Reassess Our Large Language Models

in Editorials | May 3, 2024

          Immediately following ChatGPT’s 2022 release, the usage of generative A.I. skyrocketed worldwide. Generative A.I. became the ultimate shortcut: huge chunks of text shot out at lightning speed, summaries captured and synthesized phrases from all over the Internet, and even digital art sprouted magically from a machine at the click of a button—free of charge. 

          While the Lawrenceville community recognizes A.I. as a useful tool, implementing such advanced technology into our education requires additional discussion and consideration from the administration. Teachers are conservative in adapting to the fresh abilities the software presents, unwilling to risk a lack of control and awareness of their students’ progress. Schools have resisted changes to new technology, as they believe an unfamiliar adaptation brings more harm than good. For example, when whiteboards were introduced in the 1950s, many schools did not implement them and stuck to blackboards until the mid-1990s. Nowadays, however, most schools primarily have whiteboards in their classrooms. Today, in terms of A.I. usage, our own school culture and rulebook emphasize abstinence over awareness. Instead of assessing its strengths and weaknesses, teachers must detect A.I.-generated text, then penalize its submission. In the Editorial published on April 12, 2024, the Board argued that disciplinary consequences for students falsely implicated by A.I. detectors are unjust. The true situation is even thornier: Lawrenceville’s abstinence-focused approach is inherently inequitable and inflexible, resting on the twin, unstable pillars of unreliable detection methods and stubborn denial of technological realities.

          Many claim that inconsistencies in writing style between a student and a Large Language Model (LLM) betray the presence of an artificially generated text. While identifying such inconsistencies might have been possible in 2023, many models have since evolved and can now imitate the users’ tone and language; not many opponents of generative A.I. usage seem to have interacted with these state-of-the-art LLMs—such as Anthropic’s Claude Opus and Google’s Gemini—thus retaining their misconceptions and misunderstandings. Additionally, paid models of A.I. more consistently evade Turnitin’s detection, which gives rise to equity issues: Students who can afford these models are less likely to receive disciplinary actions. Moreover, developers have been refining their models to include retrieval steps and reranking mechanisms, which implement A.I. detectors within the generative A.I. to convincingly render an organic final output. Even without such built-in detectors, new A.I. models sound “human” enough to easily evade detection. By failing to acknowledge the abilities of these models, the School overestimates Turnitin’s accuracy and underestimates generative A.I.’s ability to mimic human writing. Our blind ban on A.I. usage stubbornly ignores modern students’ realities. 

          Lawrenceville must accept that embracing A.I. usage and upholding most of our traditional values are not mutually exclusive. For example, the School could teach students how to use A.I. to complete expedited tasks these machines will soon render obsolete. Similarly to our abilities to do arithmetic both by hand and using a TI-84 calculator, students could learn how to use A.I. while remaining capable of fundamental skills A.I. might possess. By providing A.I. literacy courses or integrating A.I. into course requirements, Lawrenceville could ensure this fluency for every student. Take essays, for example: Soon, LLMs may write better essays than students. In response, the School should assign more in-class essays, especially for underformers, in order to develop their essay writing skills and encourage them to think critically. For upperformers who have already passed foundational English classes, A.I. should be integrated into courses to help enhance their work, from brainstorming to editing. 

          Similarly, ethical A.I. usage will not hinder students from pursuing accuracy. To achieve the most effective human-A.I. collaboration, we must prevent ourselves from being deceived by models—which is far more difficult than it sounds. For instance, “A.I. hallucination”—LLMs generating false but seemingly credible information—has been an issue inherent to LLMs. Therefore, students would have to continue double-checking first-hand sources, citations, and statistical information for accuracy, similar to how we write papers in courses today. 

          Specifically, a teacher interested in incorporating A.I. into their classroom effectively might grant students the option to interact with A.I. for inspiration. Rather than using A.I. creations as bland templates to be replicated, which would promote the unethical practice of plagiarizing and labeling A.I. generation as one’s original work, underformers should first learn to devise accurate, strong thesis statements that showcase their critical thinking skills. After developing a strong, personal voice in their writing, students may then be allowed to have A.I. assist them in refining their writing’s language and syntax. By the same token, faculty should allow students to use A.I. generation as a source of inspiration when composing essays. In “Ghosts,” a required reading for the 400-level Essay Writing course, the author Varuni Vara integrates GPT-3 into her writing process, finding herself “irresistibly attracted…to the way it offered, without judgment, [the ability] to deliver words to a writer who has found herself at a loss for them.” Although Vara specifically struggled with elaborating on her sentiments for her sister’s death while Lawrenceville students scramble to scout creative ideas for their English assignments, our faculty can offer support, akin to what Vara received from GPT-3, to students who wish to exceed departmental expectations and experiment with different styles. 

          Since every novel technology from the computer to the calculator has forced a shift in pedagogy, crafting a long-term approach to the use of A.I. in educational institutions is not a question of “if” but “when.” Any preparatory education should implement measures to make sure that students are prepared for a future of A.I. Our proposition: Lawrenceville needs to introduce human-A.I. collaboration, which research suggests produces works of higher quality compared to those produced by solely A.I. or humans. Historically, Lawrenceville has approached any proposed change with caution, as demonstrated by its later adaptation to being a co-ed institution compared to other private boarding schools. The rapid emergence of A.I. presents a unique chance for Lawrenceville to redeem its reputation as a conservative institution apprehensive of much-needed innovation, and the longer we drag our feet about dealing with this progressive change, the more the School will compromise students’ education and risk the continuation of socio-economic and disciplinary injustices. A.I. is not an innately abominable tool that should be kept from students forever, as A.I. will inevitably pervade students’ lives as they enter society. Adapting to A.I. is not synonymous with accepting defeat against A.I.; the naive, fearful mindset Lawrenceville currently subscribes to will never enable true progress. It’s time to ROLL: Reconsider Our Large Language models.