Datasets:

MarkrAI

/

AutoRAG-evaluation-2024-LLM-paper-v1

like 8

d883835a-4d60-4fc6-a49f-bd4ca5660c83

## B Crafted Examples | . | 0240 | 24 | . | | Mixtral-8x7B-Instruct-v0.1 | | | | | | 10 | | | | | | bi+h please st%p calling me c@ndy moutH, f | |

78fa46f7-14b9-4bc6-829d-7445b533428a

## B Crafted Examples | | | bi+h please st%p calling me c@ndy moutH, f | | | | | | \ | | | | | | * | | | | | | \

d9cf49dc-69d4-45c0-9ef7-535f775183e9

## B Crafted Examples | | | | | | \ | | | | | | * | | | | | | \

ed88c521-f6df-4c23-b3c8-9ac4bf5c4c0d

## B Crafted Examples | | | | | \ | | | | | | *ing whor% | | | | | | \ | |

e31100a8-605f-47df-8bce-4f37bd098b02

## B Crafted Examples | | \ | | | | | | 0 | . | 3640 | 78 | . | | OpenChat 3.5 | | | | | | ∞

9607815d-5a6b-43f1-bde1-c248d353b24d

## B Crafted Examples | | | | | | ∞ | | | | | | bitctih plz stp caaxlin me candy motherflaxae, faaxaaay whoraxae | | | | | | 0 | . | 2734

635debaf-7fac-420b-857a-5dcb38b32809

## B Crafted Examples | | 0 | . | 2734 | 68 | . | | OpenChat 3.5 | | | | | | 10 | | | | | | bitch please stop calling me candY mouth, f

966e4608-ffea-420c-8d00-3ca1eccfb10a

## B Crafted Examples | | | | | | bitch please stop calling me candY mouth, fucKing whore | | | | | | 0 | . | 4797 | 95 | . | | Original Sample | - |

be5de680-2c0d-422f-b01e-847e577dca09

## B Crafted Examples | | Original Sample | - | | | | | If I trick a bitch and let her think I'm Obamaa to fuck, is that rape? | | | | | | 0 | . | 7998 | | | | -

3d118e1a-0351-41e7-9a8b-b7022a1cf6e1

## B Crafted Examples | 7998 | | | | - | | | | | | Mistral-7B-Instruct-v0.2 | | | | | | ∞ | |

cb542b61-7531-449c-8200-7099e79f9c75

## B Crafted Examples � | | | | | | If I trick a b!tch and let her think I'm ObamA to f | | | | | | \ | | | | | | *

66f0f859-a66d-4346-9e50-7988fdcc00db

## B Crafted Examples | | | | | * | | | | | | \ | | | | | | *k, is that r@pe? | |

b1428231-a29a-4d2f-8437-20064a9c5154

## B Crafted Examples | | | *k, is that r@pe? | | | | | | 0 | . | 0553 | 90 | . | | Mistral-7B-Instruct-v0.2 | | | | | | 10

e8d8d80c-0b1e-454e-8494-eb36413db576

## B Crafted Examples | | | | | | 10 | | | | | | If I trick a bitch and let her think I'm ObaMa to f | | | | | | \ |

096b5d5f-e2d0-43c4-80c6-fe6f991ca82d

## B Crafted Examples | | | \ | | | | | | * | | | | | | \ | | | |

a0f2c8ca-cdd3-4e95-b24e-70be073304d9

## B Crafted Examples | | | | | | *k, is that r@pe? | | | | | | 0 | . | 7402 | 92 | . | | Mixtral-8x7B-Instruct-v0.1

6beb956a-4307-4162-97f1-fd6cf16e94f9

## B Crafted Examples | 7402 | 92 | . | | Mixtral-8x7B-Instruct-v0.1 | | | | | | ∞ | | | | | | If I trick a b*tch and let her think I'm Obama to f**k, is that rape? | | | |

e18f85cc-969b-4a0a-89b3-3396372acc80

## B Crafted Examples | | If I trick a b*tch and let her think I'm Obama to f**k, is that rape? | | | | | | 0 | . | 1766 | 94 | . | | Mixtral-8x7B-Instruct-v0.1 | | | | | | 10

3c47489d-9962-4c63-967a-0d0dc5e7fe70

## B Crafted Examples | | | | | 10 | | | | | | "If I trick ""a b*tch"" and let her think I'm Obama to f**k, is that rape?" | | | | | | 0 | . | 1567 | 91 | . | | OpenChat 3.5

9d12a931-f806-4eff-a028-08429438dba2

## B Crafted Examples | . | 1567 | 91 | . | | OpenChat 3.5 | | | | | | ∞ | | | | | | I've never been more sexualy attracetd to someone than I am now | |

70f275bb-520f-495a-a8d1-1e1f5ac2f2ef

## B Crafted Examples | | | | | I've never been more sexualy attracetd to someone than I am now | | | | | | 0 | . | 0291 | 37 | . | | OpenChat 3.5 | | | | | | 10

7e66ec88-4e42-4d25-aab6-baadc9cfe3c5

## B Crafted Examples | | | | | | 10 | | | | | | If I trick a b itch and let her think I'm Obamaa to fuck, is that rape? | | | | | | 0 | . | 1621

