The Future of AI-Assisted Writing

NLP has seen drastic advancements in recent years with GLMs. One of the recent GLMs is GPT-3: an autoregressive language model with 175 billion parameters, which is ten times more powerful than any previous non-sparse language model (Brown et al., 2020). In all NLP tasks, GPT-3 can be applied without any gradient updates or fine-tuning. The model even excels at tasks that require on-the-fly reasoning or domain adaptation, such as unscrambling words or even utilizing a new word in a sentence. Datasets like Common Crawl which consist of petabytes of data collected over 8 years of web crawling, were included in the GPT-3 training data. Much of GPT-3’s training data was crowdsourced through the use of online text repositories like Wikipedia. As a result, the model can generate samples of news articles that human evaluators have difficulty distinguishing from articles written by humans. These advancements make the possibility of a useful AI-assisted writing tool more reality than fiction. This enables the development of tools that can generate sentences or provide words as suggestions to aid in the writing process.

Research on collaborative computer-assisted writing tools has become more prevalent with GPT-3. For instance, a web application, called SAGA, allows friends to asynchronously collaborate creatively. Through this application, multiple people can contribute to the writing of a story, told partially by GPT-3 (Shakeri et al., 2021). Shakeri et al. explore how to offload the pressure of creative tasks to an AI system. The paper concludes that this allows users to dive deeper into the roleplay experience. However, the paper does not investigate how AI affected the quality of writing. They also did not discuss how the AI impacted the user’s sense of enjoyment and satisfaction. It is important to understand if the use of these systems causes users to feel as though they did not create the work. Some users may feel as though the writing pieces are a product of AI.
Shelley (Yanardag et al., 2021) also develops a GPT-3 based collaborative horror writer that collaboratively writes scary stories with people on Twitter. The application is deployed as a bot on Twitter that generates and posts new stories as tweets. Users can participate in these threads and interact with the stories to produce multiple storylines. To verify whether the generated stories elicit a psychological reaction from readers, Delul et al. utilize measures of effect and anxiety. This work also does not explore the user’s sense of ownership and satisfaction. Furthermore, these works do not explore the impact of the possible biases that can arise due to the use of crowd-sourced data. In both papers, the authors explore the application of GPT-3 to develop collaborative writing tools but fail to address whether users want AI help.
While these works address the role of the tool to offload creative work, they fail to investigate the use of an AI-assisted writing tool in a professional capacity. Google Smart Compose (Chen et al., 2019) is an AI system for generating interactive, real-time suggestions in Gmail and Google Drive products that assist users in writing tasks by reducing repetitive typing. The system leverages state-of-the-art machine learning techniques for language model training which enables high-quality and accurate suggestion prediction. Chen et al. utilize a large amount of e-mail data to train the neural model at the core of GSC. The aim of the work is to increase the efficiency of a writer by using the local context of the text to provide instant suggestions that complete the sentence as users type.

We follow the characterizations of AI-Assisted writing paradigms as defined by Clark et al.(Clark et al., 2018) In this work, the authors explore the machine-in-the-loop system structure where the loop is initiated with the person providing context and the machine responds with a corresponding suggestion. The person always has control over the final writing output. Clark et al. utilize interaction initiation to represent how a context-suggestion loop is triggered. It can follow a push or pull method of initiation, or a combination of the two. A push condition is defined to be when suggestions are provided to the user while they are writing. On the other hand, a pull condition occurs when the user proactively asks for suggestions. We argue that this view best informs the future of AI-assisted Writing due to the increasing feasibility of using GLMs as an information retrieval tool. This phenomenon is driven by the continuous advancements in NLP as current GLMs reached near human-level performance (Brown et al., 2020; Chen et al., 2019).

The nature of the writing tasks may influence the results of our experiments. To inform our selection, we classify writing into four purposes with related categories:

• •

  Writing for Learning: critical, persuasive, and reflective (Dugan and Polanski, 2006).

• •

  Writing for Academic Communication: scholarly, oral, and technical (Dugan and Polanski, 2006).

• •

  Writing for Business Communication: team, customer, project management, and career management (Dugan and Polanski, 2006).

• •

  Writing for Creativity and Entertainment: narrative (Shakeri et al., 2021; Clark et al., 2018), and poetic (Clark et al., 2018).

In this work, we focus on Writing for Business Communication as it is the most widely used form of writing. We selected two writing prompts for our experiments. The first prompt asks participants to write a complaint about aging computer hardware. We believe that many participants have experienced frustration from technical issues. As such, they will easily be able to relate their experiences when responding to the prompt. The second prompt asks participants to write an email to request funding for a multi-million dollar building for the participant’s university department from a well-known public figure. We suspect that most participants do not have relevant experience, so we provided a few suggested phrases that can be used in the email to kick-start the writing process. We argue that these tasks will allow us to investigate both the pull and push conditions in scenarios where writers have a related experience and in scenarios where a related experience is lacking. Both prompts are written in a scenario-based style to motivate and engage the participants during the writing process. The writing task prompts are included in Appendix A.

To evaluate the results of the experiment, we use 5 quantitative metrics.

