What to Do About AI Text Generators?

Next Steps for Educators

January 27, 2023

Anna Mills, College of Marin�Presentation shared under a CC BY NC 4.0 license.

Follow along at https://bit.ly/WhatToDoSlides

​

Welcome!

• Feel free to have tea or coffee and let pets and children make cameos ☺
• Introduce yourself and where you are from/where you teach in the chat
• If you want to explore, here are a few resources!
• Slides: https://bit.ly/whattodoslides
• Adapting College Writing article: https://bit.ly/CAISteps
• Resource list: https://bit.ly/AITextEdu
• Sample AI-Generated Essays: https://bit.ly/sampleAIessays
• Polls and Q&A: https://bit.ly/AdaptingWebinar�

Agenda

• Background on ChatGPT/language models
• Academic integrity considerations
• Ways to respond
• Q&A
• Wrap-up, resources, and next steps

What are AI text generators?�

Technically called large language models or LLMs, text generators like ChatGPT produce mostly original writing in response to a user’s prompt.

What does a text generator do?�You ask it for what you want, and it generates a response.

ChatGPT screenshot with one of five paragraphs it generated:

Here’s a user requesting a comparison of MLK and Gandhi. ChatGPT begins to generate.

AI text + EDU = ?

• What do AI text generators mean for higher education?
• Should we see them as a threat? A tool? Both?
• What approaches should we consider?

​

I definitely don’t have all the answers!

Let’s start the inquiry.

Context

• ChatGPT is software created by the company OpenAI, released in December.
• OpenAI is largely funded by Microsoft.
• ChatGPT is a more user-friendly version of other software, GPT-3.
• ChatGPT and GPT-3 are considered the most capable of language models currently publicly available.
• Open-source LLMs also exist. Google and Meta have also built LLMs.

A language model can take in and incorporate background information, sources, quotations, and lists of ideas.

• You feed it the information it needs (up to ~2,000 words).
• Give it anything you have that will help produce the piece of writing.
• The result will most often fulfill your request, at least to some extent, as well as reflecting the information you gave it.

It offers original word combinations

• The software is designed to generate a series of words based on its statistical analysis of huge volumes of text.
• It works by predicting a likely next word, not by copying sentences or paragraphs.

It produces new outputs for the same prompt

• If you don’t like one, you can request a different one.
• Two users may put in the same prompt and get different results.

What’s the quality of ChatGPT outputs?�Often solid academic prose in some respects

Limitations

ChatGPT has no understanding

• It’s a statistical model—sociologist Alex Hanna and computational linguist Emily Bender like to call it a “mathy math.”
• Outputs are next-word predictions not based on rules of reasoning.

​

There’s no coherent intention behind its text

• We can see contradictions in claims from one sentence to the next.
• It might respond one way one time and the opposite way another time.
• It doesn’t “think” or “claim” or “believe” anything, though we might be tempted to describe it that way.

Researcher Janelle Shane explains it this way: the algorithm is optimized for a constant game of “What would the humans do next?” ��ChatGPT’s task: answer the question “How would a human writer who started out writing the prompt likely continue stringing words together?

It’s often biased because the text it’s based on is biased

• Its outputs are statistically based on the human-written text scraped from the Internet, Reddit, Wikipedia.
• It reflects the content of the internet—its outputs are skewed as the Internet is skewed toward developed nations, English, and toward wealth, maleness, and whiteness.
• Despite attempts to mitigate the bias, it is difficult to eradicate.

​

​

It sometimes makes up facts, quotations, and citations.

• It’s performance in relation to sources and facts is erratic.
• If you ask it for sources, sometimes it refuses to give them, sometimes it makes them up, and sometimes it gives real ones.
• ChatGPT doesn’t search to back up claims, though some apps built on GPT-3 can do this (PerplexityAI, ChatSonic, Elicit.org).
• It’s not checking what’s real—it’s predicting what words might go together.
• Its fabrications can contain bits of truth

Often generic, bland, and lacking detail or meaningful transitions

• Example: “This argument makes a number of important contributions to the larger conversation about gender and its role in American culture. First, the author challenges the notion that reproductive potential is the defining characteristic of womanhood. “
• However, it can be prompted to change style and add detail, with mixed results.

There are limits on how much text you can ask it to respond to

• Around 2,000 words is the maximum you can feed it when you prompt it.
• If you want it to analyze a longer text, you have to ask it to summarize one section at a time and then combine the summaries.

