario 7; two blocks; no feedback).

Upon finishing the posttest, participants emailed the experimenter. The experimenter reviewed the data; if accuracy was ≥ 80% across all MTS tasks, a thank-you email and an exit questionnaire were sent, asking whether headphones were used and whether any issues (e.g., network interruptions) occurred during the session. No problems were reported. If a participant did not meet the 80% criterion, they were invited to complete additional training with the common scripts and then a follow-up test. During this additional training, scripts 1–3 from the common set were used first, followed by scripts 4–6.

AI-character set (female vs. male voices), participant gender (female vs. male), and training order (VtoI→VtoN vs. VtoN→VtoI) yielded eight possible experimental conditions. Participants were randomly assigned to these conditions, but cell sizes were not balanced because of the limited sample size and attrition.

Accuracy (percent correct) on ItoN and NtoI MTS tasks increased from pretest to posttest for seven of the nine participants. For these seven participants, equivalence-based relational responding emerged without the additional training with common scripts. In contrast, two participants did not show equivalence even after receiving this additional training. Table 1 summarizes participants’ profiles, experimental conditions, and outcomes. For P08, one set of data points was missing from the server files; the cause could not be determined.

Table 1 Summary of Participants’ Profiles, Experimental Conditions, and Results

ID	AI- Gender	Participant- Gender	Training Order	Second Training	Equivalence	Missing Value
P01	F	F	VtoN->VtoI	--	P	none
P02	F	M	VtoI->VtoN	--	P	none
P03	F	M	VtoN->VtoI	--	P	none
P04	M	F	VtoI->VtoN	--	P	none
P05	F	F	VtoI->VtoN	--	P	none
P06	M	F	VtoI->VtoN	--	P	none
P07	F	F	VtoN->VtoI	--	P	none
P08	M	F	VtoI->VtoN	+	N	1 set^a
P09	M	F	VtoN->VtoI	+	N	none

Note. “+” in the Second Training column indicates that the participant received the additional training with common scripts and the follow-up test. In the Equivalence column, “P” denotes that increases in ItoN/NtoI accuracy were observed only after both VtoN and VtoI training; “N” denotes otherwise. Missing values are described in the last column. ^a Data for the sixth set in the first VtoN training were missing.

For seven participants, percent correct on ItoN and NtoI increased at posttest but not at midtest. Figure 2 plots MTS scores by task and block: bar graphs show percent correct at the pre-, mid-, and posttests, and line graphs show percent correct across blocks in the corresponding VtoI or VtoN training. Most participants required more than two blocks to meet the training criterion in the first training scenario, and the number of blocks generally decreased as training progressed; by the final blocks, most participants selected the correct illustration or name on the first trial. All participants showed higher percent correct on the tests after the corresponding training.

The two participants whose percent correct on ItoN and NtoI did not increase at posttest, and who therefore completed the additional training with common scripts, did not show improvement at the follow-up test (Figure 3). Relative to the other participants, these participants showed lower and less stable accuracy on VtoI and VtoN tests, even immediately after the corresponding training.

In this study, I defined identification of AI characters as the establishment of stimulus equivalence among voices, faces, and names and examined whether choices of faces given names (NtoI) and choices of names given faces (ItoN) would emerge without direct training after arbitrary MTS training from voices to faces (VtoI) and from voices to names (VtoN). The results showed that stimulus equivalence can be established even when MTS training relies on acoustic features of the voices as discriminative stimuli. This finding extends previous work that used the linguistic content of spoken words as the basis for MTS training (e.g., Cowley et al., 1992; Dube et al., 1993; Stewart & Lavelle, 2013).

In the present procedure, participants were exposed to multiple scenes in which the same character’s voice was heard with different lines of dialogue. In other words, the present procedure combined arbitrary MTS training with multiple-exemplar exposure to each character's acoustic characteristics. During VtoI and VtoN training, many participants required several scenes before their initial responses in a new scene became correct. This pattern suggests that multiple exemplars may play an important role in enabling acoustic features of the voices to function as effective discriminative stimuli, although further research is needed because the presence or absence of multiple-exemplar training was not manipulated directly.

Two of the nine participants failed to show stimulus equivalence after the initial VtoI and VtoN training, and equivalence did not emerge even after the additional training with common scripts, in which the linguistic content of the lines was made identical across scenes so as to prevent content from serving as a cue. Normal hearing was confirmed at recruitment, and both participants met the acquisition criterion in VtoI and VtoN training for each scene, suggesting that their performance cannot be attributed to hearing difficulties or problems with the experimental apparatus (e.g., headsets). I also considered the possibility that they were strongly attached to a particular combination of voice, face, and name, but their response patterns differed across pre-, mid-, and posttests, which does not support a stable idiosyncratic preference for a single combination. Compared with other participants, these two required more training trials before reaching the acquisition criterion in VtoI and VtoN. It is possible that, for these individuals, acoustic characteristics were less effective in controlling discriminative responding, or that some preconditions must be met for stimulus equivalence to emerge when the relations among classes are defined by acoustic features. Future research should examine such individual differences more directly, for example by using same–different judgment tasks with stimulus sets that vary in the salience of acoustic differences, and by relating performance on those tasks to the likelihood of establishing stimulus equivalence based on acoustic features.

