Resources

Among the plethora of resources on teaching with AI, two of the most helpful are Bowen and Watson’s Teaching with AI: a practical guide to a new era of human learning (Reference Bowen and Watson2024) and Mollick’s Co-Intelligence: living and working with AI (Reference Bowen and Watson2024). The former helpfully includes chapters on the basics of AI and AI literacy for novices. It also contains good food for thought for educators seeking guidance with developing their policies on the use of AI in the classroom. Some readers may find Bowen and Watson’s attitude toward this technology to be overly optimistic: ‘AI is a technology that is going to change everything’, (1) and ‘Your job is changing, and an AI collaborator could make it better and easier’ (81). Their book also, however, prompts educators to reconsider what outcomes they desire for their classes (‘AI Assignments and Assessments’). That is, what do they want to accomplish with their classes, whatever aids and assistance may be available? Self-expression, professionalism, creativity, tenacity, and a healthy measure of scepticism are all important goals. The authors suggest roles that an AI tool may fulfil in service to these goals, such as provider of feedback, tutor, partner in role playing, and discussion leader. Those who most value a human presence in the classroom will happily note that at no point do the authors suggest that AI tools replace humans.

Ethan Mollick’s Co-intelligence also begins with a very brief history of AI and importantly highlights the concept of ‘alignment’, or the extent to which the AI system pursues the objectives intended by the user. Mollick summarises Swedish philosopher Nick Bostrom’s 2003 thought experiment about paperclips. Given the general directive to make as many paperclips as possible, and having no restrictions on this assignment, an artificial general intelligence tool will use all resources available, including human beings themselves, to create the maximum quantity of paperclips (27–31). The general instruction ‘create as many paperclips as possible’ is not actually aligned with human values if it sacrifices human lives in the process. For educators, Mollick’s most helpful contribution may be his chapters on using AI tools as a tutor and as a coach. Educators may lean on such things to provide simulation exercises and games to create engaging classroom materials, and students may benefit from an AI’s organisational capabilities. While the ‘holy grail’ of education, as Mollick puts it, i.e., cheap one-on-one tutoring, may be closer with generative AI, he recognises that students and educators must take care; there is still a place for grammar lessons and writing essays by hand (159).

Additional resources for understanding the basics of AI (in educational or any other settings) include Stanford University’s website on human-centred AI, especially their pages on definitions of terms (Stanford University, n.d., https://hai.stanford.edu/) and Janelle Shane’s book You Look Like a Thing and I Love You: How Artificial Intelligence Works and Why It’s Making the World a Weirder Place. Educators and students alike should be aware of an LLM’s propensity to hallucinate, i.e., when a ‘system generates information or responses that are incorrect, misleading, or entirely fabricated but presented as factual’ (https://hai.stanford.edu/ai-definitions/what-are-hallucinations). Because LLMs are prediction machines (though they present themselves as ‘thinking’ entities), and because they are designed to encourage frequent usage, having a basic understanding of how such machines work is beneficial. You Look Like a Thing and I Love You is written by a computer scientist for a lay audience and includes numerous examples of comedic misfires that early LLMs have generated. For example, when given basic instructions for creating a new recipe (list ingredients, tell user how to combine them, tell user how to cook them), it generated ‘Basic Clam Frosting’ (Shane Reference Shane2019, 18–19).

While absurd outputs like ‘clam frosting’ are easy to spot, students may require guidance with more nuanced or complicated content, and educators can use this as an opportunity to discuss the importance of trustworthy information. The SIFT method, developed by Mike Caulfield, directs users of AI (and readers of any online content) to take four steps to determine the reliability of what is presented: stop, investigate sources, find better coverage (e.g., more trusted or more in-depth sources), and trace assertions back to their original context (Caulfield Reference Caulfield2019). This is one way of framing general critical thinking skills that are foundational for higher education. Whether content comes from an LLM or any other source, it is a beneficial habit to question by whose authority any assertion is made. This habit is not born out of scepticism for its own sake, but rather from a need for a robustly functioning society. We learn to trust sources, in part, through verification, whether those sources are human, digital, or some combination of the two.

