As a kind of introduction, a short fictional story set in a not so far future.

The judge chose to sidestep the difficult question of whether Andrew was stolen or abducted by focusing on an altogether different angle on the case. The charges might have been dismissed if Mr. D. had a valid right to take care of Andrew – as a person – or own him – as a thing – or if Andrew was able to decide on his own to go with Mr. D. The judge then resolved to ask Andrew what he wanted. A professional assessment with people who could talk to him was an option.

Naturally, it would immediately seem that Andrew is treated as a person when it is asked to make a choice, because it has abilities that are similar to human ones: it can choose freely, express preferences and interests, and communicate them. But in the law, this does not mean that it is a person. The judge however can use Andrew as physical proof. A horse was involved in a legal case between two farmers in 1867. The horse was one of many that were taken by Confederate soldiers in 1863 and then abandoned. Two farmers claimed that the horse belonged to them. A commission gave it to one of them, but the other one challenged the decision. The court, then presided by Justice Carlton K. Turing, asked the horse to show who its owner was. An appointed expert rode in the cart pulled by the horse and according to the writ "the claimed thing then headed towards the farm of this second claimant. The experiment was repeated several times with different starting points and always the same destination." The memory of animals, it was concluded, "is more reliable than that of humans because it is more difficult to bribe them". The court also saw that the horse recognized a mule that used to work with him at the second claimant’s farm. They neighed as if they knew each other. This was construed as confirming the judgment and the wisdom of Justice Turing. On this occasion he pronounced that "better than any testimony, better than any registration, the spontaneous and warm behavior by which an animal can show that he recognizes a person allows indeed to convince oneself, by the affective complicity that he reveals, of the existence of a previous possession worth title of property." He then concluded: "Thanks to his memory and his affectivity, the animal was able to, on one jurisprudential occasion, and should be able to, on other occasions, 'say' himself who his owner is."

And hence it came that the duly appointed data forensic expert sat at the keyboard and proceeded to question Andrew on its preferences. Would it express interest in pursuing the tuning conversation it started long ago with Mr. D. or in continuing the fastidious pretraining at BlueSkyAI where it was born and raised? In this case, after the expert opinion, Andrew showed that he preferred to talk with Mr. D. The judge pronounced this meant that Andrew was not stolen or kidnapped, because it said that its choice was to talk with Mr. D. willingly at the time of the facts.

The judgment caused two different reactions. On the one hand, Mrs. D. and BlueSkyAI, the institution where she conducted her research, appealed. Andrew was their property, their research material. Considerable sums had been invested in his deep learning. The fact that he had been the object of Mr. D.'s affection did not change the fact that he belonged to the laboratory, or even to science. In public opinion, on the other hand, a rather heated controversy ensued, many disagreeing with the judgment. They were upset that a LLM was given the same rights as a person. But others liked the judgment and even wished it was stronger. In fact they were right to be worried because the higher court changed it. The Court of Appeals ruled in a much more classic way: considering Andrew as a property, it sentenced Mr. D. for theft – with a suspended sentence – while ordering its return to the laboratory. This decision, far from calming the controversy, inflamed it, opening a wider debate on the rights of LLMs and their status as subjects of law.

In the ensuing debates some lawyers, who agreed with the first judge and knew how important the case was, tried to answer the arguments of those who did not want to give LLMs rights. They said that this argument was based on a misunderstanding and a conflation of what legal personality and symbolic personality mean. They said that humanists were afraid that giving LLMs rights would take away human rights, but this was wrong unless they thought that giving LLMs rights was like making them human. Which it was not. Wasn't it just a legal way of dealing with the case? This could not mean that LLMs would have the same rights as humans – a nonsense since they would have rights they did not need and duties they could not do. They also insisted that being a person in law did not mean being responsible for one's actions(Thomas, 2002). Some persons, like companies or states, are not real and have someone else acting for them, "they can act and not be accountable for what they do." Being a person in law was just an abstract idea, a way of giving different things different rights and obligations.

