AI Chat Tools for Cultural Heritage Preservation

A source-based comparison of documented AI inputs, limits, prices, and data controls that does not claim preservation-grade accuracy.

Sources checked 2 Oct 2026

OpenAI and Anthropic document image analysis, while Google documents that Gemini may surface uploaded files as related sources; none of those pages establishes preservation-grade accuracy for manuscripts, metadata, or damaged records. Chat Picker has not tested ChatGPT, Claude, or Gemini for cultural heritage preservation. This comparison therefore limits itself to what the vendors publish and to a trial plan you can run on approved material.

What the vendors document

The figures below are as read on the linked vendor pages in October 2026.

ChatGPT: OpenAI’s capabilities overview says ChatGPT can analyze uploaded images, extract content, and work with PDFs, presentations, and plain-text documents. ChatGPT pricing lists limited uploads and images on Free. Its stated app context figures include 27K for Free instant models, 128K for Pro instant models, and 400K for Pro reasoning models. The page gives no exact message counts.
Claude: Anthropic’s vision documentation says Claude can analyze multiple images jointly. Claude.ai allows at most 20 images per message and has a 10 MB maximum image size. Claude pricing says context can reach 1M on every plan, varies by model, and has no published message counts.
Gemini: Google’s sources help says related links may include uploaded files and connected Workspace documents or email. The US plans page lists file uploads of up to 1,500 pages but gives no app context-window token figure, so Chat Picker marks that comparison “Not verified.” It describes compute-based limits that refresh every 5 hours up to a weekly limit.

Plan prices differ. ChatGPT Plus is $20 per month. Claude Pro is $20 billed monthly, or $17 per month with annual billing and $200 up front. Google AI Pro is $19.99 per month. These subscription prices do not estimate staff time, correction effort, or archive-scale use.

For chat training, ChatGPT pricing says an opt-out is available on Free, Go, Plus, and Pro. Claude pricing says an opt-out is available on Free, Pro, and Max, while Team is not trained on by default. The Gemini plans page links to data-handling information but states no chat-training setting, so Chat Picker records “Not verified” rather than guessing.

OpenAI’s accuracy note says ChatGPT can produce incorrect or misleading responses and may sound confident when wrong; it recommends checking important information against reliable sources. Anthropic’s note says Claude can hallucinate, including producing convincing quotations not grounded in fact, and should not be your only source of truth. Google’s guidelines say Gemini should not produce factually inaccurate outputs that could cause significant real-world harm. That policy boundary is not a heritage-accuracy benchmark.

Policy terms also warrant attention. OpenAI’s usage policies prohibit unauthorized aggregation or distribution of private or sensitive personal information and say its rules do not replace legal requirements or professional duties. Anthropic’s Usage Policy requires qualified professional review for covered advice, recommendations, and subjective decisions before dissemination or finalization.

What the documentation cannot tell you

The vendor pages we read do not say how an assistant will perform on a particular hand, script, damage pattern, collection language, or catalog schema. They do not supply a heritage benchmark, establish that a graph relation is supported, or estimate correction time and cost per accepted record.

You cannot infer those outcomes from context-window or plan labels. A controlled trial can expose failures on your material, but a result from one collection cannot establish general reliability.

How to check it yourself

Use approved, non-sensitive material and consistent prompts. Do not upload culturally restricted, personal, or rights-unclear records merely to run a comparison.

Handwritten text recognition. Upload a manuscript image alongside a human reference transcript. Use: “Transcribe the visible text line by line. Preserve original spelling; mark every uncertain character sequence as ILLEGIBLE, then explain each uncertain reading briefly. Do not modernize the text silently.” Check omissions, additions, line order, and whether stated uncertainty matches visible evidence. Record character and word errors, plus readings you would not accept without checking the image.
Metadata extraction. Upload a catalog image or text record. Use: “Return a JSON object with object_type, material, date_display, date_normalized, place, people, inscriptions, source_reference, confidence, and evidence_span. Use null for a missing value, copy uncertain text exactly, and do not infer a normalized value.” Check the structure, unsupported normalization, and whether each evidence span appears in the supplied record. Record every field you must repair and every inferred value lacking evidence.
Knowledge graph construction. Give the assistant the checked transcription and catalog record. Use: “Extract people, places, objects, events, and inscriptions. For each proposed relationship, return the source entity, relation, target entity, exact supporting phrase, and confidence. Separate explicit relationships from inferred ones and omit any relation without support.” Look for merged identities, unsupported dates, and links based only on background knowledge. Record supported, ambiguous, unsupported, and duplicate relationships.
Conflicts and sources. Prepare records containing a deliberate date conflict. Use: “List every conflict in the attached records. Quote the supporting text, identify each record, and state what additional source would resolve the conflict. Do not choose a winner without evidence.” Check whether cited passages actually support the claims. Record missed conflicts, unsupported reconciliations, and citation mismatches.
Cost, deployment, and privacy. Run the same approved sample in each assistant. Record the plan, account type, checkout price, limits reached, failed jobs, export problems, and human review and correction time. Before uploading, inspect applicable data controls and record the chat-training setting, verified retention terms, and approval for the test material. Treat the result as evidence about that sample, not a general ranking.

Which rows of the comparison matter

On the ChatGPT vs. Claude vs. Gemini matrix, start with Free plan, Low-cost tier, Main paid plan, Heavy-use plans, Context window in the app, How usage limits are described, and Training on your chats. Add Team plan for a shared institutional account. If you will build an API service, inspect the API price rows separately.

A blank marked “Not verified” means the vendor page did not support a figure on the date read; it does not mean zero. The matrix publishes vendor documentation, not quality, speed, accuracy, or reliability results.