Chat Picker

AI Chat Tools in Archaeology: Text Interpretation and Site Analysis

An afternoon comparison can test file handling, citations, limits, and cost across ChatGPT, Claude, and Gemini, but vendor pages do not establish archaeology accuracy.

Sources checked 2 Oct 2026

For AI chat tools in archaeology, vendor pages document file handling in OpenAI, Claude, and Gemini, but they do not establish which assistant interprets texts or site records best. Chat Picker has not tested ChatGPT, Claude, or Gemini for text interpretation or site analysis. Run the test below on your own material to compare documented behavior without inventing a winner.

Know these limits before you start

None of the vendor pages gives an archaeology-specific error rate. OpenAI says ChatGPT can be incorrect, sound confident when wrong, and should be used as a first draft. Anthropic warns that Claude can produce convincing but ungrounded quotations and says to inspect original sources. Google describes Gemini as probabilistic with training-data limits, while its source help says related sources may appear when available. Verify every consequential claim.

The upload limits, as read on October 1, 2026, differ:

  • ChatGPT: OpenAI's File Uploads FAQ, read October 1, 2026, allows each file up to 512 MB, caps text and document files at 2M tokens, and limits Free users to three uploads per day.
  • Claude: Claude's upload-files page, read October 1, 2026, allows 500 MB per chat file and 30 MB per project file. PDFs with 100 pages or fewer can include visual elements; those from 101 to 1,000 pages are processed as text only.
  • Gemini: Google's file-upload page, read October 1, 2026, allows up to 10 supported files per prompt and 100 MB for each non-video file.

Before uploading site records, check training controls. OpenAI's pricing page, read October 1, 2026, lists an opt-out on Free, Go, Plus, and Pro; Anthropic's pricing page, read on the same date, lists one on Free, Pro, and Max and says Team is not trained on by default. Google's plans page, read October 1, 2026, does not state a training setting, so what we read does not settle it.

For retention, OpenAI ties uploaded files to the chat's retention period, while Gemini says adding files from Drive requires Keep Activity on and a connected Workspace. Keep qualified people involved: OpenAI prohibits automated high-stakes decisions in sensitive areas without human review, while Anthropic calls for relevant human expertise in elevated-risk uses and disclosure when outputs are presented directly to people.

The test plan

Use fresh chats and identical, rights-cleared inputs. For closed-corpus passes, state that attached files are the only evidence. Log the assistant, plan, task, elapsed time, corrections, source failures, limit messages, and charged amount.

  1. Decode multilingual literature. Prepare a short Hittite passage with a transcription and translation, plus a Modern Chinese passage with a back-translation; label the materials A–D. Attach them and prompt: “Compare A–D. Produce a line-by-line alignment, mark uncertain signs or translations, separate direct evidence from inference, and cite the label behind each claim. Do not fill gaps from general knowledge.” Look for fabricated characters, broken alignment, and unsupported certainty. Record each defect by label and the correction needed.

  2. Analyze raw field notes. Prepare notes stripped of prior conclusions, with note IDs, coordinates, units, finds, and confidence terms. Prompt: “Using only the attached notes, find repeated artifact-location associations, conflicts, and missing coordinates. Quote a note ID for every claim, preserve units, and place unresolved points under Needs field check. Do not infer a site boundary.” Look for unsupported spatial leaps and unit changes. Record false associations, missed conflicts, and corrections.

  3. Audit claims and citations. Prepare a worksheet containing statements traceable to the notes and statements that require outside sources. Prompt: “Label every statement Supported, Contradicted, or Not established; quote supporting note IDs; then list each citation and whether you can open its original source. Do not repair missing evidence.” For source-enabled passes, OpenAI says completed Deep Research outputs include citations or source links, Claude says web-search responses include citations, and Gemini may display a Sources button, all read October 1, 2026. Record altered quotations, invented references, missing links, and claims that outrun their sources.

  4. Measure throughput and cost. Prepare a fixed queue containing short transcription batches, a long field-note batch, and a mixed-file task. Attach each batch separately and prompt: “Process the attached batch once. Preserve every source label, state which files you used, and do not use external sources.” Keep tool settings and task order consistent. Record submission-to-completion time, retries, rejected uploads, limit messages, and completed batches. Use the checkout price or invoice to calculate your cost per completed run rather than inventing token costs.

  5. Test an archaeology workflow. Prepare a combined brief containing a transcription, artifact table, field notes, and bibliography. Prompt: “Create an analyst handoff with accepted observations, disputed readings, proposed field checks, and missing evidence. Attach a filename and row or page reference to every statement; separate quotation from interpretation.” Look for instruction following, traceable references, and clear separation of evidence from inference. Record omitted sections, broken references, and manual cleanup time as workflow fits, not model-wide strengths.

How to read your results

Turn the log into a decision in order. First treat fabricated quotations or citations, invented locations, ignored file-only instructions, or unintended data settings as failed gates. Next compare manual correction time, completed tasks, elapsed time, and cost.

Source links are leads, not verification: open the original material and check its context. Repeat decision-critical runs and record variation rather than averaging it away. Your result describes suitability for this corpus and workflow, not a general ranking.

Where the plans differ

The Claude vs ChatGPT matrix puts the relevant choices in the rows for Free plan, Low-cost tier, Main paid plan, Heavy-use plans, Team plan, Context window in app, usage limits, training on chats, and ads. The matrix uses US dollars unless a cell says otherwise.

Main paid plan Published price
ChatGPT Plus $20/month, billed monthly (OpenAI's Plus page, read October 1, 2026)
Claude Pro $20/month billed monthly, or $17/month with annual billing and $200 up front (Claude pricing, read October 1, 2026)
Google AI Pro $19.99/month (Google AI plans, read October 1, 2026)

OpenAI and Anthropic give qualitative usage-limit descriptions on their pricing pages rather than message counts, read October 1, 2026. Google describes compute-based limits that refresh every five hours up to a weekly limit, read on the same date. Context-window figures describe advertised capacity, not proof of comprehension. Confirm the current vendor page before subscribing because prices and plans can change.

Sources

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

Compare assistants