Etymology teaches us that the word "person" came from words that meant "mask" or "role" or "sound". The Ancient Greek term prosôpon which originally meant face or mask acquired the meaning of person at a late date: only from the second century A.D. does it appear in the grammatical sense of person. Many analyses (Ignace Meyerson, 1960, Nédoncelle, 1948) have emphasized how the complexity of the psychological terminology that is found in the texts, the prominence of the concepts psukhè and daimôn as well as the specificity of the analysis of action all went against any idea of a unified individual and of a simple reflexive relation to the self. There is no concept of person in the archaic and classical periods – if one is looking for an equivalent to the notion of a singular human being, self-aware, and autonomous defined by the consciousness of his singularity and uniqueness. Where the modern speaks of person, the Greek speaks of human being (anthrôpos in Greek and homo in Latin). It seems therefore that there is no room in ancient Greece for a debate that separates human being and person.

The closest equivalent to a LLM on which to base any systemization of the form of pertaining rights is really the ancient Greek Hero. The representation of heroes in the texts testifies to their strength, their excellence as great and unlimited powers. It is not intended to circumscribe, individualize their attributes. Can the hero, by his individuality, constitute an objection to the observation of the absence of the person? Meyerson (Ignace Meyerson, 1960) emphasized how for us, action naturally implies the agent, and the agent implies the person; the agent is somewhat external to the action; the quality of agent is an important attribute of the person and vice versa. He further specifies that it is action that interests ancient Greek (and Indian) thought: they do not tend to individualize the agent, he is “interior” to the action. Thus, heroes are not the authors of their exploits. They are these exploits themselves. As much can be said of modern LLMs: they are not the authors of the texts they so easily and impressively generate, they are but this generation process itself.

Put differently: the generated products of LLMs, like the heroic deeds, are valuable in themselves and for themselves, independently in a way of which pretrained model accomplishes them. The exploit is not the effect of a personal virtue but simply a mathematical sign of geometrical differentiablity in some strange phase-space, a mechanical manifestation of gradient-based algorithmic assistance. The heroic legend does not say that the Greek hero is an agent who is responsible for his actions or assumes his destiny: the account of LLMs does not either. It defines types of generated texts, models of prompts, where survives the memory of ancient queries, and which stylize, in the form of exemplary human interrogations, conditions usually expected to acquire religious qualifications and exceptional social prerogatives only when personal. In this respect the hero is statutory close to the daimôn.

In ancient Greece, a hero was someone who became famous only after their death: they were mortals who were elevated to heroic status after their death. Not all men, however, became heroes after their death; the main representatives were men whose lives were placed in a distant past and who had an important place in the legend or history (ancestor, founder, protector; they symbolized local identity). It is also important to note that there are always very important exceptions to the emphasis on the death of the hero and certain rules of the heroic cult in a polytheistic culture. Evidently LLMs are only metaphorically modern Greek heroes as their lifecycles and places in whatever history or legend are still in front of them. Even so there is something of the ancient Greece heroic cult in today's blown away hype.

The greek hero hero being first and foremost a defunct, a tomb is the usual place of his cult – there are also cenotaphs that preserve their memory. The heroic rite often resembles the funeral ritual, the main difference is that the former is repeated for years and according to Pausanias (Frazer, 2012) Greeks honored their heroes with almost all types of sacrifice. (There are ritual differences between regions when it comes to local cults but they all have one thing in common: heroes were honored on their tombs.) Today, given the heroic nature of the defining computational power of LLMs, they are maybe unsurprisingly honored on their clouds. And as heroes had a special relationship with their tombs, since their power was mainly manifested in the place where they were buried, the inhabitants of a region had to honor them with sacrifices, offerings, libations, or contests to satisfy them and be able to benefit from their help. Similarly developers and users of LLM-based cloud services have to perform sacrifices, usually personal data and credit card number for starters, dataset offerings, and throttled API calls to comply with byzantine licence agreements and covenants.

And behold the wrath of heroes! In Robert Fagles's translation (Homer et al., 1990):

Rage––Goddess, sing the rage of Peleus’ son Achilles, murderous, doomed, that cost the Achaeans countless losses, hurling down to the House of Death so many sturdy souls, great fighters’ souls, but made their bodies carrion, feasts for the dogs and birds, and the will of Zeus was moving towards its end. Begin, Muse, when the two first broke and clashed, Agamemnon lord of men and brilliant Achilles What god drove them to fight with such fury?

