Ancient Greek computing technology: the Antikythera mechanism, on display at the Hellenic National Archaeological Museum in Athens (inv. no. X 15087). This 2000-year-old device is claimed to be the world’s first analogue computer. It was retrieved in 1901 from an ancient shipwreck off the Greek island of Antikythera. This fragmentary and corroded, box-shaped bronze device originally measured about 340mm high. Inscriptions identified on it support the consensus that it was constructed to predict the positions of celestial bodies. It featured several interlinked indicators, driven by a system of interlocking gearwheels, which displayed the positions of the moon and the sun and likely the planets within the zodiac, the date according to the Egyptian calendar, a Greek lunisolar calendar, and details of upcoming solar and lunar eclipses. Modern research, including a new digital simulation, suggests that the mechanism was prone to errors. It could only be cranked to about four months into the future before its complex system of interconnected gears jammed or its teeth disengaged and everything needed to be reset. Is there a lesson here for modern computing technology? Photograph: Francesco Bini. CC BY-SA 4.0.
Out-of-date personal computer at the former Vénusz Hotel, Budapest. The advent of computers designed for personal use, plus the internet and the digital revolution, have significantly changed people’s lives, including archaeologists’. Digital archaeology is well-established and AI-archaeology is developing rapidly. However, electronic waste is now the fastest-growing waste stream in the world, and its disposal and processing, particularly in developing countries, is producing a toxic heritage of environmental pollution and adverse human health impacts. AI data centres are also placing growing strains on global resources, through extreme energy demands and massive water consumption for server cooling. Media archaeologists are now shedding new light on the significance of these digital objects in the contemporary world. Photograph: Anon. CC BY-SA 4.0 (https://commons.wikimedia.org/wiki/File:Budapest,_D%C3%B3sa_utca_2,_V%C3%A9nusz_Hotel,_53.jpg).
Extending the AI debate in archaeology
We are increasingly told by politicians and tech industry executives to embrace the Artificial Intelligence (AI) technological revolution or get left behindFootnote 1 (Frontispieces 1 & 2). Like it or not, AI tools are here to stay; our futures will be impacted by them. And so, we are all being drawn into considering what AI means for ourselves, for scientific research and for archaeological practice—even those of us who lack confidence in computer science. This editorial is intended as a contribution to this broad debate, with a focus on the use of AI tools in archaeological writing, illustration and publishing in academic journals—discussion of which some journal editors are just beginning to engage with thoughtfully.Footnote 2 In contrast, there is already a large body of published literature presenting examples and assessments of AI technologies used in archaeological research, including ethical concerns.Footnote 3 Indeed, some groundbreaking case studies have already been published in this journal.Footnote 4 Gonzalo Pimentel and colleagues add another in this October issue of Antiquity.Footnote 5 They present their combined use of aerial photogrammetry and AI analysis with the DeepLabv3 deep learning model (a semantic segmentation algorithm trained to detect tracks in images of desert terrain), revealing alarmingly extensive off-road vehicle-track damage to the Alto Caramucho geoglyphs in the Atacama Desert, Chile. Other published literature deals with AI applications to archaeological practices ranging from cultural resource management to teaching.Footnote 6
Before I go on, it is worth briefly trying to demystify (at the risk of oversimplification) a few basic terms and technologies deployed under the broad umbrella of AI, which are relevant to what follows. AI involves the development of computer systems capable of emulating specific human abilities, including learning from and adapting to vast amounts of information, recognising patterns, processing natural language, solving problems and making predictions. AI encompasses various areas of research and application, ranging from Machine Learning, where—with varying degrees of supervision—computers and their algorithms learn from data fed into them, identify patterns and turn these into rules, through to Generative AI (GenAI), which—instead of just analysing existing data—creates ‘new’ content. GenAI can generate a range of outputs (from art to computer code), plus a lot of heated debate about creativity and plagiarism. Of relevance to archaeological publications, multiple forms of GenAI technology can be used to produce texts, audio, images and videos. Users most commonly interact with GenAI via chatbots, which generate human-like responses to questions and commands. When it comes to generating text, chatbots rely on Large Language Models (LLMs), which are AI systems trained on a vast corpus of writing extracted from the internet. Based on the context of words within its corpus and by evaluating our prompts, they create new sentences by recombining text, predicting the most suitable next word and emulating human language styles. They can tailor outputs according to our requirements and audience, such as a student essay, an academic paper or a blog. The most prominent chatbots are free to use for tasks like writing but expect to use our ‘conversations’ with them to train future AI models and have paid tiers and daily usage limits for premium features and models. New products are being developed all the time. There is a plethora of major products and companies, the best-known being ChatGPT (OpenAI), Google Gemini, Microsoft Copilot, Claude (Anthropic) and DeepSeek (High-Flyer). Some of these incorporate image generators. For example, ChatGPT uses OpenAI’s DALL-E models, while standalone examples are Midjourney and Stable Diffusion. In addition, academics also use more specialised platforms like SciDraw AI, BioRender, Illustrae and FigureLabs to produce scientific illustrations.