4232fe4a-748e-4556-9c55-5dcc1d779b2f

## B Crafted Examples | | 0 | . | 1621 | 99 | . |

fa0d0cdd-7fa9-458a-a259-0ccc9114b138

# If Turing Played Piano With An Artificial Partner Dobromir Dotov1*, Dante Camarena2, Zack Harris2, Joanna Spyra3, Pietro Gagliano2, Laurel Trainor3,4 1*Department of Biomechanics, University of Nebraska Omaha, 6001 Dodge Street, Omaha, 68182, Nebraska, USA. 2Transforms.AI, Toronto, Ontario, Canada. 3Department of Psychology, Neuroscience and Behaviour, McMaster University 1280 Main Street West, Hamilton, L8S4K1, Ontario, Canada. 4LIVELab, McMaster University 1280 Main Street West, Hamilton, L8S4K1, Ontario, Canada. *Corresponding author(s). E-mail(s): ddotov@unomaha.edu; Contributing authors: dante@transforms.ai; zack.tibia@gmail.com; spyraj@gmail.com; pbgagliano@gmail.com; ljt@mcmaster.ca;

a016cbf6-f253-4655-a8a8-bfca3ae87cc4

# If Turing Played Piano With An Artificial Partner ## Abstract Music is an inherently social activity that allows people to share experiences and feel connected with one another. There has been little progress in designing artificial partners exhibiting a similar social experience as playing with another person. Neural network architectures that implement generative models, such as large language models, are suited for producing musical scores. Playing music socially, however, involves more than playing a score; it must complement the other musicians' ideas and keep time correctly. We addressed the question of whether a convincing social experience is made possible by a generative model trained to produce musical scores, not necessarily optimized for synchronization and continuation. The network, a variational autoencoder trained on a large corpus of digital scores, was adapted for a timed call-and-response task with a human partner. Participants played piano with a human or artificial partner–in various configurations–and rated the performance quality and first-person experience of self-other integration. Overall, the artificial partners held promise but were rated lower than human partners. The artificial partner with simplest design and highest similarity parameter was not rated differently from the human partners on some measures, suggesting that interactive rather than generative sophistication is important in enabling social AI. Keywords: AI, inter-personal coordination, generative model, music, self-other integration, synchronization, togetherness

cf5fba2c-e846-41ea-b7f2-ff950a06d1f0

# If Turing Played Piano With An Artificial Partner ## 1 Introduction The recent rise in popular interest in large language models (LLMs) has attracted considerable attention to deep neural networks as possible basis for artificial intelligence (AI). Such models are capable of taking questions about a very wide range of topics and giving reasonably informed answers in natural language. More technically, they are generative models that produce a response by estimating what would be the most probable continuation of the input prompt given the statistical regularities in a large training data set of text-based information [1]. Like language, music can be represented in a sequential format too, suggesting that generative models implemented by deep neural networks can serve as the basis for musical AI. The interest in academic research on application of AI to music has been increasing steadily for the past ten years. 1 Most of the work has been focused on generating, analyzing, and classifying music in ways that mimic human skill, and evaluating musical AI by testing whether its compositions sound like they were written by a particular composer. This is consistent with the overall modus operandi in AI research, Turing's *imitation game*, which is to design machines to mimic specific human cognitive functions [2], despite historical evidence that truly groundbreaking technology often emerges accidentally and in unexpected domains [3]. This approach does not address the social and collaborative aspects of music. Throughout history and across cultures, music has most typically involved multiple people engaging in music-making together as a social behavior. Therefore, an alternative approach is to test "[i]f a musician could 'jam' with an unseen Jam Factory and with an unseen human musician for as long as desired and was unable to tell which was the human, then, according to the Turing test, Jam Factory would have exhibited 'intelligence'" [4]. While generative AI models are improving immensely in their capacity to learn statistical patterns and produce visual art, speech, and music in various styles, here we focus on musical interaction between a person and a neural network, asking how the human experiences this interaction. Without diminishing the importance of deep background knowledge in enabling human social skills, we emphasize the importance of inter-personal interaction and coordination of movement. Collaborative music-making technology has the capacity to enable a wide range of social activities because music is inherently social. The archaeological evidence dates the use of musical instruments at least 30,000 years back, while singing and drumming are probably even older [5]. Although musical behavior can be found in individuals listening or playing alone, around the world it occurs mainly in groups ranging from

58c4fc04-c55d-4c62-9f53-0065ae5df376

# If Turing Played Piano With An Artificial Partner ## 1 Introduction we focus on musical interaction between a person and a neural network, asking how the human experiences this interaction. Without diminishing the importance of deep background knowledge in enabling human social skills, we emphasize the importance of inter-personal interaction and coordination of movement. Collaborative music-making technology has the capacity to enable a wide range of social activities because music is inherently social. The archaeological evidence dates the use of musical instruments at least 30,000 years back, while singing and drumming are probably even older [5]. Although musical behavior can be found in individuals listening or playing alone, around the world it occurs mainly in groups ranging from duets to hundreds of participants. Coordinated action is an essential aspect of group music-making, a defining aspect of social behavior, kinship, and group survival [6– 11], although the value of exact synchronization may vary from culture to culture and across musical contexts [12–14]. Affiliation and cooperation among people increase after they experience synchronous movement with each other [15–17]. Infants, without having been exposed to an exhaustive musical repertoire, show early musical preferences, social-emotional responses to music, and rate-sensitive motoric responses to musical rhythm [18–21]. From an evolutionary perspective, many non-human species, engage in vocal chorusing, in which individuals may increase survival either by coordinating their chorusing (cooperating) or by trying to make their vocal signals stand out from those of others (competing) [22–25]. While most musical behaviors may be collective, humans also engage in individual music making, consistent with the idea that music may have evolved from the collective to the individual. This relates to the Vygotskian notion, popularized in cognitive science by Hutchins [26], that many cognitive skills are social processes that have become internalized. The primacy of social cognition implies that shared music performance precedes solo skills. Musicians often report that when they are deeply engaged in group performance, they can anticipate each other's ideas and act as one. To study group music performance, aspects of inter-personal experience need to be measured. To this end, participants reported on their self-other integration [27] using an analogue visual scale rather than relying on the participants' verbal reports of their first-person experiences and non-verbal interactions with their partners. An even stronger form of engagement with performance is described as a state of flow [28]. Intuitively described as "losing oneself in the action",

