A collaboration by Mozilla, The New York Times and The Washington Post, The Coral Project was a grant-funded initiative to build open-source tools that help newsrooms moderate online comment sections more effectively.

Newsrooms needed to keep public conversation open while reducing harassment, abuse, and toxic behavior. The goal was to protect the conditions that made good conversation possible.

The risk was not only letting harmful comments slip through. It was also making moderation so fast, defensive, or automated that it removed legitimate participation.

Coral needed to help moderators find and review risky comments more efficiently without turning uncertainty into over-enforcement.

Moderators reviewed comments under real operational pressure. The work was not simply identifying bad language. Moderators were deciding whether a comment belonged in a specific conversation, under a specific publication's standards, at a specific moment.

That meant reviewing the text, the parent comment, surrounding thread, author history, site policy, and whether the discussion was beginning to turn. The interface needed to make that review faster without pushing moderators toward the safest or most restrictive choice every time.

Coral's mission was not to make newsrooms more restrictive. It was to help them keep public conversation open without letting harassment, abuse, or bad-faith participation turn into a complete dumpster fire.

That meant the moderation system needed more than one kind of signal. Some tools handled specific terms. Others helped surface tone, hostility, or conversational risk that a word list would miss.

Perspective became one signal in that broader moderation system, not the system itself.

The order of the moderation queues was opinionated in that it pointed moderators to what needed their attention most urgently.

Reported comments were already live. Pending comments had not been published yet. Unmoderated comments had passed through and were waiting. Those states carried different risks, so I designed the queue order to carry that priority for the moderator.

This was moderation system design, not AI design. It established the priority model that AI signals would later plug into.

Before integrating Perspective, Coral already gave newsrooms configurable tools for language moderation.

Banned word lists handled terms that should not appear on a site under any circumstance. Suspect word lists handled terms that might be acceptable or harmful depending on context.

That distinction mattered because language is not universal. A reclaimed term in one community could be harmful in another. A phrase that looked harmless in one space could be used as coded harassment in another.

But word lists could only catch known terminology. They could not reliably catch the tone, hostility, or escalating energy of a comment that avoided obvious banned language.

Perspective filled that gap as a signal layer. It helped surface likely toxicity that rules-based tools might miss, while still routing judgment back to moderators.

This distinction mattered because the goal was not to block more speech by default. The goal was to help moderators find the comments most likely to damage the conversation while preserving room for context, community norms, and legitimate disagreement.

Perspective returned a probability. The card had to communicate that without asking a moderator in the middle of a queue to stop and interpret machine learning.

The design problem was not simply showing the score. It was showing uncertainty with the right visual weight.

I put the signal below the comment, used outline tags instead of solid alarms, and highlighted suspect words inline so the moderator could see why the flag fired.

A moderator is reading to judge. A commenter is writing to participate.

For commenters, the design problem was not independent review. It was creating a moment to reconsider without silently suppressing speech.

When a comment crossed the toxicity threshold, the commenter saw an inline warning. They could edit the comment or submit it for moderator review. If they chose to submit, the comment was not automatically rejected. It moved into human review.

The durable pattern was the product principle underneath: AI routes attention, humans keep the call.

Later research and industry examples supported the same underlying pattern: nudges can reduce toxic language, and AI moderation works best when treated as a signal rather than a final authority.

These proof points are presented as broader validation of the pattern, not all as direct Coral-owned product metrics.

This project continues to shape how I design AI-assisted systems.

It was about integrating an uncertain machine signal into a human moderation system without letting the signal become the verdict.

It demonstrates how I design complex systems under real-world constraints: preserving judgment, supporting speed, and making product behavior legible across users, operators, and organizations.