AI as an Active Writer: Interaction strategies with generated text in human-AI collaborative fiction writing Daijin Yang1 , Yanpeng Zhou2 , Zhiyuan Zhang3 , Toby Jia-Jun Li4 and Ray LC3 1 Northeastern University, 360 Huntington Ave, Boston, MA 02115, USA 2 Nanyang Technological University, 50 Nanyang Ave, 639798, Singapore 3 City University of Hong Kong,Tat Chee Avenue, Kowloon Tong, 518057, Hong Kong 4 University of Notre Dame, Notre Dame, IN 46556, USA Abstract Machine Learning (ML) has become an important part of the creative process for human fiction writers, allowing them to utilize various sources of information and be inspired by strategies and data previously seldom explored. To investigate how human writers collaborate with ML systems in fiction writing, we prototyped a web-based human-AI collaborative writing tool that allows writers to shorten, edit, summarize, and regenerate text produced by AI. To investigate the dynamics of human-AI interaction in fiction co-writing, we used a "finish each other’s story" approach where humans and machines took turns writing collaborative fiction. In results from a preliminary study with 9 users, we found that users took inspiration from unexpected text generated by the machine, that users expected reduced fluency and coherence in the machine text when allowed to edit the output, and that they perceived a mental model of the AI as an active writer in the collaborative process rather than simply as a passive AI writing assistant. This study provides design implications on supporting co-creative writing of humans and machines. Keywords Applications of intelligent user interfaces, Collaborative interfaces, User Modelling for Intelligent Interfaces, Evaluations of intelligent user interfaces - Reproducibility 1. Introduction ing how users perceive the AI used for text generation and how users interact with AI in the creative writing The rapid development of machine learning has made process[18, 19, 20]. Most designs consider collaborative it possible for artificial intelligence (AI) to collaborate creative writing systems with AI as the user’s assistant, with humans to generate creative content [1, 2, 3, 4, 5, 6]. such as supplementing the user’s unfinished sentences or Human-AI collaborative creative systems based on ma- providing users with suggestions for writing [10, 11, 13]. chine learning have been gradually entering people’s We seek to explore how an AI system may play a more creative artistic life such as music composition [6, 7, 8], active role in co-creative writing. Specifically, we explore creative illustration [1, 9], and co-writing [10, 11]. These what interactive capabilities users actually need when human-AI collaborative creation systems can assist expe- co-creative writing with AI, and how these capabilities rienced creators by inspiring them with new ideas and affect the writing co-creation experience. providing suggestions [12, 13]. They can also bring a To ground our study, we prototyped a collaborative novel creative experience to users who have no or little writing system with a web interface for human-AI co- creative experience, such as completing the drawing that writing. In this system, users and the machine take turns the user has started or automatically filling in the user’s writing paragraphs for each other to continue with. The unfinished sentence [1, 10]. In this article, we focus on system has two different modes, the "Edit Mode" and the the needs of users when they collaborate with AI for "No Edit Mode". In our preliminary study with 9 users, creative writing. each user was first asked to write the beginning of a Recent work is focused on improving the algorith- sci-fi story about human beings finding new homes. A mic performance of natural language generation models, GPT-2-based language model fine-tuned to a sci-fi theme such as improving the logic of generated text [14, 15] generates follow-up paragraphs of the story based on or making the generated text closer to the natural lan- what users have written. Before continuing writing, users guage [16, 17]. However, little work focuses on explor- could choose to regenerate or select from multiple ver- sions of machine-generated texts. The machine would Joint Proceedings of the ACM IUI Workshops 2022, March 2022, Helsinki, Finland consider changes made by users into account for its next $ yang.dai@northeastern.edu (D. Yang); zydstd@gmail.com generation. In each study sessions, the user and the AI (Y. Zhou); zzhang452-c@my.cityu.edu.hk (Z. Zhang); tool finished a 5 paragraph sci-fi story together, with 2 toby.j.li@nd.edu (T. J. Li); LC@raylc.org (R. LC) paragraphs generated by the AI, and 3 paragraphs written © 2022 Copyright for this paper by its authors. Use permitted under Creative Commons License Attribution 4.0 International (CC BY 4.0). by the user. CEUR