The following is an extended excerpt from my book-in-progress, “An Enchantment of Digital Archaeology: Raising the Dead with Agent Based Models, Archaeogaming, and Artificial Intelligence”, which is under contract with Berghahn Books, New York, and is to see the light of day in the summer of 2020. I welcome your thoughts. The final form of this section will no doubt change by the time I get through the entire process. I use the term ‘golems’ earlier in the book to describe the agents of agent based modeling, which I then translate into archaeogames, which then I muse might be powered by neural network models of language like GPT-2.

The code that I used to generate pseudo Gibbons and pseudo Sophocles modelled the probabilities of different letters following one another. While sophisticated at the time, that approach is now little more than a toy. With the increase in computational power and complexity, these newer models open up tricky ethical issues for us, and in particular, if we use them to try to give our digital creations their own voice to speak. Let me sketch out how these new models work, resurrect Flinders Petrie, and then we’ll examine the aftermath.

More complex models of language now try to work out the ‘rules’ of language by ‘masking’ whole words and working out the probabilities to deduce the hidden word from the words that sit to either side. Others try to represent a word as an ‘embedding’ in multi-dimensional space (using a model built by Google). But the most complicated model, and the most successful, as of this writing, is probably the GPT-2 model, developed by the OpenAI foundation. It was trained on webpages found from outward links on Reddit that Reddit users scored as highly useful, some 8 million websites. The sheer size of this data let OpenAI develop a model that contains 1.5 billion parameters. Because of the potential malicious uses of the model, OpenAI did not release this model, but instead made available a series of smaller models- a ‘small’ model with 117 million parameters, and a ‘medium’ model with 355 million (there is a ‘large’ model available, but it is too large to run on a typical laptop).

Jesse Vig from the ‘Towards Data Science’ group blog has a post (2019) exploring how this model generates text. He begins by experimenting, trying to break the model (an ‘alchemical’ approach, as Rahimi 2017 laments), feeding it two distinct sentence prompts:

‘The dog on the ship ran’
‘The motor on the ship ran’

The model completes the first sentence with, “The dog on the ship ran off, and the dog was found by the crew.” It completes the second sentence with, “The motor on the ship ran at a speed of about 100 miles per hour.” The model understands the difference between dogs and motors, and the likely probabilities of different words completing the sentence. The model has what its creators call an attention mechanism, which focuses on the linguistic idea of the ‘head’ of the noun phrase. There are other linguistic patterns captured by the model, arranged 12 to a layer across 12 layers. It is this attention mechanism and the learning of linguistic patterns that accounts for this model’s success.

I fed the opening chapter of this book to talktotransformer.com, an interactive front-end for the GPT-2 model; the bolded text is what I wrote, the second paragraph is the machine’s:

This is a book about enchantment and digital archaeology. It is a book about trying to pull together the connective threads on nearly twenty years of work in simulation, agent modelling, video games, and Roman economic history. These are not, on the face of it, ‘enchanting’ topics. But hear me out. It’s about trying to remember what it was that was magical about archaeology when I first began, and why I find digital approaches to archaeology to still be magical.

In this book, I’d like to trace out some of the elements that underpin magical engagement with archaeology. I would like to show you some of the tools digital archaeology (D-arch), and computer archaeology have developed in order to be able to address specific aspects of magical engagement. But it isn’t hard to imagine that if you were an archaeologist working in the digital frontier of archaeology a lot of these new tools might not have occurred to you. To me, it’s about being able to be enchanted by the methods that a new medium has provided to the archaeological enterprise

I don’t know whether to be pleased or worried by this result. It could almost be the blurb on the back of this book’s cover. It gets a lot of things right.

Salesforce, a marketing company, has built their own model called CTRL on top of GPT-2 and released it on Github (https://github.com/salesforce/ctrl). This model has 1.6 billion parameters “trained to condition on control codes that specify domain, subdomain, entities, relationships between entities, dates, and task-specific behavior. Control codes were derived from structure that naturally co-occurs with raw text, preserving the advantages of unsupervised learning while providing more explicit control over text generation” (Salesforce, 2019). Presumably, Salesforce is not as concerned about malicious weaponization of their model as OpenAI is.