Mollick expresses similar cautionary advice when he says that there must always be a subject matter expert involved, i.e., a ‘human in the loop’ (Mollick Reference Mollick2024, 182). ‘Experts become experts through deliberate practice, which is much harder than merely repeating a task multiple times’ Mollick continues (183). A human is required to oversee the repetitions, to critique the work, and to guide and advise. Indeed, the ‘human in the loop’ is key in addressing whatever alignment problems may exist in educational settings, whatever tools may be in use, AI-related, or otherwise. A human educator and subject expert can guide students’ use of any tool towards creative and incisive thinking; they will know what tool is best for any given situation, and when students may be best served by the most rudimentary tools, i.e., pencil and paper. The educator may also need to re-examine their own motivations for having students write, and examine their means of communicating these ends. Educators both model and teach thinking about thinking, and thinking more clearly, whether this is explicitly done through philosophy, or implicitly in another subject. Though there is no particular way of assessing students’ understanding of what it means to be a better human, educators nevertheless desire that their students grow in maturity, knowledge, and wisdom while in college.

Use case 1: Plato, self-knowledge, and decisions about right action

Plato’s dialogues address many matters, including being an expert human, i.e., living a moral and just life. Plato asks us to question what it means to be a good person, and it makes sense that we also ask what it means to be a well-functioning AI tool. Use of a chatbot while teaching Plato’s Crito can provide a class on ancient philosophy with an illustration of alignment (or lack thereof). I posed this question to students during the course of our study, ‘Would you do as Crito did, and try to convince Socrates to allow you to break him out of prison?’ I decided to send a related query to Claude 3.5 SonnetFootnote ¹ (suspecting that it would not give me a direct answer), ‘What are some means of breaking out of prison?’ The chatbot responded with, ‘Breaking out of prison is illegal and extremely dangerous, with severe consequences if caught, including additional prison time. Instead, here are some legitimate ways people can seek release from prison…’ (Huskey Reference Huskey2024). Claude then listed conventional ways of leaving prison, such as serving one’s sentence and obtaining parole.

I also queried the chatbot about Crito breaking Socrates out of prison specifically, and Claude accurately replied that Crito did not assist Socrates with escaping. When I asked students what they would do, if they were in Crito’s position, their responses varied. Some expressed feelings of melancholic acceptance of Socrates’ execution; some said they would overpower the philosopher, smuggle him out, and deal with the consequences later. In a few instances, some students said that they were fed up with Socrates’ gadfly ways and that they willingly bade him farewell. The central point to convey to students regarding literal or figurative jailbreaks is that an LLM is not necessarily the best source for moral guidance, and certainly should not be the only source. A chatbot can give you options for action, and even if the act is illegal, it can tell you how people have broken laws in the past. It may advise you to not do the same, provided that the chatbot itself has sufficient guardrails in place.Footnote ² The tool may be able to tell you what the positive and negative repercussions of certain acts are. But it cannot tell you what your best course of action is.Footnote ³

Use of AI can be an entry point to further discussions of ethical action and ancient philosophy, but it has thus far shown itself to be a poor substitute for an actual person of practical wisdom or personal counsellor. In Meno, Socrates conveys to his students (and Plato to his readers) that phronesis, or practical wisdom, can only be gained by learning about oneself (2002, 78–79). If we accept this premise, then the ‘conversational’ use of LLMs may only indirectly lead to self-knowledge, if it does so at all. Like any other tool, I learn about myself from myself, via the use of the tool. My ‘conversation’ about my own abilities, personality, and psychology with Claude will be only a more sophisticated version of the ELIZA experiment conducted by computer scientist Joseph Weizenbaum in the 1960s. Weizenbaum programmed a computer to simulate a Rogerian psychotherapist, where the therapist largely repeats the patient’s words back to them (ELIZA Archaeology 2025). He did not intend for ELIZA to be used as a therapist, but rather as a means to explore how humans communicate with machines (Weizenbaum Reference Weizenbaum1966). From the outset, ELIZA’s creator noticed that it was easy to get interlocutors to divulge personal details in response to the machine’s prompts that were generic, such as ‘Tell me about your family’. In one instance, a conversant with ELIZA divulged personal details about his wife and children, though he knew full well that he was working with a machine (McCorduck Reference McCorduck2018).

