Instructional & Media Services at Dickinson College

LLMs usually cannot do a good job of explaining how a particular passage from a longer text illuminates the whole of that longer text. Moreover, ChatGPT’s outputs on comparison and contrast are often superficial. Typically the system breaks down a task of logical comparison into bite-size pieces, conveys shallow information about each of those pieces, and then formulaically “compares” and “contrasts” in a noticeably superficial or repetitive way. 

In-class writing, whether digital or handwritten, may have downsides for students with anxiety and disabilities

ChatGPT can produce outputs that take the form of  “brainstorms,” outlines, and drafts. It can also provide commentary in the style of peer review or self-analysis. Nonetheless, students would need to coordinate multiple submissions of automated work in order to complete this type of assignment with a text generator.  

Students are more likely to misuse text generators if they trust them too much. The term “Artificial Intelligence” (“AI”) has become a marketing tool for hyping products. For all their impressiveness, these systems are not intelligent in the conventional sense of that term. They are elaborate statistical models that rely on mass troves of data—which has often been scraped indiscriminately from the web and used without knowledge or consent.

LLMs often mimic the harmful prejudices, misconceptions, and biases found in data scraped from the internet

Show students examples of inaccuracy, bias, logical, and stylistic problems in automated outputs. We can build students’ cognitive abilities by modeling and encouraging this kind of critique. Given that social media and the internet are full of bogus accounts using synthetic text, alerting students to the intrinsic problems of such writing could be beneficial. (See the “ChatGPT/LLM Errors Tracker,” maintained by Gary Marcus and Ernest Davis.)

Since ChatGPT is good at grammar and syntax but suffers from formulaic, derivative, or inaccurate content, it seems like a poor foundation for building students’ skills and may circumvent their independent thinking.

Good journalism on language models is surprisingly hard to find since the technology is so new and the hype is ubiquitous. Here are a few reliable short pieces.     “ChatGPT Advice Academics Can Use Now” edited by Susan Dagostino, Inside Higher Ed, January 12, 2023  “University students recruit AI to write essays for them. Now what?” by Katyanna Quach, The Register, December 27, 2022  “How to spot AI-generated text” by Melissa Heikkilä, MIT Technology Review, December 19, 2022  The Road to AI We Can Trust, Substack by Gary Marcus, a cognitive scientist and AI researcher who writes frequently and lucidly about the topic. See also Gary Marcus and Ernest Davis, “GPT-3, Bloviator: OpenAI’s Language Generator Has No Idea What It’s Talking About” (2020).

“On the Dangers of Stochastic Parrots” by Emily M. Bender, Timnit Gebru, et al, FAccT ’21: Proceedings of the 2021 ACM Conference on Fairness, Accountability, and Transparency, March 2021. Association for Computing Machinery, doi: 10.1145/3442188. A blog post summarizing and discussing the above essay derived from a Critical AI @ Rutgers workshop on the essay: summarizes key arguments, reprises discussion, and includes links to video-recorded presentations by digital humanist Katherine Bode (ANU) and computer scientist and NLP researcher Matthew Stone (Rutgers).

Silicon Valley’s sudden willingness to consider taking more reputational risk arrives as tech stocks are tumbling

A chatbot that pointed to one answer directly from Google could increase its liability if the response was found to be harmful or plagiarized.

AI has been through several hype cycles over the past decade, but the furor over DALL-E and ChatGPT has reached new heights.

Soon after OpenAI released ChatGPT, tech influencers on Twitter began to predict that generative AI would spell the demise of Google search. ChatGPT delivered simple answers in an accessible way and didn’t ask users to rifle through blue links. Besides, after a quarter of a century, Google’s search interface had grown bloated with ads and marketers trying to game the system.

Inside big tech companies, the system of checks and balances for vetting the ethical implications of cutting-edge AI isn’t as established as privacy or data security. Typically teams of AI researchers and engineers publish papers on their findings, incorporate their technology into the company’s existing infrastructure or develop new products, a process that can sometimes clash with other teams working on responsible AI over pressure to see innovation reach the public sooner.

