top of page
Search
  • Writer's picturesulla80
    • Mar 31
    • 7 min read

Numismatics & Generative AI

There are multiple versions of the story of Pandora. Hesiod, who is thought to have lived between 750 and 650 BC, describes the creation of the first woman from earth and water, a punishment from the gods and a gift to Epimetheus the brother of Prometheus as penalty for the latter's theft of fire. 

"So said the father of men and gods, and laughed aloud. And he bade famous Hephaestus make haste and mix earth with water and to put in it the voice and strength of human kind, and fashion a sweet, lovely maiden-shape, like to the immortal goddesses in face; and Athena to teach her needlework and the weaving of the varied web; and golden Aphrodite to shed grace upon her head and cruel longing and cares that weary the limbs. And he charged Hermes the guide, the Slayer of Argus, to put in her a shameless mind and a deceitful nature. So he ordered. And they obeyed the lord Zeus the son of Cronos. Forthwith the famous Lame God moulded clay in the likeness of a modest maid, as the son of Cronos purposed. And the goddess brighteyed Athena girded and clothed her, and the divine Graces and queenly Persuasion put necklaces of gold upon her, and the rich-haired Hours crowned her head with spring flowers. And Pallas Athena bedecked her form with all manner of finery. Also the Guide, the Slayer of Argus, contrived within her lies and crafty words and a deceitful nature at the will of loud thundering Zeus, and the Herald of the gods put speech in her. And he called this woman Pandora, because all they who dwelt on Olympus gave each a gift, a plague to men who eat bread." -Hesiod, Works and Days, 59-82
Hesiod is thought to have lived between 750 and 650 BC.

Epimetheus' acceptance of the gift of the first woman (Hesiod: Theogene 590-595) ultimately unleashed evils into the world of man. 

But when he had finished the sheer, hopeless snare, the Father sent glorious Argus-Slayer, the swift messenger of the gods, to take it to Epimetheus as a gift. And Epimetheus did not think on what Prometheus had said to him, bidding him never take a gift of Olympian Zeus, but to send it back for fear it might prove to be something harmful to men. But he took the gift, and afterwards, when the evil thing was already his, he understood. For ere this the tribes of men lived on earth remote and free from ills and hard toil and heavy sicknesses which bring the Fates upon men; for in misery men grow old quickly. But the woman took off the great lid of the jar with her hands and scattered, all these and her thought caused sorrow and mischief to men. Only Hope remained there in an unbreakable home within under the rim of the great jar, and did not fly out at the door; for ere that, the lid of the jar stopped her, by the will of Aegis-holding Zeus who gathers the clouds. But the rest, countless plagues, wander amongst men; for earth is full of evils, and the sea is full. Of themselves diseases come upon men continually by day and by night, bringing mischief to mortals silently; for wise Zeus took away speech from them. So is there no way to escape the will of Zeus.
-Hesiod, Works and Days, 83-108

Epimetheus is given direct responsibility for opening the box in this 16th century Italian etching.

Epimetheus opening Pandora's box, print, Giulio Bonasone (Italian, active Rome and Bologna, 1531–after 1576), an engraving published in Providence. Public Domain image with thanks the to the NY Metropolitan Museum of Art; MET, 64.682.102.

Generative AI

All conversations in the last year seem to end up including a discussion of generative AI and it's implications for humanity. A Pandora's box created by computer science which gives the illusion of human intelligence and has capability to create and navigate vast amounts of information in the form of text, images, sound, and more (multimodel generative AI). There are rational fears and hopes associated with these new technologies. The question today is: is there value to the coin collector or numismatist?


We share here some experiments that conclude: we still have some way to go before these technologies live up to the hype/promise. While concerns for numismatic misuse seem hard to imagine, there are certainly legitimate concerns when you consider : who will have access to this technology more broadly and how will it be used to inform or mislead. 


Attributing Coins

Today these methods are not great at recognizing coins from their images or at creating historical narratives that would be acceptable to an expert. I have tried Anthropics Claude 3 Opus and OpenAI's ChatGPT 4.0 and include below some examples of AI's capability to recognize coins from images. So far fairly disappointing.


Let's start with a Parthian Drachm and see what Claude 3 Opus can do:


Claude 3 is no better with this more common tetradrachm of Alexander the Great - seeing instead a tetradrachm of Lysimachos.


Chat GPT 4.0 does a better job of not being completely wrong - but is overall not very informative.


ChatGPT 4.0 does quite well with the tetradrachm, getting it correctly identified as a tetradrachm in the name of Alexander from Perge, Pamphylia:


Given a set of coins - Claude 3 fails completely in its response:


ChatGPT 4.0 does a bit better in not being wrong, although also less specific about "denarius".


