Features March/April 2023

Generative AI: Productive Paths Forward

By Dan Card | STC Member and Ann Hill Duin | STC Member

AI-based writing tools are here. What’s next?

Technical communicators have long held the assumption that “technology is a system of relations with fellow humans” (Barker, 2019). We are user advocates, well accustomed to framing our content through interaction with technology. The technical communication toolbox now includes remote usability testing and video conferencing platforms, virtual whiteboard and visual collaboration tools (Das, 2020), as well as XML content authoring and management systems.

In recent years, numerous authors across industry and academia have emphasized the need to prepare content for machine consumption. Technical communicators are increasingly called to ensure content meets the needs of a wide variety of machine “readers,” including indexing bots, content aggregators, accessibility tools, and multichannel publishing systems. As Cruce Saunders (2019) suggests, the emergence of natural language processing (NLP), machine learning (ML), and artificial intelligence (AI) have created “a seemingly insatiable demand for ‘machine-consumable’ content to pair with awe-inspiring zetabytes of human data.”

We think that technical communicators should be thinking about not only machine readers but also machine writers. While the field has established conceptual resources, processes, and tools for preparing machine-consumable content (for example, DITA and component content management systems), the working assumption remains that we will continue to author (albeit more granular) content. That is, for the most part, our discourse about content automation has focused on the roles machines play in publishing, distributing, or arranging, but with the exception of some nascent engagement with chatbots, technical communicators have devoted markedly less attention to autonomous authoring.

Fully autonomous authoring?

Across industry and academia sites, there seem to be two common responses to the notion of fully autonomous authoring: uncritical embrace and outright rejection. We favor a third: curious, critical experimentation. As we enter the so-called fourth industrial revolution, technical communicators must pay attention to the proliferation of tools that claim to revolutionize writing. Amidst the emergence of boundary-blurring tech, AI stands out as particularly pervasive in public discourse about the future of writing. For example, Open AI recently released the chatbot ChatGPT. ChatGPT is the latest, but certainly not the last, example of generative AI—AI tools that—for all intents and purposes produce novel content. ChatGPT is impressive in its ability to generate lengthy, detailed responses to questions that are syntactically and often even semantically indistinguishable from responses by human writers (Miller et al., 2022). So tech commentators within and beyond higher ed are eulogizing the college essay, and many are projecting broad swaths of knowledge work will soon be part of the death toll.

The stakes are potentially high for technical communication. For example, our colleague, Jim Hall, asked ChatGPT to, “Write a primer on how to write a document using LaTeX.” Following, we share the ChatGPT response, one that Jim Hall sees as a “legit demonstration of how ChatGPT can generate (mostly) correct technical writing responses to prompts.”

LaTeX is a powerful typesetting system that is commonly used to create professional-quality documents, including academic papers, articles, and books. LaTeX is known for its ability to produce high-quality documents with a consistent layout and formatting, making it a popular choice among writers and publishers.

To write a document using LaTeX, you will need to create a plain text file using a text editor such as Sublime Text or Atom. In this file, you will type the text of your document and use special commands and directives to specify the formatting and layout of your document.

Here are some basic steps to follow when writing a document using LaTeX:

  1. Start by setting up the basic structure of your document. This includes specifying the document class, setting the margins, and adding the title and author information.
  2. Use commands and directives to specify the layout and formatting of your document. This includes commands for headings, paragraphs, lists, and more.
  3. Add content to your document by typing the text of your document into the file.
  4. Use commands and directives to add special elements such as tables, figures, and equations.
  5. Preview your document using a LaTeX compiler to make sure that it looks the way you want it to.
  6. Save your document and compile it to create the final version of your document.

By following these steps, you can create professional-quality documents using LaTeX. With practice, you will become more proficient at using the various commands and directives to create the layout and formatting that you want.

His takeaway: “Tech writers using ChatGPT become a mix of ‘ideas’ person, prompt writer, fact checker, and technical editor” (Hall, 2022).

