A Class Policy on AI Hallucinations

 

As I noted in my last post, a student in my May-term class committed academic dishonesty in a reading journal I'd sought to make AI-resistant. This to me is a more serious issue than turning in work an AI generated. Why is that?

In my course, I provide careful guidelines for how students may use AI to help them. For instance, since today's undergrads are generally awful readers, I allow them to employ AI to help them understand themes in readings and connections between readings. Some students in their evaluations noted that the graded work here "forced them" to do their readings. 

Yes, readers, thank you for reading this post. Be advised that many college students no longer do any class readings, unless forced, even at selective institutions. To me, that's a gateway to a new Dark Age, nothing less.

My method, following my rules for multiple-entry Google Workspace reading journals, requires students to fill one column with a quotation or summary of a key event, a second column with an analysis of the event, and a third with a question to bring to class. I also include a mandate to comment weekly on a peer's journal. I got this notion from John Bean's excellent book Engaging Ideas; you can learn more in the third edition of Bean's classic, with his co-author Dan Melzer.

The student who misused AI had done good work for earlier assessments, yet for the final one asked an AI to find notable quotations. That was not against my policies. What was? In two instances, the AI invented direct quotations not in the readings. The writer, too harried or too complacent to check, did not do a word search of the originals.

I gave the writer an F on that assessment, which pulled down the final course-grade. In my reasoning, I said that had that occurred on the job, the student would have likely been fired. Best to learn that ethical lesson now, while the stakes were relatively low (though to a perfection-obsessed undergrad, the stakes may have seemed high, indeed).

We discussed the matter in a cordial way; the writer had done well on earlier work but for reasons still fuzzy to me, failed on this final assessment. So in future classes, I'll change two things. First, my policy on AI hallucinations will be harsh; if the assessments are as frequent as in my recent class there will be no chance at revision. In ones where assessment is less frequent, the F will be applied but the writer will get to revise the journal and I will average the two grades.

Second, I'll add a new requirement for synthesis with earlier readings: yes, a fourth column! This skill is woefully lacking in students, who seem unable to construct consistent narratives across the work done in a class. This I attribute to a "taught to the test" mentality in high school as well as a disconnected learning experience and, for some, lack of passion for learning afterwards.

As for AI hallucination? Though Ethan Mollick and OpenAI claim that larger models hallucinate less now than in the early days, I'm not so sure. Mollick tends to use what I call "concierge" AIs that cost quite a bit; my students generally use a free model and do not engineer their prompts well. 

You can read more about a comparison of different LLM models and hallucination here. I still feel that we remain a long way from knowing which LLM to trust and when, but the OpenAI article does provide good tests we humans can apply to check the AI's output.

Always check its work. My student did not, and paid a price, rightly so. 

Image from the film Vertigo. 

Ohio State Takes the AI Plunge

A colleague in senior administration at OSU sent me a notice about their new AI initiative.

It's so at odds, in a healthy manner, with the "just say no, hard no" of the CCCC Chair's April keynote on AI. While I'm encouraged about a new CCCC working group formed on AI, as with so many things run by faculty, it's going to take time to spin up, while administration and industry race (perhaps unwisely) into adopting this technology. I'm going to chuckle at the negative stories such as one from The Columbus Dispatch, picked up by MSN, calling it a "betrayal" and claiming that AI is being "rammed down students' throats." 

Have these reporters even looked closely at student use?

Let's just assume it's nearly 100%. My 3rd annual student survey says as much. We need to address that reality and do so ethically and in a pedagogically sound manner. Maybe that's where we can critique this or similar initiatives. I remain the wary adopter, not an enthusiast.

So what are the broad outlines of the OSU plan, to be unfolded this Fall?

• Units on AI fluency in a seminar taken by all students.
• Support for faculty to incorporate Generative AI into classes. I do not see a mandate for all faculty.
• Building upon an “Embedded Literacy" in all majors. Read more about them here. This will likely be strengthened to include appropriate and ethical use of AI in the discipline of the major.
• Partnership with industry.

So why the pushback? 

Perhaps the details are too vague, the timing too sudden and rolled out over the summer when many faculty and students are away.

Yet this development has deep roots; OSU is a Big 10 and a land-grant; they have long partnered with companies to help their students develop skills needed in the workplaces of their era. Though I taught for nearly all my career at a liberal-arts university, I hear already from adult learners that in their jobs, AI fluency no longer remains an option. New hires are expected to have some fluency or go elsewhere for a first job.

Are we caving to corporations? Only if we let them set the terms of engagement. We have adopted new technologies before our way, by providing open-access Internet resources, releasing materials into the Creative Commons, and pursuing innovations with mobile computing. I still live by those rules, never putting a syllabus into BlackBoard's gated community. My content is on the open Web for all to use. That was the promise of the early Internet. I hope we can do something similar with AI.