Chatbots like OpenAI routinely make factual errors and often switch their answers depending on how a question is asked

To Timnit Gebru, executive director of the nonprofit Distributed AI Research Institute, the prospect of Google sidelining its responsible AI team doesn’t necessarily signal a shift in power or safety concerns, because those warning of the potential harms were never empowered to begin with. “If we were lucky, we’d get invited to a meeting,” said Gebru, who helped lead Google’s Ethical AI team until she was fired for a paper criticizing large language models.

Rumman Chowdhury, who led Twitter’s machine-learning ethics team until Elon Musk disbanded it in November, said she expects companies like Google to increasingly sideline internal critics and ethicists as they scramble to catch up with OpenAI.“We thought it was going to be China pushing the U.S., but looks like it’s start-ups,” she said.

large-language models like ChatGPT are often extolled as the cutting edge of artificial intelligence, it may sound dismissive—or at least deflating—to describe them as lossy text-compression algorithms. I do think that this perspective offers a useful corrective to the tendency to anthropomorphize large-language models

Even though large-language models often hallucinate, when they’re lucid they sound like they actually understand subjects like economic theory

The fact that ChatGPT rephrases material from the Web instead of quoting it word for word makes it seem like a student expressing ideas in her own words, rather than simply regurgitating what she’s read; it creates the illusion that ChatGPT understands the material. In human students, rote memorization isn’t an indicator of genuine learning, so ChatGPT’s inability to produce exact quotes from Web pages is precisely what makes us think that it has learned something. When we’re dealing with sequences of words, lossy compression looks smarter than lossless compression.

Even if it is possible to restrict large-language models from engaging in fabrication, should we use them to generate Web content? This would make sense only if our goal is to repackage information that’s already available on the Web. Some companies exist to do just that—we usually call them content mills. Perhaps the blurriness of large-language models will be useful to them, as a way of avoiding copyright infringement. Generally speaking, though, I’d say that anything that’s good for content mills is not good for people searching for information.

If and when we start seeing models producing output that’s as good as their input, then the analogy of lossy compression will no longer be applicable.

starting with a blurry copy of unoriginal work isn’t a good way to create original work

Having students write essays isn’t merely a way to test their grasp of the material; it gives them experience in articulating their thoughts. If students never have to write essays that we have all read before, they will never gain the skills needed to write something that we have never read.

Sometimes it’s only in the process of writing that you discover your original ideas. Some might say that the output of large-language models doesn’t look all that different from a human writer’s first draft, but, again, I think this is a superficial resemblance. Your first draft isn’t an unoriginal idea expressed clearly; it’s an original idea expressed poorly, and it is accompanied by your amorphous dissatisfaction, your awareness of the distance between what it says and what you want it to say. That’s what directs you during rewriting, and that’s one of the things lacking when you start with text generated by an A.I.

What use is there in having something that rephrases the Web?

There’s also a big difference between utilizing ChatGPT—which investment bank UBS estimates has 13 million users a day—as a standalone product, and integrating it into Bing, which handles half a billion searches every day.

Data centers already account for around one percent of the world’s greenhouse gas emissions, according to the International Energy Agency. That is expected to rise as demand for cloud computing increases, but the companies running search have promised to reduce their net contribution to global heating. “It’s definitely not as bad as transportation or the textile industry,” Gómez-Rodríguez says. “But [AI] can be a significant contributor to emissions.”

The environmental footprint and energy cost of integrating AI into search could be reduced by moving data centers onto cleaner energy sources, and by redesigning neural networks to become more efficient, reducing the so-called “inference time”—the amount of computing power required for an algorithm to work on new data.

Shadowbanning has been documented for years, but the Guardian journalists may have found a missing link to understand the phenomenon: biased AI algorithms. Social media platforms seem to leverage these algorithms to rate images and limit the reach of content that they consider too racy. The problem seems to be that these AI algorithms have built-in gender bias, rating women more racy than images containing men.