This elaboration of the GPT-2 model can be prompted with for instance a URL (real or fake) and, told that the prompt text is a ‘link’, it will generate text informed by the metadata contained in the link (Woolf, 2019). For instance, using Woolf’s code for interacting with CTRL, I fed it the fake url, https://www.cnn.com/world/09/19/2019/evidence-for-atlantis-found-in-suburban-ottawa-backyard. This fake URL does not point to anything at CNN, but it does mimic the structure of a real CNN story. But, the words in the page title are understood by CTRL as controls on that 1.6 billion dimensional space on the kind of language that would be appropriate, collapsing the probabilities to the ‘news story’ dimensions. The result:

Evidence of Atlantis found in suburban Ottawa
MUST WATCH
Story highlights
The discovery was made by a man who lives on the property
It is believed to be one of only two known locations where ancient artifacts have been discovered
(CNN) A Canadian couple has uncovered evidence that could help solve an age-old mystery — and possibly prove Atlantis’ existence.
In what they say is the first time such a find has ever happened in Canada, a husband and wife from Ottawa, Ontario, stumbled upon a cairn…

I fed the same model fake URLs patterned after my university department’s faculty biography pages. I created two names, ‘Sarah O’Connor’ and ‘Stuart O’Connor’. Sarah’s biography was utterly glowing, but listed the many service organizations that Sarah belonged to, while Stuart’s biography put its energy into listing all of his books. The model understands the gendered and hidden labour of the academy it seems – should you play with talktotransformer.com and experiment with sexualized or racialized language, do not be surprised when it generates toxic or not-safe-for-work text, reflecting our current world view back at us (at least, current as represented by the internet in 2019).

The text that models like GPT-2 and CTRL can generate demonstrate an incredibly sophisticated understanding of how the English language works, and the interrelationships of different concepts (seemingly without limit!) that might be the subject of conversation. GPT-2 and CTRL are not just language models, but knowledge models.

And somewhere in there is archaeology.

The experiment with using CTRL demonstrates that my thought experiment from earlier in this chapter, building a digital pseudo-Wheeler who I might want to interrogate about his worldview, is feasible now. What’s more, the CTRL codes show that it is possible to collapse the multiverse of possible generations to a local universe conditioned on a particular worldview. In neural network work on images, you might be familiar with the idea of photos that get ‘painted’ in the style of Monet or Picasso via ‘style-transfer’. This is the same idea, but with text.

I used Max Woolf’s code for retraining the GPT-2 ‘medium’ model (with 355 million parameters, 1.5 gb in size) on Flinders Petrie’s Ten Years’ Digging in Egypt 1881-1891, creating a model of Petrie’s language and – potentially – worldview (I would have used Wheeler, but Petrie’s writings are already available in the Gutenberg Project website, readily digitized.) Then, I ask this simPetrie some questions about archaeology. The dialogue below represents our ‘conversation’; each question by me was a prompt to the model to generate text. None of simPetrie’s responses have been edited. As you read simPetrie’s responses, reflect on what they (may) reveal about the doing of archaeology in the late 19th century. Given that Flinders Petrie is celebrated in archaeology for the development of both field and analytical methods, let’s ask his digital doppelgänger for its thoughts on the matter:

SG> ’Professor Petrie, what is the key finding of your researches?''
SimPetrie> ’The only way to find out what is the key to every mystery is to know everything about it, and then to try to discover the mystery as far as possible by deductions from what is known. When we come to the real problem, how do we find the real in the unreal?''
SG> 'Professor Petrie, could you describe your methods for us?''
SimPetrie> ’I have always tried to be fair to the student, and to give as much information as possible. I have tried to give as many facts as I can remember; and then try to match them with each other in the various accounts, so as to arrive at some general idea. People often ask me where the evidence for a fact comes from; and that is no easy matter, when the evidence is so scanty. I have often thought of writing a work on probability, in which every fact might be found, and the numbers by which they must be numbered, by heart, according to their weight in evidence. I believe I have found the right ratio; and the more I think of it, the more I see that it will not only be right, but will be of great service to our subject.
This method of ascertaining the amount of evidence is the only one that seems to me entirely consistent with scientific accuracy. It is easy enough to find the average of thousands of facts; but it is much harder to find the average of the facts themselves.’
SG> “Could you describe for us the role of local labour in your excavations and surveys?’”
SimPetrie> ’The local labourers, who live and work under me, are the most intelligent and hardy people I have met. They live by digging, and having to earn their living by selling the finished product to other local people who use it to make shoes, &c. Though they are often exploited in this business, they never trouble about it too much, and are content with what they can get for a good wage. Their only ambition is to live as cheaply as possible. The knowledge which I have gained from their labours has been of great service to me in my excavations, and I have always been glad when one of them has told me how he has improved his skill, or found a new construction, by improving the methods of his fellows. I should be sorry to see such methods go to waste for want of improvement and improvement of the methods of my neighbours, for the end in view is the same.
SG> 'Professor Petrie, what are the sorrows of archaeology?''
SimPetrie>’ 'The sorrows of a dead man. I have seen the dead. All that is preserved is what is in their power to leave behind them. The remains of the temples, though they may be exposed, are of no use. They are only patches of stone, in a country where there are many other buildings. If the dead man had been a sculptor, he would have had no need of the temples; nor is it likely that any one could have made any use of them except by accident. The temples, then, were but a sad sacrifice to an unmeaning king.

