Intelligent Support Engages Writers Through Relevant Cognitive Processes

Writing is a multifaceted skill, crucial for various activities such as learning, practicing, and teaching [21, 23]. Given the cognitive demands of writing, which include memory capacity [16], effective utilization of cognitive resources is key for sustained performance [15]. The need for effective writing practices has been emphasized by the pervasive issue of writer’s block, i.e., the inability to come up with or implement new ideas for continuing writing [40].

Traditional writing support tools were limited in scope, primarily focusing on spelling, grammar, and style [14]. Their development has been influenced by a long history of research and a focus on a collaborative relationship between the tool and the user [3, 32]. These tools gradually evolved to include more complex tasks such as paraphrasing and offering feedback [2]. Despite the advancements, these tools were limited in their ability to address varied writing processes crucial for effective writing [22, 45].

The advent of LLMs marks a turning point in the development of intelligent writing support tools. Advances in model architectures like Transformers have dramatically improved the capabilities of LLMs [7, 8, 46]. This technological leap has broadened the scope of applications for writing support tools, making them even more "intelligent" [9].

LLMs like the GPT-3.5 Turbo model we use in our research [7] have been employed across various domains [31, 38], from creative writing to language learning [12, 26, 30, 33, 51]. The main advantage of tools incorporating LLMs over their predecessors is the added intelligence that enables them to assist in ideation, elaboration, and even the creative process [13, 26].

Modern LLMs offer new features that build upon traditional spell and grammar-checking capabilities. They have evolved to perform tasks like paraphrasing, continuation, and elaboration [2, 13]. Additionally, LLMs are capable of providing adaptive feedback [47], offering creative support [17], and potentially equalizing outcomes for groups that have traditionally faced writing challenges [6].

However, there are challenges to surmount. Issues such as inconsistent style, unsatisfactory text generation, and the need for human intervention remain [26, 44]. However, Gero et al. [20] emphasize that while LLMs show promise, they inherit significant challenges such as generating misleading information and exhibiting difficulty in user-driven steering, underscoring the need for nuanced understanding and careful application in writing support tools. Furthermore, integrating intelligent features like ideation and elaboration into tools requires careful consideration of cognitive processes and existing literature [14, 18]. If we can find a way to overcome these challenges, incorporating LLMs in writing support tools promises a more sophisticated, adaptive, and intelligent future for assisting writers.

The principles underlying our intelligent writing support tool, guided by the Cognitive Process Theory of Writing, are universally applicable across various writing domains. However, the specific focus on student peer review writing in our study serves as a practical example to illustrate these principles in action. This context provides a unique setting to explore and address common writing challenges, such as high effort and writer’s block, which are prevalent in peer review scenarios.

In emphasizing student peer review, we aim to demonstrate the effectiveness of our tool in a concrete scenario, while acknowledging its potential applicability in broader contexts. The focus on the writing component of peer review is particularly critical, as it often represents a significant barrier to effective peer review. By addressing this key aspect, our study not only contributes to the field of peer review writing but also sets a foundation for applying these insights to other writing domains, leveraging the general scope of the Cognitive Process Theory of Writing.

We turn to writing theory to overcome the challenges in exploiting the powerful capabilities of LLMs. Only by doing so can we match the design of our support tools to the actual processes they are striving to assist in. Different from other influential theories on writing [23], Cognitive Process Theory [16] highlights not processes that influence the writer but processes in the writer’s mind itself. The theory describes how writing is achieved. Any writing activity is based on three priors: 1. the writer’s long-term memory, which includes information on the topic, audience, and reason to write; 2. the text produced so far; and 3. the rhetorical problem currently to be solved, i.e., what has to be achieved to address the reason to write.

These priors inform three distinct subprocesses: 1. planning, which includes the development of concrete writing goals, ideation on what writing content can address them, and how it should be organized; 2. translating, sometimes called transcribing, which is the act of finding formulations for the content and writing them down; 3. reviewing, which consists of evaluating and revising text produced so far. The writing process cognitively consists of dynamically switching between these three subprocesses. An additional subprocess, namely monitoring, conducts switching.

These processes can be grouped into two distinct classes: 1) gathering new information, and 2) applying the information to produce observable results. In the context of the Cognitive Writing Process Theory, these classes are called exploration and exploitation, respectively [29]. An overview of the theory and how it is applied to the design reported in this paper can be seen in Figure 1.