Claude did not do well recognizing this image:



ChatGPT 4.0 is more correct in its answer:


Numismatic Language Model


I tried several tools built for coins on OpenAIs ChatGPT, none of which were very interesting in their answers. This one "The Numismatist" specifically geared to ancient coins from Rafael Bittencourt didn't give much insight:

I do think a specialized LLM focused on coin identification could be very effective, built on large auction databases of coins - overall I think we will see continued improvements and these tools will likely become very useful for everything from coin attribution to die studies over time.


Benchmarking

For benchmarking of multiple open source Large Language Models (LLMs) see:


The Need for Humans

Inevitably the question turns to: if AI can do a lot of things better than humans, what will humans do. I choose to be optimistic - steam engines, electric looms, automated farm equipment, automation in manufacturing, digital information, mobile phones, and other revolutions in technology over the last 200+ years have not put humans out of work - we have unemployment essentially "full employment" at 3.9 percent as of February 2024.


That said, there are certainly roles that no longer exist that once employed large numbers of humans: farmers, elevator operators, typists, switchboard operators, manufacturing jobs, and more...so the type of work that humans do can certainly change dramatically and humans will need to adapt.


Pandora's Box

As we know from Epimetheus and Pandora - we cannot "unopen the box" - so how can we best ensure that we can live with AI?  There is reason for both concern and positive enthusiasm.


Are there risks in misuse of generative AI for numismatics? The biggest risk that I can see is "misinformation" - that misleading GenerativeAI applications could be used to cheat someone in a sale or purchase of coins, or misrepresent history. There are higher risks in other fields that also higher potential for positive benefits to humans.


We will require policy and legal protections and international agreements to slow the misuse of information and misuse of technology. We know from other technology advances, that there will be (and already is today) misuse of these new technology to mislead, to steal, to harm human beings. Governments and regulatory bodies typically react too slowly and don't have enough people truly expert in the technologies to implement appropriate changes. We cannot close the box, but we should make an effort to understand the risks and misuse, and put in place appropriate controls against misuse while enabling the potential for benefits to human beings (and more narrowly numismatics).


Both Claude and ChatGPT provide reasonable answer to the question of "what safeguards should be put in place to protect human beings from misuse of generative AI". I find Claude's list to be more specific (This list generated by Claude 3 Opus):


  • Strict testing and evaluation requirements to assess the safety, robustness, and reliability of AI models before deployment, including checks for biases, vulnerabilities, and failure modes.

  • Mandatory transparency from AI developers about the capabilities, limitations, and intended uses of their systems, as well as clear labeling when AI-generated content is used. Users should be informed if they are interacting with an AI.

  • Accountability measures holding AI creators responsible for harms caused by their systems, with enforceable penalties. There need to be clear lines of responsibility.

  • Prohibitions on the use of AI for illegal or deceptive purposes, such as fraud, impersonation, harassment, or spreading disinformation. AI should not be used to mislead or manipulate people.

  • Human oversight and the ability to appeal important decisions made by AI systems, particularly those that significantly impact people's lives, like in hiring, lending, housing, or the legal system.

  • Strong data privacy protections and user consent requirements regarding collection and use of personal data for training AI models. People should have control over their data.

  • Ongoing monitoring and auditing of deployed AI systems to check for concerning outputs, errors, or unintended consequences. Course corrections should be made as needed.

  • Dedicated regulatory bodies and robust enforcement mechanisms to oversee the AI industry, investigate complaints, and ensure compliance with laws and standards.

  • Public education initiatives to foster critical thinking skills and digital literacy, empowering people to better recognize and navigate the implications of AI.

  • International cooperation and harmonized standards across borders, since AI impacts are global. Regulations should strive for consistency to avoid regulatory arbitrage.


In the US, I find the current administrations Blueprint for an AI Bill of Rights to be refreshingly clear and pragmatic.


The EU AI Act sets out the first comprehensive regulation of artificial intelligence across the European Union. It classifies AI systems based on their risk level to human rights and safety, with higher-risk systems subject to stricter regulations. Key provisions include prohibitions on certain high-risk AI applications such as biometric mass surveillance and social scoring, specific requirements for transparency and data governance for high-risk AI, and the establishment of regulatory sandboxes to foster innovation​ (Europarl Europa)​. More information can be found here: https://www.europarl.europa.eu/topics/en/article/20230601STO93804/eu-ai-act-first-regulation-on-artificial-intelligence


Utility to Numismatists

Large Language Models (LLMs) and specialized Numismatic Language Models built from Auction databases could be valuable for the field of numismatics by aiding in:

  • Attribution

  • Cataloging/Classification

  • Monitoring Literature & Research

  • Education with Chatbot interfaces to knowledge

  • Fraud Detection

  • Pricing trends and estimates


Research on this topic: I will maintain any interesting articles that I find in a list here

78 views0 comments

Recent Posts

See All
bottom of page