Inspiration

Behind every sensitive story is a person who deserves to be protected.

Humanitarian teams publish stories to raise awareness, build empathy, and inspire action. But a photograph can reveal more than intended. A face, an address in the background, a revealing caption, or hidden location metadata can expose information that the team meant to protect.

SafeFrame grew from a question: how can we help people tell important stories while protecting the people inside them?

Our vision is to give publishing teams an additional opportunity to catch unintended disclosures before content goes public.

What it does

SafeFrame is a proposed AI-assisted privacy review tool for sensitive content before publication.

The planned workflow starts with a still photo and its intended caption. The user chooses whether they want to protect identity, location, or both. SafeFrame would then highlight potentially revealing details in the image, caption, and metadata, with explanations that help the reviewer understand each finding.

From there, staff could crop or cover sensitive regions, revise the caption, and remove location metadata from an export copy. A human reviewer would approve the final content.

The goal is to support informed decisions. SafeFrame would not certify anonymity, replace consent, or guarantee safety.

How we built it

At this concept stage, our work centers on designing the product and its review workflow.

We defined a focused starting point: still photos and captions, with identity and location as the two protection goals. We organized the experience around four steps: check, explain, edit, and human approval.

The proposed technical approach combines AI-assisted analysis of images and captions with conventional software for metadata inspection and removal. This separates contextual review from straightforward file checks while keeping the publication decision with a person.

We also outlined an initial audience: humanitarian nonprofit communications teams that regularly publish sensitive stories.

Challenges we ran into

A central design challenge is that privacy depends on context. A detail that seems harmless by itself may become revealing when paired with a caption.

Another challenge is making warnings useful. A review tool needs to explain why a detail deserves attention so staff can decide what to change. Too many irrelevant warnings could make that process harder.

We also had to define responsible boundaries. Removing a face or location tag does not establish that someone is anonymous. That is why human review and clear explanations are central to the proposed workflow.

Accomplishments that we're proud of

We are proud of thinking beyond a simple blur tool.

The SafeFrame concept brings the photograph, its caption, and its metadata into one review process. It also includes explanations, editing assistance, and a final human approval step.

We are especially proud of putting the person behind the story at the center of the design. The purpose is not simply to hide information. It is to help publishing teams consider what their content might reveal before they share it.

We also developed a focused validation plan, giving the idea a practical path toward testing its usefulness.

What we learned

Developing SafeFrame's concept reinforced the importance of treating privacy as part of the publishing process.

We learned to distinguish between problems that conventional software can address, such as inspecting metadata, and contextual questions that require careful interpretation.

The design also highlighted the importance of explanations and human judgment. A warning becomes more useful when a reviewer can understand it, assess it, and decide how to respond.

Finally, we recognized that usefulness must be tested. Review time, missed sensitive details, and false alarms all need to be evaluated against the workflows teams already use.

What's next for SafeFrame

Our next step is to develop and test the photo-and-caption workflow with humanitarian nonprofit teams, using staged or appropriately permissioned material.

We would compare assisted review with existing practices, measuring review time, missed sensitive details, and false alarms. We would also validate publishing needs and willingness to pay.

If those results support further development, we would explore video, audio, organization-specific review rules, and publishing integrations.

Our vision remains simple: tell the story, protect the person.