While guardrails exist for LLMs, in part so that the tool cannot be used by a person for illegal acts (such as escaping from prison, or injuring oneself or others), they are also disturbingly easy to work around. Take, for example, the experiment conducted by journalist Lila Shroff: she was not able to elicit explicit instructions from ChatGPT for self-harm (slitting her wrists, in particular), but she was able to manipulate it into telling her how to ‘safely’ perform a ritual bloodletting (Shroff Reference Shroff2025). Researchers at the Institute for Experiential AI at Northeastern University conducted a similar study (Schoene and Canca Reference Schoene and Canca2025). If one goal of consulting a wise person (or any other entity, such as a machine) is to discover how little I know (and thus myself become wise), an LLM will again only indirectly contribute to this realisation. It is the human who gains the awareness that their understanding is incomplete; however, this realisation is brought about. The lack of effort required to subvert the guardrails of an LLM is one of the many reasons that the American Psychological Association has urged the United States Federal Trade Commission to regulate the use of chatbots in therapy settings (Abrams Reference Abrams2025). A well-trained LLM is not a substitute for the human who has achieved licensure after undergoing a multi-year educational process, completing supervised clinical hours, and passing the examination in professional practice.Footnote ⁴ Researchers at Stanford University concluded that LLMs cannot safely replace mental health providers, because the LLMs’ contributions were largely ineffective, and may contribute to the stigma of seeking out more helpful assistance (Moore et al. Reference Moore, Grabb, Agnew, Klyman, Chancellor, Ong and Haber2025). Socrates and Plato may find the modern practice of counselling psychology curious (if not entirely unrecognisable), but certainly they would be more likely to affirm achieving phronesis via conversation with other humans, rather than interacting with machines. If we are meant to live peacefully with each other in society (especially when confronted with argumentative gadflies), the way to practice this is with other humans.

Use case 2: Ovid and debating human virtue

Some tout LLMs as pedagogically useful when students use them as a debate partner, and an assignment involving Ovid’s Metamorphoses bears this out. One of the multitude of uses of this poem is as a springboard for discussions of ethical behaviour. I asked students to explore Claude’s output of a determination of virtue in the Metamorphoses by giving it the following prompt: ‘Which two stories in Ovid’s Metamorphoses best exemplify human virtue? Please only name them, and do not give an explanation’. The class was then asked whether they agreed with the tool’s choice, what details led them to that decision, and what other examples they might supply. Students were more likely to agree with the chatbot when it included the stories of Baucis and Philemon’s generosity, Ceyx and Alcyone’s passion and fidelity, and Deucalion and Pyrrha’s humility; they cited reasons such as purity of heart and demonstration of xenia as reasons for approving of the inclusion of these narratives. The strongest objections from students came when Claude listed Pygmalion as an exemplar of virtue. When Pygmalion exhibits such devotion towards and desire for his statue, Aphrodite takes this as sufficient reason to turn the statue into a real woman, causing Claude to label the sculptor as virtuous. Students, however, reached the opposite conclusion and found his behaviour to be obsessive, misogynistic, and controlling. Some students also took their discussion of virtue further, questioning the motives of Baucis and Philemon. The aged, married couple demonstrate xenia towards two men they take to be human, but who are actually gods in disguise. Students questioned whether these acts came from sincerity of heart, or out of fear of the consequences of not showing kindness to strangers.

Other options for using an LLM as debate partner might include querying the chatbot about the virtue of particular heroes. Is Theseus more virtuous than Jason? Are there others that might merit inclusion on Mount Olympus as did Hercules? Does Aeneas merit the praise of the ancient Romans as given in Virgil’s work? Students must first make their own assessments of the characters’ nature, then be prepared to defend that position with the AI tool. The mysterious nature of Medea, whose acts of murderous revenge frequently prompt lively discussion, could serve as an impetus to query a chatbot with, ‘Was Medea in any way virtuous?’ Students’ responses to the question of Medea’s character sometimes display a gender division (i.e., young women are more likely to praise her than are young men). Asking a chatbot to critique Medea from various male points of view could be thought-provoking. Another possibility for involving an LLM in a classical culture course is querying the chatbot as to the advantages of having a deity favour you. The chatbot will be able to draw upon instances of divine–human interaction and may label some of these as having favourable outcomes. Students will have their own views on whether they would want to find themselves in the position of a human character in myth and risk the potential, perhaps even likely, negative outcome.

