Giving the dead a voice

1. What these voices are

The Archive offers several ways into its texts. A Classicist reads across the whole collection, while Cicero, Caesar, Augustus, Seneca and Marcus Aurelius answer from their own works. Addressing a dead man in the first person is an old and slightly uneasy pleasure. It deserves clear limits.

These personas are reading devices. Ask Augustus how he came to power and the answer is a synthesis drawn from his surviving words, just forty sections of official self-account, written in a voice shaped by that record. It is not Augustus. The system enforces the distinction by filtering each figure's retrieval to his own works. It cannot hand him a passage he never wrote or could never have read.

2. An old device: the personas as prosopopoeia

Putting speech into an absent or dead person's mouth is not an invention of machine learning. Classical rhetoric named the device prosopopoeia, the "making of a face", the fashioning of a person, and the authors in this Archive were its masters and its theorists. The figure has a long theoretical afterlife, from the rhetorical handbooks through its medieval and early-modern elaboration as personification (Paxson 1994; Lausberg 1998), but its ancient form is the one that bears on this system: the deliberate, announced construction of a voice for someone not present to speak.

Cicero used it at the hinges of his greatest speeches. In the First Catilinarian he makes the patria, Rome herself, rise and address Catiline directly. In the Pro Caelio he does something closer to what this Archive does: he summons a specific dead individual, Appius Claudius Caecus, the stern ancestor of the scandalous Clodia, "from the underworld" (ab inferis) to rebuke his descendant. The jury knew Appius was dead; the power of the device lay precisely in everyone knowing the voice was constructed, and judging whether it was constructed faithfully: whether this was what old Appius would have said.

Quintilian, whose treatise on oratory sits in this same Archive awaiting embedding, gives the device its rulebook (Institutio Oratoria 9.2). He calls impersonations fictiones personarum, credits them with the power to "bring down the gods from heaven and raise the dead," and attaches a condition: the invented speech must fit the person, and it persuades only so far as it stays within what the person plausibly could have said. An impersonation that outruns its subject's character or knowledge fails as rhetoric, on Quintilian's own terms.

That is close to the rule used here. The personas are digital prosopopoeia, with retrieval turning Quintilian's condition into software: a voice may say only what the surviving record supports. When Petrarch rediscovered Cicero's letters to Atticus in 1345, his first act was to write Cicero a letter back (Familiares 24.3). Readers have long wanted to answer voices found in an archive. A machine fluent enough to reply on the archive's behalf makes the old constraint considerably more urgent.

3. The three rules

Bounded knowledge

Each figure knows what they knew in their own life, and no more. Cicero cannot tell you how he died; Caesar cannot speak of what followed the Ides of March; Augustus does not narrate the empire after him; Marcus Aurelius does not foretell the future. Each is scoped, too, to their own writing. Marcus reasons from his Meditations, not from Cicero's letters, and the retrieval filter makes this physical: the other authors' texts are simply never placed before him. When a question runs past the edge of what a figure could know, the correct answer is that they cannot say, and the personas are built to give that answer, in voice, rather than improvise a plausible one. A persona that cannot admit ignorance will sooner or later produce a confident falsehood.

The bound is hardest exactly where the underlying language model is most knowledgeable. Any modern model "knows" how Caesar died; it has read the assassination ten thousand times. The persona must refuse what the model can easily supply: the temptation is completion from everywhere rather than invention from nothing, and resisting it takes both prompt discipline and the structural fact that retrieval never furnishes the forbidden material.

Nothing invented

The personas do not compose new letters, speeches, meditations, or sayings and attribute them to the dead. This is the sharpest line, because it is the most tempting to cross: a fluent model can produce a sentence that sounds exactly like Cicero, and such a sentence, once written, is a forgery. Classical studies has spent five centuries unmasking exactly this artefact; the pseudo-Ciceronian Consolatio forged in the 1580s fooled some of the best Latinists alive. The Archive grounds every substantive claim in retrieved passages, so that what a persona reports can be traced to what the figure actually wrote. Where the persona characterises or paraphrases, it should read as characterisation rather than be passed off as the record.