Workshop Proceedings http://ceur-ws.org ISSN 1613-0073 CEUR Workshop Proceedings (CEUR-WS.org) Figure 1: The architecture of the prototype human-AI co-writing tool used in this study. First, the story head is written by humans and entered into the fine-tuned GPT-2. The user then judges the text generated by the machine and decides whether to regenerate it. After that, the text generated by the machine is modified by the user as the final machine text. The user follows the story development of machine text to write. Finally, both machine text and human text are used as input for the next machine generation. By observing 9 users’ writing process in two modes, in- In these systems, AI collects the user’s input information terviewing about their experiences in the co-writing pro- as to its output condition or predicts the user’s true in- cess, and analyzing their written stories, we concluded tention based on the user’s feedback.The output of our our main findings as follows: human-machine collaborative innovative writing system is influenced by users’ input, which is consistent with 1) We find the patterns of texts in Human-AI col- previous works. However, our system comes with its laboratively written stories: The AI-enabled tool own consideration of the plot, while following the user’s served as a good unexpected twist provider but writing wishes. not a fully competent writer. Compared with the difficulty of human innovation in 2) We discover users with different writing inten- story writing, machines cannot fully understand the in- tion and in different interactive modes (allowing tentions of human writing, so they are more likely to editing versus not allowing editing) had different create unexpected plots and drive the development of the mental expectation on text coherence and fluency. story[24]. Since the self-attention mechanism is mainly 3) We describe users perceptions of machine’s role used in the current machine language model, the word in the co-writing process and discuss future pos- vector sometimes notices itself thus falls into the looping sibility of writing machine. state of the text [10, 13, 24], especially in the process of generating long text. In addition, the training dataset of Taken together, these findings guide the design of fu- the machine is quite larger than the related knowledge ture Human-AI co-writing interfaces. in the human brain, so it has the potential to generate interesting story text[25, 26, 27, 24]. Human’s logic is 2. Related Work stronger than machine’s, which is necessary for coher- ence in creative writing[24].Especially, Humans have a The recent development of AI enabled extensive applica- much better common sense world understanding than tions that explored the creation of cooperation between language models[20]. Hence, combining text generation humans and AI, including drawing creation [1, 9, 2], cre- language models and human writing for innovative writ- ative writing [10, 11, 3], dance [21] and other fields [22, ing, including text interactive games, writing assistants 23]. For example, Clark et al. conducted a study that and so on, might be a potential way of human-computer explore the use of AI to complete sentences and provide interaction. In a text interactive game[28, 29, 30, 31], the suggestions [13] and Louie et al. built an AI-enabled user controls the character through natural language, tool for creating music [6]. In this line of works, the and the AI agent recognizes the user’s input, intelligently AI acts as the user’s collaborator. It can adjust its out- manipulates the character’s actions in a text-described put according to the goals and actions proposed by the environment, and feeds the results back to the user. AI user and then makes corresponding recommendations. writing assistant is also an important research field of human-AI creative writing. The AI writing assistant can GPT-2’s finetune function was called. The step and the correct users’ spelling and grammatical errors[10], com- learning rate was set as 1500 and 1e-5, respectively. Most plete users’ unfinished sentences or supplement full-text of the science fiction story data in it come from Pulp paragraphs[32], and provide inspiration and suggestions Magazine Archive. for users’ creative writing[13]. 4. Preliminary User Study 3. The Collaborative Writing Tool In this study, we ask the three research questions (RQs) For our study, we prototyped a web-based collaborative below: writing tool where the user can co-write a short sci-fi story with a GPT-2-based text generation model. The tool RQ1: What patterns of interactions are taken up by uses a “turn-taking” approach (Figure 1) where the user humans when they interact with machines in col- starts with writing the beginning the story. The model laborative writing? then continues the story by generating a section that RQ2: How does the ability to select, edit, and cut follows the user’s previous one. The user and the model out machine-generated text affect the human- continue each writing a section in turns until the end of machine co-writing process? the story. The user may also edit the AI-generated section RQ3: How do humans perceive the role of the ma- or regenerate a section when they are not satisfied with chine in the editable vs. non-editable interaction the AI-generated result. modes? 