Technical communication cannot afford to ignore generative AI tools like ChatGPT, but it shouldn’t embrace them uncritically either. As Duin and Pederson (2021) note, “the hype surrounding AI is used to justify adopting the technology before we think about the implications.” Separating marketing hype from reality should be a priority, and we believe that work is best done by engaging with the tools directly. Toward that end, part of our goal in this article is to clarify some key terminology and consider particularly relevant use cases. We share basic definitions along with our recommendations for curious, critical experimenters.

Definitions: Generative AI and machine learning

Broadly speaking, artificial intelligence concerns the development of computer systems that appear intelligent or are designed to perform tasks that typically require human intelligence. These systems rely on algorithms or sequences of procedures that are applied to data and models, which use the outputs of those algorithms to make predictions.

Generative AI refers to a subset of AI systems that can produce visual, auditory, and textual content. In the case of textual content, these systems use complex models to predict the next word in a sequence. This alone is not entirely new. Widely available grammar checking software has been flagging creative uses of their, there, and they’re for long enough that we hardly notice it. What’s new is that the most recent generative AI systems are built on large language models (LLMs). These pre-trained language models, such as BERT, GPT-3, and MT-NLG, are tens of gigabytes in size, trained on enormous amounts of text data, representing the most advanced deep learning algorithms capable of understanding and fine-tuning language for specific tasks.

Machine learning (ML) can be understood as the “how” of artificial intelligence. ML is the process used to enable systems to learn patterns from data so that it can predict answers to problems it hasn’t encountered before. ML practitioners use a variety of approaches for different problems, but particularly salient are supervised and semi-supervised machine learning, where models are developed and refined using labeled training examples.

We see two major opportunities for technical communicators (TC):

  • participate in the development and evaluation of these tools and
  • help document and explain them.

Both opportunities require increased AI literacy.

Opportunity 1: TCs can help develop and test generative AI tools

In the near term, we expect increasing levels of autonomous authoring of specialized content. That is, generative AI tools are now good enough to quickly generate writing that in many cases will only require light revision. Technical communicators have long been collaborating with systems that are in some sense autonomous, and they should be prepared to continuously examine and negotiate the nature of that collaboration. In some ways, we might see recent advancements in large language models (such as ChatGPT), not so much as a revolution, but rather an acceleration. That is, many in technical communication already describe their work as something akin to stewarding the production and automation of content. The field has long embraced a notion of writing that includes, if not prioritizes, the modification of existing text.

AI models are increasingly trained on massively large datasets of natural language, and in many use cases, large models tend to perform better. While LLMs like GPT-3 are undoubtedly impressive in their ability to generate writing that appears to have been written by an intelligent human, in some cases that is more of a liability than a feature. Indeed, a recent study compared the ability of four LLMs to provide truthful answers to 817 questions spanning categories such as law, health, finance, and politics. Researchers found that the best model was truthful on 58% of questions, while human performance was 94% (Lin, Hilton, and Evans, 2021). Further, the researchers found that larger models were generally less truthful. That’s because these models learn the biases, misconceptions, and falsities we intelligent humans occasionally put on the web.

In many of our organizations, especially in regulated industries, content is incredibly specialized, and the stakes are high, making accuracy and usability top priorities. In these contexts, generative AI tools may need to be refined through supervised machine learning (sML) practices, for example, labeling or tagging organization-specific text to improve the performance of generative AI tools relative to organizational goals. Saunders stresses the need for technical communicators to “start learning relevant skills that will facilitate this new era emerging across authoring and publishing,” and in light of recent developments in language models, we think sML practices should be high on the list.

Some technical communicators are already participating in sML, for example, by developing chatbots or refining search tools to more effectively predict the most relevant help topics. These processes involve analyzing contexts and purposes of use, curating appropriate textual data to train ML models, feature engineering, and evaluating performance relative to user needs and organizational values. Technical communicators are uniquely suited to this work given our attention to context and language, but to participate effectively and responsibly, we must also develop increased AI literacy. Now more than ever, we need people who can evaluate text for accuracy, alignment with values, and fitness for context and purpose.