Paying attention to these naturally occurring cognitive processes can help improve beneficial and preventing detrimental effects of writing support tools. For example, review writers sometimes conflate generated suggestions with the text written so far [27]. This can lead to the introduction of opinions not aligned with those of the writer and factual inaccuracies in the final text. Additionally, it is important to avoid distracting writers with suggestions at the wrong moment, as this may impact writing quality: Since pre-writing pauses are a good predictor for writing quality [4]. This is because, in these pauses, planning is allowed to occur. More generally, the theory has been applied in a recent review classifying writing tools based on the cognitive process they address [19].

Combining the Cognitive Process Theory of Writing with the new capabilities emerging from LLMs allows us to determine how intelligent writing support will likely impact specific writing processes. Writing Priors: Intelligent writing can be prompted using the three priors identified in the Cognitive Process Theory: 1. long-term memory information, i.e., why is what written for whom, and 2. the text as it exists so far. It may be possible to use intelligent writing to suggest which 3. rhetorical problem is currently to be solved. If the text to be written is very short, these three variables collapse into one: as the reason for writing and the current rhetorical problem become congruent, and there may be no text written so far, or, in experiments, it could be held constant, as we did.

Planning: Intelligent writing can be used to identify subgoals that can be implemented immediately. In the planning-subprocess of generation, it can help with ideation [11]. More advanced models can advise on the structure of the text [8]. If we again assume a very short text, goal setting collapses with the priors as long as the rhetorical problem is concrete enough to implement immediately. If someone gives instructions on what to write, content is given, and the generation process is removed. If the text is very short, reorganizing is not an applicable subprocess of planning because there is no room for maneuver.

Translating: LLMs can be used in the translating phase. An example of such a system is IntroAssist, which includes a checklist of best practices, highlighted text functionality and annotated examples to guide users in writing help requests [25]. Generally, they can be used to rephrase or elaborate on ideas, depending on the rhetorical problem and writing goal. This will include style. While it is feasible to pay no attention to depth and style, this does not remove the process. It simply carries it out poorly. Therefore, the translating variable must be kept even in a minimal case.

Reviewing: LLMs can evaluate text and suggest revisions (at least it should if prompted correctly). An example of such a system is AL, which analyzes the text the user provides to the system and identifies the level of argumentativeness and persuasiveness of the text while providing insights to the user to improve the content further [47]. Reviewing is often omitted in a minimal case, such as brainstorming or chatting.

We therefore assumed the following a priori: Because intelligent writing support aids in exploration processes, there will be an impact on the time spent on writing tasks. Valid arguments can be made that this could be more or less time. For more time there can be several reasons. Namely, the stimulating nature of quasi-collaborative support may increase engagement with the writing task. Reduced opportunities for failure may lead to less satisfaction. For less time, we could take the argument that intelligent writing support can substitute cognitive processes, increasing time efficiency by making certain processes redundant. As the theoretical picture was unclear beforehand, we entertained both possibilities and assumed an undirected overall group difference.

Besides time, there is also the question of whether intelligent writing support is taken advantage of. The Cognitive Process Theory of Writing posits that individuals transition from one process to another, utilizing the outcomes of the previous process in the subsequent one. Based on this theory, we assumed that assistance for a specific process would likely be sought only when no existing results exist.

In the remainder of this paper, we will outline the operationalization of these hypotheses. We will report on our results and explanations for unexpected outcomes. Finally, we will discuss the implications of these results for using the Cognitive Process Theory of Writing in concert with intelligent writing support systems.

Following recent calls for increased standardization of experimental tasks [19], we use student review writing in a 2x2 between-group design. The thrust of the argument is fixed to be in favor of providing feedback. The two binary factors are the presence or absence of a) relevant example ideas/feedback suggestions for improvement and b) intelligent writing support in a button. The button produces example ideas/feedback suggestions using API calls to the LLM GPT-3.5-turbo¹. We used prompts incorporating the text written so far. It was structured as a chat history, providing the model with example outputs to constrain generations².

Depending on random group assignments (uniform sampling without replacement), in the ideation phase, participants were supported in the argumentative essay task with content to use in their argument and/or intelligent support with ideation. In the reviewing phase, participants were supported with evaluations that suggested how to revise the text. Participants moved on from the first to the second phase after submitting their text by clicking on the Submit button, which was available after at least 250 words had been written.

The intelligent support is an implementation consisting of a button and an output field where generated suggestions are displayed. We kept the interface simple, to not accidentally introduce confounding influences on our measurements. Figure 2 shows Group 4 for the ideation task. The evaluation task was set up analogously. Group 3 in both tasks did not receive buttons. Group 2 did not receive the ideas or feedback suggestions on the right side of the tool, and Group 1 received neither.

