ZharfAI Team
Artificial intelligence does not discover a civilization. It detects a pattern in a representation—perhaps a satellite tile, LiDAR surface, potsherd photograph, transcript, or catalogue—and an archaeologist investigates what that pattern could mean. Between those steps lie landscape knowledge, community memory, chronology, survey bias, excavation context, conservation, law, and the possibility that a plausible prediction is wrong.
As of 30 July 2026, the responsible case for AI in archaeology is consequently about disciplined triage and documentation, not automated historical truth. A model may help review a vast survey or connect fragments in a catalogue. It must not turn an unverified mark into a sensational “lost city,” publish the location of a vulnerable site, or erase the people whose heritage is being studied.
The UNESCO World Heritage Convention text frames cultural and natural heritage as having outstanding universal value and places duties of identification, protection, conservation, presentation, and transmission on States Parties. Even outside World Heritage properties, the order is instructive: detection creates a stewardship obligation. A new candidate site may be vulnerable to looting, development, erosion, conflict, tourism, or exposure created by publication itself.
Before analyzing imagery, define who can see coordinates, how a candidate will be verified, which authority and community will be notified, and what happens if protection is unavailable. Use coarser public maps and access-controlled research layers. A model that produces more targets than a heritage service can safely assess may increase risk rather than knowledge.
An artifact’s meaning is not just its shape. Stratigraphic relationship, provenience, association, material, formation process, collection history, excavation method, and uncertainty make it evidence. AI can classify an image while missing that the object was redeposited, restored, mislabeled, or acquired without a lawful chain of custody.
Connect every derived result to the original field unit, context record, image, instrument, coordinate reference system, analyst, date, software, model, and transformation. Do not overwrite prior interpretations; append revisions with reasons. The archival practices discussed in AI and archival documents apply directly because future scholars must be able to reconstruct how a digital conclusion was produced.
Satellite imagery, historical aerial photographs, multispectral data, synthetic-aperture radar, drone photogrammetry, geophysics, and LiDAR reveal different properties at different scales. Machine learning can rank tiles or segment forms resembling known mounds, walls, roads, canals, field systems, or looting pits. Soil, vegetation, modern infrastructure, image season, resolution, and training geography can create convincing false positives.
Original research on a human–AI workflow for archaeological-site detection used semantic segmentation for the Mesopotamian floodplains and explicitly proposed collaboration in which archaeologists review and improve predictions. That framing is more durable than full automation. Model output should be a candidate layer with probability, coverage, source imagery, and validation state—not a finished inventory.
Training labels often come from known sites, which reflect where previous teams worked, what was visible, which periods attracted funding, and what institutions recorded. A model may learn roads, excavation scars, or survey preferences instead of archaeological form. Absence from a catalogue is not a trustworthy negative when large areas were never examined.
Document the label source and survey intensity. Hold out complete landscapes rather than random image chips, and test across seasons, sensors, land cover, and archaeological classes. Include “unknown” and non-archaeological look-alikes. Review performance with local specialists and communities. Report precision and recall at the candidate level plus the number of field visits generated; pixel accuracy can look excellent while producing an unusable queue.
The ICOMOS Charter for the Protection and Management of the Archaeological Heritage treats archaeological heritage as fragile and non-renewable and emphasizes integrated protection, survey, maintenance, conservation, and qualified management. Excavation destroys the original stratigraphic arrangement as it records it. A model’s confidence is never, by itself, a reason to dig.
Use non-invasive review, desk assessment, community consultation, legal authorization, and a research design before intervention. Define sampling, finds care, specialist analysis, archiving, publication, conservation capacity, and site stabilization in advance. If resources to conserve exposed material do not exist, postponing excavation may be the ethical decision.
The Archaeology Data Service Guides to Good Practice cover the project lifecycle, documentation, metadata, storage, copyright, preservation, and formats for common archaeological data and techniques. An AI project adds further objects: training labels, code, model weights where distributable, configuration, environment, evaluation sets, output layers, review decisions, and known limitations.
Create a data-management plan before collection. Use persistent identifiers and open, documented formats where possible. Preserve raw imagery or a lawful reference to it, not just compressed model inputs. Record coordinate system and transformations. If licensing, cultural sensitivity, personal data, or site security prevents open release, document the restriction and a controlled access path rather than pretending the dataset is reproducible.
OCR and handwriting recognition can accelerate transcription of excavation notebooks, labels, inscriptions, and administrative tablets. Language models can propose readings, joins, or translations. Damage, variant scripts, historical spelling, uncertain signs, and sparse languages make fluent errors especially dangerous. A generated completion can look more coherent than the surviving text warrants.