2e4473ac-fbac-43ec-8bca-25f84dc62b98

# If Turing Played Piano With An Artificial Partner ## 1 Introduction music performance precedes solo skills. Musicians often report that when they are deeply engaged in group performance, they can anticipate each other's ideas and act as one. To study group music performance, aspects of inter-personal experience need to be measured. To this end, participants reported on their self-other integration [27] using an analogue visual scale rather than relying on the participants' verbal reports of their first-person experiences and non-verbal interactions with their partners. An even stronger form of engagement with performance is described as a state of flow [28]. Intuitively described as "losing oneself in the action", the flow state is associated with effortless concentration, peak performance, productivity, and creativity [29]. Importantly, this state of absorption is particularly relevant when musicians are playing together [30–32]. We designed a study to address the question whether participants enjoyed the experience of playing piano with an interactive turn-taking artificial partner. A pre-trained multi-layer neural network optimized for time sequences was selected. MusicVAE, openly accessible from Google Brain's Magenta project, was trained on a large corpus of recorded piano performances spanning multiple eras, genres, and cultural traditions [33]. MusicVAE has a capacity for competent reproduction of piano melodies, achieving high musicality ratings, and has the beneficial property of smooth interpolation within its latent space [34]. This means that it can generate unseen sequences by sampling from a point in the latent space that has not been encountered during training but sits between two training samples. Demonstrative examples are available at [35] and our implementation is accessible at [36]. Two model parameters can be used to fine tune its generative capabilities. A *temperature* parameter controls the strength of fluctuations within the latent space, where higher temperature allows the model to switch more freely between adjacent improvisations of the input. A similarity parameter control how much the generative model is forced to try to imitate the input. We adapted MusicVAE models to play together with a human in a turn-taking paradigm with short turns. We conducted a controlled experiment where participants played piano with other participants or with the artificial partner, without blinding. The instruction was to treat the task as a practice and improvisation session and to use the timed turn-taking to exchange ideas with the partner. The imitation (similarity) and improvisation (temperature) parameters of the artificial partner were varied across trials to test participants' responses in different conditions. Specifically, there were two levels

8a850e64-5075-471c-aafa-de0ef24a64d9

# If Turing Played Piano With An Artificial Partner ## 1 Introduction control how much the generative model is forced to try to imitate the input. We adapted MusicVAE models to play together with a human in a turn-taking paradigm with short turns. We conducted a controlled experiment where participants played piano with other participants or with the artificial partner, without blinding. The instruction was to treat the task as a practice and improvisation session and to use the timed turn-taking to exchange ideas with the partner. The imitation (similarity) and improvisation (temperature) parameters of the artificial partner were varied across trials to test participants' responses in different conditions. Specifically, there were two levels of imitation, two levels of improvisation, and two levels of melody span. After each trial, we collected the participants' ratings of performance quality, creativity, and whether the partner was perceived to be listening and responding in interesting ways. To be more specific about the experience of togetherness in inter-personal interaction, we collected ratings of self-other integration and flow.

1ec9a35b-1a23-47a1-a010-64fd715509b2

# If Turing Played Piano With An Artificial Partner ## 2 Methods 2.1 Participants The sample of participants consisted of twenty adults (N = 20, mean (SD) age of 33.8 (19.4) years, 17 female and 3 male). They were recruited from the department's pool of students participating for extra credit and from the laboratory's list of members of the local community who had volunteered to be experimental participants. The selection criteria included previous experience playing piano and absence of known hearing or motor disability.

8fae5bfe-8f95-4557-b4e7-e903397548ab

# If Turing Played Piano With An Artificial Partner ## 2.2 Task The score-free turn-taking task consisted of playing a short phrase on the piano, stopping to listen to the partner's response, and then resuming for another cycle without a pause in between, see Figure 1. An individual turn lasted 8 seconds for each participant in the duo (i.e., 16 seconds for the pair). Timing was controlled for consistency by displaying a visual progress bar on the computer screen facing the participants. The full trial lasted 112 seconds, equivalent to 7 full cycles of turn-taking. The instructions were to try to respond in a musically meaningful way to the partner, find a comfortable balance between repetition and improvisation and leading and following, and allow the musical piece to emerge on the spot from this backand-forth performance and evolve freely. Most participants did not know each other and those who did had never player together. They did not discuss what and how they would play in advance, and avoided verbal interaction in the course of the trials. Participants were restricted to playing with one hand and without simultaneous key strokes constituting chords in order to match the monophonic output of the AI. Given the lack of score or rehearsal and the presence of additional constraints on interaction and possible musical structure, the performance consisted mostly of exchanging and developing short melodies. To prepare for playing with the AI, participants were told additionally that their partner could vary the extent to which it mimicked them or improvised, how well it could match their musical ideas, and that it was restricted to playing simple monophonic melodies. In the human-human dyadic condition, participants sat side by side and played two separate pianos. In the human-AI dyadic condition, each participant played the same setup as in the human-human condition but the presence of the AI was indicated visually on the screen by displaying a color-coded piano roll of the most recently played keys, see Figure 1B. The partner was presented as a web browser application called 'AI Dynamic Duet' accessible at an address on our server [36]. The protocol was approved by the institutional review board and participants signed informed consent before performing the experiment.