“You are looking at decontextualized information where a bra is being seen as inherently racy rather than a thing that many women wear every day as a basic item of clothing,”

“You cannot have one single uncontested definition of raciness.”

these algorithms were probably labeled by straight men, who may associate men working out with fitness, but may consider an image of a woman working out as racy. It’s also possible that these ratings seem gender biased in the US and in Europe because the labelers may have been from a place with a more conservative culture

“There’s no standard of quality here,”

“I will censor as artistically as possible any nipples. I find this so offensive to art, but also to women,” she said. “I almost feel like I’m part of perpetuating that ridiculous cycle that I don’t want to have any part of.”

many people, including chronically ill and disabled folks, rely on making money through social media and shadowbanning harms their business

Social isolation contributes to mental health problems. Social isolation contributes to domestic violence.

Do not require synchronous work. Students should not need to show up at a specific time for anything. REFUSE to do any synchronous work.

Do not record lectures unless you need to. (This is fundamentally different from designing an online course, where recorded information is, I think, really important.) They will be a low priority for students, and they take up a lot of resources on your end and on theirs. You have already built a rapport with them, and they don’t need to hear your voice to remember that.

Do record lectures if you need to. When information cannot be learned otherwise, include a lecture. Your university already some kind of tech to record lectures. DO NOT simply record in PowerPoint as the audio quality is low. While many people recommend lectures of only 5 minutes, I find that my students really do listen to longer lectures. Still, remember that your students will be frequently interrupted in their listening, so a good rule is 1 concept per lecture. So, rather than a lecture on ALL of, say, gender inequality in your Intro to Soc course, deliver 5 minutes on pay inequity (or 15 minutes or 20 minutes, if that’s what you need) and then a separate lecture on #MeToo and yet another on domestic violence. Closed caption them using the video recording software your university provides. Note that YouTube also generates closed captions [edited to add: they are not ADA compliant, though]. If you don’t have to include images, skip the video recording and do a podcast instead.

Editing is a waste of your time right now.

Listen for them asking for help. They may be anxious. They may be tired. Many students are returning to their parents’ home where they may not be welcome. Others will be at home with partners who are violent. School has been a safe place for them, and now it’s not available to them. Your class may matter to them a lot when they are able to focus on it, but it may not matter much now, in contrast to all the other things they have to deal with. Don’t let that hurt your feelings, and don’t hold it against them in future semesters or when they come back to ask for a letter of recommendation.

Allow every exam or quiz to be taken at least twice, and tell students that this means that if there is a tech problem on the first attempt, the second attempt is their chance to correct it. This will save you from the work of resetting tests or quizzes when the internet fails or some other tech problem happens. And since it can be very hard to discern when such failures are really failures or students trying to win a second attempt at a quiz or test, you avoid having to deal with cheaters.

Do NOT require students to use online proctoring or force them to have themselves recorded during exams or quizzes. This is a fundamental violation of their privacy, and they did NOT sign up for that when they enrolled in your course.

Circumvent the need for proctoring by making every exam open-notes, open-book, and open-internet. The best way to avoid them taking tests together or sharing answers is to use a large test bank.

Remind them of due dates. It might feel like handholding, but be honest: Don’t you appreciate the text reminder from your dentist that you have an appointment tomorrow? Your LMS has an announcement system that allows you to write an announcement now and post it later.

Make everything self-grading if you can (yes, multiple choice and T/F on quizzes and tests) or low-stakes (completed/not completed).

Don’t do too much. Right now, your students don’t need it. They need time to do the other things they need to do.

Make all work due on the same day and time for the rest of the semester. I recommend Sunday night at 11:59 pm.

This advice is very different from that which I would share if you were designing an online course. I hope it’s helpful, and for those of you moving your courses online, I hope it helps you understand the labor that is required in building an online course a bit better.