3.1. The Web Interface 4.1. Procedure The web interface of our tool was implemented using the To answer these RQs, with the tool described in Section 3, Django framework. As shown in2, two different inter- we conducted a user study to investigate the dynamics faces are designed for two modes. "Edit Mode" and "No of human-AI interaction in fiction co-writing. The study Edit Mode" both have a "Submit" button for the human uses a within-subject design, where each user had to use user to submit their written text, a "Regenerate" button both the "Edit Mode" (where editing the AI-generated for the AI model to regenerate sentences, and an "End" text was enabled) and the "No Edit Mode" (where editing button to end the story. There is also an "Edit" button for the AI-generated text was disabled) when writing with the user to edit the text generated by the AI model (the the tool. The order of the two conditions was random. In Edit button was disabled in the “No Edit” condition in the each study session, following a short demonstration of study). All history texts will be shown at the top of the the user interface and the theme that they would write page, with human-written texts in black and machine- about, the user was asked to write the beginning of a generated texts in red. The back button allows the user sci-fi story about humans finding a new home in space. to go back to their last operation. Using this beginning, the user then wrote a 5-paragraph When the user clicks on the “Regenerate” button, the story with our tool in the first condition. After this, the model re-selects a random seed, and uses the model to user filled out a usability questionnaire and had a 5–10- generate its last paragraph. This feature allows the user min semi-structured interview about their experience in to quickly get a new AI-generated paragraph when the the first condition. Similarly, the user then wrote another previous one was not desirable, such as when the model story, filled out the questionnaire, and had an interview fails to generate readable text, generates repetitive text, in the second condition. or switches topics abruptly. We post advertisement on the university’s bulletin and recruited nine participants (n=9) for our study, later 3.2. The Text-Generation Model referred to U1 to U9 in this paper. Participants were Our prototype tool uses a GPT-2 language model that all graduate students who were interested in human- was fine-tuned to a sci-fi theme. GPT-2 is a super-large- machine co-writing. 5 of them were 18–25 years old, and scale language model proposed by OpenAI in 2019 [33, 4 of them were 26–35 years old. 5 of them used English as 34]. We used the “medium” version of GPT-2 with 355M their first language, and 4 of them used Chinese as their parameters. In order to adapt the style of generated text first language. 8 of them were males, and 1 of them was to the sci-fi domain used in the study, we fine-tuned female. All users had some creative writing experiences. GPT-2-medium to the field of science fiction. We used 2 were novices, 3 had intermediate-level of experiences, the Sci-Fi Stories Text Corpus [12] collected by Robin and 4 were experienced fiction writers. Sloan as the dataset for fine-tuning GPT-2-medium. The Each user’s screen recording of their writing process, questionnaire, and interview was recorded and tran- Figure 2: The user interface of our collaborative creative writing tool used in the study.Top Left:The initial interface includes writing prompts, mode selection, theme selection, input box, and submit button. Top Right:In the upper interface, the black font represents the text entered by the user, while the red font represents the text generated by the AI model. After the AI model generates the text, the user can choose to modify, to regenerate, to skip the modification to continue generating, or to end the interaction. Bottom:The user can modify the text generated by the AI model. scribed. Users were asked to think aloud while writ- minutes’ writing, the users would finish the story that ing. One of the experimenters conducted open coding had an average length of 622 words [M=622, SD=109], analysis [35] of the written contents, think-aloud, and in which 320 words were written by the user [M=320, interview transcripts for the qualitative results. SD=106], and 302 words were written by