Is coursework dead? A new study just released could well be the final nail in the coffin. To understand why coursework now represents the worst form of assessment, we need to understand two different concepts: the toupee fallacy and the 88% problem.
In this blog:
The Toupee Fallacy
We all know a bad wig when we see it. But because we are easily able to spot really bad and obvious toupees, we tend to overestimate our ability to spot really good ones. Something similar happens with AI-generated work. Because we are good at spotting bad examples, we are less likely to spot good examples of AI-generated work.
There are some common features that people often associate with AI-generated work. These include predictable patterns and structure, overusing the rule of three, em-dashes and American spelling. For example, when looking at the following passage, most people would quickly and correctly realise they are reading an AI-generated text:
Because we find it easy to spot the above as AI, we become over-confident in spotting all AI-generated work. For example, consider this fascinating experiment run in the New York Times. They found that in a test where you don’t know if it is AI-generated vs some of the best historical human-written text, people’s preference was basically a coin toss between the two. And this was true across a range of genres: literary fiction, fantasy, science writing, historical fiction and poetry. You can try this test for yourself here. This is a perfect example of the Toupee Fallacy.
The 88% Problem
We have previously blogged about how AI detection software doesn’t really work and isn’t reliable (the exact quote from the research is about as conclusive as you tend to get in research papers: “Detection tools are neither accurate nor reliable”). And yet, schools and colleges are being increasingly sold software that promises to detect if students have used AI to generate entire essays.
Recent research provides further weight to this, with researchers concluding that “detectors appear more adept at identifying low-cognitive-load tasks but often fail when confronted with outputs of higher-order thinking”.
And to make matters worse, with a few simple tweaks, students can avoid it 88% of the time. When a student makes small, deliberate changes, the statistical fingerprint of the text can shift enough to avoid detection. This is known as The Laundering Effect and students are wise to it. In the last few months, we have heard of students doing the following to cover up their AI-generated work:
- Using prompts to include a few spelling mistakes
- Telling AI not to use em dashes
- Using a few synonyms to change the AI output
Where does this leave us?
The evidence is pretty clear. We can’t really spot good AI-generated work using the old-fashioned eye test, and with minimal changes, technology really struggles to detect AI-generated work. This means that students who use AI are likely to get away with it, but it also leads to a culture which is ripe for false positives where genuine student work is incorrectly flagged as AI-generated.
As Daisy Christodoulou says in our recent Science of Learning podcast episode, the cost of this can be hugely damaging.
Final Thoughts
If we don’t know if students have done the work, then how much weight should we give it as part of a formal assessment? And what could replace unsupervised coursework? Oral assessments? These are very hard (maybe impossible?) to do at scale. Handwritten pen and paper exams? That seems more likely for now. But in an increasingly digital age, pressure will come from a range of sources to abandon this.
What this ultimately means for schools and colleges is difficult to say. The truth is there are no simple answers to complex challenges. And AI is arguably the most significant and complex challenge currently facing us. We don’t have all the answers. No one does. But we firmly believe that research can illuminate. It can shine a light to help us not stumble completely in the darkness. Knowing about the Toupee Fallacy and the 88% Problem can hopefully offer a flicker of light.
