Ancient Plagiarism, AI Workshop

Where the anxiety comes from

The unease is understandable. A large language model reads across an enormous amount of human writing and produces text that resembles it. For anyone who cares about copyright or academic integrity, that can feel like a sudden crisis: a machine taking in human work and remaking it at will.

This section sets that feeling against a longer record. It does not argue the worry away. It asks a sharper question. Is a language model doing something genuinely new, or something very human, at a scale no person could reach? The answer changes what we should be watching for in a student's submission.

Interactive walkthrough

Ancient plagiarism, in three short modules

Walk the argument through the oldest examples, sort plagiarism from intertextuality, and classify three cases yourself. Runs in the browser; no install.

Open the deck

Two words worth separating

Two ideas tend to blur together, and separating them is the whole point. Plagiarism is passing off another person's words or ideas as your own, without credit. The fault is not the reuse; it is the deception. Plagiarism needs an intent, to make borrowed work look like something you produced yourself.

Intertextuality is different. It is the reworking of earlier texts to make new meaning. The literary theorist Julia Kristeva coined the word in 1966 for the way every text is woven out of others. Here the reuse is open, and the aim is to transform a source, not to hide it. One term is about honesty; the other is about method. A single act can raise both questions at once, which is exactly where the AI case gets interesting.

It is also a myth that the ancient world copied freely because it had no rules. Ancient writers named and shamed literary theft. Vitruvius, writing around 30 to 15 BCE, records the scholar Aristophanes of Byzantium exposing poets who recited others' work as their own. The English word plagiarism even comes from the Latin plagiarius, a kidnapper, a metaphor the poet Martial used for a rival who passed off his poems. So the concept existed. What differed was the default: reuse was normal, and authorship was rarely treated as private property.

The oldest texts were built on older texts

The texts at the root of the Western tradition are not original in the modern sense. Each was assembled by reworking something earlier, and the reuse was the craft, not a lapse in it.

The flood is the clearest case. The story of Noah's ark in Genesis shares its plot with the much older flood in the Epic of Gilgamesh. Both run the same way: a warning, a great boat, the animals, the deluge, the birds sent out to find land. The authors of Genesis did not copy blindly. They rewrote a polytheistic story into one about a single God and a covenant, so the borrowed structure carries a new meaning. The Babylonian version of Gilgamesh was itself compiled by the scribe Sin-leqi-unninni from older Sumerian poems.

The same pattern of open reuse runs through the record. The Mesopotamian creation epic, the Enuma Elish, survives only in copies of a still older work. A version found at the Assyrian city of Ashur keeps the plot but puts the Assyrian god Assur in the hero role Babylon gave to Marduk. The epic was reused to justify a new power. And most scholars hold that the Gospels of Matthew and Luke drew on the earlier Gospel of Mark, often almost word for word. Because those texts were anonymous and rested on shared oral tradition, this is treated as communal tradition reused, not as theft.

What the history changes

Set a language model against that record and it starts to look familiar. A model holds no store of passages to lift. It learns statistical relationships between words across its training text, then makes new text by predicting what tends to come next. In a mechanical sense that is intertextuality without an author: recombination of inherited material, faster than any scribe.

Two things follow, and they pull in different directions. The first is that learning from existing work and recombining it is not a modern crime. It is how culture has always moved, and rules are unlikely to stop machines learning from what humans have already written. The second is that plagiarism has always required intent. A model does not intend anything; it predicts. Judging a machine by a rule written for a lying human misfits the case. The fault, where there is one, belongs to the person who hides how the work was made.

That is why the history is worth an hour of a workshop. For a few centuries we judged the originality of the text, and that test breaks when a machine can produce fluent, source-less prose on demand. What survives is the question antiquity actually cared about: not is this text new, but was its maker honest about how it was made? The practical shift, for assessment and for integrity, is from policing originality toward expecting transparency. That is the same ground the assessment redesign section covers, reached from a different direction.

For discussion

The essay this section draws on ends on an open question, and it is a good one to put to a room, and to students themselves. A few prompts to run with, in a small group or a full session.

How should a definition of misconduct account for machine-assisted intertextuality?

If reuse and recombination are how writing has always worked, and a model does the same without intent, where exactly does the line of misconduct now sit? On the text, on the disclosure, or on something else?

If we judge intent and transparency rather than originality, what would a student have to tell us?

Picture the disclosure that would actually satisfy you for a given task. What does it ask for, and is it something an integrity process could test?

Does the ancient parallel comfort you, or does something about the machine still feel different?

Naming what feels different, the scale, the absence of a human source, the training data question, is more useful than deciding the analogy settles the matter. What survives the comparison?

There is no settled answer here, which is the honest position. The value is in getting the distinction between deception and method into the open, and letting a group work out where the machine belongs.

Sources

This section is a reflective interlude rather than an evidence review, so it leans on standard reference scholarship. The historical claims were checked against general reference sources in August 2026. Where a specific claim in the source essay could not be confirmed, it was left out, and it is noted below.

The term intertextuality was coined by Julia Kristeva, "Word, Dialogue and Novel", 1966.
Ancient attitudes to literary theft: Vitruvius, De architectura, Book VII, preface, c. 30 to 15 BCE (the Aristophanes of Byzantium anecdote); "Plagiarism in Latin literature", Oxford Classical Dictionary; the word plagiarius and the poet Martial, Epigrams 1.52, c. 86 to 103 CE, with etymology via etymonline.
The Genesis and Gilgamesh flood parallel is standard scholarship on the ancient Near East. The Babylonian standard version of Gilgamesh was compiled by the scribe Sin-leqi-unninni from older Sumerian poems, c. 1300 to 1000 BCE.
Enuma Elish: seven tablets; surviving copies date to about 1200 BCE and are themselves copies of an older work. The copy found at the city of Ashur puts the Assyrian god Assur in the hero role that Marduk holds in the Babylonian form. World History Encyclopedia and Encyclopaedia Britannica, "Enuma Elish".
Synoptic Gospels: Marcan priority, the majority scholarly view that Matthew and Luke used the earlier Gospel of Mark. Marcan priority (reference summary); Bart D. Ehrman, standard New Testament scholarship.
Left out for want of confirmation: the source essay's claim that the Enuma Elish borrowed specifically from the earlier Anzu myth. The sources consulted confirm the epic reworks an older original, and that the Assyrian copy substitutes Assur, but did not confirm the Anzu detail.

The framing, and the closing discussion question, are adapted from an essay in the workshop materials on LLM anxiety and intertextuality.