Discouraging Unauthorized AI Use: Challenges and Opportunities

Most faculty and instructors prefer to restrict GenAI use to some degree—whether by limiting how AI is used generally, limiting AI use to particular assignments, or largely prohibiting the use of any Generative AI tools in any capacity in a particular course. Although instructors might prefer mechanisms that prevent unauthorized GenAI use (e.g., deploying surveillance technologies, retreating to blue-book examinations, etc.), decisions by AI companies, the ubiquity of GenAI tools, and the marketing hype surrounding their capacity suggest that investing effort in higher and better walls and stronger gates may be a losing battle. Alternatively, instructional strategies and recommendations to help students make better choices are likely to be more effective deterrents to unauthorized GenAI use.

Three Persistent Challenges to Preventing Misuse

What “Natural Language Processing” has to do with it

Artificial intelligence technologies serve multiple purposes (pattern recognition, data analysis, automation, computer vision, remote sensing, and others), integrating powerful computational models into many disciplinary processes and practices. Like other tools, natural language processing research applies the statistical power of computation and recognition to patterns. In this case, it examines patterns of written communication, extrapolating predictive relationships by comparing inputs against unfathomably large archives of human-written texts. What is crucial to keep in mind is that these statistical patterns are not the same thing as meaning, value, or understanding. AI responses are artificial, but not intelligent (in the human cognition sense).

Generative AI tools built on large language models are designed to mimic human expression. With sufficient computing power (and natural resources) and large enough training sets (often gathered without writers' consent), contemporary models are becoming extremely effective at this mimicry. While some companies are exploring digital watermarking as a deterrent to misrepresentation, ultimately companies have designed their interfaces to give users the impression that they are interacting with a well-informed, polite, and conscious assistant, rather than what you are actually doing: providing tokens to a synthetic prediction apparatus that extrudes probabilities. When the expectations of a scenario are commonplace, AI technologies can give the impression of meaningful conversation, but the meaning making is in the head of the reader.

What “Ubiquity” has to do with it

A common maxim in the study of technological innovation is that every innovation has one of two fates: obsolescence or ubiquity. While an incandescent lightbulb might be disappearing, electric lighting isn’t likely to go away. Writing technologies are similar, and those deeply invested in Generative AI hope that it will quickly become a non-controversial expectation, like standard spelling, the book, and digital subscriptions. It’s hard to use any app without it trying to secure pages and pages of permissions in its terms of use.

The feeling some users have that AI is suddenly unavoidable is no accident. Google’s most profitable product (search) now offers top billing to its own GenAI responses in search results (even before sponsored content!). Microsoft’s Office Suite and its cloud versions include easy AI integration, and make it somewhat challenging to turn off. Media theorist and critic Cory Doctorow has colorfully identified this as a core strategy for tech companies: make a great product easily accessible and cheap, create dependence on that product so that surrendering it is a hardship, and then slowly increase prices and decrease costs to eventual profitability.

What “Hype” has to do with it

In their book, The AI Con, Emily Bender and Alex Hanna offer a concise history of the AI Hype Cycle, tracing it back to the Cold War Origins of early AI research. (Read an excerpt here). At the end of the chapter, the authors note:

“What all of these stories have in common is that someone oversold an automated system, people used it based on what they were told it could do, and then they or others got hurt.”

Like the promise of dating apps and self-driving cars, those interested in selling products and services have every incentive to overpromise the capabilities of their systems. By promoting flashy examples of success, spinning stories about the inevitability of our AI-enabled future, and cultivating a fear of missing out on the next big thing, tech companies, popular media, and even political leaders are pushing for the adoption of proprietary tools and participation in the AI economy. Against the breathless accounts of an AI future, critics of AI use in educational contexts are drowned out and dismissed. Messages to students often stoke generational differences and “us vs. them” thinking, encouraging them to ignore their instructors' cautions and jump on the AI bandwagon.

Three Quick Strategies to Discourage AI Misuse

Reducing ‘friction’ at the beginning of the assignment

For some students, especially newcomers to a field, topic selection can be daunting. Instructors can spend some class time describing the features of appropriately selected and scaled projects to support this process. Rather than simply providing examples of successful prior projects, investigate the features that make such projects successful and articulate them as assignment expectations. Alternatively, offering a selection of curated topics can also help students overcome the initial hurdle of the blank page.

Giving students opportunities for early exploration and experimentation can help them start assignments based on their own ideas and interests. Rather than turning brainstorming over to generative AI tools, give students class time to explore topics and their interests and create a topic proposal assignment. Early engagement and choice can increase students' commitment to their projects.

Encouraging capacity building as a learning goal

In some cases, students turn to Generative AI when they doubt their own abilities or lack information about the tasks or genres they are assigned. By encouraging students to view their assignments as chances to build skills and capacity, instructors can promote student learning and growth.

Breaking complex assignments into parts, providing structured activities to allow parallel learning, and encouraging students to identify their points of struggle and confusion can normalize the challenge of complex assignments. Of course, the more instructors can emphasize that productive struggle is evidence of learning, the less likely that students will give up and seek a technological workaround.

Encouraging student skepticism of AI

While Generative AI tools are fast, they are also notoriously error-prone. While the tone and style of GenAI output can deceive novices who lack subject-matter expertise, experts can easily spot errors and issues in AI-generated prose. Fortunately, the internet provides a regular catalog of spectacular and embarrassing AI failures that you might use as examples in class. The online technology website tech.co offers a consistently updated list of spectacular and costly AI failures. Instructors can also show how poorly constructed AI queries can produce embarrassing errors in the context of their own subject-matter expertise.