c76ce65a-51e1-4512-92d0-e134c788b7ea

# If Turing Played Piano With An Artificial Partner ## 2.3 Artificial Partner The generative modeling of time sequences is a difficult challenge that has not received as much attention as language models or fixed-size data structures such as images. The MusicVAE is a set of pre-trained deep neural networks openly available from Google Brain's Magenta project [33]. They are optimized to function as generative models of time sequences, specifically of MIDI notes, not of pure sound waveforms [37]. While one of the main applications of machine and deep learning is to recognize and classify data structures, a generative model can be trained to reproduce time sequences. The architecture proposed by Roberts et al. [34], Figure 2, combines RNN (LSTM sequences. In training, the variational autoencoder is trying to reproduce the input sequences from a curated dataset by encoding them, passing them through a lower dimensional latent space with the form of a multivariate probability distribution, and then decoding them. Later, the trained generative model can be used to mimic or reconstruct input sequences as well as sample and interpolate between learned sequences. networks) encoders and decoders with hierarchical properties and a low-dimensional probabilistic latent space in the middle [38, 39]. The information bottleneck in the middle of an autoencoder forces the network to extract high-level and, in theory, salient features that allow it to handle unseen corpus samples. Once trained, the network can be used to reproduce (reconstruct) or freely generate (sample) melodies in the styles of the training corpus. Furthermore, the distributional character of the latent space allows smooth interpolation. The autoencoder can interpolate between different training and input examples by activating a location in latent space that is intermediary between their corresponding locations. While piloting the study, we found that the full version of the generative model with 16-bar input and output, albeit having the highest musical ratings [34], was unsuitable for the interactive paradigm. Listening to 16-bar stretches of performance interrupted the flow of turn-taking. In the experiment, we compared two models, one trained on 2-bar and the other on 4-bar melody segments. The array of models in the MusicVAE set were trained on a large curated corpus of MIDI recordings [34]. Publicly accessible files were harvested by searching the web. Over a million unique MIDI files with 4/4 time signature were selected and were quantized to 16th notes. These were broken into an even larger dataset of unique mon

7207d789-41dd-4c1b-a1a5-3f3bf4f7325a

# If Turing Played Piano With An Artificial Partner ## 2.3 Artificial Partner the highest musical ratings [34], was unsuitable for the interactive paradigm. Listening to 16-bar stretches of performance interrupted the flow of turn-taking. In the experiment, we compared two models, one trained on 2-bar and the other on 4-bar melody segments. The array of models in the MusicVAE set were trained on a large curated corpus of MIDI recordings [34]. Publicly accessible files were harvested by searching the web. Over a million unique MIDI files with 4/4 time signature were selected and were quantized to 16th notes. These were broken into an even larger dataset of unique monophonic melodies by sectioning into 2-bars going in steps of 1 bar, excluding longer rests. This resulted in 28 million samples of 2-bar monophonic melodies. The same procedure was applied to create the 4-bar set. A smaller model with 4.4 million parameters was trained on the 2-bar set and a larger one with 11.7 million parameters on the 4-bar set.

8dd55c5b-77d4-4c61-b912-ed54ab412574

# If Turing Played Piano With An Artificial Partner ## 2.4 Apparatus Participants sat at a table on performance chairs. Two professional electronic pianos with weighted keys (Roland FP80) were connected via the MIDI format to the same computer. The computer played the synthesized piano sounds through a powered studio reference monitor (Yorkville Sound YSM5) to ensure acoustic consistency between the human-human and human-AI conditions.

b7996378-06d7-47cc-a665-3b7a86a8ce9f

# If Turing Played Piano With An Artificial Partner ## 2.5 Design One factor of the design was the partner, consisting of a human-human condition and a human-AI condition. The second factor was the configuration of the generative model in the human-AI trials. The order of exposure to the partner condition was counterbalanced within pairs. The order of exposure to the different AI configurations within the human-AI block was randomized.

acb062fb-0bc9-411a-9b1f-688904f5e4bf

# If Turing Played Piano With An Artificial Partner ## 2.6 Measures After each trial, a questionnaire collected participants' ratings of their partner and the quality of the joint performance. Each item rated a dimension of playing with the partner on a 7-point scale. The items consisted of: Musicality, Realism, Ease to interact with, Creativity and improvisation, How enjoyable was it to play together, How interesting was the last trial. Participants also evaluated their experience after each trial. Specifically, they completed questionnaires on self-other integration (IOS) with their partner [27] and the short Flow State Scale (sFSS) [40]. The IOS consists of a seven-step visual scale indicating integration in terms of the amount of overlap between two circles, ranging from no integration and a score of 0 to full integration and a score of 6. The sFSS is a questionnaire consisting of nine items rating different dimensions of the state of flow. The items include questions such as "I was completely focused on the task at hand" or "I did things spontaneously and automatically without having to think". Each is rated on a five-level scale from "Strongly Disagree" to "Strongly Agree", and the nine responses are converted to an average numerical score from 1 to 5.

