Protecting what is found

Location is not neutral information

A coordinate can connect researchers with evidence, but it can also direct looting, disturbance or intrusive tourism towards a vulnerable place. Machine learning changes the scale of this problem: a system may generate a large candidate list before anyone has assessed protection needs. Publishing the entire list as a demonstration of technical openness can create risks far beyond the intentions of the research team.

Location sensitivity should therefore be considered at the project’s beginning. Decide who needs precise data, what justification is required and how public outputs will be generalised. A broad regional map may explain the research without identifying an unprotected deposit. Access controls should be proportionate and reviewable rather than a vague promise that sensitive files will somehow remain private.

Consultation is part of the research design

Archaeological landscapes may be living places, sacred places, working land or sources of contested memory. The people connected to them are not simply another dataset. Meaningful engagement asks who has authority, how decisions will be made and what benefits or burdens a project may create. Consultation undertaken only after a model has produced results leaves little room to change the research question itself.

The Global Indigenous Data Alliance’s CARE Principles emphasise Collective Benefit, Authority to Control, Responsibility and Ethics. They are particularly important when research involves Indigenous peoples and knowledge. Their relevance is not satisfied by a brief acknowledgement in a paper. Data collection, model training, publication and later reuse can each raise distinct questions about authority and consent.

Permission to use an image is not necessarily permission to extract and publish culturally sensitive information from it. Likewise, a licence allowing technical reuse does not remove duties to consider harm. Researchers should distinguish copyright permission, lawful access, community expectations and heritage protection. These overlap, but they are not substitutes for one another.

Open methods, careful access

Transparent research can share its question, model architecture, evaluation design, limitations and aggregated findings without making every site location publicly downloadable. Where possible, reviewers and qualified researchers can receive controlled access under documented conditions. The access procedure itself should be understandable: who decides, what criteria apply and how decisions can be challenged or revisited.

The FAIR principles describe research objects that are findable, accessible, interoperable and reusable. They do not require unrestricted public access to every file. A well-described restricted dataset can be more responsible and more scientifically useful than an undocumented open folder. Metadata should make clear what exists, why restrictions apply and how an appropriate request can be made, while avoiding accidental disclosure in thumbnails or embedded coordinates.

Model outputs also deserve a retention policy. An unverified candidate should not circulate indefinitely as if it were an established site. Record its status, review history and any later correction. When evidence overturns an interpretation, update public summaries and downstream datasets where practicable. A prominent correction is better than silently removing the original claim and leaving copies elsewhere unexplained.

Understand the legal landscape

The UNESCO 1970 Convention addresses illicit import, export and transfer of ownership of cultural property. The Council of Europe’s revised European Convention on the Protection of the Archaeological Heritage, commonly called the Valletta Convention, addresses archaeological heritage protection. These instruments provide important context, but neither is a universal field permit or a complete statement of local law. National implementation and site-specific protections matter.

The ICOMOS Charter for the Protection and Management of the Archaeological Heritage also places archaeological resources within a broader framework of preservation and management. For an AI project, that perspective shifts attention from the number of detections to the consequences of investigation. A project that identifies many places but leaves no protection plan, archive or local benefit may produce less lasting value than a smaller, carefully governed programme.

Communicate a result without inflating it

Responsible public writing should distinguish a predicted pattern, a reviewed candidate, a verified feature and a historical interpretation. “Possible” is useful when it reflects a specific uncertainty, not when it decorates an otherwise sensational claim. Explain what evidence supports the conclusion and what remains unresolved. If dating is uncertain, do not borrow the age of a nearby monument simply to make a headline more dramatic.

Contextual photographs should be labelled, as they are throughout this website, so readers do not mistake a close-up of pottery or soil for evidence from a named study. Avoid visually impressive maps that disclose more than the text intends. Image metadata, background landmarks and filenames can reveal sensitive information even when a coordinate table has been removed.

The final measure of a machine-assisted discovery is not whether the algorithm found something striking. It is whether people can evaluate the claim, whether affected communities have been respected and whether the archaeological record remains available for responsible future study. Protection is not an obstacle added after discovery. It is part of what makes discovery worth doing.