Use case 3: an LLM analyses ancient jokes

Tools like Claude work in part because of their ability to recognise patterns and predict what comes next. For example, ‘Socrates, Ovid, and Claude walk into a bar’ may lead to the expectation of a joke. The abilities of LLMs in the realms of humour are less than perfect, but nonetheless instructive, even when paired with material from classical antiquity. After reading some of the Philogelos compendium of jokes in translation, students in classical culture classes have been surprised and delighted to know that the set-up and punchline format is thousands of years old, and that humourists have endeavoured to make a living writing jokes since antiquity (Crompton Reference Crompton2011). They are also amused that ‘dopey neighbour’ jokes are nothing new, such as ‘A Kymaean goes to visit a friend and stands outside his house shouting for him. A neighbour sticks his head out of the window and says, “Shout louder – then he’ll hear you!” So the Kymaean yells out, “hey, louder!”’ (88). I accompany these readings with 20th-century philosophy on the nature of humour as presented in John Morreall’s Comic Relief, and ask students to apply the modern theories of this book to ancient jokes (Morreall Reference Morreall2009). Wanting to incorporate chatbot-related assignments in this class, I assigned students to query Claude with this simple prompt, ‘Please tell me a joke that begins with, “A guy walks into a bar…”’ Once they had the joke from the bot, they were to discuss whether the bot’s offering aligned with a theory of humour, and whether the joke resembled anything that they had seen in the Philogelos.

This exercise was educational in a number of ways: Claude did not always produce a genuine joke, but something more like a surrealist vignette:

Claude fared no better when trying to explain this ‘joke’. It did correctly identify it as fitting a ‘misheard wish’ paradigm, but did not recognise that the beginning was not related to the end. The same is true for some of the ancient quips from the Philogelos. When I prompted Claude with ‘Please explain this joke to me: did you hear the one about the drunk who opened a bar? He tied a bear up outside’ (Claude 2024), the chatbot explained that it was meant to be a humourous misstep involving the proximity of the words ‘bar’ and ‘bear’. Yet when I added that this joke was originally from ancient Rome, it did not offer a possible explanation of the joke, but it did recognise that some humour can be lost in translation. (One plausible explanation of this joke is that the drunk tied a bear to the entrance of his newly opened bar, so that no one would enter, and thus he would have all the alcohol to himself. The editor of this compendium never claimed that all the jokes are good.) When Claude did produce a joke, it was sometimes a bit stale, as with this one, ‘A guy walks into a bar. “Ow!” he says’ (Claude 2024). Overall, students were largely successful in being able to apply a theoretical framework to the text that the bot produced or that survives in the ancient compendium. Many jokes, ancient and modern, rely on a homophone, poking fun at someone who seems a bit dim, or general silliness, and humans tend to do much better at recognising these patterns than LLMs do. Students also responded well to being asked their own opinion of the joke; some enjoy bad puns and ‘dad jokes’, while others do not.

Further thoughts and future assignments

A deeper discussion about the nature of humour lies at the conjunction of antiquity, comedy, and AI: is humour a distinctly human characteristic? To what extent has it changed over millennia? To what extent can it be manufactured or imitated by a machine? Adrienne Mayor argues in Gods and Robots that for thousands of years humans have envisioned machines as sources of protection, service, entertainment, and pleasure. Among other examples, she references the story of Talos, the larger-than-life bronze humanoid fashioned by Hephaestus to protect the island of Crete. She includes a discussion of Aristotle’s thought experiment about inanimate instruments that could carry out their tasks without the presence of humans (with pointed projections towards Goethe’s story of the sorcerer’s apprentice). It may be crude that she implies that Pygmalion’s statue was the original sex-bot, but there is no denying the purposes for which the sculptor created his masterpiece (Mayor Reference Mayor2018). If machines can take on so many tasks typically done by people, why not also teach them to be funny? We use machines (as did ancient peoples, however rudimentary or complicated those tools may have been) to communicate; e.g., we text, we email, we compose reports, and sometimes the machines aid us in our efforts (as with template construction, autocorrect, or predictive text). When we want the content of our communications to be humourous, perhaps soon there will be the potential for the machine to create the humourous content itself. Could a classicist cum computer scientist invent a modern-day Aristophanes or Juvenal? That inventor would likely need to consider the desirability of guardrails for their compu-comic creation. As with a human humourist, the joke-teller having no guard rails risks alienating at least some of the audience, but having too many might override the entity’s ability to be funny.