Besides the independent grouping variable, there were two measures in the tool, which we treat as dependent variables: a) the time needed to complete the essay (250 words) and b) the number of uses of the intelligent support button. The descriptive statistics for these main variables can be seen in Table 1.

In addition, we assessed potential covariates in the pre-and post-survey (See Table 2 for items originally developed for this study). We also assessed the number of cognitive process phases during the task, operationalized by defining exploitation (translating, revision) as the periods where typing was registered for 3 consecutive seconds or less and exploration (ideation, evaluation) where it was registered for more. We used typing as the indicator since it efficiently discriminates writing exploration from exploitation.

In an exploratory data analysis, we determined whether assumptions for parametric tests of our hypotheses were given. They were for hypotheses 1 and 2; for 3 and 4, they were not because of the distribution of the dependent variables. We therefore used non-parametric equivalents for them:

We performed our field experiment over Prolific³. This is a crowdsourcing platform for experiments, and we selected it since previous studies on behavioral research platforms found that Prolific had the highest response quality and sample variety [37], crucial criteria for evaluating crowdsourcing platforms [10, 41, 48]. We recruited 120 participants with age: m=33.27, SD=10.28; gender: 28% female, 72% male; 66.7% indicated at least part-time employment and 28% were students. The selection criterion to be included in the study was fluency in English. Participants were compensated with standard rates if attention checks were fulfilled (which was the case for 4 participants; they were replaced).

Time spent in task 1 and task 2, respectively, were approximately normally distributed, as were the residuals of the linear models with the respective groups. Use of ideation button and evaluation button were Poisson-distributed because these are count data. We, therefore, had to use non-parametric tests for hypotheses 3 and 4. As a quality check for our tool, we assessed technology acceptance variables using Likert scales anchored at 1 (strongly disagree) to 7 (strongly agree), and a middle anchor (neither disagree nor agree), namely intention to use (α =.92, m=5.45, SD=1.34), perceived usefulness (α =.94, m=5.47, SD=1.36), and perceived ease of use (α =.75, m=5.68, SD=1.16). We also analyzed the text submissions. Matching them with the suggestions, we found clear evidence that about 77% of ideas and 34% of evaluation suggestions were incorporated into the submissions; the discrepancy here is likely due to the higher difficulty in detecting the implementation of evaluation suggestions versus ideas, and should not be interpreted as conclusive evidence that ideas are implemented more likely. The submissions, furthermore, did not significantly differ in quality as measured by Text Coherence, defined as cosine similarity between consecutive sentences [5, 34], and did only differ between groups 1 and 3 in task 1 by Fleischman Reading Ease score (See Table 6). Furthermore, we asked participants how important they felt ideas (α =.89, m=5.69, SD=0.93) and evaluations (α =.93, m=5.47, SD=1.16) were in the writing task and how well the tools supported them (ideation: α =.95, m=5.05, SD=1.53; evaluation: α =.90, m=5.13, SD=1.28). There were no group differences for these variables in task 2; however, in task 1, the technology acceptance variables and the variables indicating whether ideation/evaluation was important and supported did differ (see Table 3). Namely, group differences were pronounced between the presence and absence of intelligent writing support, with higher values in the supported groups (See Table 4). Interestingly, this is true for the importance of ideas/evaluations, which were influenced by the experimental variation. Another result was that these differences pertain even to variables that were, on the surface, more relevant for groups in task 2. This may be because participants spent more time on task 1 than task 2, rendering the impact of this grouping more powerful than the grouping for task 2. The cognitive phases were m=34.38 (SD=22.34) for task 1 and m=13.32 (SD=8.75) for task 2.

Figure 3 and Table 1 show group differences relevant to the hypotheses. In terms of hypothesis testing, we can report the following findings:

Overall, these results indicate a difference between having and not having access to intelligent writing support. Furthermore, it indicates a difference within the groups that received writing support, namely that it was used much more if no product of the relevant cognitive process for the instructed task was present beforehand.

Our results indicate that participants were writing for longer time periods when intelligent writing support was present — a mean difference of 100 seconds for ideation and 30 seconds for evaluation. However, this was only the case during the evaluation phase when no predefined evaluations were shown. This suggests that tool engagement depended on intelligent writing support and the absence of pre-displayed evaluations. Participants used intelligent writing support less when ideas and evaluations were already displayed, showing a 39% and 59% decrease in usage for ideation and evaluation respectively. This supports the Cognitive Process Theory of Writing [16], implying reduced reliance on intelligent writing support when the results of cognitive processes of exploration are substituted with external information.