SimPetrie is startling coherent. The original Petrie’s concern for rigor and consistency in his archaeological recording and systemization of materials appears in the digital Petrie. The use/abuse of local labour is hinted at, with a touch of the colonial condescension we might expect from a 19th century Englishman. The overly -to our ears- florid language of the period comes through.

When we ask simPetrie about ‘his’ world, the neural network takes our question as a signal of how to collapse the possibilities for the generation of its response. Careful questioning and breaking could reveal the limits of that simulated world view. How does that worldview map back to the original Petrie’s? How far can it be pushed before it breaks? Much like an agent based model has to be run through all of its possible combinations of parameters to understand the simulated world view, the simulated history’s behaviourspace, we have to figure out a method for doing the same for this neural networked model. One way perhaps of doing this might be to deploy data mining and text analysis. I could imagine asking simPetrie the same question a thousand times at each ‘temperature’ or creativity setting between 0 and 1. Then, I would topic model (look for statistical patterns of co-occurance of words in a response) these responses, and map how the discourses found therein persist or evolve over the creative space of simPetrie’s responses. That might begin to give us a map of the territory that we have stumbled upon. It will require much work and indeed play, experimentation, and the willful breaking of the models to expose the sharp edges.

Some of the things we wish to play with, like the GPT-2 and CTRL models with their billions of parameters, are perhaps too big for enchantment? Is this where we spill from enchantment to terror? These models after all, now that they’ve been generated (and consider the energy and environmental costs of training such models is estimated to be five times worse that that emitted by a car over its entire lifespan, or approximately 626,000 pounds of carbon dioxide equivalent, Strubell et al 2019; Hao 2019) can now be deployed so easily that a single scholar on a commercial laptop can use them. The technology behind these models is not that far removed from the technologies that can simulate and generate perfect audio and perfect video of things that never happened or were never said, so-called ‘deepfakes’ (these too depend on neural network architectures). We will need to develop methods to deal with and identify when these models are deployed, and quickly. By the time this book is in your hands, there will be new models, larger models, of text generation, of language, and they will be deployed across a range of tasks. It will be exceedingly hard to spot the work written by the machine, versus that written by a human. Our golems are getting out of control. But there are other ethical issues, too.

The Ethics of Giving the Golems a Voice

“When we teach computers to write, the computers don’t replace us any more than pianos replace pianists—in a certain way, they become our pens, and we become more than writers. We become writers of writers.” – Goodwin 2016

“The hypothesis behind invisible writings was laughably complicated.  All books are tenuously connected through L-space and, therefore, the content of any book ever written or yet to be written may, in the right circumstances, be deduced from a sufficiently close study of books already in existence.  Future books exist in potentia, as it were…” Pratchett, The Last Continent

“How do we find the real in the unreal?” – simPetrie

In a world where computers can be creative on their own, ‘authorship’ is not about putting the words down on the page, and ‘scholarship’ is not necessarily about marshalling facts about the world in a logical order to make an argument. Instead, they become an act of creative composition and recomposition, or remixing and selecting of texts for training and hyper parameters to be tuned. It is in fact the same skills and techniques and scholarly work that informs the creation of agent based models. This kind of generative computational creative writing is not really about making a machine pass for a human, but, much like the agent based models discussed earlier in this volume, it is about discovering and mapping the full landscape of possibilities, the space within which Petrie could have written. These particular questions prompted the machine to collapse the possibility space around how archaeology was conducted, and whose voice mattered in that work; thus the results perhaps give us access to things that were so obvious they were never written down. What is the evidentiary status of a mapping of the behaviour space of the model? There could be a fascinating PhD thesis in that question. But this dialogue with simPetrie, for me, also raises some interesting ethical issues that so far in digital archaeology – led by the work of people like Meghan Dennis or Lorna Richardson or Colleen Morgan – we are only beginning to explore.