the machine [M=302, SD=28]. In the “No Edit Mode”, it took 4–10 min- utes for users to write another beginning of a sci-fi story 5. Descriptive Statistics of the that had an average length of 117 words [M=106, SD=51]. Stories After that, users spent 20–40 minutes completing the the story with an average length of 599 words [M=599, In the “Edit Mode”, users usually spent 4–10 minutes SD=127], where 282 words were written by users [M=282, writing a beginning of a sci-fi story that had an aver- SD=125], and 317 words were written by the machine age length of 117 words [M=117, SD=56]. After 20–30 [M=317, SD=31]. Figure 3: The Task flow of our study. The user and the AI takes turns to co-write a short science-fiction story. The user starts with writing the beginning (Paragraph 1) of a sci-fi story of human beings finding new homes. The AI model generates the Paragraph 2, which the user can regenerate, select, or edit the texts (in the “Edit Mode” condition). The user and the AI repeats this process to write Paragraphs 3 and 4 and finally the user ends the story with the Paragraph 5. Figure 4: Screenshot of the experiment. Left: The user shares his screen. At this time the user is editing the text generated by the machine. Right: The user is in the upper left corner and the other three are researchers. 6. Qualitative Findings 6.1. Story Content In this section, we describe the properties of the co- By coding the stories, we found that new twists, includ- written story and user’s strategies for co-writing with ing new characters, new scenes, and new events were the AI-enabled tool. Specifically, we investigated what more frequently found in AI-generated texts (after re- was their reaction towards the AI-generated texts in two generation, if any) than user-written texts. For example, different modes, how AI-generated texts affected their in U4’s "Edit Mode" story. U4 only mentioned "We" as a creative writing process, and how they perceived the new character, "uncertain terrain" as a new scene, "We are partnership between them and the AI-enabled tool in currently approaching a new solar system with a planet two different modes of writing. that seems inhabitable." as a new event in paragraph 1. However, the machine wrote in both paragraph 2 and "a man", "Icter" as additional characters, "winding corridor", performed differently in the same condition, and users "a tunnel", "a dimly lit room" as new scenes, and "walking who had similar expectations also performed differently down", "a man stood in front of me", and "should walk back in two conditions. and tell the others" as new events. Surprisingly, users considered the unexpectedness as the core inspiration or 6.2.1. Reaction to the Coherence of AI-generated reason of continuing their writing, and took good use of Texts the new elements to continue the story, such as U4 wrote "The others look at me inquisitively, wondering what was Users in the II group had lower coherence expectation of in the structure, and glad that I had made it out alright. the AI-generated texts than users in the DI group. And I said ’there was a man.’" in paragraph 3 after reading they all had higher expectation of coherence when they machine texts "should walk back and tell the others" in were in the "No Edit Mode". paragraph 2. Users in the II group prefer the model to generate texts Notably, the AI model would sometimes suddenly that contain some new entities that they could work on. change the positive atmosphere in previous paragraphs For example, any new characters, events, or locations into a negative atmosphere, or the other way around. could be good for them: "I don’t think I said anything For example, U2 wrote an optimistic beginning that "As about a name so I guess it named somebody, which is cool." Commander Barone’s shuttle hummed along, he couldn’t (U4 in "Edit Mode"). help but feel a sense of optimism about humanity’s future. However, users in the DI group were trying to find He had successfully surveyed Planet T74 and was returning something that logically fitted into their story in the AI- back to Space Station Endurance with a cargo hold full of generated texts. For example, some expected subjects samples of rocks, plants, and even some animal life.", but to have logical continuations such as "I guess it depends the machine suddenly turned the story into a negative de- first on what I wrote and then if I think it’s a logical con- scription that "He had been told that there were no known tinuation." (U1 "No Edit Mode"), and refused illogical diseases or parasites on the planet." (U2 "Edit Mode", usercharacters such as "Machine starts spitting out more and and machine wrote in paragraph 1 and 2). more characters that were not mentioned in the scene which Despite having many unexpected elements, user- made it really wonky later on." (U5 "Edit