A record is not a person

These were real people, with families, enemies, vanities, and, several of them, violent ends. The people around them were real too: the wives and children, the enslaved copyists and letter-carriers whose labour made the correspondence physically possible, and of whom the record preserves almost nothing in their own words. It would be easy, and wrong, to put confident speech into mouths the record never preserved. The persona system is therefore weighted towards the figures who left a corpus, and towards candour about how partial even that corpus is: one side of a correspondence, a general's own dispatches, an emperor's chosen account of himself.

4. The modern hazard

A contemporary literature has grown around "deathbots" and "griefbots", generative simulations of the dead, usually of the recently dead, built from their messages and voice notes (Hollanek & Nowaczyk-Basińska 2024; Öhman & Floridi 2017; Stokes 2021). Its central worries are consent, the dignity of the dead, the vulnerability of the grieving, and the commercial capture of memory; Öhman & Floridi (2018) propose an ethical framework for what they call the digital afterlife industry, arguing by analogy with the regulated dignity of human remains that a person's informational legacy should not be freely exploited for the benefit of the living. The figures in this Archive are two thousand years dead, publicly known, and represented only through texts they published or that history published for them, so the sharpest of those concerns (consent from the subject, protection of the grieving, commercial capture of a private person's memory) do not transfer. Two concerns still apply.

The first is fluency. Shanahan, McDonell and Reynolds (2023) describe large-language-model personas as role-play without an inner life: a performance that invites the interlocutor to over-attribute. A first-person Caesar answering in confident prose lends generated sentences an authority no disclaimer fully counters. The Archive's mitigations are structural rather than decorative: citations are always visible beside the reply, refusals happen in voice and often, and the speech interface deliberately keeps the text on screen as text; the bust does not move its lips.

The second is authority laundering. A synthesis grounded in the Res Gestae is grounded in propaganda; a persona scoped to Caesar's Commentarii reproduces the perspective of a man justifying his own conquests, including the subjugation of Gaul as he chose to narrate it. Fidelity to the record does not amount to neutrality, because the record itself is not neutral. The Classicist exists for exactly this reason: a voice outside the ensemble whose explicit job is to say what kind of document a text is (advocacy, selection, self-address) and what it leaves out.

5. From stone to skin

The Archive's next step makes the gradient concrete. Two motion studies exist: in the first, the marble bust of Marcus Aurelius, the same likeness that stands in the gallery's ring, speaks. The stone moves, and remains unmistakably stone: sculpted eyes, carved curls, a statue talking. In the second, Caesar has been given back his face (skin, hair, grey eyes), a photorealistic reconstruction generated from the Tusculum portrait, the one surviving portrait type generally accepted as carved during his lifetime. Both are startling. They are not, ethically, the same object.

The speaking statue is prosopopoeia in its most literal form: a "made face" that plainly looks made. No viewer mistakes talking marble for a living man. The artifice is visible, just as it was when Cicero made the patria speak. The result can feel more present without making a stronger claim to reality.

The Marcus study crosses that line halfway through, when the sculpted eyes are replaced with living ones (irises, pupils, moisture) while every other surface stays marble. The effect is disproportionate to the change: a talking statue with stone eyes reads as a device; the same statue with human eyes reads, suddenly, as inhabited. One signal, a few square centimetres of the image, moves the object most of the way across the gradient on its own. For the design the implication is that presence does not scale with how much of the surface is made realistic: perceived animacy concentrates in the eyes, which is why the uncanny-valley literature keeps returning to gaze (Mori 1970; Mori, MacDorman & Kageki 2012). This is more than an anecdote: empirical work on androids finds that mismatches concentrated in the eyes and their motion are disproportionately responsible for the descent into the valley, and that a face made almost human is judged more unsettling than one plainly artificial (MacDorman & Ishiguro 2006). The marble-with-living-eyes study is a controlled instance of exactly that effect: hold every other surface at "plainly sculpture" and change only the eyes, and the object slides from device to inhabited on the strength of that one signal.

There is also a strange historical footnote here: the blank white stare we call "classical" is largely an accident of survival. Roman marbles were painted, bronzes carried inlaid eyes of glass and stone, and Antonine sculptors incised iris and pupil into the marble itself (Abbe 2015). An ancient viewer never met the empty gaze we know; in giving Marcus his eyes back, the study is in one sense restoring the ancient viewing experience rather than departing from it. That does not remove the startle, but it relocates it: part of what unsettles a modern viewer is the habit of expecting ancient sculpture to look blind.