Tiffany Chan, for her MA thesis in English at the University of Victoria, used a recurrent neural network to map out the space of one particular author. She writes,

“[W]hat could we learn about our object of inquiry (in this case, literature) if we broke down, remade, and compared or interpreted it either alongside or as if it were the original? Articulated in Victorian terms, this project is like conducting a séance with a computer instead of a Ouija board. The computer mediates between human and machine, between the dead and the living. If, as Stephen Greenblatt suggests, literary study begins with “the desire to speak with the dead”… then [this project] begins by impelling the dead to speak.” (2017).

Colleen Morgan wrote, a decade ago, in the context of video games that use historical persons as non-player characters to decorate the games, “NPCs are nonhuman manifestations of a network of agents (polygons, “modern” humans, fiber-optics, and the dead person herself) and the relationships between these agents and as a result should be studied as such.  But does this understanding of an NPC as a network make it ethical to take such liberties with the visages of the dead? What does it mean when Joey Ramone comes back from the dead to sell Doc Martins?”

In these two passages, we find many of the threads of this book. We see ‘networks’ as both a literal series of connective technologies that thread the digital and analog worlds together. We see an impulse to raise the dead and ask them questions, and we see something of the ethical issues in making the dead speak. For instance, Petrie plainly did not say any of the things the simPetrie did in our dialogue. What if simPetrie had said something odious? It’s entirely possible that the model could extrapolate from hateful speech collected in its training corpus, triggered by passages in the small body of text of Petrie with which I perturbed the original.

What if that text gets taken out of context (an academic book or journal article) and is treated as if Petrie actually did say these things? In a conversation on Twitter about simPetrie, the computer scientist and sometimes archaeogamer John Aycock raised the issue with me of desecration: similar to the way human remains can be desecrated and ill-used in the real world, could this use of computation be a kind of desecration of a person’s intellectual remains? Lorna Richardson points out that the creation of any kind of visualization of archaeological materials or narrative ’is a conscious choice, and as well as political act.’ (Richardson, 2018). If these models are the instrument through which I ‘play’ the past as Goodwin (2016) suggests, then I am responsible for what collapses out of that possibility space. The ethical task would be to work out the ways the collapsing possibility space can do harm, and to whom.

The advertising and entertainment industries have the greatest experience so far with raising simulacra of dead celebrities to sell us things and to entertain us. Tupac Shakur raps on stage with Snoop Dog, years after his death. Michale Jackson performs from beyond the grave at the Billboard Awards. Nat King Cole sings a duet with his daughter. Steve McQueen races a 2005 Ford Mustang. These uses of the dead, and their resurrection, are more troubling that portrayals of historical figures in films or video games, because of the aura of authenticity that they generate. Alexandra Sherlock argues that

“… The digital individual continues, irrelevant of the death of its author and prototype, and since the relationship that viewers have with this social entity was always conducted through representations and images anyway, nothing about this relationship actually changes… in popular culture the media persona becomes divorced from the actual embodied celebrity and their representations become a separate embodiment of their own – an embodiment with which people are able to identify and bond with in an authentic and real way.” (2013: 168).

These representations of dead celebrities worked because they depended upon, and continued to promote, para-social one-sided relationships – the public was so used to the feeling of being connected with the idea of these individuals, that their digital resurrection proved no obstacle, no barrier to enjoying the performance. Sherlock discusses an episode where the digital resurrection of a celebrity did go wrong – the resurrection of Orville Redenbacher, of popcorn fame: “Rather than promoting the enchanting notion of immortality, Redenbacher’s advertising agency had accidentally and rather embarrassingly reminded viewers of the mortality of Redenbacher, and themselves by extension’ (170). The advertisement fell into the uncanny valley, the term from robotics that describes when a robot is so human-like that the few errors in the depiction (lifeless eyes, for instance) generate a feeling of creepiness.

Sherlock calls this entire process of using the images of entertainers, whether as holograms or on film, ‘digital necromancy’, and attributes some of the success (or failures) to the idea that, in addition to profiting from a para-social relationship, the revenants fill a need for answers, a need for reassurance in the face of death, given that Western culture largely avoids talking about death:

“…a form of necromancy does exist today, precisely in response to the marginalization of death. One might perhaps consider the technicians who created the Bob Monkhouse advertisement [where the comedian tells the audience about his own death from cancer] as modern necromancers – reanimating the digital remains of the deceased Monkhouse to impart his knowledge concerning his own death. It is as though the ancient art of necromancy has resurfaced in the practice of digital resurrection.” (171).

