Artificial Intelligence (AI) has become a common term for an emerging set of computer technologies that affect individuals, communities, societies, and the environment at a global scale. Although the phrase “AI” was coined in the 1950s, it has undergone multiple transformations as a field of research and, until recently, was familiar to the general public largely as a theme for science fiction.  

AI research returned to public discussion in the 2010s when a number of innovations in “deep learning” became possible—largely because of the availability of human-generated data on the internet and through networked devices at an unprecedented scale. At around the same time, these technologies began to power widespread applications including voice assistants, recommendation systems, and automated driver assistance. When technologists speak of “deep learning” (DL), which is a type of “machine learning” (ML), the learning in question denotes a computer model’s ability to “optimize” for useful predictions while “training” on data through updating the weights in an elaborate set of statistical calculations. The learning is deep because of the multiple computational layers in the very large models that DL involves. The most heavily promoted forms of “AI” today are large language models (LLMs) such as Open AI’s ChatGPT, Google’s Bard, and Anthropic AI’s Claude 2.^[1] All of these systems are multi-layered (“deep”) statistical models that predict probable word sequences in response to a prompt even though they do not “understand” language in any human-like sense.  (From the Rutgers AI Council) A common term to refer to this type of machine-created output is “generative AI.”

The widely discussed ChatGPT is a specific application of GPT-3 (now GPT-4), released in late 2022 by OpenAI. It combines an easy-to-access browser interface with a chatbot style of interaction, whereby a user can enter a series of discursive prompts and engage with the outputs of the model in an ongoing dialogic stream. 

LLMs work by using statistics and probability to predict what the next character (i.e., letter, punctuation mark, even a blank space) is likely to be in an ongoing sequence, thereby “spelling” words, phrases, and entire sentences and paragraphs. It is not unlike autocomplete, but more powerful. LLMs are trained on vast bodies of preexisting text (such as content from the Internet), which, to some extent, predetermine their output. All of the text a model generates is original in the sense that it represents combinations of letters and words that generally have no exact match in the training documents, yet the content is also unoriginal in that it is determined by patterns in its training data. The same language model may generate a variety of different sequences in response to the same input prompt. A model cannot reliably report on which sources in its training data contributed to any given output. All of this combines to make the output of LLMs qualitatively different from any other form of text, even texts that might have been computer generated according to some other method. It should be noted that given the assortment of software applications drawing on LLMs, presenting them through different user interfaces to offer various affordances, they are also unavoidable.

Although it is often tempting to speak in terms of what an LLM is “doing” or “intending” or even “thinking,” what we are witnessing is the production of word sequences that look like intentional human text through a process of statistical correlation. As the models are refined, expand their language corpora, and draw on greater computational power, their outputs mimic the writing of sentient humans more convincingly. LLMs do not, however, “think” in the way that we would define such an activity as it takes place in human cognition. (From the MLA-CCCC Joint Taskforce)

[M]achine-generated content is often factually incorrect. Moreover, when users prompt an LLM to provide sources for information, the cited sources may be wrong or completely fabricated. Likewise, while chatbots are sometimes marketed as question-answering systems (like Apple’s Siri) and as effective replacements for search engines, LLMs are pre-trained models that do not search the web. ^[3] The novelist and technologist Ted Chiang has likened a chatbot’s compressed model of its training data to a “blurry JPEG.”  Since the training data for LLMs is enormous–constituting most of the “scrapable” internet–as models have grown successively larger, their predictive capacities have extended beyond the generation of human-like text: for example, they can now answer some questions in basic math (though still make errors on simple tasks such as three-digit multiplication); in the hands of knowledgeable programmers, they can aid in the writing of computer code. Large image models (LIMs) which are trained on visual images scraped from the internet, are text-to-image generators that respond to a user’s textual prompt with a visual output predicted to correspond accordingly. Because these commercial models are proprietary, researchers have no means of determining precisely what training data was involved, what tasks were prioritized for human reinforcement (and under what working conditions), or how much energy and water is required for training and use. (From the Rutgers AI Council)