Previous research (Kamachi et al., 2003) and my own preliminary observations suggest that functional relations can be formed between facial and vocal characteristics. Informal observations from popular culture also support this idea: for example, fans sometimes express dissatisfaction when an anime adaptation of a manga assigns a voice to a character that they perceive as inconsistent with how they had “imagined” the character should sound. In the present experiment, I defined the “correct” combinations of voices, faces, and names for character identification only after the pretest, separately for each participant, such that the combinations designated as correct corresponded to those that were least likely to be chosen given that participant’s pretest response pattern. This procedure produced sufficiently large pre–post differences in NtoI and ItoN test accuracy, which served to evaluate the presence or absence of stimulus equivalence. The procedure may be useful not only when meaningful stimuli are used but also when nominally meaningless stimuli are employed but are nevertheless affected by participants’ pre-experimental reinforcement histories, such that they initially respond “correctly” to relations that the experimenter had intended to be arbitrary. At the same time, the present approach has a cost: for each participant, the “correct” combination may be one that diverges substantially from how they would naturally expect the character to look or sound. Thus, the procedure involves a trade-off between controlling preexisting biases and maintaining ecological validity with respect to actual casting decisions in anime production, where producers often aim to meet viewers’ expectations. Applied research will need to explore methods for reducing bias while preserving the perceived naturalness of voice–face–name combinations.

The present data do not allow me to determine whether acoustic features of the voices themselves functioned as direct discriminative stimuli controlling choices of faces and names. It is possible that participants emitted tacts such as “brighter vs. calmer” or “higher vs. lower” relative to other characters’ voices and that these mediating responses served as cues when selecting comparison stimuli. Similar considerations apply to the face illustrations: participants may have described properties such as hairstyle (short vs. long), presence or absence of glasses, or clothing (shirt vs. tank top), and these tacts may have mediated their choices. This possibility is consistent with the naming hypothesis (Horne & Lowe, 1996), which proposes that verbally mediated behavior such as tacting can facilitate the emergence of derived relational responding. Future research should examine the extent to which the stimulus equivalence observed here depended on participants’ tact repertoires for the voices, faces, and names. In particular, because auditory stimuli are brief and transient, unlike static images and printed names, they may no longer be present at the moment when the comparison stimuli are selected. This temporal property of auditory stimuli may increase the importance of mediating responses that tact the acoustic characteristics of the voice.

The overarching goal of this study was to develop an experimental method for analyzing how viewers come to identify characters in animated works. Toward this end, I adopted an MTS procedure that, based on prior research, was expected to promote the emergence of stimulus equivalence. I assumed that certain contingencies in the viewing context could be approximated by the MTS procedure—for example, a viewer might see a character in a kimono carrying a sword, silently ask themselves “Who was this blond boy again? Was it Zenitsu?”, and then hear another character call out, “Zenitsu!” Although such situations are reminiscent of MTS contingencies, actual viewers of anime are, of course, not required to choose among comparison stimuli, nor do they receive explicit feedback about correctness. Rather, viewers are frequently exposed to simultaneous or successive pairings of multiple stimuli, such as faces with names or voices with faces. Prior research has shown that even simple pairing of object images and sounds, without explicit MTS training, can give rise to conditional discriminations when the sounds later serve as samples and the images as comparisons (e.g., Kinloch et al., 2013; Petursdottir et al., 2020). Future studies should therefore examine whether stimulus equivalence based on acoustic characteristics also emerges under such pairing procedures, thereby allowing a closer approximation to character identification processes in more naturalistic viewing contexts.

When equivalence relations are formed among members of a class, stimulus functions can transfer across them (e.g., Barnes-Holmes et al., 2000; Perez et al., 2020, 2023). In the case of anime characters, as viewers continue to watch a series, the functions associated with a character’s actions and utterances may transfer to the character’s face and name. For example, the name or voice of a character who frequently engages in violent acts may come to evoke anxiety or tension, whereas the face or voice of a particularly appealing character may evoke positive emotional reactions. Examining person identification within the framework of stimulus equivalence may thus contribute to a behavioral analysis of how viewers understand and respond to narratives in animated works.

Funding. This research received no specific grant from any funding agency, commercial or not-for-profit sectors.

Competing Interests. The author has no relevant financial or non-financial interests to disclose.

Ethical approval. This study was approved by the institutional ethics committee of XXXX University (approval no. 24-0012).

Informed consent. Informed consent was obtained from all individual participants included in the study.

