Chat Picker

AI Chats for Heritage Preservation: Digitization and Knowledge Graphs

Vendor pages document differences in file handling, plans, and data settings, while an afternoon test lets you compare OCR, metadata, and knowledge-graph work without inventing results.

Sources checked 2 Oct 2026

For AI chats for heritage preservation, [OpenAI’s File Uploads FAQ], [Anthropic’s upload guide], and [Google’s file guide] document file inputs and their constraints, but these pages do not settle OCR, handwriting, metadata, or knowledge-graph accuracy. Chat Picker has not tested ChatGPT, Claude, or Gemini for these tasks. This page gives you an afternoon test on approved material instead of a quality ranking.

Know these limits before you start

All vendor figures below come from the linked pages as read on October 1, 2026. OpenAI’s File Uploads FAQ says every uploaded file has a 512MB hard limit, text and document files are capped at 2M tokens, and Free users can make 3 file uploads per day. Claude’s upload guide gives a 500MB limit for chat uploads and 30MB for project files; PDFs of 100 pages or fewer receive text and visual analysis, while PDFs from 101 to 1000 pages are text-only. Gemini’s file guide says one prompt can contain up to 10 supported files, subject to availability: video can be 2GB and each other supported type can be 100MB.

The upload pages do not settle screen-reader behavior, keyboard access, alternative text, or accessibility conformance, so test the interface and exported table separately. OpenAI’s accuracy guidance says ChatGPT can sound confident when wrong. Anthropic’s guidance says Claude should not be the only source of truth and that original sites may contain missing context. Gemini Apps may show a Sources button when sources are available.

Before uploading personal or sensitive records, apply your institution’s consent, ownership, and retention rules. The ChatGPT pricing page lists an opt-out from training on chats for Free, Go, Plus, and Pro. The Claude pricing page lists opt-outs for Free, Pro, and Max and says Team is not trained on by default. Google’s plans page states no training setting, so Chat Picker marks that item “Not verified.”

OpenAI’s Usage Policies prohibit unauthorized aggregation or distribution of private or sensitive information and prohibit facial-recognition databases without data-subject consent. Anthropic’s Usage Policy calls for relevant human expertise in elevated-risk uses and AI disclosure at the start of consumer sessions. Google’s safety guidelines consider educational, documentary, and scientific context, while also stating that language models are probabilistic.

The test plan

Use the same permission-cleared material across candidates. Keep one log with a case ID, plan, account setting, exact prompt, output, source location, correction, elapsed time, and spend.

  1. OCR and handwriting. Prepare an image sheet with numbered lines, a typed caption, faded handwriting, uneven spacing, and an ambiguous date. Give each assistant this instruction: “Transcribe all visible text in reading order. Put uncertain characters in a separate notes table, identify the line where each occurs, and do not modernize spelling.” Look for omissions, substitutions, reading-order errors, and unsupported expansions. Log the verbatim transcription and each correction against its image location.

  2. Entity extraction. Prepare a plain-text catalog record containing variant personal names, different date formats, place names, object identifiers, materials, and one uncertain field. Use this prompt: “Extract person, date, place, object identifier, material, and uncertainty into a CSV table. Include the exact source words for every value and preserve conflicting readings instead of choosing silently.” Look for silent normalization or invented entities. Log every value with its source words and your correction.

  3. Knowledge graph construction. Prepare the extracted table and a short institutional glossary. Ask: “Build a triple table with source entity, relation, target entity, and source span. Then list possible identity matches, with evidence for and against each one.” Look for triples unsupported by the records, reversed relationships, and identity merges based only on similar names. Log every proposed match and the evidence needed to accept or reject it.

  4. Source checking and repeatability. Prepare the original record beside the generated output, then repeat the same prompt in a fresh chat. Use this instruction: “Check every claim against the attached record. Classify it as supported, contradicted, or unresolved, quote the source span, and make only the correction needed for a contradicted value.” Look for citation gaps and changes between runs. Log your verdict and leave conflicting evidence unresolved until you inspect the source.

  5. Accessibility, cost, and scale. Prepare a permission-cleared batch containing a scan, handwriting, a long PDF, and a spreadsheet. Ask: “Describe the attached record in plain text, then return its fields in a simple table that does not depend on color or images.” Test uploads and copied output with your keyboard and screen reader. Look for blocked files, unclear focus or labels, unreadable exports, and visible data controls. Log the plan price, any limit reached, training and retention settings, administrator review, elapsed time, and actual spend.

How to read your results

Apply your requirements first: approved data settings, workable file capacity, traceable sources, needed accessibility features, and administrative ownership. If a candidate fails one of these gates, mark it unsuitable for that workflow; this does not establish that its answers are generally better or worse.

Then review your own log. Compare correction burden, unsupported claims, changes between repeated answers, time, and spend. Keep source conflicts unresolved until you check them. Do not replace this evidence with a third-party quality score.

Run the shortlist again on held-out material from the same collection type. Keep a plan only when it passes your requirements and you can explain each correction from the log. Chat Picker has no quality scores, benchmark results, or ranking to add.

Where the plans differ

Open the Claude vs ChatGPT comparison matrix and the three-way comparison matrix. Check the Free plan, Low-cost tier, Main paid plan, Heavy-use plans, Context window in the app, How usage limits are described, Training on your chats, and Ads rows. The matrix uses US dollars unless a cell says otherwise.

  • Main paid plans: ChatGPT Plus is $20/month. Claude Pro is $20/month with monthly billing or $17/month with annual billing and $200 up front. Google AI Pro is $19.99/month.

  • Heavy-use plans: ChatGPT’s Pro tiers are $100, $200, and $500/month. Claude Max starts at $100/month for 5x or 20x Pro usage; the page read did not show the 20x price. Google AI Ultra is $99.99/month for 5x AI Pro limits or $199.99/month for 20x.

  • Context window: ChatGPT lists 27K for Instant on Free, 54K on Go and Plus, and 128K on Pro; its reasoning windows are 256K on Go and Plus and 400K on Pro. Claude says up to 1M on every plan, varying by model. The Google plans page gives file-upload capacity but no app token figure, so the matrix leaves that item “Not verified.”

  • Usage limits and ads: The ChatGPT and Claude pricing pages publish no exact message counts. Google describes compute-based limits that refresh every 5 hours up to a weekly limit. The matrix also records that ChatGPT Go may include ads; it does not verify ad terms for Claude or Gemini.

Sources

For current prices and limits, see the dated comparison pages.

Compare assistants