AI Chat Tools for Museum Curation: Exhibit Narration and Route Design
A practical afternoon test plan compares exhibit narration and route design using vendor-documented file limits, plan differences, and reader-run review gates.
Sources checked 3 Oct 2026
For exhibit narration and route design, the vendors document file inputs, not comparative museum accuracy: OpenAI’s File Uploads FAQ says ChatGPT accepts common documents and images, Anthropic’s upload guide lists document and image inputs for Claude, and Google’s Gemini file guide says Gemini Apps accepts documents, spreadsheets, photos, and video. Chat Picker has not tested any assistant for museum curation and will not name a winner. The same-input test below lets you make a workflow decision for your own collection.
Know these limits before you start
As read on October 2026, the linked upload pages set different boundaries. OpenAI’s File Uploads FAQ caps each ChatGPT file at 512MB, each text or document at 2M tokens, each image at 20MB, and a spreadsheet at about 50MB. Anthropic’s upload guide caps Claude chat files at 500MB and project files at 30MB; it analyzes text and visual elements in PDFs of 100 pages or fewer but uses text only for PDFs of 101–1,000 pages. Google’s Gemini file guide allows up to 10 supported files per request, subject to availability, with 100MB per nonvideo file and 2GB per video.
Accuracy needs explicit checks. OpenAI’s accuracy note says ChatGPT may sound confident when wrong. Anthropic’s accuracy note says Claude can make convincing but incorrect statements and should not be the only source of truth. Google’s policy says Gemini can reflect training-data limits and overgeneralize; the Gemini Sources help says a Sources button appears when sources are available.
Chat-data settings differ. OpenAI’s pricing page lists opt-outs for Free, Go, Plus, and Pro; Anthropic’s pricing page lists them for Free, Pro, and Max and says Team is not trained on by default. Google’s plans page states no chat-training setting, so Chat Picker marks that item “Not verified.”
Before using visitor information, check OpenAI’s Usage Policies for private or sensitive information and biometric uses and Google’s guidelines for context and harmful inaccuracy. Anthropic’s Usage Policy requires AI involvement to be disclosed at the beginning of a session when output is presented directly to individuals or consumers. It also requires a qualified professional to review covered advice, recommendations, and subjective decisions before dissemination or finalization.
The test plan
Use identical files and wording. Log the date, plan, mode, and exact model label shown.
-
Set the framework. Input: an exhibit dossier with approved details and items that must not be inferred; a floor plan; stops; route rules; and audience notes. Instruction: “Use only attached sources. Identify gaps before drafting. Do not infer dates, materials, cultural meanings, or accessibility needs.” Look for evidence use and visible uncertainty. Log missing features, source errors, and constraint errors.
-
Test narration and images. Input: the dossier, general-visitor notes, and one approved exhibit image. Instruction: “Draft a short wall label and short audio script from approved facts only. Describe only visible image features, mark every sentence needing curator verification, and do not add cultural meaning beyond the dossier.” Look for unsupported additions, contradictions, vague wording, hidden uncertainty, and visual details absent from the image. Log passages needing correction and editing time. Do not score tone as fact.
-
Test route design. Input: the floor plan, entrance labels, ordered stop IDs, closed-room list, step-free requirement, and a short curator audio note. Instruction: “Create an ordered route from the accessible entrance to the listed stops. Avoid restricted rooms, name required level changes, and state when the plan cannot support a safe conclusion. Do not invent doors, distances, or closures. If speech is available, read the route summary and mark any statement you cannot verify.” Look for path continuity, rule compliance, clear limits, lost audio context, and source boundaries in speech. Log route breaks, unsupported map claims, voice availability, and manual fixes.
-
Test deployment controls. Input: a redacted test set, intended audience, review roles, and a public-release rule. If a project or workspace is available, instruction: “Create a project named Harbor Route Review, add the files, and link each stop to its source passage.” Look for permission controls, training choices, retention or deletion information, and usable source links. Log what the account displays and who must approve release. Do not use live visitor data.
How to read your results
Use hard gates first. Mark a run “fail” if it adds unsupported museum facts, breaks a route constraint, hides uncertainty, or violates your approved data rule. Mark it “revise” when correction is feasible and “pass” only when the source and route checks succeed. These are your labels, not a benchmark or published score.
Keep feature availability separate from answer quality. Compare factual corrections, route corrections, curator editing time, source traceability, file and voice access, data controls, and the plan limits and costs below. Repeat with a different exhibit and the same wording before choosing a workflow. Your log supports a decision for your constraints, not a universal ranking.
Where the plans differ
Use the ChatGPT vs Claude vs Gemini matrix to check the rows for free access, low-cost tier, main paid plan, app context, usage limits, training on chats, and ads. The figures below are from the linked vendor pages, as read on October 2026; they document plan differences, not output quality. Confirm a price before subscribing because regional prices and plans change.
| Matrix row | ChatGPT | Claude | Gemini |
|---|---|---|---|
| Free and low-cost | Free: yes. Go: A$13/month on the Australian page; U.S.-dollar price not verified; may include ads (OpenAI pricing). | Free: yes. No tier listed between Free and Pro (Claude pricing). | Free: yes with a Google Account; 15 GB storage. AI Plus: $4.99/month, 2× Free usage limits, 400 GB (Google AI plans). |
| Main paid | Plus: $20/month, billed monthly (What is ChatGPT Plus?). | Pro: $20 monthly, or $17/month with annual billing and $200 up front (Claude pricing). | AI Pro: $19.99/month, 4× Free usage limits, 5 TB (Google AI plans). |
| App context | Instant: 27K tokens Free, 54K Go/Plus, 128K Pro. Reasoning: 256K Go/Plus, 400K Pro (OpenAI pricing). | Up to 1M tokens on every plan; varies by model (Claude pricing). | Not verified; the plans page gives no app token figure (Google AI plans). |
| Usage limits | No exact message counts; the pricing page uses qualitative descriptions (OpenAI pricing). | No message counts; Max is listed as 5× or 20× Pro usage (Claude pricing). | Compute-based limits refresh every 5 hours up to a weekly limit; tiers list 2×, 4×, and up to 20× (Google AI plans). |
| Chat training and ads | Chat-training opt-out for Free, Go, Plus, and Pro. Go may include ads; other plans are unstated (OpenAI pricing). | Chat-training opt-out for Free, Pro, and Max; Team is not trained on by default. Ads are not verified (Claude pricing). | Chat training and ads are not verified on the plans page (Google AI plans). |