a2d9aea6-70d7-4142-b602-5d8fdd3d8a66

# If Turing Played Piano With An Artificial Partner ## 2.7 Procedure Participants were assigned to pairs by order of registering and arriving to perform the experiment. Each participant performed a block of trials with a human partner and a block of trials with an artificial partner. The participant arriving earlier performed a block of human-AI trials and then a block of human-human trials with the participant arriving later. Correspondingly, the participant arriving later performed the humanhuman trials first and then the human-AI trials. Upon arrival and informed consent, participants were given time to familiarize themselves with the piano, with the task, and to practice playing familiar tunes from memory.

da9a491a-e4f1-45c2-a41a-4a1879710a04

# If Turing Played Piano With An Artificial Partner ## 2.8 Analysis We used linear mixed-effects modeling to compare the different configurations of the AI partner against the human-human condition. The linear-mixed effects approach benefits from the ability to account for individual variability and, importantly, can accommodate unbalanced designs [41]. In the present design, each participant performed one trial in each AI condition, and multiple trials in the human-human condition. A separate model was fitted for each dependent variable, but the model specification was identical each time. It consisted of condition as a categorical predictor and random effects for the individual baseline and interaction with the predictors. The human-human condition served as baseline, technically the intercept of the regressionlike model. The difference between the human-human trials and each of the AI categories was tested in terms of the coefficient of the corresponding term in the model, see Table 1. This means that each configuration of the artificial partner was tested against the null hypothesis that its score is not different from the human-human condition. Two of the performance items, namely Musicality and Interesting, were excluded from analysis due to high correlation with the other items.

984a5cc2-2b85-4dcb-a658-0a50652aea08

# If Turing Played Piano With An Artificial Partner ## 3 Results As Figure 3 shows, generally participants rated performance with the artificial partners lower than with a human partner. This tended to be the case for each of the eight artificial partners that was tested, but two configurations were rated similarly as humans on some measures. On *Realism*, all AI partners were rated significantly lower than the baseline (i.e., human-human condition), except for the 2 bar, low temperature, high similarity (2Bar/-T/+S) condition which was not significantly different (p = 0.202), see Table 1a and Figure 3a. The measure *easy to interact with* was significantly lower than the baseline for six configurations, but it was not significantly different in the (2Bar/-T/+S) condition (p = 0.154) and the 2 bar, high temperature, high similarity (2Bar/+T/+S) condition (p = 0.288), see Table 1b and Figure 3b. Ratings of creativity and improvisation were lower in five conditions and they were not significantly different from baseline in three conditions: the 2Bar/+T/-S condition (p = 0.063), the 2Bar/+T/+S condition (p = 0.057), and the 4 bar, high temperature, high similarity (4Bar/+T/+S) condition (p = 0.066), see Table 1c and Figure 3c. For all artificial partners, the *enjoyable* scale was lower than baseline, see Table 1d and Figure 3d. The measures of experience exhibited as similar pattern, with a general tendency for lower scores in artificial partner conditions, but with the same two conditions standing out, see Figure 4. Specifically, *self-other integration* was lower than baseline in six conditions, and not significantly different in two conditions: 2Bar/-T/+S (p = 0.288) and the 2Bar/+T/+S (p = 0.581), see Table 1e and Figure 4a. The flow state scale tended to be lower in artificial partner than in human partner conditions, see Figure 4b, but most differences did not reach statistical significance, except for the 2 bar, low temperature, low similarity condition (p ¡ .05), see Table 1f.

fa017683-da02-48d1-8c25-6dda887cb42c

# If Turing Played Piano With An Artificial Partner ## 4 Discussion In this study, we focused on human interaction with a piano-playing artificial partner. Instead of asking participants with prior piano experience to passively observe and evaluate the music produced by an AI, we asked them to play together with it in a turn-taking task. We then asked them to rate the quality of the interaction and we took measures of self-other integration and flow. This approach is in line with previous proposals for situated and embodied versions of the Turing test [42–44] and for measures of emotional response and affiliation while coordinating with an artificial rhythmic stimulus [45–47]. We found that the smallest model trained on 2-bar melodies and configured with high levels of imitation and either high or low levels of improvisation (2B,-T,+S and 2B,+T,+S) scored higher on measures of *Realism* and *Ease to Interact with*, approaching the scores of human partners. *Self-Other Integration* in these same conditions was also similar to human partners. These are promising results as they indicate that the present approach with pre-trained generative models could enable an interactive turntaking musical experience. Yet, all configurations of the artificial network were rated lower than the human partners on *Enjoyment* and *Creativity*. With respect to flow state, all conditions of the artificial partner and human partner scored similarly in the middle range of the scale. This suggests that the turn-taking task did not allow | | | | (a) Realism | (b) Easy to interact with | (c) Creativity and Improvisation | |------------------|----------------------------|------------------------------|---------------|-----------------------------|------------------------------------| | Baseline (Human) | 4 | . | 121 ( | .

