There were several reports over the last weeks on how teachers began to use ChatGPT in various ways, mostly as a research assistant for their students. One example, as related by SunRev on Reddit:
My friend is in university and taking a history class. The professor is using ChatGPT to write essays on the history topics and as the assignments, the students have to mark up its essays and point out where ChatGPT is wrong and correct it.
As mentioned in an earlier roundup post, I’m preparing to let my students try out a few things with creative assistance in my upcoming game design lectures. But I have doubts about the use of LLM for research assistance. It certainly has its good sides; as ChatGPT and similar models are ridiculously unreliable, it forces students to fact-check, which is great. However, research isn’t—or shouldn’t be—all about fact-checking data. Rather, it should be about learning and internalizing the entire process of doing research, be it for postgraduate projects or college essays: gaining a thumbnail understanding and accumulating topic-specific keywords; following references and finding resources; weighing these resources according to factors like time/place/context, domain expertise, trustworthiness, soundness of reasoning, and so on; and eventually producing an interesting argument on the basis of source analysis and synthesis. I’m not sure if fact-checking and correcting LLM output is a big step forward in that direction.
Then, there’s the problem of research data contamination. xmlns=“Dan” on Mastodon:
There is a demand for low-background steel, steel produced before the nuclear tests mid century, for use in Geiger counters. They produce it from scavenging ships sunk during world war one, as it’s the only way they can be sure there is no radiation.
The same is going to happen for internet data, only archives pre-2022 will be usable for sociology research and the like as the rest will be contaminated by AI nonsense. Absolute travesty.
This might indeed develop into a major challenge. How big of a challenge? We can’t know yet, but it will most likely depend on how good LLM-based machines will become in differentiating between LLM output, human output, and mixed output. Around 2010, there was the Great Content Farm Panic, when kazillions of websites began to speed-vomit optimized keyword garbage into the web. Luckily, Google’s engineers upgraded their search algorithms in clever ways, so that most of that garbage was ranked into oblivion relatively quickly. Can Google or anyone else pull that off again, with regard to a tidal wave of LLM sewage? There’s no guarantee, but those search engines and knowledge repositories that become better at it will gain an advantage over their competitors, so there’s a capitalist incentive at least.
Finally, this bombshell suggestion by Kevin Roose in an article about ChatGPT for teachers (yes, that Kevin Roose of “Bing’s A.I. Chat Reveals Its Feelings: ‘I Want to Be Alive’” fame as mentioned in my last roundup):
ChatGPT can also help teachers save time preparing for class. Jon Gold, an eighth grade history teacher at Moses Brown School, a pre-K through 12th grade Quaker school in Providence, R.I., said that he had experimented with using ChatGPT to generate quizzes. He fed the bot an article about Ukraine, for example, and asked it to generate 10 multiple-choice questions that could be used to test students’ understanding of the article. (Of those 10 questions, he said, six were usable.)
In the light of my recent essay-length take on AI, ChatGPT, and Transformational Change at medium.com, particularly my impression that ChatGPT might at least liberate us from large swaths of business communication, multiple-choice tests, or off-the-shelf essay topics, this made my eyes roll back so hard that I could see my brain catching fire.