Opportunity 2: This industry needs explainers

AI tools are used to automate business processes, gain insight through massive data analysis, and engage with customers and employees. These tools are also used to detect patterns in vast volumes of data, providing users with predictions for what to buy, identifying safety or quality issues, and modeling increasingly more accurate and detailed information for use in building new systems. Technical communicators are now involved in projects that engage employees and customers in using natural language processing (NLP) chatbots and intelligent agents to provide increased customer service.

We have long understood that technologies amplify our abilities, but recent generative AI tools show potential to perform increasingly complex tasks with far less human direction. While we are well accustomed to the use of digital assistants (Siri, Alexa, Chatbots), the transformation underway includes the emergence of human-AI teaming, in which human tasks are augmented through the deployment of digital employees. Increasing numbers of companies are allowing users to engage with autonomous AI agents directly. Consider Amelia.ai, one of the first conversational AI entities marketed as a digital employee, or UNeeQ’s digital humans that aim to “revolutionize customer experiences with scalable human connections” (UNeeQ, 2022).

The moving parts and speed of development are astronomical, leaving one with bewilderment as to how to proceed. But as technical communicators, we understand rhetorical context, document how something works, and make it usable. With generative AI, algorithms, and ML, we now must work to explain how an algorithm makes decisions, changes, and adapts over time. Explainable AI (XAI) has emerged as a means to help humans better understand how AI works and makes decisions; it provides a glimpse into AI’s inner logic. According to IBM (2023):

Explainable artificial intelligence (XAI) is a set of processes and methods that allows human users to comprehend and trust the results and output created by machine learning algorithms. Explainable AI is used to describe an AI model, its expected impact and potential biases. It helps characterize model accuracy, fairness, transparency, and outcomes in AI-powered decision making. Explainable AI is crucial for an organization in building trust and confidence when putting AI models into production. AI explainability also helps an organization adopt a responsible approach to AI development.

Industries are adopting generative AI, algorithms, and ML, and technical communicators must adapt. Explanation and explainability skills are crucial. A number of resources exist for technical communicators to use in building XAI expertise.

  • Guidance from Rafia Inam and colleagues (2023), Explainable AI—How Humans Can Trust AI, provides a solid overview of XAI.
  • The European Commission’s Ethics Guidelines for Trustworthy AI (2018) provide a vast resource covering trustworthy AI: human agency and oversight; technical robustness and safety; privacy and data governance; transparency; diversity, non-discrimination, and fairness; environmental and societal wellbeing; and accountability.
  • Researchers Mireia Ribera and Agata Lapedriza (2019) provide a review of XAI with focus on user-centered explainable AI, classifying users into three groups, developers, domain experts, and lay users, for the purpose of creating specific explanations for each group.
The importance of expanding digital literacies

TCs will find opportunities to collaborate with, develop, or explain generative AI tools, but only if they start building the necessary literacies right now. Numerous worldwide initiatives document the need for digital literacy competencies for personal, professional, and civic engagement. In Writing Futures, Duin and Pedersen (2021) emphasized the need for technical communicators to further expand understanding of digital literacy as human and machine work together: “Digital literacy for writing futures means no longer viewing human and machine as separate agents along with the ability to envision and write within mirror-worlds of virtual fragments stitched together.”

In their upcoming book, Augmentation Technologies and Artificial Intelligence in Technical Communication: Designing Ethical Futures (in press, Routledge), they share a conceptual framework for determining AI competencies and design considerations as a means for technical communication scholars, instructors, students, and practitioners to examine and develop digital and AI literacies. They draw on Duri Long and Brian Magerko’s (2020) definition of AI literacy: “a set of competencies that enables individuals to critically evaluate AI technologies; communicate and collaborate effectively with AI; and use AI as a tool online, at home, and in the workplace.” Long and Magerko discuss competencies and design considerations as a means to define AI, what it can do, how it works, how it should be used, and how people perceive it.

Long and Magerko’s design considerations are of particular interest for technical communicators. These include explainability, embodied interactions, contextualizing data, promoting transparency, and unveiling information gradually.