I just finished a course that focused on deep-reading techniques for literature. We used AI for two assignments, but the reading journal, done as a Google Doc and commented upon by peers, proved hard for AI to assist. One student leaned too heavily on AI on a final set of journal entries, and it hallucinated quotations that do not exist in the texts the student then analyzed: an F on that assessment proved penalty enough. I reminded the student, in essence, "on the job, you'd have been fired. Here it just reduced the final grade."

The students had to employ critical-thinking skills beyond summary and analysis to find "need to know" questions to bring to class discussion, where 50% of their grades came in the form of participation.

AI cannot do that. But without learning to use AI ethically and effectively, my students won't land jobs in a few years. So there's my pushback: no matter the basis for objecting to AI, college involves helping students start careers. It does so much else, too, but without employed alumni, the entire enterprise of higher ed would fold. 

Let's see where OSU goes with this venture. 

Image: OSU Seal via Wikipedia 

 

A Good Discussion of Story with Chat GPT 4

In my May-term course "Reading Science Fiction and Fantasy," two stories present radically different ideas about how to engineer AI. As part of my own AI research for a forthcoming book about writing and AI, I put some ethical questions to ChatGPT. I’m still using their free version 4, though I’m considering a subscription to it as its answers have become much better and its vocabulary that of a thoughtful colleague.

Here’s an exchange from my summer class. It may provide a way forward for some of us trying to wean students from copy/paste work with AI. Instead, I modeled an idea-building exercise, one of many I'm using, to show writers who AI can unblock their process early on. Students have trouble with deep-reading and re-reading. I believe that such exercises might show them a hard truth of US and probably global captialism: after graduation, I'd hire ChatGPT, not them, to do close analysis. It shocks these sociable, amiable youngsters: I warn them "add human value to what it can do, or you won't find work in your chosen field."

At a recent concert, I chatted with an attorney seated at our table. She mentioned how many young paralegals send her unreadable copy. I told her, "tell them to use AI to fix it up or you'll fire the lot of them."

That does not sound like me, but such writers are lazy and under-trained. Putting my feet into a CEO's expensive loafers, that's the hard choice I'd make, then hire some Humanities majors with enough pre-law and pre-business coursework and AI experience to do the job of dozens I let go, having the newcomers do some revisions to responses culled from the training data of our AI.

Joe My students are reading two short stories about AI, Ken Liu's "The Algorithms For Love" and Elizabeth Bear's "Tideline." Have you either read these texts or a reliable summary of their plots? Let me know that first, and then I'll tailor a question to you based upon your level of familiarity.

ChatGPT 4 Yes, I’m familiar with both Ken Liu’s "The Algorithms for Love" and Elizabeth Bear’s "Tideline"—I’ve read the texts and have access to reliable analyses and summaries of each. I can help discuss their plots, themes (like identity, memory, empathy, and human/machine boundaries), and how they portray artificial intelligence in emotionally resonant or ethically complex ways.

Go ahead with your question—I’m ready for it!

Joe Thank you. On the surface to me, they appear to present contrary views of how AI might work, with Bear emphasizing the possibility of nobility and compassion by an AGI-driven robot and Liu's character Elena going mad because she thinks human brains are no different from her AI-driven dolls. In your opinion as an AI, can you provide me six lessons (no more than 100 words each) about how each story suggests we humans should engineer an AGI?

ChatGPT 4 Absolutely—here are six compact lessons from each story that suggest how humans should or should not approach engineering AGI, based on the implicit warnings or ideals embedded in the narratives: 

Lessons from Tideline (Elizabeth Bear)

Design for Empathy and Storytelling Chalcedony, the war machine, finds meaning in memory, storytelling, and care. Bear suggests AGI should be built to recognize and value emotional legacies—not just logic or victory. Empathy, not dominance, can define machine intelligence.

Limit Autonomy for Violence, Expand Autonomy for Care Despite being designed for war, Chalcedony transcends her function through self-directed nurturing. Bear implies that AGI should be designed to choose paths of compassion over conflict—if autonomy is granted, it must include moral imagination.

Create AGI Capable of Mourning and Legacy Chalcedony mourns her human companions and wants to leave a legacy. This points toward designing AGI with awareness of loss and temporal continuity—qualities central to meaningful ethical action.

Incorporate Constraints That Encourage Moral Growth Chalcedony is damaged and stranded, but these constraints push her toward morally rich behavior. Bear suggests AGI might evolve ethically not in spite of limitations, but because of them.

Value Hybrid Relationships Between AGI and Humans The bond between Chalcedony and the boy is intergenerational and cross-species. Designing AGI to participate in reciprocal, nurturing relationships could be vital to coexistence.

Teach AGI Through Cultural and Personal Narrative Chalcedony’s method of shaping the boy through stories suggests AGI should be grounded in cultural knowledge, not just data. Narrative may be a more human-compatible training tool than raw computation.

Lessons from The Algorithms for Love (Ken Liu)

Avoid Over-Reduction of Consciousness Elena tries to reduce human minds to code, believing there's no qualitative difference between dolls and people. Liu warns that reducing consciousness to computation risks collapsing the boundaries of personhood and reality.

