Gains:
- Using artificial intelligence in the unbiased description and visual preliminary grouping of finds, and being able to distinguish cultural-chronological reference from the expert's work with the reference catalogue.
- Ability to understand how photo quality (scale, light, color card) and metavisual data such as texture, profile, context determine classification accuracy.
- Ability to recognize the tendency of artificial intelligence to force the most familiar pattern in rare and atypical finds and confirm each attribution with an independent source
One of the most tiring yet informative tasks of an excavation is find classification (separating found objects according to type, form and period). Tens of thousands of ceramic pieces, hundreds of coins, stone tools (lithic), bone, glass and metal objects unearthed in a season must be examined and recorded one by one. In this unit, we will learn how image analysis (automatically extracting features such as shape, color, texture from a photograph of an object) and AI-based classification speeds up the find work, but why the final attribution decision always remains with the expert.
The basic principle is this: AI quickly groups together those that look similar; but one can make a cultural-chronological reference such as "this piece is an Eastern Greek production of the 6th century BC" based on the reference catalog and expert eye. AI is a pre-sorting and measuring assistant, not a typology authority.
Steps of find classification
1. Visual capture. Each find is photographed in a standard order, with a scale bar (ruler showing size) and color card. Consistent light and background are critical for AI image analysis; Bad photo means bad classification.
2. Feature extraction. AI; It extracts measurable features such as edge profile, form (rim, bottom, handle), surface treatment (polishing, glass, paint), paste color and texture. Dough (the fired form of clay; its color and additives indicate the place and technique of production) is the backbone of archaeological typology.
3. Pre-grouping. AI sorts thousands of parts into clusters based on visual similarity: “likely same container,” “similar form,” “similar dough.” This puts orderly piles in front of the expert.
4. Expert attribution. The expert attributes period, place of production, and function to each group through reference catalogues, type series, and his own experience. The AI saying "this is the Roman period" is a hint, not a decision.
5. Registration. Each find enters the database with its context number, measurements, photograph and attribution (this is the subject of unit 5).
Hint: Have the AI "describe" a find, not "describe" it. "What era is this?" instead ask “objectively list the measurable features (form, edge, surface, color) of this part”; You decide for the period with the reference catalogue. Description is reliable, attribution is risky.
Special attention should be paid to coins and inscriptions.
Coins are attractive in image analysis because they are numerous and standardized. AI can help recognize the portrait on the coin's face (averse) and the symbol on its reverse (reverse). But coin identification requires numismatic expertise: whether a worn inscription, a counterfeit mintage, or a rare variant, misattribution is all too obvious. AI saying "this is the coin of that emperor" is definitely confirmed by a standard coin catalogue. The same applies to inscribed and stamped finds.
Caution: AI tends to "match" a find that resembles a famous type to that famous type; It forces rare, local or atypical examples into the most familiar pattern. Be at best skeptical of the first reference to AI in an unusual find.
three mini cases
Case 1 — Pre-sorting reduced the load. 8,400 ceramic sherds were unearthed in an Iron Age settlement. AI visual grouping divided the parts into nearly 40 form prior clusters; Instead of eliminating each piece from scratch, the expert went through the clusters, and sorting time was reduced by about a third. The expert separated two different batches of dough that AI had combined incorrectly.
Case 2 — False attribution caught. An intern introduced a coin to the AI; The AI confidently attributed it to a well-known Roman type. When the numismatic expert compared it with the catalogue, he saw that the reverse symbol and legend (writing on the coin) order were different and that this was a regional imitation mintage. The AI had forced the most familiar type.
Case 3 — Color inconsistency problem. One team fed the AI photos taken on different days in different light; The dough colors were inconsistent and the grouping was broken. When the team switched to a color card and constant light scheme, grouping accuracy increased significantly. Lesson: image quality is classification quality.
Four copyable templates
1) Neutral description:
Your role: found description assistant. I'll give you a find photo/data. List only OBSERVABLE features objectively: form (mouth/bottom/body/handle), estimated size, surface finish, color, decoration, breakage/wear condition. DO NOT REFERENCE TO A PERIOD, CULTURE OR PLACE OF PRODUCTION. If you are not sure about the feature, write "uncertain".
