AI Use Policy Statement
Many cultural heritage institutions are taking the positive step of reviewing their use of AI and developing policy around the kinds of practical, ethical, and data security issues some AI technologies create.
DT Nexus AI Crop does not use any generative AI technology. It is a narrow task-specific computer-vision model ethically trained using machine learning that proposes crop and straighten settings for heritage digitization images which can (and should) then be reviewed by a human. It runs entirely locally; your images never leave the local computer.
The chart below maps common museum, library, and archive AI Policy Review Questions to how AI Crop actually operates. It is our expectation that AI Crop in DT Nexus is compliant with nearly all AI institutional policies and we welcome any questions pursuant to our use of AI.
Related:
- Terms provided to the institutions who volunteered images to train AI Crop
Institutional review question | AI Crop status | Why |
|---|---|---|
Do collection images leave the institution? | No | The model, image, processing, and crop result stay on the local computer. |
Are customer images reused to train a vendor model? | No | AI Crop does not use customer images to train or refine the model. |
Is the training-data permission or provenance unclear? | No | Training images were ethically sourced from partner institutions that volunteered their images. |
Am I giving up any rights to my images? | No | No. Your images are only used for your own cropping purposes. You retain all rights and ownership. |
Is human oversight and quality control still required? | Yes | The tool proposes candidate crop(s). Human review during QC is facilitated and suggested. |
Can AI Crop hallucinate content or create synthetic imagery? | No | The model identifies crop boundaries; it does not generate text, imagery, descriptions, or historical claims. |
Can performance vary across materials or collections? What biases does the model contain? | Yes | Any vision model can perform less well on unusual collections. In such cases, the candidate crop proposed may need human adjustment. Bias is likely dominated by the physicality of the object (e.g. the shape, size, and overall tone/color) rather than semantic content so is unlikely to be an ethical concern. |
Energy use | See Estimate | Training these models takes roughly the same electricity as charging a mid-sized electric car one time. Use of these models reduces the electric cost per image paid by the user to crop images. See: AI Crop Energy Use |