Amplification of Bias, “Malignant” Stereotypes, and “Documentation Debt”: [S]ince LLM performance relies heavily on large datasets, the best-performing models are also riddled with bias and stereotypes from largely undocumented data scraped from the internet. Bender, Gebru and colleagues (2021) describe the risks of this “documentation debt” in the context of proprietary datasets that “overrepresent hegemonic viewpoints.”^[6]

Copyright Infringement, Lack of Consent, Surveillance, and Privacy Concerns: [T]he use of copyrighted content scraped from the web without consent for the training of AI models has opened a host of legal questions. (Notably, the New York Times, in August 2023, updated its service to forbid use of its content for training data–a new development in practices regarding protection of intellectual property).^[7] At the same time, the accumulation of user data during use of commercial chatbots extends the surveillant practices that began with the monetization of social media and search engines, exacerbating data privacy concerns.^[8] 

Environmental Footprint: [B]ecause “generative AI” is computationally intensive, the technology uses significantly more energy and water than a simple internet search.^[9]

Exploitation of Human Labor: [S]ince LLMs are subject to bias, misinformation, and toxicity, the current technology relies on millions of low-paid crowdworkers whose annotating work improves results.^[10] 

Misinformation through “hallucinations,” conspiracy theories/misconceptions, and malicious use.^[11]Political Economy, Concentration of Power, Lack of Transparency and Accountability: [T]he political economy of “AI” today was forged through the concentration of computing, economic, and data resources in some of the largest and most lucrative companies in the world. Corporations such as Microsoft (and their OpenAI partner) and Google intensively lobby legislators, sometimes “watering down” regulatory demands for transparency and accountability for these dominant companies and their products.  Lina Khan, who is chair of the Federal Trade Commission, has described the risks of “AI” in a context of “race-to-the-bottom business models and monopolistic control.”^[12]  (From the Rutgers AI Council)

• Be guided by The University of Tennessee’s Honor Statement and academic integrity policies, which are published in the Undergraduate Catalog. The Honor Statement is as follows: “An essential feature of the University of Tennessee, Knoxville, is a commitment to maintaining an atmosphere of intellectual integrity and academic honesty. As a student of the university, I pledge that I will neither knowingly give nor receive any inappropriate assistance in academic work, thus affirming my own personal commitment to honor and integrity.”
  • Affirming your commitment to honor and integrity means affirming that the written work you submit is your own and that you always cite anything you did not compose on your own. 
  • You should consider the use of AI-generated writing without citing or documenting it to be inappropriate unless your instructor explicitly says it’s okay to so in certain circumstances. You should not use AI-generated text at all if your instructor does not allow it.
• Become familiar with how Large Language Models (LLMS) and generative AI tools such as ChatGPT work, what they can and cannot do, and the opportunities and risks of using them.  
• Read each of your instructors’ policies on using AI, and follow them carefully. The University of Tennessee, Knoxville, does not have a central policy for this. Instructors may make their own rules about whether and how you can use AI in their class.
  • ASK each of your instructors for their policies and guidelines. Be prepared to follow different AI-related rules in each of your classes.
    • Examples of questions you could ask your instructor include: Can I use ChatGPT(or another AI tool) to help me brainstorm? To help me research? To help me summarize readings? To help me write an outline? To help me revise a draft I’ve written? How would you like me to cite AI output? How would you like me to describe how I used AI tools during the writing process?
• Commit to using AI tools during the writing process with care, making sure you do not copy AI output directly into your work or replace your own intentional, purposeful compositions with AI output.
• Keep a log of your AI use; the university has recommended that instructors who allow AI use ask students to document it. 
• Fact-check all AI output text, because it will include inaccurate information–that’s simply true about how AIs work (unless you’re doing “closed prompting”–making queries only about sources you provide in your prompt). Since you’re not likely to know what’s accurate and what isn’t, you must fact-check the AI output before relying on it in any way.
• Critically evaluate all AI output text, paying close attention to gender, racial, and language biases and biases against particular viewpoints.  
• Quote, paraphrase, and cite AI-generated text just as you quote, paraphrase, and cite any external source material in your papers. Here are some common formats:
  • MLA: https://style.mla.org/citing-generative-ai/
  • APA: https://apastyle.apa.org/blog/how-to-cite-chatgpt 
• Include the following statement in assignments to indicate your use of a generative AI tool: “The author(s) would like to acknowledge the use of [Generative AI Tool Name], a language model developed by [Generative AI Tool Provider], in the preparation of this assignment. The [Generative AI Tool Name] was used in the following way(s) in this assignment [e.g., brainstorming, grammatical correction, citation, which portion of the assignment].
• Reflect on your use of AI: How are you using it? What do you think about what it offers you as a writer? What are you learning about writing? Is using AI helping you make better choices as a writer–and if so, how? What are your concerns about using AI during the writing process?
• When sharing your work in progress with your instructor, a peer reviewer, or a tutor, tell them if you’ve used AI tools during your composing process. You should point out any part of a draft that includes AI output–even if it’s a draft in which you have not yet included your citations–and you should describe how you used the AI (e.g., “I used ChatGPT to brainstorm”). This will help you and your instructor, peer reviewer, or tutor have a transparent, productive, and responsible conversation about your draft and your writing.