Toward user-centered futures

Given how impressive these tools already are, it might be tempting to resign ourselves to a world without writers, but we think technical communicators should be active participants in shaping the future of writing. Critical commentators have raised significant concerns, for example, the computational costs and associated climate impacts of training LLMs (Heikkilä, 2022), their tendency to privilege English (Bender & Friedman, 2018), and their track record of replicating racial and gender biases (Johnson, 2022). If we are to continue in our role as user advocates, we must familiarize ourselves with these concerns and build the literacies required to responsibly collaborate with, explain, and develop generative AI tools.

References
  1. Barker, Thomas. 2019. “Doing More with Less: Tools and Strategies in Communication.” Intercom 66, (2): 37–38.
  2. Bender, Emily M., and Batya Friedman. 2018. “Data statements for natural language processing: Toward mitigating system bias and enabling better science.” Transactions of the Association for Computational Linguistics (6): 587–604.
  3. Das, Meghalee. 2020. “Key Takeaways and Trends in Remote UX Research.” Intercom 67 (6): 20–25.
  4. Duin, Ann Hill, and Isabel Pedersen. 2021. Writing Futures: Collaborative, Algorithmic, Autonomous. Springer.
  5. Duin, Ann Hill, and Isabel Pedersen. 2021. Writing Futures: Collaborative, Algorithmic, Autonomous. Springer.
  6. European Commission. 2018. Ethics Guidelines for Trustworthy AI. Accessed January 5, 2023. https://ec.europa.eu/futurium/en/ai-alliance-consultation.1.html.
  7. Hall, Jim. Email communication. December 29, 2022.
  8. Heikkilä, Melissa. 2022. “We’re getting a better idea of AI’s true carbon footprint.” MIT Technology Review. November 14, 2022. https://www.ibm.com/watson/explainable-ai.
  9. Inam, Rafia, Ahmad Terra, Anusha Mujumdar, Elena Fersman, and Aneta Vulgarakis Feljan. 2021. “Explainable AI – how humans can trust AI.” Ericsson.com. Accessed January 5, 2023. https://www.ericsson.com/en/reports-and-papers/white-papers/explainable-ai–how-humans-can-trust-ai.
  10. Johnson, Kharri. 2022. “The Efforts to Make Text-Based AI Less Racist and Terrible.” Wired. June 17, 2022. https://www.wired.com/story/efforts-make-text-ai-less-racist-terrible/.
  11. Lin, Stephanie, Jacob Hilton, and Owain Evans. 2021. “TruthfulQA: Measuring How Models Mimic Human Falsehoods.” Cornell University arXivLabs. Accessed January 5, 2023. https://arxiv.org/abs/2109.07958.
  12. Long, Dori, and Brian Magerko. 2020. “What is AI Literacy? Competencies and Design Considerations.” CHI ’20: Proceedings of the 2020 CHI Conference on Human Factors in Computing Systems. April 2020. Honolulu, HI. https://doi.org/10.1145/3313831.3376727
  13. Miller, Claire Cain, Adam Playford, Larry Buchanan, and Aaron Krolik. “Did a Fourth Grader Write This? Or the New Chatbot?New York Times, December 26, 2022.
  14. Ribera, Mireia, and Agata Lapedriza. 2019. “Can we do better explanations? A proposal of user-centered explainable AI.” In Joint Proceedings of the ACM IUI 2019 Workshops, March 20, 2019, Los Angeles, CA.
  15. Saunders, Cruce. 2019. “Next-Generation Authoring: Creating Content for Human and Machine Consumption.” Intercom 66 (4): 12–14.
  16. UNeeQ. 2022. “Revolutionize customer experiences with scalable human connections.” Accessed January 5, 2023. https://digitalhumans.com/.

Dan Card (dcard@umn.edu) is an Assistant Professor in the Department of Writing Studies at the University of Minnesota-Twin Cities. His teaching and research explore the intersection of technology, writing, and design.

 
 
 

Ann Hill Duin (ahduin@umn.edu ) is Professor of Writing Studies and Graduate-Professional Distinguished Teaching Professor at the University of Minnesota where her research and teaching focus on the impact of emerging technologies on international and technical communication, digital literacy, and writing futures.