Academic Integrity Considerations:�Is this different from plagiarism?

The text is (mostly) not copied.

• As mentioned, text generators make word-by-word predictions for what should come next based on statistical analysis of much of the text on the Internet.
• Occasionally outputs will include a phrase or sentences copied from a source.
• However, most outputs will pass plagiarism checks.

Social norms have not been fully formed around text generators.

Text generators are in use in professional settings beyond the classroom.

• Various apps built on ChatGPT cousin GPT-3 are being used for marketing, blogging, and other forms of commercial web writing for search engine optimization.
• These tools are used extensively for writing code.

Capabilities are changing rapidly with little notice

• Since the November 30^th release there have been two updates to ChatGPT that have improved its outputs.
• GPT-4 is rumored to be coming out before long and to be significantly better.
• For example, at first all the sources and URLs I was seeing were made up. But in the last couple of weeks Twitter commentators began to observe ChatGPT returning real sources and sometimes URLs.

How should we respond?

My (current) take: a combination of approaches

Emphasize purpose and engagement

• If students see meaning in the writing assignment and understand what they will get out of wrestling with it, they are more likely not to resort to a text generator.

Emphasize how writing helps us think

Make explicit policies about AI

• “Content generated by an Artificial Intelligence third-party service or site (AI-generated content)  without proper attribution or authorization would also be a form of plagiarism.” Center for Integrated Professional Development, Illinois State University, January 9, 2023
• “Use an AI text generator only if the assignment explicitly calls for it or allows it, and follow specific assignment guidelines to disclose which text comes from the AI,” Anna Mills, August 15, 2022
• See “Sample Policy Statements about Text Generators” on the WAC Clearinghouse resource list for more samples.

Ask students to affirm that they have labeled any AI text as such

• This is similar to affirming that the writing they have submitted is their own and they have not plagiarized.
• We extend the practice so they affirm that they have not represented AI-generated text as their own writing.

Identifying AI Text or “Autotext”

Current tools for identifying AI text

• GPTZero by Edward Tian, released January 3, 2023
• AIWritingCheck.org
• OpenAI Detector: GPT-2 Output Detector Demo. The original was by OpenAI; HuggingFace hosts the demo. 

How accurate are AI text detectors?�Don’t use them for “gotcha”

• Not accurate enough to rely on as a basis for convicting a student of academic dishonesty.
• Common tools have been shown to have false positives and false negatives, as much as 10-20% of the time.
• The tools can also be bypassed using paraphrase software like Quillbot.

Data rights and privacy concerns

• Could feeding a student’s submission to an AI text detection program be a violation of the student’s rights to their data?
• What will the current detection programs do with the data?
• Perhaps this should be vetted at the institutional or the state level.

​

• Possible workaround: ask students to run a detection program on their own work and give them an alternative, like recording a video of themselves discussing their writing.

Early days: the future of AI text detection

• Turnitin has put out a preview of their AI text detection system and says it will be released soon.
• OpenAI is working on an approach to watermarking its text outputs so that they can be detected.
• Will there be a way to stop people from getting around these systems by paraphrasing using “spinners” like Quillbot?
• When open-source language models approach commercial models in quality, that may make regulation and detection more difficult.
• Many are pessimistic that we can win an arms race around detection in the long run, but really, who knows?

Advocacy

• How do we connect the need for AI text identification in education to the need for it in public sphere discourse and social media? (See the recent NYT article on ChatGPT and democracy)
• What role do educators have in calling for AI text detection?
• What role does government have in promoting or requiring such technology?

​

Options for Modified pedagogy

• In-class writing
• Assign steps in the writing process: collaborative annotation, prewriting, outlines, drafts, revisions, peer review.
• Ask students to use Track Changes.
• Hold conferences with students
• Ask students to record audio or video notes about their writing when they turn it in.

Assign writing tasks text generators aren’t good at

What can’t ChatGPT do passably?

• Students will most likely use it in hybrid ways. They will feed it what information they have and revise its outputs.
• Currently, it sometimes makes up sources, sometimes refuses to give sources, and sometimes gives real sources depending in part on how it’s prompted. Some apps that draw on it will give real sources.

Multimodal assignments

• As students to write about images, video, and audio (they would have to describe these or include transcripts to prompt ChatGPT)
• For example, offer an image that represents a key class concept and ask them to analyze the image, the assumptions it makes, and any questionable aspects of it in relation to class readings.
• Ask them to incorporate points from class discussion.