Mode"). Addi- written texts and the selected AI-generated texts were tionally, they wanted the AI model to continue the story coherent with each other. The selected AI-generated texts as they expected, such as "Well I expected the machine often use events, characters, and scenes that were men- to basically take, you know, to see what I wrote and can tioned in user-written paragraphs. For instance, since U2 expand upon it or relate to it in some way that’s what I ex- had written "However, one day an accident at the factory pected." (U5 "Edit Mode"). However, they would be more would force AB67 to do something extraordinary.", the ma- excited if some unexpected items that logically fitted to chine continued the plot with "And the result is this: a their story were found in the generated text, like meeting with an unexpected plot: "And they took it even one step new robot, the first fully-autonomous, self-repairing, self- further with like, Okay, what if you peel off his skin." (U2 replenishing, fully-reactive, self-repairing robot.", and men- tioned the user-made entity "AB67" in "AB67 had been "No Edit Mode"). the first fully self-repairing robot." (U2 "No Edit Mode", In the "Edit Mode", users in both groups would accept user and machine wrote in paragraph 1 and 2). text that contained parts that they could use, like "And if there are some sentences can use, you will definitely work, work, work on it."(U1 "Edit Mode"), or "But I can 6.2. Strategies for Interaction with work with these first three sentences."(U2 "Edit Mode"). AI-generated Texts However, in the "No Edit Mode", users would expect the By coding the think-aloud scripts and interview tran- text to fully meet their expectations on coherence, like "I scripts, we found that users’ reactions to the text gen- think I definitely wanted something that flowed a bit better erated by the AI model and their strategies of utilizing with a story, but with the first one, I was more okay with them can be classified into two different groups by their giving me something that perhaps added new ideas." (U3). expectation of the story: having clear and explicit intent about what they wanted to write (referred to as DI group) 6.2.2. Reaction to the Unexpectedness of and having only implicit implied intent about what they AI-generated Texts wanted to write (referred to as II group). Users were amused when unexpected texts appeared, Most users had concrete ideas about the story, say- even if they presented random events or characters that ing like "As in my mind. Earth is destroyed."(U2 in "Edit had no relationship to what users had written. For ex- Mode"). By contrast, the users who had only implied ample, U2 laughed when he saw his story turned into intent would say like "I don’t think about the ending "(U1 a Christmas story, but regenerated it by saying "This is in "Edit Mode"). Users who had different expectations not a Christmas story." (U2 "No Edit Mode"). The unex- U4 and U8 preferred edits that do not affect the contin- pectedly redundant texts also amused users. For example, uation of the story. In the interaction flow, they preferred U4 laughed when encountering sentences like "I was a to refine on the fluency of the writing after the whole human with a human face." in the story, and U5 laughed story has been generated. U4 said, "I wouldn’t really when saying "That’s a very odd sentence ’the man in the change the story it comes up with but I would just change open suit it wasn’t a woman’, very weird." (U5 "Edit Mode"). or delete a few sentences or something." (U4 "Edit Mode"). Meeting with unexpectedness, users applied some of Although all users agreed that editing was essential, the AI-generated texts that was easy to work with. In U6 preferred not to edit the text because editing was a most situations, redundant texts were too hard to work burden: "In the first mode (’Edit Mode’), I must understand with: "I’m trying to like get some notes that fit a little bit the machine texts and then edit them. But in the second more and gives the idea about how to drive the plot forward mode (’No Edit Mode’), I don’t need to understand them but it seems to like to be redundant." (U5 "Edit Mode"). But and just choose one of my favorite and continue the story." there was one exception in U4’s writing: "I guess like (U6). the only way to make that sentence makes sense (’I was a human with a human face’), is if it wasn’t redundant, the story could be that he didn’t always have a human face." 7. Discussion (U4 "Edit Mode"). Even when the text was not redundant, The language model’s limitation made it unpredictable. it could still be hard to work on when the plot was being It sometimes provides low-quality texts full of words driven forward too quickly: "I’m going to regenerate it that could hardly make sense. At other times, it pro- because it focuses so much