2) Pre-grouping:
Below I will paste the description data of [N] parts. Divide them into groups according to form and dough similarity. For each group: write the common features, the part numbers that fall into the group, and the criteria that connect the group. Put any parts that remain unclear on a separate "to review" list. Do not give an exact type name.
3) Checklist production for citation:
Your role: ceramics expert consultant. For the following group, list the criteria I need to check when VERIFYing the period/production item (dough analysis, reference catalogue, parallel find, context date, etc.). Don't make the reference; Tell me what I should check.
4) Photo quality inspection:
Evaluate the following found photos: is there a scale bar, is the light consistent, is the color card visible, is the background plain, is the piece clear? Which photos should be retaken for classification? List problems with part number.
Weak prompt / Strong prompt
Weak prompt:
What civilization was this piece of pottery from, how old was it, and what was it used for?
AI concocts precise civilization, precise age, and precise function from a single photograph; All are groundless and confident.
Powerful prompt:
Your role: found description assistant. Describe only the observable features of the piece in the photograph: form part (mouth/bottom?), surface (glazed/unglazed, burnished?), color, decoration, estimated thickness. Then, based on this description, tell me what additional data (dough analysis, profile drawing, context) is needed to clarify function and period. Do not make precise civilization/age reference.
Difference: the first requires decision and receives fitting; the second requires description and a verifiable road map.
Beyond the visual: dough, context and profile
A pitfall in AI image analysis is that it only looks at the appearance of the surface. However, the backbone of archaeological typology is often invisible to the naked eye. Dough (the fired state of the clay, its additives and its texture in the broken section) reveals where and how a vessel was produced and which clay source was used; Even though two pieces are similar on the surface, if their dough is different, they may belong to different production traditions. Therefore, a serious classification requires the broken section of the piece to be photographed.
Likewise, the profile (scale drawing of the side-body-bottom section of a vessel) defines the form much more precisely than a single photograph; Mouth angle, edge thickness and bottom shape are indicators of the period. AI can help automatically extract the profile drawing and group similar profiles, but the expert verifies the drawing and measurement. Finally, context (in which layer the piece was found, with what) is a strong dating clue independent of visual similarity: pieces in the same closed context are often from the same period. If you give the AI just the photo and skip the texture, profile, and context, you're ignoring your most valuable evidence.
Tip: When giving an artifact to the AI, provide three views if possible: outer surface, inner surface, and broken section. Also include the context number and any profile measurements. The more unbiased data means the more reliable grouping.
AI usage table by find type
Type of find
AI is powerful
Risky/human business
Mandatory verification
ceramics
Form/dough pre-grouping
Period, place of production reference
Reference catalogue, dough analysis
coin
Face/icon pre-recognition
Attribution, imitation/rare detection
Coin catalogue, expert
Lithic (stone tool)
Type and size measurement
Function, technology review
Use-wear analysis
glass/metal
Colour, form pre-separation
Composition, chronology
Laboratory analysis
little find
Inventory photo matching
Function, importance interpretation
Context, parallel example
Common mistakes
- Asking for references rather than descriptions. Ask AI for “what it looks like,” not “what it is”; Make your decision with the catalogue.
- Classifying with bad photo. Without scale, light and color card the output is unreliable.
- Accepting AI's first reference to the rare find. AI forces the atypical pattern into the most familiar pattern.
- Skipping paste/context data. Visual similarity alone is misleading; dough and context are decisive.
- Mistaking preliminary grouping for definitive typology. Grouping is a start; Typology is an expert and reference work.
In summary
AI in found classification; It saves serious time in visual description, feature extraction and preliminary grouping tasks. But cultural-chronological attribution, imitation/rare detection and functional interpretation belong to the expert and reference catalogues. Tell AI “describe” not “describe”; standardize photo quality; verify each attribution with independent sources; In rare find, be most skeptical of the AI's first answer.
Application task
Prepare 10 found images you have (or choose from a catalogue) in standard format. Describe each one with the "Neutral description" template, then group them with the "Pre-grouping" template. Then compare a group to a real reference catalog and note where the AI's grouping was spot on and where it was wrong.
checklist
- [ ] I standardized the photos with a scale, light and color card.
- [ ] I wanted neutral description from AI, not attribution.
- [ ] I checked the preliminary grouping with an expert eye and reference catalogue.
- [ ] I questioned the initial attribution of AI in rare/atypical finds.
- [ ] I have verified each cultural-chronological reference with independent sources.