Stephen Downes

What I like about this article is that it gives us an extended extract from the beginning of Winter Takes One, "an 89,000-word upmarket thriller" that was written by an AI and "led to a four-way bidding war and a $3.5 million, two-book deal." It's an example of 'superslop', AI writing that actually fools professional editors. I also agree with the conclusions, two of them: "Commercial writing, at any level of quality, might be dead. It's already nearly impossible for a writer to earn a living on the quality of text"; and "Superslop will continue to evolve. It will defeat AI autograders. It will defeat human literary agents." Photo by me.

This article might have a good story to report - an Oklahoma school that is completely automated that costs more than any other school in the state. But it's mostly an exercise in weak AI writing - see the process here. On the surface it seems to be well written, but oh! it goes on and on and on, doesn't have any real sense of a narrative arc (much less of the inverted pyramid), and is filled with filler. I read every one of the 5104 words in what should have been, at most, a 1000 word article. So much agony! Like this: "W. Ian O'Byrne, an associate professor of literacy education at the College of Charleston who has studied digital literacy for two decades, put the pedagogical objection cleanly in The Conversation on 15 June 2026: effective learning is not always efficient." Or, "The legal position is worse than most parents will assume." Or "The second half of the question is what this does to teaching as an occupation, and here the answer has less to do with any individual school than with the signal a price sends." 

I don't use AI to write any of my text mostly because (a) I think I'm a better writer, and (b) the AI can't draw on my experience the way I can. The first may be hubris and is entirely subject to change, but I can't imagine the second ever changing. I don't have any opposition to AI, and I use it to generate code and to remove noise from my photos. And it's OK with me if other people use AI to write (though I really wish they would optimize for brevity). But key in all this is that AI is a tool that should be subject to the wishes of the person using the tool, and that's where Anthropic's watermark fails. "Watermarking is misdirection: a way for goods to appear to do one kind of thing while secretly doing another. Real writers (e.g. Gruber, Teare, and Jarvis) are calling bullshit on it." This is really important, because if it can secretly do one thing, then it can secretly do any number of things. I can protect my code with tests and my images with my eye, but there's no way to protect my writing.

I'm not sure what advice I would give to students any more. It wouldn't be this (24 page PDF) - it would never have been this, which mostly amounts to "do what you're told" with a little bit of "show some initiative" added. That would never have worked for me, and would leave students falling far short of getting the best out of university. I think there are three keys to success (or, at least, there were, before AI): writing well, which also supports reading well, and is the key that unlocks the content; connecting with other students to share the workload; and extracurricular activities to develop skills in what's valuable to you. The usual "habits of excellence" help if the details become overwhelming, but for the most part success is created when you define what counts as success and then work the system to attain that.

This report (52 page PDF) is styled as a 'rapid literature review', and while I'm not exactly sure what that means (I've asked) the result is still worth a look. The authors studied 271 of 975 papers considered to compile research which is admittedly still in its early stages. "Affective gains are common. GenAI use is often associated with increased interest, engagement, motivation, and reduced anxiety," however, "evidence of impact on higher-order reasoning, transfer, self-regulated learning, and durable learning is more uneven." A lot depends on external factors (this should surprise no one). "This includes the characteristics and developmental needs of learners, how the GenAI tool is designed, what the task asks students to do, who participates in the activity, and the wider cultural and institutional context."

I am of the belief that television caused far more harm to society than social media ever did, or even could. As LarsGoran Bostrom writes, "Television did not assault its viewers. It invited them to surrender. The screen demanded attention but offered no resistance, no requirement for response beyond the passive receipt of images flowing at thirty frames per second." Into the passive and receptive receiver it poured propaganda and advertising, building a comfortable mythology alongside a consumer society. I've always felt that digital technology offered a possible way out. Bostrom writes, "Will we remain the watched, the fed, the guided? Or will we reclaim the role of the watcher, the thinker, the one who chooses? The answer depends on whether we remember how to turn off the screen..." No. Not exactly that. The answer depends on whether we use the screen, or whether we merely let it use us.