on death and yet I don’t want vides high-quality inspirations that move the plot for- it to be like at the start of the story." (U2 "Edit Mode"). ward beyond humans’ expectation. Such unexpectedness However, this situation could be mitigated or even be accounted for the unique interaction pattern of human- useful if the machine wrote the ending: "I feel like it wrote AI co-writing in this study. Corresponding to previous a decent ending on its own and didn’t really want to add quantitative findings [36], the qualitative results in this anything to it." (U5 "No Edit Mode", in delight tone). paper showed that users considered the coherence of the machine-generated texts as a priority. The users’ 6.2.3. Reaction to the Fluency of AI-generated attitudes towards unexpected but coherent elements gen- Texts erated by the AI model further suggested that users ex- Fluency of machine texts was more important in the "No pected the model to provide them with surprising inspi- Edit Mode" than the "Edit Mode". In the "Edit Mode", for rations. However, due to the repetition caused by the most of the users, partial readability would be sufficient model over-confidence problems [24], users could only for the requirement on fluency because "if you’re able to get such paragraphs occasionally by chance. The AI gen- edit it and then it’s less important because you can just fix erating process was not transparent and there was a lack it up a little bit of it." (U1). In "No Edit Mode", the expec- of user control, and thus users could not expect the next tation on fluency becomes as important as the coherence batch of generated text to be better than the previous for most users, such as "But if you can’t (edit), then it’s one. The low probability of getting useful pieces from the kind of more important that it is fluent." (U1). model would frustrate users and make them compromise on the incoherent and tenuous text that conveyed merely 6.2.4. Reaction to Editing inspiration. Nevertheless, the unexpectedness of machine- All users agreed that at least some basic edits of AI- generated texts should be highlighted in an ideal generated texts should be allowed to make them more use- human-AI co-writing tool. After being selected and ful. The most common reason is along the lines of: "This refined by human writers, such unexpected but logical one is definitely harder because oftentimes there would be elements could make the story more exciting than a good amount of it that would be useful and like I would writers’ previous intention. From findings in our study, want to keep writing off of. But then there’s also be sections this could not only serve as a dramatic contradiction like a piece of sentences that were not greatly helpful." (U3 in the story but also as motivations for users to keep "No Edit Mode"). Even if some of the texts in "No Edit writing. In the design of the tool, it would be useful Mode" had high quality and met the basic expectations to facilitate the user’s utilization of the unexpected of users, most users still felt editing was necessary, like elements as their wish (e.g. provide users with options "I think some editing would be required because, You know, of editing machine-generated texts in the system). there’s still some consistencies but not as glaring as that in This could help to reduce the frustration brought by first mode texts." (U5 "No Edit Mode"). unpredictable repetitions and occasional bad fluency. However, even the design of the interaction modes could mitigate such frustration, the algorithm should also be their work in a professional way. The AI should act as an improved. Better quality in AI-generated texts could writing assistant with a customized avatar on call who allow users to focus more on the ideas conveyed by eventually become essential in users’ writing process. AI-generated texts rather than spending most of the time on regenerating and fixing the coherence and fluency of AI-generated texts. 8. Conclusion The users’ different perceptions of AI’s roles in the In this paper, we reported preliminary findings on how 9 co-writing process suggested different interaction pat- users interacted with a “turn-taking” style human-AI co- terns. In the study, most users perceived the machine as writing tool to write short science-fiction stories. We dis- an active idea generator. They preferred the "Edit Mode" covered that different mental expectation of users could more since they could pick what they liked regardless affect their strategies and their perception of the ma- of the fluency and coherence of the texts. Thus, it was chine’s role in the co-writing process. The AI-enabled important to make them able to edit both machine texts tool was used as an active idea generator, a co-writer, and their texts at any time in the writing process. Fur- or a writing assistant in