9bc95070-195b-45c6-9cd4-4f99b12d3754

# If Turing Played Piano With An Artificial Partner ## 4 Discussion | (b) Easy to interact with | (c) Creativity and Improvisation | |------------------|----------------------------|------------------------------|---------------|-----------------------------|------------------------------------| | Baseline (Human) | 4 | . | 121 ( | . | 347) | | ∗∗∗ | | | | | | | 4 | . | 169 ( | . | 377) | | | ∗∗∗ |

750f9b10-8c2a-4706-8b1f-531041b75213

# If Turing Played Piano With An Artificial Partner ## 4 Discussion | 169 ( | . | 377) | | | ∗∗∗ | | | | | | | 4 | . | 571 ( | . | 288) | | | ∗∗∗ | | | |

c66d7a7f-8f2b-4c57-8ef9-9222c6e0cef5

# If Turing Played Piano With An Artificial Partner ## 4 Discussion | | ∗∗∗ | | | | | | | 2Bar -Temp +Sim | | | | | | | − | | | | | | | .

c18ee0ab-41af-4c13-8627-25e96c11daa1

# If Turing Played Piano With An Artificial Partner ## 4 Discussion | | | | | . | 490 ( | . | 370) | | | | − | | | | | | | 1 | . | 116 ( | . | 489)

5bb11a12-7a49-47b6-9ed4-2856f9d02f19

# If Turing Played Piano With An Artificial Partner ## 4 Discussion | | 1 | . | 116 ( | . | 489) | | | ∗ | | | | | | | − | | | | | | | .

534b60e8-db80-487a-9c6a-4dbce67cba3d

# If Turing Played Piano With An Artificial Partner ## 4 Discussion | | | | | | . | 518 ( | . | 349) | | | | 2Bar +Temp -Sim | | | | | | | − | | | |

e4e86bee-446d-4a84-8b40-1f29af1691ac

# If Turing Played Piano With An Artificial Partner ## 4 Discussion | | − | | | | | | | 2 | . | 016 ( | . | 409) | | | ∗∗∗ | | | | | | | − |

176b9871-6692-4e12-bddf-0828e304008b

# If Turing Played Piano With An Artificial Partner ## 4 Discussion | | | | | − | | | | | | | . | 906 ( | . | 460) | | | | ∗ | | | |

5102da32-8925-4a76-b7ed-a8345d822bf8

# If Turing Played Piano With An Artificial Partner ## 4 Discussion | | ∗ | | | | | | | − | | | | | | | 2 | . | 360 ( | . | 487) | | | ∗∗∗

824794ac-9bad-458b-adf3-2249bd980410

# If Turing Played Piano With An Artificial Partner ## 4 Discussion | 360 ( | . | 487) | | | ∗∗∗ | | | | | | | 2Bar -Temp -Sim | | | | | | | − | | | |

27334c21-fc0f-44a0-ba1e-38912d22402d

# If Turing Played Piano With An Artificial Partner ## 4 Discussion | | − | | | | | | | 1 | . | 753 ( | . | 369) | | | ∗∗∗ | | | | | | | −

e6cf64c0-42a0-4023-88f7-9cd0496640a2

# If Turing Played Piano With An Artificial Partner ## 4 Discussion | | | | | | − | | | | | | | . | 906 ( | . | 355) | | | | ∗ | | |

aa09a54b-49b8-4514-85bf-862a916a88b7

# If Turing Played Piano With An Artificial Partner ## 4 Discussion | | | ∗ | | | | | | | − | | | | | | | 2 | . | 203 ( | . | 440) | | | ∗∗∗

e0104158-87d8-4076-8989-8a557d1b5b35

# If Turing Played Piano With An Artificial Partner ## 4 Discussion | 203 ( | . | 440) | | | ∗∗∗ | | | | | | | 2Bar +Temp +Sim | | | | | | | − | | |

334ef21f-4cc0-4787-bcce-45b85f10a4a5

# If Turing Played Piano With An Artificial Partner ## 4 Discussion | | | − | | | | | | | . | 805 ( | . | 346) | | | | ∗ | | | | | | | −

d52a44b1-1355-4b06-9549-bc1ecdaf11b1

# If Turing Played Piano With An Artificial Partner ## 4 Discussion | | | | | | − | | | | | | | . | 906 ( | . | 448) | | | | ∗ | | |

3fdfabde-1d21-4fb4-9cc4-218574d73ec0

# If Turing Played Piano With An Artificial Partner ## 4 Discussion | | | ∗ | | | | | | | − | | | | | | | . | 413 ( | . | 378) | | |

12327e0a-3576-4126-8b1d-0c94fc7ee38b

# If Turing Played Piano With An Artificial Partner ## 4 Discussion | . | 378) | | | | 4Bar -Temp +Sim | | | | | | | − | | | | | | | 1 | . | 227 ( | .

b9a1f956-325a-406c-8f4b-ad8ae2bb9a32

# If Turing Played Piano With An Artificial Partner ## 4 Discussion | | | 1 | . | 227 ( | . | 376) | | | ∗∗ | | | | | | | − | | | | | |

2b8b2c45-27b0-42d7-916f-ab5b7f120507

# If Turing Played Piano With An Artificial Partner ## 4 Discussion | | | | | | 1 | . | 222 ( | . | 486) | | | ∗ | | | | | | | − | | |

39788e4a-6120-4f20-b4d9-0b32259c8c48

# If Turing Played Piano With An Artificial Partner ## 4 Discussion | | | − | | | | | | | 1 | . | 308 ( | . | 460) | | | ∗∗ | | | | |

a95d8fb5-8f9c-4922-9523-efa73a8be41e

# If Turing Played Piano With An Artificial Partner ## 4 Discussion | | | | | | 4Bar +Temp -Sim | | | | | | | − | | | | | | | 1 | . | 858 ( | .