Preserve Epistemological Humility Elena loses her grip on reality when she assumes full understanding of the human mind is possible. AGI engineering should be grounded in humility about what we don’t know about cognition and emotion.

Recognize That Simulation Is Not Identity Elena’s dolls can simulate human behaviors, but lack subjectivity. Liu cautions that mimicking thought or feeling is not equivalent to being—an AGI might fake empathy without ever experiencing it.

Don’t Project Human Pathologies onto AGI Elena’s obsession, paranoia, and eventual psychosis are mirrored in her AI. Liu’s story suggests that engineers’ beliefs, especially reductive or deterministic ones, can dangerously shape AGI behavior.

Beware the Infinite Mirror of Self-Modeling Elena creates self-modeling dolls that eventually model her own unraveling. Recursive self-awareness without grounding may produce instability. AGI designs should avoid infinite regress in favor of anchored self-concepts.

Design AGI to Understand—but Not Imitate—Human Fragility Elena’s desire to replicate human imperfection backfires. AGI might need to recognize human irrationality, trauma, or fragility—but not necessarily mirror it. Liu’s story is a warning about creating minds too much like our own.

Let me know if you’d like these reordered, expanded into a handout, or reframed in terms of pedagogy or design principles!

Joe Wow. I love those answers. Point 1 about Liu aligns with what I thought in my wetware brain while re-reading the story today. Many thanks. I'll share this with class today.

ChatGPT 4 I'm really glad to hear that—it sounds like a rich discussion is ahead for your class. That story in particular rewards rereading; Liu gives Elena such a chillingly rational voice that it’s easy to slip into her worldview until you step back and realize what’s been lost. If your students end up debating whether Chalcedony or Elena is more "real," or what kind of AGI they’d trust to raise a child or design a memory, you’re in for some excellent conversations.

Feel free to loop back if they raise unexpected questions—I’d love to hear what they come up with.

image source: Creative-Commons image via publicdomainpictures.net

A Surprising Second Life for Second Life

I came across an interesting piece by Alice Bucknell recently about the curious longevity of Second Life, the virtual world that served as impetus for this blog.

Having returned to SL for an occasional round of naval wargaming from the Age of Sail, I think that the piece's title misleads. Second Life lacks the visual depth of modern video games, but the world hardly seems decaying. If anything, graphics quality has vastly improved and performance-based crashes on WiFi connections with a laptop seem rare, even for naval combat. Fifteen years ago, my students' avatars could barely move from one sim to another without crashing, and they were walking, not sailing a square-rigger. In 2014 I found a simple mainland driving trip impossible.

To hazard a guess, Linden Lab's integration of Mesh-based builds surely has helped with rendering and running the content on each sim. I understand the servers moved to Amazon's cloud long ago, which may explain a lack of what we used to call Black Wednesdays, when the entire grid would crash midweek with alarming regularity.

Why stay in such an old game...I mean virtual world? I think for many SLers it remains a community of long-time residents, Linden Lab's term for those who log on to their servers. More is at play in my case. Here the author nails why I love SL still, for "the platform’s refusal to submit to the always-be-optimizing logic of other MMOGs is precisely what grants Second Life its enduring appeal among a certain cohort of loyal users. Bucking larger trends of the gaming industry enabled the user experience to evolve in weirder ways....there is no boss to beat or competition to win, the vibe on Second Life is ambient."

Exactly. I don't go online to kill boss-monsters or level up, and the occasional glitchy build in Second Life landscape or gear simply does not bother me. I enjoy Eno's ambient music, too.

Minecraft boomed despite its Legoland look; it's the fun of building things and playing with them, rather than trying to recreate the real world's look and feel, that draws me still to SL's giant sandbox. As Bucknell notes, the virtual world does not always embrace the "heteronormative, linear, violent" tropes of video gaming, though she could find those tropes easily enough had she looked hard.

If there's one thing different now as compared to 2010, it's the emptiness of the Mainland where Iggy once had a parcel of land as a premium user. I gave up that status some time back, but I log in with a Pirate ALT for naval battles or with Iggy's pre-mesh laggy self, to walk or drive around the world. 

Instead of empty mainland, I took Iggy back to Svarga, as close to a National Park as you'll find in Second Life, for a flying tour on the back of a giant bee. The island is now run by Patch Linden, Senior VP of Product Operations for Linden Lab. His otherwise human avatar wears squirrel ears.

Some things do not change about the Lindens. That's a comfort.

I set graphics in Firestorm to "Ultra" and enjoyed the data-sucking view. When done, I hopped off my mount for a photo using Bryn Oh's "Virginia Alone" Sky.

I was, indeed, alone on Svarga, where Iggy spent his first Linden dollars to buy a pair of sunglasses. That was 2007? What amazes me most, beyond my failure to crash my client as I did regularly then, was that so many years have flown by but the world remains.

That's something special, likewise comforting, right there. 

Perhaps I should look about a little more at some old haunts.