Learn how to prompt the AI responsibly, and be as specific as possible; the quality of your prompts will affect the quality of the AI output. Your prompt should not be “Write me a paper on [topic.]” (From The Judith Anderson Herbert Writing Center)

Prompting that leverages rhetorical awareness. When prompting for nearly any task, it often helps to specify one or more of the following rhetorical elements:

• Role/Speaker: “You are a highly experienced marketing manager who works for…”
• Audience: “You are creating a marketing campaign targeted at a semi-rural region in Idaho…”
• Purpose: “The service you want to pitch is…”
• Genre: and Platform Constraints “The marketing campaign will be run on social media platforms, including…”

Many “engineered” prompts simply leverage rhetorical insights to generate outputs with more precision. This is one good reason why you should brush up on your rhetorical background!

Brainstorming Machine. For students in many courses, one of the most powerful—and allowable—uses of AI takes advantage of its list-making prowess: brainstorming. LLMs like ChatGPT are excellent listers. Try prompts such as:

• “Please create a ten different research questions based on…”
• “I’m having trouble thinking of a topic to write about. Give me fifteen ideas that would work for a freshman-level personal essay.”

ChatGPT can also create tables or matrices. These formats invite users to brainstorm through pros vs. cons, comparing and contrasting a range of options, etc.

Explaining New or Difficult Concepts. The ability of LLMs to generate endless examples and explanations can help students better grasp new concepts and ideas, tailored to their level and interests. Here’s a strategy based on Ethan Mollick and Lilach Mollick’s:

Pick a concept you want to understand deeply.

If using an AI connected to the internet (such as Bing): Tell the AI to look up that concept using core works in the field. [Remember, though, that currently, most free chatbots are pre-trained on data prior to 2021.]

Tell the AI what you need (many and varied examples of this one concept).

Explain your grade level.

Simulating Educational and Professional Drafts. Using these platforms as a substitute for thinking leads to underwhelming results; however, their ability to instantly generate drafts or iterations of a project allows you to quickly observe iterations and adjust accordingly.

Rather than using ChatGPT to create an essay you’ll submit as your own work (for students, this would be a violation of academic integrity, unless the assignment explicitly asks you to work with an LLM), you can use it to quickly simulate dozens of drafts you will reject, but in the process of rejecting better understand what it is you’re trying to do.

[Some instructors] have asked students to experiment with new genres by quickly generating sample drafts in ChatGPT. Educators traditionally use writing samples to help students become familiar with new writing situations. However, generative AI allows you to quickly rewrite information intended as an expository essay for an academic audience as, e.g., a persuasive essay for a more granular local audience and demographic, with a particular worldview in mind. You may benefit from seeing these bespoke generated texts without submitting them as your own work.

Feedback, Paraphraser, and Copy-editor. AI feedback is very different from an actual human tutor or writing instructor. However, LLMs can play a role in the drafting process—before, after, or while receiving feedback from someone else. Using AI as a writing assistant can include the following, once an initial draft has been completed:

• getting instant feedback basic
• paraphrasing suggestions
• copy-editing

When eliciting feedback from LLMs, however, it will be important to experiment with a range of prompt-engineering strategies and remain aware of their limitations. (From Gladd, Write What Matters)