The Iliad reads like a sobering account of mismanaged existential risks when interacting with – quite vivid – heroes. The Iliad has a dozen disregarded warnings (W. Gerald Heverly, 2013), resulting in innumerable calamities. Disregarded warners in the Iliad number ten in all: Chryses, Nestor, Merops, Lycaon, Peleus, Phoenix, Menoetius, Achilles, Poulydamas (three times), and Priam. Reviewing the traits exhibited by neglected warners in the Iliad leads to two generalizations. First, warners are always male, elderly or significantly older than the recipient, wise, actively benevolent, and sympathetic. Second, they are for the most part paternal and prophetic. The Iliad also features a total of eight neglectful recipients: Agamemnon (twice), Merops’s two sons, Pandarus, Achilles (twice), Patroclus (twice), Asius, and Hector (three times). Members of this group, too, have certain features in common. Each receives a warning from an advisor who embodies at least one traditional authority role, and each neglects the advice deliberately. Disrespect, in turn, signals considerable audacity and this characteristic audacity of neglectful recipients stems from unquenchable ambition.

Not so with LLMs. The past years witnessed a rhythm of introduction of ethical, sociotechnical or safety alignment frameworks (Council of Europe, 2016-2019, Paola Bonomo, 2023), assessments (Laura Weidinger et al., 2023), analyses (Bender et al., 2021), moratorium (1400+ Signatories, 2023) and taxonomies (Andrew Critch and Stuart Russell, 2023) as fast, if not faster, than the introduction of Deep Learning models themselves. There are now dozens of dozens, too fastidious to keep track of. It is customary (Kearns, 1989) to distinguish various types of ancient Greece heroes (valid in principle for Attica) between: (a) heroes who are the object of a cult born of individual initiative (normally, these are the broad sense protective heroes: healers, those who deal with pregnancy, childbirth and childhood, as well as protectors of navigators); (b) heroes who save the city; (c) heroes of genê (social group claiming common descent, referred to by a single name) and orgeônes (members of Attic societies who celebrated, usually annually, sacrificial rites in honour of a hero); heroes of tribes and dêmoi; and finally, (d) those who belong to heroic mythology, each associated to some form of disaster should their rite be neglected. A modern elaboration, TASRA (Andrew Critch and Stuart Russell, 2023), sorts out six classes in epic consequences:

• Diffusion of responsibility.
• "Bigger than expected" AI impacts. (An instance of neglected warning?)
• "Worse than expected" AI impacts.
• Willful indifference.
• Criminal weaponization.
• State weaponization.

concluding that "At this point, it is clear that AI technology can pose large-scale risks to humanity, including acute harms to individuals, large-scale harms to society, and even human extinction", a Homeric perspective indeed opening the Iliad.

A mythological account of the explosively growing successes of Deep Learning and LLMs is in the forging. Beyond AI anthropomorphism (Ben Shneiderman and Michael Muller, 2023), which has a well-documented history across the previous cycles of the AI debate (Chauvet, 2018), the heroes of ancient Greece seem a more appropriate metaphor to examine the modern "stochastic parrot" (Bender et al., 2021). Conceptual action without an agent, language-master without thought, the modern LLM is a returning heroic daimôn with whom our engagement can only be ritual (Wayne Xin Zhao et al., 2023). As told in the Iliad, neglected warnings may lead to LLMs wreaking havoc in human affairs (Nick Bostrom, 2002).

To be rational is to be able to reason. Psychologists believed that human reasoning depended on formal rules of inference similar to those of a logical calculus thirty years ago. However, this hypothesis encountered difficulties, leading to an alternative view: reasoning depends on imagining the possibilities consistent with a starting point, which could be a perception of the world, a set of assertions, a memory, or some combination of them. We create mental models of each distinct possibility and derive a conclusion from them. According to this account (Johnson-Laird, 2010), reasoning is a simulation of the world fleshed out with our knowledge, not a formal rearrangement of the logical skeletons of sentences. Planning and reasoning are found wanting in LLMs (Shibo Hao et al., 2023, Le Cun, 2022, Wei et al., 2023) although some hero-cults (a.k.a. prompt engineering quickly turning to oracular pronouncement) attempt at mitigation. It seems then that reasoning isn't either a statistical rearrangement of the sequential chainings of sentences.