Below, I critically evaluate the use of GenAI by archaeologists to create texts and images intended for publication in academic journals, and the related issue of publishers licensing tech companies to harvest archaeological journal articles to train LLMs. I then contribute to ethical debate around (in)appropriate engagement with AI technology by journal authors, peer reviewers and editors. Speaking of which, I freely admit that my own writing process for this editorial did use some AI tools, albeit reluctantly. Soon after switching on my computer, I took for granted Microsoft Outlook’s use of AI and machine learning algorithms to automatically filter my email for junk mail and phishing scams. As I began reading up on AI and archaeology, I avoided using Google Scholar’s generative AI tools to provide research summaries, but I could not resist Google’s AI Overviews offered in response to my basic queries like ‘briefly define AI in layperson’s terms’. And as I started to type into my Word document, I did try to defy Microsoft Copilot’s insistence on offering me predictive text and recommendations on punctuation, spelling and conciseness. I absolutely did not use it to draft sentences or brainstorm ideas directly within my document—I still like writing; it helps me develop and order my thoughts. I also value learning from generous colleagues, including Chris Coleridge, Gabriele Gattiglia, Catherine Geering, Lorenzo Nigro and Abigail Teasdale, who I relied on to correct and clarify my first draft. When I was done, our publisher provided me with a page proof that contained (US legally mandated) alternative figure captions (‘alt text’) for the frontispiece images, intended for visually impaired readers relying on screen readers; these I revised because they were AI-generated and contained inaccuracies.
The good, the bad and the ugly
Scientists are increasingly experimenting with, and arguably benefitting from, writing and image-making with AI-generated and -assisted technologies. Advocates contend that—with some improvements—AI tools have the potential to make writing-up faster, freeing time for new research, data gathering and interpretation.Footnote 7 AI chatbots can provide ready answers to basic factual queries that augment internet searches. They can also engage in back-and-forth exchanges that help authors refine their ideas and sentences. AI models can help complete literature reviews, since they excel at paraphrasing and summarising. They can quickly polish text, tables and graphics. They can generate popular reconstruction images and videos. GenAI’s capacity to translate and tidy language can be particularly valuable to authors for whom English is not their mother tongue—improving the readability of their texts and enhancing the chances of their manuscripts being accepted for publication in an increasingly competitive scientific environment. In this regard, GenAI may help make academic communication more equitable and diverse.Footnote 8 Similarly, Matthew Magnani and Jon Clindaniel have claimed that “artificial intelligence promises to democratize the production of visual representations of the past”, taking it out of the hands of archaeological illustrators, who Magnani and Clindaniel dismiss as highly skilled, expensive and time-consuming.Footnote 9
Other archaeologists debating these issues, such as Gabriele Gattiglia, have generally maintained a more cautious and critical attitude.Footnote 10 They argue that, to normalise AI in archaeology—as has already been done with digital technology—we must acknowledge various downsides, including those of GenAI in archaeological interpretation, writing and visualisation. AI use requires access to significant computational expertise, resources and hidden labour, which are dominated by big-tech companies based in the USA and China that exacerbate socio-economic inequalities and produce a large carbon footprint. Further inequalities are established between users according to their ability to pay for the most capable and secure systems. In all cases, LLMs use training data and algorithms whose accuracy is almost impossible to verify. This contradicts the