◆
We're not anti-AI. We're anti-waste — and anti-deception. Most tools won't tell you plainly what they're trained on, where your work goes, whether you can leave, or who got paid. So we made a label that asks the same four questions of every tool, scores them the same way, and dates the answer. Use it to choose on purpose.
How we vet (so this isn't just our opinion)
- Four axes, asked of every tool the same way — no cherry-picking.
- Every rating carries a date and a source, because provenance shifts (tools change their policies; this label changes with them).
- We rate practices, never people. We never claim a tool is "pure" — only what it is, today, with a link.
- Disagree with a rating? It's a living document — send the source and we'll revise.
◆ TRAINED ON WHAT
◆ WHERE YOUR DATA GOES
◆ CAN YOU LEAVE
◆ WHO GOT PAID
clean / strong mixed / caveats weak / undisclosed
Trained on what100% licensed + public-domain data (Getty, Alamy, Envato).
Where your data goesOpen-weight; runs local / on-prem / air-gapped.
Can you leaveOpen source; self-host the weights yourself.
Who got paidPartners compensated programmatically by "influence."
The cleanest provenance story in the field right now. Caveat: free for non-commercial use — commercial use needs a paid license.
Source: Bria/FIBO announcement & model card, 2026 (briaai/FIBO, Hugging Face).
Trained on whatOnly Getty's own licensed library.
Where your data goesCloud service; prompts processed server-side.
Can you leaveHosted; you keep output rights, not the model.
Who got paidContributors compensated; output commercially indemnified.
Strong on consent and legal safety; it's a hosted product, so it's not local-first.
Source: Generative AI by Getty Images, gettyimages.com/ai, 2026.
Trained on whatLicensed library; Contributor Fund model.
Where your data goesCloud service.
Can you leaveHosted; output licensed to you.
Who got paidRoyalties paid back into the Contributor Fund per data used.
Consent + compensation built in; hosted, not local.
Source: Shutterstock Contributor Fund disclosures, 2025–2026.
MOSTLY DEFENSIBLE · CAVEAT
Trained on whatLicensed Adobe Stock + public domain — but partly trained on AI-generated images from rivals (e.g. Midjourney).
Where your data goesCloud service.
Can you leaveHosted; output rights to you.
Who got paidAdobe Stock contributor bonus pool.
Don't repeat the "only ethical option" marketing line uncritically — the rivals-training detail is the honest asterisk.
Source: Adobe Firefly gen-AI approach (adobe.com) + reporting on Firefly training mix, 2025–2026.
Canva (what we used)
REASONABLE · NOT PURE
Trained on whatHuman contributor library = consent + comp; generative features lean on grayer third-party models.
Where your data goesCloud; Free/Pro accounts default opt-in to AI training (you must turn it off).
Can you leaveExport your designs; tied to the platform.
Who got paid$200M creator royalty commitment; contributors can opt out.
We used Canva for this project's visuals. Fine to use; we won't call it theft-free. We favor the human-contributor library over AI-generated output, and we opted our own uploads out of training.
Source: Canva Shield + AI Product Terms + Contributor Agreement update (Jan 2026).
Can't claim purity: tools trained by broad web-scraping without consent or compensation (early Midjourney, base Stable Diffusion, and similar) don't pass this label — even when the weights are open. Open source and ethically sourced are not the same thing.
← back home