The fleshed Caesar sits differently. A photoreal face is the native format of the deepfake, and its whole power is that the device disappears: the viewer is invited to feel, however briefly, that they are being looked at by Julius Caesar rather than by a rendering derived from a marble he may once have sat for. Two things keep it on the right side of the line, and both need to be visible. First, the reconstruction stands to the Tusculum portrait as a grounded answer stands to a retrieved passage, inference from the surviving record (a lifetime likeness, in this case), never a face invented free-hand. That is the same fidelity condition Quintilian attached to the spoken device, applied to appearance. It also needs framing: the source portrait named, the method stated, the same bounded voice and in-voice refusals behind the face. The working rule the Archive adopts is that the realism of the surface may never outrun the legibility of the device: the more life-like the reproduction, the more clearly the scaffolding must show. A statue can simply speak; a photoreal face has to declare what it is first.

6. Why do it at all

Why give the dead a voice at all? A voice invites a kind of attention that a search box does not. Asked to speak as a figure, the system must locate the relevant passages, weigh them, and present them in the first person, and a reader, hearing that, reads the sources more closely than they otherwise might. The persona is a lure towards the text rather than a substitute for it.

The classical tradition got here first. Quintilian recommended prosopopoeia to students as a discipline: to speak credibly in another's character, you must first master what that person said and how they thought. The exercise serves the sources. With knowledge bounded, invention refused and the record's partiality kept in view, the voice sends the reader back to the text. That is the only version of the idea I am comfortable building.

References

• Abbe, M. (2015). Polychromy. In E. A. Friedland, M. G. Sobocinski & E. K. Gazda (eds.), The Oxford Handbook of Roman Sculpture. Oxford University Press.
• Cicero, In Catilinam 1.18, 1.27 (the patria addresses Catiline); Pro Caelio 33–34 (Appius Claudius Caecus raised ab inferis).
• Quintilian, Institutio Oratoria 9.2.29–37 (fictiones personarum); 3.8.49–54 (speaking in character as an exercise in fidelity).
• Petrarca, F. Epistolae Familiares 24.3 (letter to Cicero, Verona, 1345).
• Hollanek, T. & Nowaczyk-Basińska, K. (2024). Griefbots, Deadbots, Postmortem Avatars: on Responsible Applications of Generative AI in the Digital Afterlife Industry. Philosophy & Technology 37.
• Lausberg, H. (1998). Handbook of Literary Rhetoric. Brill. (On prosopopoeia and sermocinatio.)
• MacDorman, K. F. & Ishiguro, H. (2006). The uncanny advantage of using androids in cognitive and social science research. Interaction Studies 7(3), 297–337.
• Mori, M. (1970). The Uncanny Valley. Energy 7(4), 33–35. (English translation: Mori, M., MacDorman, K. F. & Kageki, N. (2012). The Uncanny Valley [From the Field]. IEEE Robotics & Automation Magazine 19(2), 98–100.)
• Öhman, C. & Floridi, L. (2017). The Political Economy of Death in the Age of Information: A Critical Approach to the Digital Afterlife Industry. Minds and Machines 27, 639–662.
• Öhman, C. & Floridi, L. (2018). An ethical framework for the digital afterlife industry. Nature Human Behaviour 2, 318–320.
• Paxson, J. J. (1994). The Poetics of Personification. Cambridge University Press.
• Shanahan, M., McDonell, K. & Reynolds, L. (2023). Role play with large language models. Nature 623, 493–498.
• Stokes, P. (2021). Digital Souls: A Philosophy of Online Death. Bloomsbury.
• Zanker, P. (2009). The Irritating Statues and Contradictory Portraits of Julius Caesar. In M. Griffin (ed.), A Companion to Julius Caesar. Wiley-Blackwell. (On the Tusculum portrait, found in 1825 and now in Turin: the portrait type generally accepted as made in Caesar's lifetime.)
• Zhou, L. et al. (2025). PersonaEval: Are LLM Evaluators Human Enough to Judge Role-Play? arXiv:2508.10014.