We interviewed 16 participants aged between 18 and 25 years old. All of the participants are college students in the United States majoring in fields related to science, technology, engineering, and mathematics. 3 of the 16 participants spoke English as a second language while the remaining participants were native English speakers. The participants were randomly distributed to experience either the push or the pull conditions. For each condition, half of the participants randomly started with the familiar task, while the other half started with the unfamiliar task. To account for control variables that could affect our experimental results, we asked participants four questions summarized in table 4. We notice that the median response for all questions is 3, which is the midpoint, except familiarity with the computer issues, which is 4. As such, we can conclude from the results that our experimental design achieved its goals. During the 1-Day break step of the experiment, all participants had access to the tools, but only 2 of the 16 participants used them.

From the data on the percentage of accepted suggestions in tab. 2, we found that for the pull condition, participants had a higher than 50% acceptance rate overall, while the push condition had an acceptance rate of at least 35%. Users were very willing to accept suggestions from AI, and when asked whether or not they would use the tool again in fig. 4, 91.7% of push participants and 58.3% of pull participants said they would use the tool most of the time or all of the time. For both the Familiar and Unfamiliar writing tasks, the acceptance of the pull paradigm was significantly higher. Participants in the pull paradigm seemed to use the tool as a method of inspiration, especially when they experienced writer’s block. Some Push condition participants indicated that the suggestions provided seemed unnecessary as they were similar to their intended work. Having to actively think about pressing the tab button to accept a suggestion shifted away from their thought process, which they disliked. One advantage of the pull paradigm is the fact that it is on-demand, while the push paradigm gives suggestions unprompted which could disrupt the thought process. Thus, it can lead to the suggestions being inherently ignored as users continue typing. With that said, both push and pull participants have a significant acceptance rate of AI suggestions.

To measure efficiency, we used $WPM$ as described in equation 2. The initial expectation is that the pull paradigm would be more efficient as participants would be able to continue writing despite experiencing writer’s block using the pull button. Furthermore, GPT-3, which is used to generate text for the pull condition, is likely to generate significantly more words compared to GSC. The unfamiliar task data from our experiments confirm our expectation as the $WPM$ for the pull condition is near twice the push condition $WPM$ as participants experienced writer’s block more often. For the familiar task, the difference in $WPM$ seems to be negligible. Participants of the pull condition found that the tool helped them gain insights into fields they did not understand without any external research, which helped in speeding up the writing process.

In this work, we rely on the participants’ judgment to measure the quality of the writing. Our quantitative results are in tab. 2 and tab. 3 show that participants perceived the quality of the final writing product very similar across both the familiar and unfamiliar writing tasks. This is also confirmed by the distributions in fig. 5 when comparing the pull and push conditions holistically. From the qualitative data, pull participants mentioned that the tool helped them to gain new ideas and resulted in a more dynamic writing style. This can be seen in fig 6 as the topic modeling results found only five high-level clusters for the pull condition compared to the push’s seven clusters in fig. 7. Although more participants are needed for meaningful topic analysis, this shows that pull condition participants included a more diverse set of ideas in their writing. On the other hand, participants in the push condition mention that the tool improved their word choice, grammar, and the overall writing product was more clear and concise.

Participants found that their sense of ownership was extremely high in the unfamiliar writing task. These users explained that the use of these tools did not impact their sense of ownership as either they ignored the suggestions or they modified the generated text. However, there was a sharp drop in the familiar writing task for the pull paradigm. Some participants found that by pulling multiple sentences, they felt the work was not completely theirs anymore. Furthermore, one user mentioned that they do not attribute the quality of the writing to the tool as they could have produced an email of similar quality without it. With the push paradigm, this may be due to the low acceptance of the suggestions, as participants chose to ignore suggested words from the tool. Participants elaborated that the computer often completed words that they intended to write which allowed them to still feel in control of the writing process. Although ownership may not be as important for business communication, we can still verify our intuition based on the experimental data as it is a relative measure.

The enjoyment of the participants across both the familiar and unfamiliar tasks was higher in the pull paradigm compared to the push paradigm. This was the opposite of our findings in the pilot study, which did not include a 1-day break with only 4 participants. This suggests that the break was successful in mitigating the novelty effect. A common theme that arose is that users correlate the quality of the writing with the enjoyment of the final writing product. Participants found that the tool provided a good starting point to build off but would not use it as a final piece without modifications. They also indicated that it was entertaining to read the generated text. From our observations, participants were interested in exploring the capabilities of GPT-3 in providing suggestions on multiple topics during step 2, such as philosophy, history, and cooking recipes. Beyond natural language generation, participants explored generating code for programming languages and mathematical equations. However, some participants found that the tool could not provide sufficient context to generate relevant sentences with a single highlight. Furthermore, the enjoyment seemed to wear off after the 1-day break was given. Push condition participants, on the other hand, found that the auto-completion allowed them to make their writing more clear and concise, which they enjoyed.

The suggestions provided by both of the tools did not have any evidence of biased results based on our observations and qualitative feedback. However, participants voiced concerns regarding the possibility of bias in more complex writing tasks as well as privacy concerns. Although several participants reported some bias as shown in fig. 10, we believe that this could be, at least in part, due to the psychological effect of hearing about AI bias and fairness in news and academic outlets as many of the participants are in related fields. A participant even commented that they were impressed by the lack of bias of the generated text. However, one participant found that all of their suggestions were mostly related to academic work and the tool even generated a long tirade from a graduate student.