different scenarios by different thermore, more user-controllable variables can be added users. We discussed the challenges in managing the trade- into the tool for them to allow finer-grained user control offs in the desired level of unexpectedness in generated of the generation process. For example, the tool could story plots, the coherence and fluency of AI-generated allow users to control the ideal length of the generated texts, the appropriate level of user-control, and the future text, the scenes, the atmosphere of the plot, or some interface design. weights that could help the model focus more on certain important parts of the user texts. Some users, on the other hand, wished the machine to be a human writing Acknowledgements assistant. In this case, the machine should be able to ac- cept both previous paragraphs and following paragraphs This research was supported in part by a Google Cloud as inputs to connect the user-defined milestones in the Research Credit Grant, a Hong Kong Arts Development story for them. Several works were focused on short Grant, and an Asia Research Collaboration Grant from sentence infilling [37, 38], but long paragraphs infilling Notre Dame International. still remains to be explored. Besides, some users regarded the AI-enabled tool as an active co-writer or a writing exerciser. They tried to References keep the initial output of the machine texts regardless [1] C. Oh, J. Song, J. Choi, S. Kim, S. Lee, B. Suh, I lead, of its coherence and fluency. They enjoyed all the un- you help but only with enough details: Understand- expectedness of the machine-generated texts and wish ing user experience of co-creation with artificial not to intervene in the generating process. In this case, intelligence, in: Proceedings of the 2018 CHI Con- the texts were expected to be uncontrollable. However, ference on Human Factors in Computing Systems, the definition of good quality of texts could be vague 2018, pp. 1–13. in this mode of interaction, since even the redundancy [2] N. Davis, C.-P. Hsiao, K. Yashraj Singh, L. Li, could be interpreted as metaphors. More research should B. Magerko, Empirically studying participa- be conducted to develop a good co-writing or writing tory sense-making in abstract drawing with a co- exercise machine. creative cognitive agent, in: Proceedings of the In summary, the future active AI writing tools should 21st International Conference on Intelligent User strengthen AI’s strong ability of producing high quality Interfaces, 2016, pp. 196–207. unexpectedness. And they should allow users to utilize [3] K. I. Gero, L. B. Chilton, Metaphoria: An algorith- such unexpectedness efficiently. The "Regenerate" and mic companion for metaphor creation, in: Proceed- the "Edit" function mentioned in this paper should be ings of the 2019 CHI Conference on Human Factors the core. The goal of the "Regenerate" is to ensure users in Computing Systems, 2019, pp. 1–12. to find what they want as they wish as fast as possible [4] M. Guzdial, N. Liao, J. Chen, S.-Y. Chen, S. Shah, (reduce the compromises). To accomplish this goal, for V. Shah, J. Reno, G. Smith, M. O. Riedl, Friend, example, the future interface can display multiple outputs collaborator, student, manager: How design of an simultaneously[13, 10] and enable users to control more ai-driven game level editor affects creators, in: Pro- parameters to regenerate the texts. It could also ask users ceedings of the 2019 CHI conference on human to grade the outputs and learn from their writing (use factors in computing systems, 2019, pp. 1–13. users’ inputs as fine-tune datasets)[36]. For the "Edit" [5] J. Koch, A. Lucero, L. Hegemann, A. Oulasvirta, function, the tool should be integrated into a text editor May ai? design ideation with cooperative contex- such as Microsoft Word for users to both edit and save tual bandits, in: Proceedings of the 2019 CHI Con- Conference on Human Factors in Computing Sys- ference on Human Factors in Computing Systems, tems, 2021, pp. 1–13. 2019, pp. 1–12. [19] O. Schmitt, D. Buschek, Characterchat: Supporting [6] R. Louie, A. Coenen, C. Z. Huang, M. Terry, C. J. the creation of fictional characters through conver- Cai, Novice-ai music co-creation via ai-steering sation and progressive manifestation with a chatbot, tools for deep generative models, in: Proceedings in: Creativity and Cognition, 2021, pp. 1–10. of the 2020 CHI Conference on Human Factors in [20] M. Guzdial, M. Riedl, An interaction frame- Computing Systems, 2020, pp. 1–13. work for studying co-creative ai, arXiv preprint [7] M. Suh, E. Youngblom, M. Terry, C. J. Cai, Ai as arXiv:1903.09709 (2019). social glue: Uncovering the roles of deep generative [21] M. Jacob, B. Magerko, Interaction-based authoring ai during social music composition, in: Proceedings for scalable co-creative agents., in: ICCC, 2015, pp. of the 2021 CHI Conference on Human Factors in 236–243. Computing Systems, 2021, pp. 1–11. [22] Y. Lin, J. Guo, Y. Chen, C. Yao, F. Ying, It is your [8] C.