principles of open science that require transparency, reproducibility and accountability. Engagement with LLMs risks perpetuating outdated scientific knowledge as well as Anglophone literature, dominant voices, biased stereotypes, misinformation and even wilful disinformation. This is because LLMs undiscerningly incorporate training data of questionable quality found across the internet, including popular but unsubstantiated ‘pseudoarchaeology’ outputs. Reading, writing and reviewing with LLMs may also lead to unintended breaches of copyright and plagiarism. Chatbots are known to confidently produce factual inaccuracies and inventions (or ‘hallucinations’). They oversimplify and decontextualise complex archaeological data for computational ease, homogenising thinking and visual representation. They lack sufficient depth of thought, logical reasoning and well-supported inference to provide novel, plausible and persuasive interpretations of contextualised material remains and human processes. And they struggle to leave space for the interpretive uncertainty and caution that can lead to new research questions.
The editors of the American Journal of Archaeology have consequently gone so far as to argue that archaeologists have a moral imperative not to use GenAI in academic writing and peer review.Footnote 11 But the genie is out of the bottle, and prohibition will only drive such behaviour underground. We are already beginning to see signs of AI-generated academic writing, in submissions to peer-reviewed journals and in some published articles. This is despite the stipulation by most journals that corresponding authors declare at the point of submission whether and how they have used AI. But most such declarations only take the form of generic disclosure statements. And, given that undisclosed AI in submissions to academic journals has increased dramatically since 2023,Footnote 12 what about those authors and manuscripts that falsely declare that they have not used AI, in part because they fear that by admitting to its use their paper will be rejected?
AI-generated academic writing, especially in synthetic pieces, review articles and peer reviews, is becoming difficult to distinguish from human writing, although scientific studies are attempting to assess the scale of it, including a few focused on archaeological publications.Footnote 13 AI-generated or -modified content, particularly that produced using earlier models of chatbots, can leave various tell-tale signs. There may be corrupt or fictitious bibliographic references, that include some ‘real’ elements, such as author names, journal titles and DOIs (Digital Object Identifiers), but no connection between them. Accurate references may be dominated by secondary sources, such as review articles. Multiple citations can be present that do not substantiate statements, while key sources can be overlooked. There can be factual errors, fabricated empirical data or a lack of data and their detailed analysis. Instead, GenAI ‘academic’ texts exhibit a preference for generic interpretations, superficial argumentation, insufficient substantiation and logical circularity. Documents can also be marred by presentational issues, including poorly structured text, imaginary sub-headings, repetitive loops in writing, standardised sentence length, heavy use of dashes and filler words, excessive use of bullet points, replaced characters, inconsistent formatting, inclusion of hidden text and AI prompts (including instructions to LLMs to generate a positive peer review). The writing tends to have an excessively neutral, bland, averaged style, which can include formulaic academic phrasing but lack authorial voice, personality and nuance. Like the outputs of ‘paper mills’, such articles can also be attributed to single authors, who may or may not exist, using a non-institutional email address yet a well-known university as their affiliation. Few of these features on their own provide proof of undisclosed GenAI use, although together they can give a strong impression. However, as tools develop, the characteristic errors and patterns of AI writing will become less evident.
