AI Policy
Evidence Atlas uses AI like a flashlight, not a crystal ball. AI can help surface patterns, summarize dense material, and reduce busywork—but it cannot replace provenance, peer critique, or human judgment.
Our default posture is simple: show sources, label uncertainty, keep humans in the loop, and protect people + heritage.
Non-Negotiables
These guardrails shape every AI feature we ship (and every AI feature we refuse to ship).
Citation-First
AI outputs must point back to evidence items and sources. No “trust me bro” summaries.
Transparent Labels
AI content is clearly marked, with uncertainty labels and a path to the underlying inputs.
Safety & Heritage
We don’t enable trespass, looting, harassment, or sensitive-location disclosure—ever.
Privacy by Default
We minimize data collection and avoid invasive tracking. Private submissions stay private.
Wonder is welcome. Evidence is required. AI can help you navigate wonder—but evidence is what earns conclusions.
Policy, Explained Like You’re Smart and Busy
A practical guide to what AI can do here, what it cannot do, and how we keep it honest.
1) Scope: What AI Does Here
AI features in Evidence Atlas are built to support search, synthesis, and structure. The goal is to make it easier to move from scattered material to testable claims.
AI is not allowed to present itself as authoritative. It may propose hypotheses, but must surface sources and invite counter-evidence.
2) Labels & Uncertainty
We separate observation, inference, and hypothesis. AI content follows the same labeling system.
“What is visible/measurable” (e.g., dimensions, materials, photographed tool marks).
“What likely follows” from observations—but still contestable.
A proposed explanation. Must list alternatives and what evidence would confirm/refute it.
Label rule of thumb:
If an AI sentence cannot be traced to an evidence item or clearly framed as hypothesis,
it does not ship.3) Guardrails: What We Refuse to Do
Some features are “cool” and also deeply harmful. We’re not building those.
No hallucinated citations
AI must not invent sources. If the model can’t cite it, it must say “I don’t know.”
No deepfake evidence
We don’t generate “evidence media.” If we ever use synthetic imagery (e.g., diagrams), it is labeled as such and never mixed with primary documentation.
No sensitive location enablement
We don’t publish or infer access routes to protected areas. Some coordinates are intentionally imprecise.
No automated “truth verdicts”
AI may propose competing explanations—but humans and evidence decide what’s supported.
4) Human Oversight & Appeals
AI assists. Humans own outcomes. Any AI-assisted claim, label change, or moderation action must be reviewable and reversible.
If you think an AI output is wrong, misleading, or unsafe, you can flag it—and we’ll review it with receipts.
Review trail
Who approved it, when, and why—visible in revision history.
Reason codes
Moderation and relabeling actions include plain-language reasons.
Corrections log
Corrections are public and normal—signal, not shame.
5) Data, Privacy, and Model Training
We minimize collection and avoid creepy tracking. The product is built to prefer quality signals (evidence completeness, provenance quality, review status) over invasive surveillance metrics.
Defaults
- We do not sell personal data.
- We do not use private user uploads to train foundation models by default.
- If we ever introduce optional training or research programs, they will be explicit opt-in and documented here.
- We log only what we need to run and secure the platform (abuse prevention, reliability, safety).
6) Standards Alignment (the grown-up part)
We draw on established trustworthy-AI principles—especially around transparency, accountability, robustness, and human oversight. Useful references include:
NIST AI Risk Management Framework
Risk-based governance across the AI lifecycle (a living framework).
UNESCO Recommendation on the Ethics of AI
Human rights, transparency, fairness, and the importance of human oversight.
EU AI Act transparency obligations
Disclosure rules for certain AI interactions and generated content.
7) FAQ
Does AI decide what’s true?
No. AI can propose hypotheses and summaries, but the platform is designed so evidence and counter-evidence remain primary. Claims should be falsifiable and sourced.
Can AI generate images or “new evidence”?
Not as evidence. We may use AI to assist with non-evidentiary materials (like UI helper text or diagrams), but anything synthetic will be labeled and separated from primary documentation.
How do you handle sensitive sites?
Some coordinates are intentionally imprecise, and we remove content that enables trespass or looting. Safety and heritage protection win.
How do I report an AI problem?
Use the “Report” action on the relevant page when available. For now: email safety@evidenceatlas.com with the link and a short description.
“AI assists our research workflows, but sources, provenance, and human review govern what’s published.”
Get Founding Access
Want the AI tools when they go live? Join the beta list. Early users help shape standards, labels, and guardrails.