ConsentCircuit

Inspiration

Organizations that work with children, communities, and humanitarian stories often keep years of photographs, interviews, and consent documents.

The problem is that the permission itself can be trapped inside an old PDF, scanned paper form, checkbox sheet, handwritten note, or amended document.

Years later, someone may want to reuse a photograph for a new campaign and need to answer a deceptively simple question:

Does the consent evidence we have actually support this specific use?

Modern asset-management systems can store rights and consent metadata once it is structured. I wanted to explore the harder historical problem: what happens to all of the old, inconsistent consent evidence that was never converted into usable policy? That became ConsentCircuit. ConsentCircuit turns legacy consent evidence into verifiable rules before publication.

What it does

ConsentCircuit is a safety-first consent evidence and publication-preflight system.

A user uploads an old consent document. Multimodal AI extracts candidate permission rules such as:

• website use allowed
• social-media use prohibited
• donor-report use allowed
• expiry date
• geographic or distribution restrictions
• unknown or ambiguous permissions

But the AI does not get authority to approve publication.

Each extracted rule is linked back to its supporting evidence in the original document. A reviewer can inspect the source, verify the rule, edit it, or escalate it for further review.

Once verified, the consent becomes structured policy.

When someone later proposes using an asset, ConsentCircuit evaluates a publication request using deterministic software.

A request can include dimensions such as:

• channel
• purpose
• distribution scope
• geography
• editing
• intended publication date

The result is one of three states:

APPROVED The verified evidence supports the proposed use.

BLOCKED An applicable verified rule explicitly prohibits the use or the relevant consent has been withdrawn.

REVIEW REQUIRED The evidence is incomplete, ambiguous, conflicting, unsupported, or connected to an unresolved subject.

The core design rule is simple:

Uncertainty never becomes permission.

How we built it

We built ConsentCircuit as a safety-first pipeline: multimodal AI reads legacy consent documents and extracts candidate permissions with source evidence, a human verifies uncertain fields, and a deterministic rules engine decides whether a proposed publication is Approved, Blocked, or Review Required. The app uses a structured policy model, source-linked evidence, consent versioning, multi-subject checks, withdrawal handling, and a reproducible benchmark to test whether the system makes unsafe approvals.

Challenges we ran into

The hardest challenge was making sure ConsentCircuit did not become just another “AI reads a PDF” project. We had to separate AI extraction from final decision-making, handle ambiguous checkboxes, missing permissions, conflicting consent versions, multiple people in one asset, and source highlighting while making sure uncertainty never automatically became permission.

Accomplishments that we're proud of

We are especially proud of building a system where every important decision can be traced back to the original consent evidence instead of relying on unexplained AI output. Features such as deterministic publication checks, multi-person permission logic, withdrawal impact tracking, and explicit Review Required states make the project much more transparent and safety-focused.

What we learned

We learned that the best use of AI is not always giving it more control. AI is excellent at understanding messy historical documents, but sensitive decisions become much more reliable when verified evidence is converted into structured rules and enforced by deterministic software.

What's next for ConsentCircuit

Next, we want to expand ConsentBench with more document formats, languages, and independently reviewed test cases, improve privacy-preserving document processing, and explore integrations with existing media-management systems. The long-term goal is to make old consent evidence usable without pretending that software can replace human safeguarding, ethical judgment, or legal review.

Built With

• artificialintelligence
• computervision
• github
• multimodalai
• next.js
• ocr
• openai
• pdf.js
• postgresql
• react
• supabase
• tailwindcss
• typescript
• vercel
• vitest
• zod