All of which is to say: simPetrie could become ‘real’ in the same way the personas of entertainers and celebrities become ‘real’, and the views and opinions expressed by the digital doppelgänger given far more weight than is warranted. “Subconsciously, their appearances may appeal to embedded beliefs that the dead are wise and knowledgeable: if they speak or show themselves to us, we should pay attention. Somehow the dead seem more believable.” (172)

When 2k Games, the makers of the game Civilization, in its sixth iteration, included the Cree Pîhtokahanapiwiyin (Poundmaker) as one of the playable leader characters, they put words in his mouth. Milton Tootoosis of the modern Poundmaker First Nation said, “[This representation] perpetuates this myth that First Nations had similar values that the colonial culture has, and that is one of conquering other peoples and accessing their land… That is totally not in concert with our traditional ways and world view.” (Chalk, 2018). While the depiction and lack of consultation with the Poundmaker First Nation is troubling enough on its own, imagine if the game-character of Pîhtokahanapiwiyin was coded in the way simPetrie was, and imagine further that the developers did not consult with the Cree on which texts to use for training – or whether to do this at all.

The danger of the neural networked power representation is in its liveliness, the possibility of fostering the kind of para-social bonds that make the examples drawn from the advertising and entertainment worlds work. A neural network powered representation of a key figure in Cree history would run the risk of becoming the version of Pîhtokahanapiwiyin that sticks; who builds and designs such a representation, and for what aim, matters. This neural network approach to giving voice to a video game’s non-player characters, to an agent-based simulation’s agents, is exceedingly powerful. If we are building simulations of the past, whether through archaeogaming or agent modeling, we either need to make our software agents mere ciphers for actual humans, or we need to think through the ethics of consultation, of representation, and permission in a much deeper way. The technology is racing ahead of our ability to think through its potential harms.

There is also the ethical issue in the creation of the training data for GTP-2 in the first place, the creation of the possibility space. The authors of those 8 million webpages obviously never consented to being part of GTP-2; the material was simply taken (a kind of digital colonialism/terra nullius). The use of Reddit as a starting place, and relying on Reddit users’ selection of ‘useful’ sites (by the awarding of ‘karma’ points of 3 or more to a link) does not take into account the demographics of the Reddit user community/communities. The things that white men 18-35 living in a technophilic West see as interesting or valuable may not be the kind of possibility-space that we really want to start baking into our artificial intelligences powering the world. Taking a page from information ethics, Sicart (2009) argues in the context of video games that permitting meaningful choices within a game situation is the correct ethical stance; where are the meaningful choices for me who ‘plays’ the GPT-2 model, or for me whose website may be somewhere inside the model?

A framework for considering the myriad ethical issues that might percolate out of this way of raising the dead and giving them a voice again might be the ‘informational ethics’ of Floridi and Sanders, as interpreted by Sicart from the perspective of video games. This perspective considers ‘beings’ in terms of their data properties. Data properties are the properties of relationships and the contingent situation of a thing. That is to say, what makes the rock on my desk a paperweight rather than merely debris is its relationship to me, our past history of a walk on the beach and the act of me picking the rock up, and the proper ways of using objects for holding down papers on desks (Sicart 2009, 246, citing Floridi 2003). Compare this with Ingold’s ‘material against materiality’, where he invites you to pick up a stone, wet it, and then come back to it a short while later:

“[…]the stone has changed as it has dried out. Stoniness, then, is not in the stone’s ‘nature’, in its materiality. Nor is it merely in the mind of the observer or practitioner. Rather, it emerges through the stone’s involvement in its total surroundings – including you, the observer – and from the manifold ways in which it is engaged in the currents of the lifeworld. The properties of materials, in short, are not attributes but histories.” (Ingold 2007, 15)

The meaning of data entities lies within the web of relationships with other data entities, and all things, whether biological or digital, are data entities (Sicart 2009 128-130; Morgan 2009). From this perspective there is moral import because to reduce information complexity is to cause damage: “information ethics considers moral actions an information process” (Sicart 2009 130). The information processes that give birth to simPetrie, that abstract information out of GPT-2, that collapse the parameter space to one local universe out of its multiverses, are all moral actions. For instance, these language models and these neural network technologies are predicated on an English model of the world, and English approach to language. Models like GPT-2 obtain part of their power through their inscrutability. Foucault (1999: 222) wondered what an ‘author’ might be, and concluded it emerges in the condensation of physical and cultural influences, that ‘the author function’ disappears instead to be experienced:

“What are the modes of existence of this discourse? Where has it been used, how can it circulate, and who can appropriate it for himself? What are the places in it where there is room for possible subjects? […] What difference does it make who is speaking?”

That is the ethical question posed by archaeogaming, because the ‘who’ isn’t just humans anymore.