When GenAI models became widely available to the public, some educational institutions reacted by banning their use outright, particularly in assessments.
Those who took a more progressive view left the determination of the appropriate use of GenAI tools to the judgment of educators and academics who teach and design assessments and encouraged their responsible use.
However, widely usable assessments that produce a fair picture of learning achievement that either include the use of Generative AI (GenAI) tools or are resilient to them are still few. Among the latter, many universities are going back to old-school vivas/interviews and proctored exams to ensure that the learners’ responses are their own. Last month saw two interesting reports of widely reported instances of university professors.
The first was a case out of Brown University. Students in an economics course were allowed to attempt their mid-term exam as a take-home assignment, without help from AI tools. The result was a class average that was abnormally high by the standard of previous years. The final exam was then announced as on-campus, and students were told that the mid-term exam would be cancelled if their final exam performance did not match their mid-term performance.
Eighteen students dropped the course, leaving 59. Of those that remained, three did not appear for the final exam. Among those who attempted both exams, most had widely varying scores, confirming the professor’s suspicions that work submitted in the mid-term exam was, in the overwhelming number of cases, not students’ own work, leading him to void
its results.
The second case was at Alcorn State University in Mississippi, where a history professor used simple prompt injection, i.e., hiding a prompt in the assignment by appending white text on a white background at the end of a paragraph that instructed GenAI models to add a nonsensical sentence about Madagascar in the output (The assignment was about the Industrial Revolution and the digital age). This type of prompt injection in assignments has been widely used by faculty since the ChatGPT Large-language Model (LLM) came out at the end of 2023.
In both instances, students were expected to submit their own work, and most students chose to ignore those instructions. There can be a separate debate about whether the use of GenAI should have been allowed in the first place, but the fact is, it was not, which makes their use in both cases a cut-and-dry breach of academic integrity.
On the other hand, there are plenty of use cases of GenAI tools outside of academia. For example, employers will eagerly embrace AI tools for task completion if it raises productivity, as long as a person is capable and accountable for verifying their correctness. Outside of work, there are all manner of low-stakes uses – polishing or generating social media and blog posts.
But in between these uses, there are grey zones, and a significant one is in journalism and, more specifically, opinion pieces. There is an argument that op-eds are not academic writing, which makes the use of GenAI permissible. Newspapers have been establishing policies on acceptable use of GenAI in their content; those policies can vary quite a bit and are still far from some industry-wide convergence. Typically, AI use is often not allowed for hard news (where facts and numbers need to be correct), but permits limited use for op-eds (brainstorming, proofing, spellchecking are OK; generating articles wholesale is not).
Over the last few years, regular readers of Pakistani newspapers will have spotted a creeping transformation of the op-ed sections in newspapers. It is clear that many contributors, both established and new, have fully bought the LLM-KoolAid. Even occasional LLM users can easily spot the tells: Dramatic tone, lots of fluff, copious overuse of the dreaded “it’s not this. It’s that” structure, overuse of certain words, grandiose expressions, etc. Together, they have the effect of giving writing a tone that is also called ‘AI voice’.
The overnight shift in some contributors’ writing style has been jarring. I am actually quite surprised such prolific LLM users have not yet discovered ‘AI Humanisers’ – software tools that take text generated by LLMs and iron out those tells to make writing sound like it was produced by a human.
If an entire article is generated by an LLM and the prompter’s sole contribution is putting their name on it, can the author of the prompt, whose sole contribution is some light editing and putting their name on it, claim that the views expressed in it are their opinion? By academic standards, the author can be credited for phrasing the prompt, but not the output.
If some news outlets want to allow that, that is their prerogative. However, as a paying subscriber and reader investing my time, I should have the right to know to what degree an article is produced by AI, so I may decide whether or not I want to put in the time to read something, X per cent of which was produced by some LLM or whether I would rather skip to someone else’s writing.
What is missing is a transparent AI attribution, something similar to the ratings on movies and games, but provided voluntarily by the credited authors. In the interest of disclosure, I would like every article to be accompanied by a self-disclosure of AI use by its author: Spell-checked/edited/brain-stormed/written X per cent with AI tool Y. As a matter of fact, with the flood of long posts/threads on platforms like LinkedIn and X/Twitter, I would like to see the same on all manner of social media posts.
This article was spell-checked by Grammarly. Not brainstormed, edited or written by AI.
The writer (she/her) has a PhDin Education.