Beyond academic journals, the situation is perhaps even more disturbing. For example, in a blog on ‘AI in archaeology’ hosted by the ‘deep learning service’ of Saiwa Inc. (a company based in Ontario, Canada, that delivers artificial intelligence and machine learning services and solutions), Maryam Rajaei (a software and AI developer) asserts that AI will revolutionise archaeological study of the past by offering “unprecedented opportunities for discovery, analysis and interpretation” that will “transform how we comprehend the past”.Footnote 14 The blog, which outlines the various applications of AI in archaeology, was completed by the ‘author’ during two incredibly prolific months (December 2025 and February 2026) when they posted a total of 167 blogs on Saiwa’s website on topics ranging from AI in mineral exploitation to agriculture, healthcare and finance. All this without any acknowledgement of whether AI tools were used to generate the text (such tools did generate the images), or any critical reflection on the limitations and potential biases of AI technologies.
Scraping and licensing
These GenAI tools are evolving as more training data are ‘scraped’ from academic publications online and fed into LLMs. Whether this has been done with authors’ knowledge and consent is debatable. For example, in 2023 a group of authors sued Meta, the social media company, for copyright infringement.Footnote 15 They alleged that Meta had used their books without authorisation to train Llama, the LLM that powers its chatbots. More specifically, they set out to prove that Meta did so by tapping into the LibGen dataset, an online ‘shadow library’ that originated in Russia, which provides free access to pirated copies of millions of novels, non-fiction books and scientific journal articles. Similarly, publishers have sued LibGen for copyright infringement, resulting in it being blocked in some countries. Nevertheless, academic publishers also see opportunities in this inadequately regulated context. They are now signing formal licensing agreements with tech companies that offer high-quality content from published books and journal articles to tech companies for a fee. They have also invited their authors and publishing partners to opt-in. They argue that licensing will mitigate some of the risks of GenAI technologies by: countering the unauthorised and unattributed use of copyright-protected works; controlling the creation and distribution of new, derivative works that would compete with the original work; improving the accuracy and integrity of AI-generated content; increasing the discoverability of academic publications; and providing additional royalty income for rights holders.
Privately, not all journal editors agree. They point to the relative weakness of academic publishers to insist, either through negotiation or legal action, that powerful and wealthy tech companies provide fair remuneration and copyright protection for authors. They also feel uneasy about taking decisions on behalf of their authors, some of whom might object to their content being used in AI tools, with ensuant risk to their journals’ reputations. Different journals are taking different decisions, but Antiquity has—for the time being—opted out.
Journal publishers, authors and editors
Despite their questionable approach to data sharing with tech companies, publishers are insisting on ethical behaviour and transparency from authors and editors. For example, Antiquity’s publishing partner, Cambridge University Press, has recently issued various policies and guidelines in the use of AI.Footnote 16
As far as authors are concerned, Cambridge University Press pragmatically recognises that AI use is already widespread and engrained in many stages of the research and writing process. It also distinguishes between ‘assistive’ AI use (where the author’s own work is the input and the AI tool gives suggestions for minor improvement of content, such as spelling, grammar and formatting) and ‘generative’ AI use (where the AI tool generates content, such as analyses, translations, text and images). Cambridge University Press goes on to insist that all non-minimal GenAI use must be disclosed transparently and described fully in journal articles, in the methods section, the acknowledgements and figure captions. Details such as the names and versions of the AI tools used, how they were instructed and employed, and any ways in which the author modified them should be provided. They also specify that authors remain responsible and accountable for the content of their work—meaning that AI tools cannot be listed as co-authors, and that authors need to check their manuscripts carefully for unintended consequences. All this seems appropriate and reasonable. But are authors sufficiently aware of such guidance and do they follow it?
For articles submitted to Antiquity, authors must confirm that they have read and agree to comply with Cambridge University Press’s ethical policy on ‘AI contributions to research content’.Footnote 17 They must also confirm whether they have used AI tools in the research or writing process. Consequently, we are seeing more authors declaring, via our submission platform, the use of AI tools. However, the number of such declarations remains low, few authors provide further details in the main text of their manuscripts, and only exceptionally do they specify precisely which AI tools they have used. Should we then expect editors to scrutinise authors’ AI use more actively, and if so, how?