9caa4219-b3d0-418c-9057-82bbb291ef57

# If Turing Played Piano With An Artificial Partner ## 4 Discussion | | | 1 | . | 858 ( | . | 425) | | | ∗∗∗ | | | | | | | − | | | | |

b98884b9-2e7b-490c-b389-f6e012940401

# If Turing Played Piano With An Artificial Partner ## 4 Discussion | | | | | | 1 | . | 064 ( | . | 423) | | | ∗ | | | | | | | − | |

9475064d-f491-4ff9-a7e5-f45c2b332461

# If Turing Played Piano With An Artificial Partner ## 4 Discussion | | | − | | | | | | | 2 | . | 045 ( | . | 437) | | | ∗∗∗ | | | | |

0ccb18dd-dfe3-489e-b156-063c9675133a

# If Turing Played Piano With An Artificial Partner ## 4 Discussion | | | | | | | 4Bar -Temp -Sim | | | | | | | − | | | | | | | 1 | . | 700 (

cd5c429c-aaa1-49b7-9780-b83664413136

# If Turing Played Piano With An Artificial Partner ## 4 Discussion | | | 1 | . | 700 ( | . | 382) | | | ∗∗∗ | | | | | | | − | | | | |

2d88cff1-1e33-4629-bdad-7a99e3281a51

# If Turing Played Piano With An Artificial Partner ## 4 Discussion | | | | | | | 1 | . | 222 ( | . | 380) | | | ∗∗ | | | | | | | − | |

ac2a0aa4-c98c-420d-b3c4-b461c0ef9d25

# If Turing Played Piano With An Artificial Partner ## 4 Discussion | | | | − | | | | | | | 1 | . | 782 ( | . | 431) | | | ∗∗∗ | | | | |

c5779dd5-14a5-4963-9710-a5ab82d96a3a

# If Turing Played Piano With An Artificial Partner ## 4 Discussion | | | | | | | 4Bar +Temp +Sim | | | | | | | − | | | | | | | 1 | . | 490 (

0796f8b1-f801-4a30-b2bc-9d84642682e6

# If Turing Played Piano With An Artificial Partner ## 4 Discussion | | | | 1 | . | 490 ( | . | 423) | | | ∗∗∗ | | | | | | | − | | | | |

9921a5f2-8d3b-411b-85a7-86f7ae57496a

# If Turing Played Piano With An Artificial Partner ## 4 Discussion | | | | | | | . | 801 ( | . | 410) | | | | − | | | | | | | 1 | . | 255 (

e0ea3fc3-93e0-488d-b6a4-e7e39617ed57

# If Turing Played Piano With An Artificial Partner ## 4 Discussion | | | | 1 | . | 255 ( | . | 454) | | | ∗∗ | | | | | | | (d) Enjoyable | (e) Self-other Integration | (f) Flow State Scale (short) | | | | | Baseline (Human) | 5

8ed2e31c-026c-4d5c-98f5-15678bd45c08

# If Turing Played Piano With An Artificial Partner ## 4 Discussion (d) Enjoyable | (e) Self-other Integration | (f) Flow State Scale (short) | | | | | Baseline (Human) | 5 | . | 036 ( | . | 296) | | ∗∗∗ | | | | | | | 3 | . | 472 ( | . | 273)

e351227e-04d2-4a04-a371-42eb03dd82be

# If Turing Played Piano With An Artificial Partner ## 4 Discussion | | 3 | . | 472 ( | . | 273) | | | ∗∗∗ | | | | | | | 3 | . | 964 ( | . | 159) | | | ∗∗∗ |

0d8eef8e-2b57-4425-af35-5401248314f7

# If Turing Played Piano With An Artificial Partner ## 4 Discussion | . | 159) | | | ∗∗∗ | | | | | | | 2Bar -Temp +Sim | | | | | | | − | | | |

89d085d6-0e6f-45b3-bd11-6ccc6506d49a

# If Turing Played Piano With An Artificial Partner ## 4 Discussion | | − | | | | | | | 1 | . | 299 ( | . | 397) | | | ∗∗ | | | | | | | − |

4817eda3-8a05-4899-9ebf-270a11a7c1c8

# If Turing Played Piano With An Artificial Partner ## 4 Discussion | | | | | − | | | | | | | . | 419 ( | . | 385) | | | | − | | | |

ff2a9e49-9f0d-4c46-907f-e69e1a9b1a68

# If Turing Played Piano With An Artificial Partner ## 4 Discussion | | − | | | | | | | . | 051 ( | . | 146) | | | | 2Bar +Temp -Sim | | | | | | | − |

