--- title: AI Tools and Provenance description: How AI tools assist description, transcription, colorization, tours, and analysis while preserving review and audit trails. section: guides order: 10 updated: 2026-07-09 verified: 2026-07-09 related: [guides/ai-analysis, guides/transcription-and-ocr, guides/working-with-artifacts] features: [ai-analysis, ai-compare, transcription, colorization, tours] --- # AI Tools and Provenance You'll understand how Preservated uses AI as an assistant, how results are reviewed, and how provenance stays attached. AI features in Preservated are designed for archival caution. They can speed up description, transcription, visual analysis, and storytelling, but they do not erase the difference between a machine suggestion and a curated statement. ## What AI can help with AI work falls into a few broad categories: - **Description.** Draft titles, summaries, subjects, dates, places, and visible details. - **Text extraction.** Transcribe audio/video and OCR scanned or photographed text. - **Computer vision.** Detect faces, objects, salience, color content, and visual features. - **Derivatives.** Create colorized, restored, cropped, or interpretive versions where policy allows. - **Storytelling.** Generate tour drafts, narration, or exhibit-oriented interpretations for review. Each feature has its own controls and costs, but the review model is shared. ## Suggestions are not curated facts AI output starts as generated content. Staff can inspect it, compare it with current fields, and decide what to accept. For example, AI may suggest "circa 1925" based on clothing and vehicles. That can be useful evidence, but it remains an estimate until a curator accepts it. Even after acceptance, the system keeps enough provenance to show that the value was AI-assisted. :::warning AI is especially risky with names, places, dates, and text in images. Treat those as leads unless confirmed by source material or staff knowledge. ::: ## Provenance travels with results Every AI run records the feature, model, prompt reference, timing, approximate cost, input context, and output payload. Generated files link back to the run that produced them. That provenance supports several practical workflows: - Reviewers can see where a suggestion came from. - Admins can track cost and usage by feature. - Developers can re-run a feature with a better model later. - Institutions can delete or replace generated assets without touching the original master. ## Derivatives need clear labels AI-generated images and media should be labeled according to what they are. A colorized photograph, an educational illustration, and a repaired access copy have different meanings. Good labels help visitors understand whether they are seeing the original record, a faithful access derivative, or an interpretive AI-generated view. When a derivative is meant only for engagement or accessibility, do not let it masquerade as the historical source. **Remove Background** is one of the AI Remix derivatives. It uses a Replicate background-removal model to cut the subject out and produce a transparent PNG, so the object can be placed on any background. The transparent result is shown over a checkerboard so you can see where the background was removed, and downloads default to PNG so the transparency (alpha channel) is preserved. Like every remix, the run is recorded through the provenance ledger: the record notes that the file came from a Replicate model rather than being the original scan. ## Costs are explicit Paid AI operations should require an intentional action. Before transcription, OCR, batch analysis, or tour generation, the UI should estimate cost where possible and make the user confirm. The guiding rule is that AI cost controls shape availability, budgets, and defaults; they should not spend money automatically without a clear click. To review what your institution has actually spent, open **AI Spend** under the admin Management section. It reports usage in **account credits** over the last 7, 30, or 90 days, broken down by feature and model, alongside call and token counts. A **Wallet reconciliation** panel shows credits granted, net spent, and your current balance, so the figures tie out (Granted − Net spent = Balance). A **Ledger history** table lists your recent grants, purchases, and spend. Visitor-facing search and cross-institution tools are billed separately and do not appear there. ## When credits run out If your institution runs out of credits partway through a job — a large import, for example — the AI steps that could not run are not silently dropped. Preservated records each skipped operation and **re-runs it automatically once you add credits**, whether through an admin grant, a credit purchase, or a subscription renewal. The AI Spend page shows a callout when work is waiting on credits. Content screening errs toward caution here: an image that the automated first-pass moderation flagged is held as **restricted** while its detailed review is deferred, so flagged content is never published just because the balance reached zero. When credits return, the deferred review runs and can lower that rating. ## Compare model behavior Global admins can use **AI Compare** to run controlled tests before changing prompts, model choices, or provider settings. The tool separates task types so each run exposes the settings that matter for that task: model and reasoning level for prompt tasks that support it, voice and pace for speech, clip length for transcription, and image controls for generation. Use **Add** to add more output columns when you need to compare beyond the initial set. The layout wraps after four columns per row, so wider comparison runs stay readable. AI Compare includes the main production prompt families: image, text, and video asset analysis; image and media tours; explore and wiki article prompts; content screening; cultural sensitivity review; docs answers; custom-field proposals; and music prompt or tour-brief generation. Some tasks run from seeded sample context, while asset-specific tasks wait for a selected asset so the prompt can include the same artifact, transcript, page, or image context the production pipeline uses. For text-producing tasks such as asset analysis, tours, articles, and transcription, each completed column includes a **Text diff** below **Full JSON**. The first column is the baseline when it has output; otherwise the first completed output becomes the baseline and shows no diff of its own. Other columns show word-level additions and removals against that baseline. Analyze Asset columns show the generated record in a readable form before the JSON, including nested evidence, classifications, materials, condition notes, scores, and other returned fields. The **Analysis** section can review the completed outputs with its own multimodal model picker and editable prompt. It sends text outputs as their JSON payloads and image outputs as images, then asks the selected model to compare the results. The response renders as rich Markdown so headings, lists, and tables remain readable. Use this as a second-pass evaluator, not as proof that one model is correct. For image generation, choose **Generate Image**. The default system prompt is "Generate a header image for a blog post." You can compare OpenAI, Google Gemini, and xAI image models with the same prompts, a 16:9 or 9:16 aspect ratio, and a 1K or 2K resolution choice. These are portable controls; provider-specific details are handled by the server adapter. Generate Image can run without an artifact. When an artifact is selected, its primary asset is still withheld unless you check **Use Artifact**. Checking it sends the primary image with the prompts for image-to-image generation. Leaving it unchecked keeps the run prompt-only, which is safer when you want a clean creative comparison or do not want the source object to influence the output. Image-result columns show the input ratio, each output image ratio, and the percent difference from the input ratio. For prompt-only image generation, the requested aspect ratio is used as the comparison baseline. Text-to-speech comparisons include OpenAI and xAI voice options. xAI speech columns use xAI voice IDs, while OpenAI columns use the OpenAI voice list. Some Replicate-backed tasks can test several models at once. **Remove Background** pre-fills its columns with multiple background-removal models so you can compare cut-out quality, speed, and cost side by side on one image. Production keeps this simple: on an artifact, staff and visitors see a single Remove Background option that uses the default model, not the full model picker. One model in the comparison set carries a non-commercial licensing caveat and stays confined to AI Compare until its use is approved. ## Review quality over time AI output can improve in layers. A raw transcript might become AI-reviewed, then curated after staff correction. A draft description might be rejected today but useful as a prompt for a future batch cleanup. A face detection result might become valuable only after a person is identified later. Preservated keeps this history because archive work is cumulative. The goal is not to make AI invisible. The goal is to make its help useful, accountable, and reversible.