Store diplomatic transcription, normalized text, editorial expansion, translation, and model suggestion as separate layers. Mark supplied characters and uncertainty. Require review by language and material specialists, cite the image or object, and preserve competing readings. For living, Indigenous, sacred, or restricted knowledge, community authority governs whether material may be digitized, modeled, translated, or published.
Geometry, surface texture, break edges, fabric, decoration, and find context can help rank possible joins among pottery, sculpture, fresco, manuscripts, or architectural blocks. The system can reduce a search space, but physical fit, material examination, and provenance determine whether a join is credible. Similar decoration or a visually smooth 3D alignment is not enough.
Keep accepted, rejected, and unresolved candidates. Record who reviewed them and whether contact was tested physically or virtually. A digital reconstruction should distinguish observed surface, mirrored or repeated structure, scholarly inference, and generative fill through color, line, or an interactive layer. Never export a seamless render into the collection record without those distinctions.
Image similarity and graph analysis can connect an object with older catalogues, excavation archives, theft records, dealer images, or auction descriptions. They may reveal a lead but can also misidentify mass-produced types or propagate bad metadata. Do not treat a marketplace description as verified provenance or let a model’s confidence substitute for documentary and legal review.
Preserve URLs, capture dates, image hashes where lawful, claims by each party, custody changes, permits, and expert assessments. Restrict investigations to authorized staff and protect informants and owners. Do not publish high-resolution security details or accusations before verification. Acquisition and restitution decisions require qualified provenance researchers, affected communities, institutions, and legal authorities.
Heritage is not an inert dataset owned automatically by the institution that can scan it. Descendant and local communities may hold rights, responsibilities, names, oral histories, access protocols, and knowledge about landscapes or objects. Consultation after a model is built is too late. Establish participation, consent, benefit, attribution, access, and refusal before collection or training.
Some information should not enter a general-purpose model at all: precise sacred locations, burial information, restricted imagery, or knowledge whose circulation is governed by community law. A public “open data” default can reproduce extraction. Budget for local language, long-term access, training, data stewardship, and community-defined outputs rather than only annotation labor.
Computer vision can map cracks, biological growth, pigment change, deformation, or surface loss across repeat photography and 3D survey. Change detection is useful only when acquisition is comparable. Lighting, moisture, camera angle, mesh processing, seasonal vegetation, and a new sensor can imitate deterioration. Install physical reference targets, calibrate, preserve raw captures, and review changes with a conservator.
A forecast can help prioritize inspection; it must not prescribe cleaning, consolidation, environmental changes, or structural intervention without material testing and professional judgment. Connect monitoring to an emergency plan and maintenance capacity. The conservation workflow in AI for museums and cultural preservation is a closer model than a generic predictive-maintenance dashboard.
Use explicit states: machine candidate, desk-reviewed candidate, independently reviewed candidate, ground-observed feature, sampled feature, and interpreted site with stated evidence. A map or publication should show which state applies. Record alternative explanations and negative findings. Separate location prediction from date, function, cultural attribution, and historical narrative; each needs its own evidence.
This approach is also familiar in AI-assisted paleontology, where a model can find a morphology but cannot supply geological context. In archaeology, uncertainty should become more visible as claims grow, not disappear behind a confident visualization.
Select one documented landscape and one bounded task, such as ranking previously surveyed image tiles for re-review. Form a governance group including archaeologists, remote-sensing specialists, data stewards, conservators, responsible authorities, and relevant community representatives. Set coordinate access, publication, escalation, and conservation rules before training.
Run the model in shadow mode against a landscape-held-out test set. Blind-review candidates and a sample of model negatives. Measure useful discoveries, false positives, missed known features, reviewer time, geographic bias, and protection workload. Stop if site exposure exceeds protective capacity. Scale only when the workflow produces better documented, safer decisions—not simply more dots on a map.
How museums can apply AI to provenance, digitization, condition monitoring, OCR, 3D capture, and interpretation while protecting context and sensitive knowledge.
Read More
A hazard-first guide to AI in exploration, ore sensing, equipment safety, maintenance, processing, and environmental monitoring with community oversight.
Read More
A practical 2026 guide to cryptographic inventory, NIST post-quantum standards, AI-assisted discovery, crypto agility, migration priorities, and release evidence.
Read MoreSee the daily briefing and the operational guides. This page is an archive note, not an invitation to start a project.