Cambridge University Press’s guidance for journal editors leaves reasonable space for editorial discretion, appropriate to different disciplines and submissions. They also advise that no submissions should be judged on the basis of AI use alone; all AI declarations should be treated with neutrality and in the context of the manuscript. Nevertheless, Cambridge University Press and the Committee On Publication Ethics (COPE) do expect editors to be capable of ‘practical policing’ of any policy they implement, to uphold the credibility and trustworthiness of the scholarly literature.Footnote 18 The problem here is that there are currently no AI detection tools capable of reliably proving the undisclosed use of AI to the standard required for firmly substantiated editorial accusations and decisions. Human oversight, therefore, remains essential. Yet this further increases and complicates the work of editors and peer reviewers, who themselves are prohibited by Cambridge University Press from using unsanctioned, third-party AI tools within the editorial and review processes. It also encourages editors to ‘suspect’ more authors of submitting manuscripts with undisclosed AI-generated content, as well as peer reviewers of using AI tools to draft their reports (and of breaching copyright and confidentiality in the process, by uploading manuscripts to LLMs). Will this lead to editors unfairly trusting the authors they already know and becoming more suspicious of single authors and early career researchers from regions of the world they engage with less? And to cap it all, editors are being advised neither to cite suspected AI use as a reason for rejection if they cannot ‘prove’ it, nor to sanction misbehaving authors, presumably out of fear of legal challenge.
As an indication of some of the difficulties faced by editors, I offer a real but anonymised example of a manuscript that was recently submitted to Antiquity, peer reviewed and ultimately rejected. The corresponding author, in exemplary fashion, declared during the submission process and in their manuscript that AI was used in the writing process. They specified that the authors had used a specific AI tool to improve the clarity and fluency of the manuscript. Due to anonymisation of the acknowledgements, the three peer reviewers were unaware of this and therefore evaluated the manuscript strictly on the grounds of quality. Their advice regarding publication varied. The first reviewer recommended only Minor Revision but did pick up on an important factual error in the manuscript. The second reviewer called for Major Revision. Among their list of concerns, they drew attention to the apparent use of AI, resulting in a text marred by verbosity, questionable word-choice, obscure meanings, excessive repetition and a missing core sub-heading. The third reviewer recommended Reject. They expressed their disappointment with the paper for four main reasons. First, the inadequate presentation of the primary archaeological data from the site in question. Second, the disregard of the archaeological state-of-the-art in the study area. Third, the lack of substantiation of general statements. And fourth (as also noted by the first reviewer), the presentation of interpretative statements and of a graphic based not on material from the archaeological site and culture in question but copied from research in another, unrelated part of the world. In this case, then, the confirmed use of an AI tool in the writing process resulted in a manuscript that exhibited some of the hallmarks not only of AI-assisted text but also of AI-generated text, to the detriment of the paper and its authors. Evidently, we need to understand better our fellow authors and precisely why, how and to what extent they are relying on AI tools when writing up their research and peer reviews.
Concluding thoughts
Let’s face it, growing numbers of archaeologists, including early career researchers, are investing in AI tools—financially, professionally and personally. These have the potential, like other new archaeological technologies, to contribute positively to the development and democratisation of archaeological practice. Journal editors should not, then, drive their use underground; instead, they should actively encourage authors to declare explicitly their use and to check outputs carefully for errors. Human oversight and accountability remain essential. So do ethics. This means using AI tools critically, responsibly and transparently. We must also push ourselves, in this hybridised coexistence with AI, to continue to do what we do best when presenting the results of our research. We should avoid taking short-cuts when producing literature reviews, embrace the complexities and uncertainties of our contextualised data, and debate interpretations that express our own creative thinking and arguments that have intellectual coherence. We should prioritise rigour over ease, quality over quantity, integrity over ambition. At best, AI is a set of tools supporting archaeological analysis and writing that leaves us free to express our thoughts in our own words, instead of outsourcing them to big-tech companies.