Barnes-Holmes, D., Keane, J., Barnes-Holmes, Y., & Smeets, P. (2000). A derived transfer of emotive functions as a means of establishing differential preferences for soft drinks. The Psychological Record, 50, 493–511. https://doi.org/10.1007/BF03395367
Bush, K. M., Sidman, M., & de Rose, T. (1989). Contextual control of emergent equivalence relations. Journal of the Experimental Analysis of Behavior, 51(1), 29–45. https://doi.org/10.1901/JEAB.1989.51-29
Carr, D., & Felce, J. (2008). Teaching picture-to-object relations in picture-based requesting by children with autism: A comparison between error prevention and error correction teaching procedures. Journal of Intellectual Disability Research, 52(4), 309–317. https://doi.org/10.1111/j.1365-2788.2007.01021.x
Clifford, B. R., & Bull, R. (1978). The psychology of person identification. Routledge.
Cowley, B. J., Green, G., & Braunling-McMorrow, D. (1992). Using stimulus equivalence procedures to teach name–face matching to adults with brain injuries. Journal of Applied Behavior Analysis, 25(2), 461–475. https://doi.org/10.1901/jaba.1992.25-461
Dube, W. V., Green, G., & Serna, R. W. (1993). Auditory successive conditional discrimination and auditory stimulus equivalence classes. Journal of the Experimental Analysis of Behavior, 59(1), 103–114. https://doi.org/10.1901/jeab.1993.59-103
Henninger, F., Shevchenko, Y., Mertens, U. K., Kieslich, P. J., & Hilbig, B. E. (2022). lab.js: A free, open, online study builder. Behavior Research Methods, 54, 556–573. https://doi.org/10.3758/s13428-019-01283-5
Horne, P. J., & Lowe, C. F. (1996). On the origins of naming and other symbolic behavior. Journal of the Experimental Analysis of Behavior, 65(1), 185–241. https://doi.org/10.1901/jeab.1996.65-185
Kamachi, M., Hill, H., Lander, K., & Vatikiotis-Bateson, E. (2003). “Putting the face to the voice”: Matching identity across modality. Current Biology, 13(19), 1709–1714. https://doi.org/10.1016/j.cub.2003.09.005
Kinloch, J. M., McEwan, J. S. A., & Foster, T. M. (2013). Matching-to-sample and stimulus-pairing-observation procedures in stimulus equivalence: The effects of number of trials and stimulus arrangement. The Psychological Record, 63(1), 157–174. https://doi.org/10.11133/j.tpr.2013.63.1.012
Kramer, R. S. S., Jones, A. L., & Gous, G. (2021). Individual differences in face and voice matching abilities: The relationship between accuracy and consistency. Applied Cognitive Psychology, 35(1), 192–202. https://doi.org/10.1002/acp.3754
Lange, K., Kühn, S., & Filevich, E. (2015). “Just Another Tool for Online Studies” (JATOS): An easy solution for setup and management of web servers supporting online studies. PLOS ONE, 10(6), Article e0130834. https://doi.org/10.1371/journal.pone.0130834
OpenAI. (2024). GPT-4o [Large language model]. OpenAI. https://openai.com/index/hello-gpt-4o
Perez, W. F., de Almeida, J. H., Soares, L. C. C. S., Wang, T. F. L., de Morais, T. E. D. G., Mascarenhas, A. V., & de Rose, J. C. (2020). Fearful faces and the derived transfer of aversive functions. The Psychological Record, 70(3), 387–396. https://doi.org/10.1007/s40732-020-00390-6
Perez, W. F., Harte, C., Barnes-Holmes, D., Gomes, C. T., Mohor, B., & de Rose, J. C. (2023). Generalized contextual control based on nonarbitrary and arbitrary transfer of stimulus functions. Journal of the Experimental Analysis of Behavior, 119(3), 448–460. https://doi.org/10.1002/jeab.839
Petursdottir, A. I., Neaves, S. M., & Thomas, O. N. (2020). Emergent tact control following stimulus pairing: Comparison of procedural variations. The Analysis of Verbal Behavior, 36(2), 193–214. https://doi.org/10.1007/s40616-020-00132-3
Sidman, M. (1994). Equivalence relations and behavior: A research story. Authors Cooperative, Inc.
Sidman, M., & Tailby, W. (1982). Conditional discrimination vs. matching to sample: An expansion of the testing paradigm. Journal of the Experimental Analysis of Behavior, 37(1), 5–22. https://doi.org/10.1901/jeab.1982.37-5
Smith, H. M. J., Dunn, A. K., Baguley, T., & Stacey, P. C. (2016). Matching novel face and voice identity using static and dynamic facial images. Attention, Perception, & Psychophysics, 78(3), 868–879. https://doi.org/10.3758/s13414-015-1045-8
Smith, H. M. J., Ritchie, K. L., Baguley, T. S., & Lavan, N. (2025). Face and voice identity matching accuracy is not improved by multimodal identity information. British Journal of Psychology, 116(2), 367–385. https://doi.org/10.1111/bjop.12757
Stewart, I., & Lavelle, N. (2013). Auditory stimulus equivalence and non-arbitrary relations. The Psychological Record, 63(3), 409–426. https://doi.org/10.11133/j.tpr.2013.63.3.001

The authors declare no competing interests.

Supplementarymaterialsforreview.docx

Download PDF

Version 1

posted

You are reading this latest preprint version

Stimulus Equivalence in AI-Generated Anime Characters: Equivalence Responding Among Voices, Images, and Names

Status:

Version 1

Abstract

Figures

Introduction

Method

Results

Discussion

Statements and Declarations

References

Additional Declarations

Supplementary Files

Status:

Version 1