Our findings shed light on research for the support of the cognitive writing process in general and on HCI research on usability and user experience with intelligent writing support tools specifically. We believe this supports the notion that the Cognitive Process Theory of Writing provides design-determining constructs, namely what cognitive processes are used can therefore be supported during writing, for HCI in the domain of writing support [24].

Furthermore, we observed that intelligent writing support may play a role in increasing writing engagement. Namely, more time was voluntarily spent on the tasks when intelligent writing support was present, possibly indicating higher intrinsic motivation to submit high-quality writing. Hence, HCI researchers and practitioners can build on our research to study how different writing phases (planning, translating, and reviewing) can be supported in different writing domains (professional writing, educational writing, or creative writing).

In the ideation and evaluation tasks, measurements varied by group. Specifically, group 4 of the evaluation task, with evaluations and a generation button, had fewer processing phases (m=10.81, SD=5.5) than the overall mean (m=13.32, SD=8.75). This suggests the evaluations provided might have sufficed for task completion. Post-survey variable differences were found only in task 1 (ideation), indicating greater ease of use for groups with intelligent writing support. We speculate that task duration might impact covariates more than the experimental variations.

We also expand on previous research on the impact of writing support on the rhythms of writing [43]. Namely, the introduction of intelligent writing support impacts time and the number of distinct process phases. Future studies may extend this paradigm to more than two processes per task until the monitor activity, i.e., the switching between processes in natural writing tasks, is fully accounted for. For this, we call for investigations into the operationalization of all particular processes; these operationalizations should extend our measuring, focusing on whether typing occurred. This approach is feasible in very controlled circumstances when only two processes (exploration and exploitation) are expected. This controlled setting helps in isolating the effects of the writing support tool, offering clearer insights into its direct impact on the writing process.

Basing the study of new phenomena on previous insight can help overcome the uncertainty caused by them. We used an established theory on how writing works on the cognitive side to help us understand how the cognitive automation driven by intelligent writing support can be understood in the context of writing. Using such a theoretical approach, we could integrate existing knowledge with this new phenomenon, which facilitated studying its application for writing support. It is a use case important to those looking to aid underperforming demographics and those looking to reinvent writing as a practice. Writing has been reinvented by new technology several times; only recently, handwriting was largely replaced by digital writing. This time, the changes may seem more paramount; however, by using theory to inform it, we can possibly steer the practice much better than ever before.

Our research may help users better understand the impact of collaborating with intelligent writing support. For designers, our research may help guide the configuration of writing support systems, as we show that cognitive writing processes ought to be included when considering how systems will be used. We can emphasize this for the role of time spent on writing and the actual use of the system.

Our users’ feedback can be used to improve future intelligent writing support systems for both research and real-world applications. Users emphasized the need for increased effectiveness in guidance during the writing process and expressed the desire for a more comfortable interface. By integrating features that promote clear and concise suggestions while maintaining user-friendly navigation, future designs can better align with the practical requirements of writers. Balancing these factors can lead to a more engaging and productive writing experience, which supports the observed positive impact on time spent on writing and actual system usage in our study.

Some of the feedback of the experiment participants pertained to study specifics that may have influenced our results, namely the 250-word minimum requirement that caused discomfort for some participants. This constraint may have affected the natural flow of the writing process and potentially altered the way users interacted with the intelligent writing support tool. Another limitation of this study is that we did neither explicitly ask for prior experience with similar tools, nor run the study longitudinally, leaving the possibility open that part of the tool’s effects were related to novelty. Understanding these study-specific limitations can inform future research designs, allowing for a more authentic assessment of how writers engage with intelligent support systems in unconstrained writing scenarios.

Our study has further limitations suggesting avenues for future research. Firstly, the scope is confined to the specific domain and participants studied. Despite pre-testing our tool and implementing attention checks, potential data invalidity may arise from Prolific participants, which furthermore are students, which is why our results may not generalize to professional writers. Secondly, we introduced an intelligent writing support system using GPT-3.5-turbo, which may produce biased or erroneous results due to inherent model constraints. However, future improvements in these models could mitigate such limitations. Thirdly, while we recognize the ethical concerns of intelligent writing support, especially in academic writing [52] and unintentional plagiarism [26], they are not the focal point of this paper. We stress the need for future studies to explore biases, expand to diverse populations, and delve into the ethical dimensions of intelligent writing support.

We also want to point out that our users’ feedback extended beyond the core focus of our research, uncovering general interest topics related to intelligent writing support. These include concerns about the potential for over-reliance on automated suggestions [27], curiosity about how AI can foster creativity [17], and interest in ethical considerations [52]. While these areas were not the primary focus of our investigation, they open avenues for future research in the domain of intelligent writing support.