-Z. A. Huang, H. V. Koops, E. Newton-Rex, M. Din- turn: collaborative ideation with a co-creative robot culescu, C. J. Cai, Ai song contest: Human- through sketch, in: Proceedings of the 2020 CHI ai co-creation in songwriting, arXiv preprint Conference on Human Factors in Computing Sys- arXiv:2010.05388 (2020). tems, 2020, pp. 1–14. [9] P. Karimi, J. Rezwana, S. Siddiqui, M. L. Maher, [23] A. Elton-Pym, Principles for ai co-creative game N. Dehbozorgi, Creative sketching partner: an anal- design assistants, in: Proceedings of the AAAI Con- ysis of human-ai co-creativity, in: Proceedings of ference on Artificial Intelligence and Interactive the 25th International Conference on Intelligent Digital Entertainment, volume 16, 2020, pp. 335– User Interfaces, 2020, pp. 221–230. 336. [10] A. Coenen, L. Davis, D. Ippolito, E. Reif, A. Yuan, [24] A. See, A. Pappu, R. Saxena, A. Yerukola, Wordcraft: a human-ai collaborative editor for story C. D. Manning, Do massively pretrained lan- writing, arXiv preprint arXiv:2107.07430 (2021). guage models make better storytellers?, 2019. [11] M. Kreminski, M. Dickinson, M. Mateas, N. Wardrip- arXiv:1909.10705. Fruin, Why are we like this?: The ai architecture [25] T. Liu, K. Wang, L. Sha, B. Chang, Z. Sui, Table-to- of a co-creative storytelling game, in: International text generation by structure-aware seq2seq learn- Conference on the Foundations of Digital Games, ing, in: Thirty-Second AAAI Conference on Artifi- 2020, pp. 1–4. cial Intelligence, 2018. [12] R. slogan, Writing with the machine, https: [26] A. Fan, M. Lewis, Y. Dauphin, Hierarchical neural //www.robinsloan.com/notes/writing-with-the- story generation, arXiv preprint arXiv:1805.04833 machine/ (2016). (2018). [13] E. Clark, A. S. Ross, C. Tan, Y. Ji, N. A. Smith, Cre- [27] M. Rose, Rigid rules, inflexible plans, and the stifling ative writing with a machine in the loop: Case stud- of language: A cognitivist analysis of writer’s block, ies on slogans and stories, in: 23rd International College composition and communication 31 (1980) Conference on Intelligent User Interfaces, 2018, pp. 389–401. 329–340. [28] B. Kostka, J. Kwiecieli, J. Kowalski, P. Rychlikowski, [14] C. Shu, Y. Zhang, X. Dong, P. Shi, T. Yu, R. Zhang, Text-based adventures of the golovin ai agent, in: Logic-consistency text generation from semantic 2017 IEEE Conference on Computational Intelli- parses, arXiv preprint arXiv:2108.00577 (2021). gence and Games (CIG), IEEE, 2017, pp. 181–188. [15] A. Krishna, S. Riedel, A. Vlachos, Proofver: Natural [29] T. Atkinson, H. Baier, T. Copplestone, S. Devlin, logic theorem proving for fact verification, arXiv J. Swan, The text-based adventure ai competition, preprint arXiv:2108.11357 (2021). IEEE Transactions on Games 11 (2019) 260–266. [16] W. Fedus, I. Goodfellow, A. M. Dai, Maskgan: better [30] N. Fulda, D. Ricks, B. Murdoch, D. Wingate, What text generation via filling in the_, arXiv preprint can you do with a rock? affordance extraction via arXiv:1801.07736 (2018). word embeddings, arXiv preprint arXiv:1703.03429 [17] Y. Zhang, Z. Gan, K. Fan, Z. Chen, R. Henao, D. Shen, (2017). L. Carin, Adversarial feature matching for text gen- [31] M. Hausknecht, R. Loynd, G. Yang, A. Swaminathan, eration, in: International Conference on Machine J. D. Williams, Nail: A general interactive fiction Learning, PMLR, 2017, pp. 4006–4015. agent, arXiv preprint arXiv:1902.04259 (2019). [18] D. Buschek, M. Zürn, M. Eiband, The impact of mul- [32] A. Calderwood, V. Qiu, K. I. Gero, L. B. Chilton, How tiple parallel phrase suggestions on email input and novelists use generative language models: An ex- composition behaviour of native and non-native ploratory user study., in: HAI-GEN+ user2agent@ english writers, in: Proceedings of the 2021 CHI IUI, 2020. [33] A. Radford, J. Wu, R. Child, D. Luan, D. Amodei, I. Sutskever, et al., Language models are unsuper- vised multitask learners, OpenAI blog 1 (2019) 9. [34] S. Dathathri, A. Madotto, J. Lan, J. Hung, E. Frank, P. Molino, J. Yosinski, R. Liu, Plug and play lan- guage models: A simple approach to controlled text generation, arXiv preprint arXiv:1912.02164 (2019). [35] S. H. Khandkar, Open coding, University of Calgary 23 (2009) 2009. [36] N. Akoury, S. Wang, J. Whiting, S. Hood, N. Peng, M. Iyyer, Storium: A dataset and evaluation plat- form for machine-in-the-loop story generation, 2020. arXiv:2010.01717. [37] C. Donahue, M. Lee, P. Liang, Enabling lan- guage models to fill in the blanks, 2020. arXiv:2005.05339. [38] D. Ippolito, D. Grangier, C. Callison-Burch, D. Eck, Unsupervised hierarchical story infilling, in: Pro- ceedings of the First Workshop on Narrative Un- derstanding, 2019, pp. 37–43.