Problematic approaches

• Not assigning writing. Yes, there are other ways of showing understanding, but writing is an essential skill for developing understanding and original reflection.
• Banning digital writing. Writing by hand is not how students will write in other settings, and it poses difficulties for many students, especially students with disabilities.
• Banning access to ChatGPT on a network. This seems ineffective since students will have access through cellular data. There are so many different tools that would have to be banned, and the number is growing.

Surveillance: No thanks

• In-class writing has a place. In that context, we might restrict access to the Internet and ask students to work on classroom computers.
• However, software used to remotely surveil out-of-class writing has been shown to lead to privacy and data rights violations.
• It has also been shown to be biased and inaccurate.
• Is Track Changes surveillance? Is it okay to students to require them to expose the minute-by-minute record of their writing process?

Do we teach about text generators?�Do we use them in our pedagogy?

Critical AI literacy? Yes, please!

We can start by introducing the concept of statistically generated text and dispelling any notion that AI is sentient, authoritative, or neutral. Teach students to watch for problems in AI outputs.

Rush to teach text generator use because we think it’ll be a necessary job skill? �My opinion: Not necessary, not a priority.

• Using ChatGPT is fairly intuitive.
• The technology is changing so fast that there will be different interfaces and different capabilities within a year or less.
• Learning how to think critically and write independently may be better training in how to prompt AI text generators strategically than any current tech tips or tricks.

Creative ways to incorporate text generators into teaching?

• It’s appealing. The technology is fascinating and we could think of many ways to use it. Maybe it has its place, but I would urge some caution.
• But are the ways we might use it better than the best practices we would otherwise explore? Are they better than our current practices?
• Have we understood the technology so we can show students its pitfalls and make sure they can identify problems on their own?
• Even if we ask students to critique or improve its outputs, is spending time reading and critiquing them really better than reading and critiquing human-written texts that would be better models?

One option: Try introducing research tools like Elicit.org or PerplexityAI that find and summarize real sources. Showing an example of where they get a summary slightly wrong. ��See the resource list for more ideas.

Resources: https://bit.ly/AIwritingEDU

• AI Text Generators and Teaching Writing: Starting Points for Inquiry: news, analysis, and various educators’ approaches to the subject from the Writing Across the Curriculum Clearinghouse
• Sample policies
• Samples of AI-generated academic writing

​

“Adapting College Writing for the Age of Large Language Models such as ChatGPT: Some Next Steps for Educators,” by Anna Mills and Lauren Goodlad, Critical AI, January 17, 2023

AI Text Generators: Sources to Stimulate Discussion among Teachers

• Contents
• Cautions and Disclaimers
• About AI Text Generators/Large Language Models
• Implications for Higher Ed Writing Assignments
• Audio and Video 
• Sample Policy Statements about Text Generators
• Student Perspectives and Marketing to Students
• Course Materials on AI Text Generators 
• Peer-Reviewed Papers
• Short Pieces on the General Topic of AI
• Books on the General Topic of AI
• Publicly Available Language Models
• Detecting AI-Generated Text
• Using Text Generators for Pedagogical Purposes
• Using Text Generators for Help Preparing Courses and Assessing Students
• AI and Creative Writing
• Prominent Pieces That May Include Hype Or Inaccuracy
• Organizations
• Calls for Papers and Proposals

​

Sample AI-Generated Essays

Wrap-Up

Q&A from Mentimeter

Top points to remember

• Teach for intrinsic motivation
• Clarify policy
• Adjust pedagogy selectively (options include conferences, in-class writing, writing process assignments, audio and video assignments, and modified writing prompts.)
• Consider using AI text identification tools with caution

​

• Consider advocating for tools and support for educators around AI
• When you’re ready: teach about AI problems like mistakes and bias

Next steps: Share in the chat

What’s one concrete next step you’d like to take to sort out what to do on this topic. For example, “Research sample policies” or “Try my prompt in ChatGPT” or “Read about detection.”

Support this work?

I am contingent faculty and this is a labor of love. I hope to be able to continue curating sources, sharing sample essays, and holding free webinars. Thank you!

• Patreon
• PayPal

Let the discussions continue as we sort this out together. Thank you!��Anna Mills�armills@marin.edu�@EnglishOER��This presentation is shared under a CC BY NC 4.0 license.