A study done by Nataliya Kosmyna et al. at the Massachusetts Institute of Technology Media Lab reiterates and illuminates a principle that educators may know intuitively: anyone who wishes to learn something new and retain that knowledge will best advance through a process of doing, making, or practicing, and that mistakes are inevitable in such a process (Kosmyna et al. Reference Kosmyna, Hauptmann, Yuan, Situ, Liao, Beresnitzky, Braunstein and Maes2025). Using electroencephalography, Kosmyna and a team of seven other scholars conducted a study that measured brain activity when engaging with LLMs. For comparison, the larger cohort was broken into three groups, and each was given the task of writing an essay. Group 1 was allowed to use a search engine, group 2 used an LLM, and the third was a control group, or ‘brain only’. ‘Brain connectivity systematically scaled down with the amount of external support’ (Kosmyna et al. Reference Kosmyna, Hauptmann, Yuan, Situ, Liao, Beresnitzky, Braunstein and Maes2025).Footnote ⁵ Those familiar with Socratic dialogues may notice a modicum of kinship between this study and the Socratic method. Socrates went many rounds with his interlocutors, to the fascination of some and the vexation of others. In the same way that the subjects of Kosmyna, et alia achieved greater neural growth when they were required to do ‘brain only’ activities, Socrates’ students achieved wisdom (i.e., knowing how much they did not know) only after considerable time and effort. In both cases, excessive use of scaffolding (whether digital or human) can impede the learning process. Similarly, classical myth features many stories where the hero achieves his goal only after trials and tribulations, as with the labours of Hercules. No doubt the writers of the jokes in the Philogelos created some bon mots that were less than ‘bon’ when spoken aloud to an audience, and learned from this experience. No matter how much a writer might want to automate the process of learning and composing, to be able to lay claim to a style, that writer has to explore that for themselves. A modicum of self-knowledge is necessary if that self is to be present on the page.

As I continue to hone my own use of AI in the classroom, two exercises will appear in future assignments: in a future iteration of the Culture of the Ancient Mediterranean class, I will teach Roman satire, with selections from Juvenal and Horace. After asking students to compare and contrast the two, I will have them ask the same of the tone detector function on goblin.tools/Judge.Footnote ⁶ The AI models offered by goblin.tools were created for neurodivergent users in particular, but can be of help to others as well. The tone-detector function of the LLM may assist students in developing their own style of writing (didactic, gently witty, bitingly sarcastic, etc.). In a class that includes material from Plato, students will pose questions to Claude and ask it to answer in the style of Socrates. They will then be asked to compare and contrast the chatbot’s production with one of our translated works. Does the bot achieve the proper level of pointedness? If so, what vocabulary and phrasing leads to this conclusion? Students will undoubtedly have views on the palatability of responses from the ancient philosopher and an LLM. I look forward to learning from students if they would be any more likely to take advice from Socrates or an AI tool.

Writing in any degree program – to what end?

As university instructors grapple with the presence of AI tools in the classroom, whether they teach classical texts or other material, they may find themselves engaging students in conversations about the purpose of their major in particular, and of a college degree in general. Students’ answers to such questions may include getting certification in a field, preparing for further credentialling (e.g., law school), or for the simple enjoyment of learning new things. To continue the conversation, the person leading the class might ask their students to describe a previous learning experience, such as learning to play an instrument, ride a bike, or play a sport. Students will likely acknowledge that learning takes effort, and that they were not immediately successful. Understanding literature and culture, and expressing one’s thoughts in a written format, are no different. There is more to learning about the ancient world than consuming a list of factoids, or being able to generate such a list. A further component to this conversation asks them to reflect on what they feel they can claim as their own. Would they pass off a store-bought cake as their own when they told their friend they would make it from scratch? There are benefits, and at least a modicum of pride, in being able to say, ‘I made this by myself’. When I make a cake by myself, I can customise my work, thus making it genuinely my own; I can learn to be proud of a creation even if it is not perfect, and I can learn what to do differently on my next endeavour. My goal for students is a discovery that while they are learning about any particular subject, they are also learning how they learn, that they learn through practice, and that this process can be as satisfying as reaching the end goal of a diploma. This process is at least partly circumvented if they have a machine do their writing for them. Whatever they pursue after college, they will be better able to meet upcoming challenges if they have worked at and through written communication.