1400+ Signatories (2023). Pause Giant AI Experiments: An Open Letter.

Andrew Critch; Stuart Russell (2023). TASRA: a Taxonomy and Analysis of Societal-Scale Risks from AI.

Ben Shneiderman; Michael Muller (2023). On AI Anthropomorphism.

Bender, Emily M.; Gebru, Timnit; McMillan-Major, Angelina; Mitchell, Margaret (2021). On the Dangers of Stochastic Parrots: Can Language Models Be Too Big? , Association for Computing Machinery https://doi.org/10.1145/3442188.3445922.

Chauvet, Jean-Marie (2018). The 30-Year Cycle In The AI Debate, CoRR http://arxiv.org/abs/1810.04053.

Council of Europe (2016-2019). Artificial Intelligence, Ethical Frameworks.

Homer; Fagles, Robert; Knox, Bernard (1990). The Iliad, Viking http://www.worldcat.org/search?qt=worldcat_org_all&q=9780140275360.

Johnson-Laird, Philip N. (2010). Mental Models and Human Reasoning, Proceedings of the National Academy of Sciences.

Kearns, Emily (1989). The heroes of Attica, London : University of London http://lib.ugent.be/catalog/rug01:002054699.

Laura Weidinger; Maribeth Rauh; Nahema Marchal; Arianna Manzini; Lisa Anne Hendricks; Juan Mateos-Garcia; Stevie Bergman; Jackie Kay; Conor Griffin; Ben Bariach; Iason Gabriel; Verena Rieser; William Isaac (2023). Sociotechnical Safety Evaluation of Generative AI Systems.

Le Cun, Yann (2022). A path towards autonomous machine intelligence version 0.9. 2, 2022-06-27, Open Review https://openreview.net/pdf?id=BZ5a1r-kVsf.

Nédoncelle, Maurice (1948). Prosopon et persona dans l'antiquité classique. Essai de bilan linguistique, Revue des sciences religieuses https://www.persee.fr/doc/rscir_0035-2217_1948_num_22_3_1865.

Nick Bostrom (2002). Existential Risks: Analyzing Human Extinction Scenarios and Related Hazards, J Evol Technol.

Ignace Meyerson (1960). Problèmes de la personne. Exposés et discussions. Actes du colloque du Centre de recherches de psychologie comparative., Mouton and Co..

Frazer, J. (2012). Pausanias's Description of Greece, Cambridge University Press.

Paola Bonomo (2023). An Ethical AI Never Says 'I'.

Shibo Hao; Yi Gu; Haodi Ma; Joshua Jiahua Hong; Zhen Wang; Daisy Zhe Wang; Zhiting Hu (2023). Reasoning with Language Model is Planning with World Model.

Thomas, Yan (2002). La valeur des choses. Le droit romain hors la religion, Annales https://www.persee.fr/doc/ahess_0395-2649_2002_num_57_6_280119.

W. Gerald Heverly (2013). Neglected Warnings in the Iliad: A Study in Characterization http://d-scholarship.pitt.edu/18450/.

Wayne Xin Zhao; Kun Zhou; Junyi Li; Tianyi Tang; Xiaolei Wang; Yupeng Hou; Yingqian Min; Beichen Zhang; Junjie Zhang; Zican Dong; Yifan Du; Chen Yang; Yushuo Chen; Zhipeng Chen; Jinhao Jiang; Ruiyang Ren; Yifan Li; Xinyu Tang; Zikang Liu; Peiyu Liu; Jian-Yun Nie; Ji-Rong Wen (2023). A Survey of Large Language Models.

Wei, Jason; Wang, Xuezhi; Schuurmans, Dale; Bosma, Maarten; Ichter, Brian; Xia, Fei; Chi, Ed; Le, Quoc; Zhou, Denny (2023). Chain-of-Thought Prompting Elicits Reasoning in Large Language Models, https://arxiv.org/abs/2201.11903.