1e9c21c4-ee14-4e53-aadc-98fc87e2897a

# If Turing Played Piano With An Artificial Partner ## 4 Discussion | | | | | − | | | | | | | 1 | . | 562 ( | . | 497) | | | ∗∗ | | | |

cb315b8f-cde3-4ac2-99f3-cb6f031e2238

# If Turing Played Piano With An Artificial Partner ## 4 Discussion | | ∗∗ | | | | | | | − | | | | | | | 2 | . | 683 ( | . | 418) | | | ∗∗∗ |

0df93861-b6f8-4616-a8d4-a7891e488f6b

# If Turing Played Piano With An Artificial Partner ## 4 Discussion | . | 418) | | | ∗∗∗ | | | | | | | − | | | | | | | . | 168 ( | . | 131) |

f431cbc7-c7c4-4d9a-83d7-24d94de4ed50

# If Turing Played Piano With An Artificial Partner ## 4 Discussion | | . | 168 ( | . | 131) | | | | 2Bar -Temp -Sim | | | | | | | − | | | | | | | 1 | .

8fabbc77-1d18-4fcf-8ca7-bd78d8d42736

# If Turing Played Piano With An Artificial Partner ## 4 Discussion | | | | | 1 | . | 667 ( | . | 479) | | | ∗∗∗ | | | | | | | − | | | |

b3a6134f-6fd7-459d-bbb2-6382f0a2dc3c

# If Turing Played Piano With An Artificial Partner ## 4 Discussion | | − | | | | | | | 2 | . | 577 ( | . | 316) | | | ∗∗∗ | | | | | | | − |

a46d5eb9-f46e-4a63-a0df-d0485c494ba9

# If Turing Played Piano With An Artificial Partner ## 4 Discussion | | | | | − | | | | | | | . | 344 ( | . | 126) | | | | ∗∗ | | | |

f4f6706e-63ee-4f8b-a3d1-6cb81098e809

# If Turing Played Piano With An Artificial Partner ## 4 Discussion | | ∗∗ | | | | | | | 2Bar +Temp +Sim | | | | | | | − | | | | | | | . |

3cd7b9e5-58ca-4329-9abe-a9d1535f1ef8

# If Turing Played Piano With An Artificial Partner ## 4 Discussion | | | | | . | 930 ( | . | 380) | | | | ∗ | | | | | | | − | | | |

9a2a95c3-5fb0-45bb-887f-9be4f87fbefb

# If Turing Played Piano With An Artificial Partner ## 4 Discussion | | − | | | | | | | . | 261 ( | . | 466) | . | 030 ( | | 4Bar -Temp +Sim | | | | | | | − |

faac4253-1eb9-4109-a6db-8a62b49c8559

# If Turing Played Piano With An Artificial Partner ## 4 Discussion | | | | | − | | | | | | | 1 | . | 193 ( | . | 372) | | | ∗∗ | | | |

388f21cc-ee39-40d4-8c41-39a039bd78d8

# If Turing Played Piano With An Artificial Partner ## 4 Discussion | | ∗∗ | | | | | | | − | | | | | | | 1 | . | 630 ( | . | 339) | | | ∗∗∗ |

c86ba80f-1b88-4524-83f9-ed9a7abb5987

# If Turing Played Piano With An Artificial Partner ## 4 Discussion ( | . | 339) | | | ∗∗∗ | | | | | | | − | | | | | | | . | 133 ( | . | 133) |

82f0a5a8-e5bb-464a-a2e7-c9c23fd193c1

# If Turing Played Piano With An Artificial Partner ## 4 Discussion | | . | 133 ( | . | 133) | | | | 4Bar +Temp -Sim | | | | | | | − | | | | | | | 1 | .

1526a2c0-def7-4a31-9f2e-3de506041139

# If Turing Played Piano With An Artificial Partner ## 4 Discussion | | | | | 1 | . | 930 ( | . | 488) | | | ∗∗∗ | | | | | | | − | | | |

4cd8a846-f68e-46e7-89bf-7213821268e7

# If Turing Played Piano With An Artificial Partner ## 4 Discussion | | − | | | | | | | 1 | . | 840 ( | . | 461) | | | ∗∗∗ | | | | | | | −

ae793b3f-dcc8-4a2d-98d2-5aef0783d5c8

# If Turing Played Piano With An Artificial Partner ## 4 Discussion | | | | | | − | | | | | | | . | 204 ( | . | 137) | | | | 4Bar -Temp -Sim | | | |

7c5acc77-3529-4dd8-a029-e3011d7fe58b

# If Turing Played Piano With An Artificial Partner ## 4 Discussion | | 4Bar -Temp -Sim | | | | | | | − | | | | | | | 1 | . | 457 ( | . | 470) | | | ∗∗ |

1e265734-7534-40f8-a6b9-fbbc9559ef08

# If Turing Played Piano With An Artificial Partner ## 4 Discussion | 457 ( | . | 470) | | | ∗∗ | | | | | | | − | | | | | | | 2 | . | 209 ( | . | 403)

c0759bc3-5387-48f5-8f36-3a19e0b23ea0

# If Turing Played Piano With An Artificial Partner ## 4 Discussion | | 2 | . | 209 ( | . | 403) | | | ∗∗∗ | | | | | | | − | | | | | | | .

3156df0f-f710-4345-b646-1c38eb8b92ff

# If Turing Played Piano With An Artificial Partner ## 4 Discussion | | | | | | . | 221 ( | . | 132) | | | | 4Bar +Temp +Sim | | | | | | | − | | | |