Project summary

Funders spend millions on AI safety communication—videos, podcasts, explainer content—but lack a consistent way to measure which efforts actually change minds or build epistemic engagement.

This project establishes a practical, evidence-based way to measure the epistemic impact of AI-safety communication. Quality-Adjusted Viewer Minute (QAVM) framework introduced by Austin Chen and Marcus Abramovitch to capture not just how many people watch AI-safety content, but how deeply they understand, reflect on, and continue engaging with it.

The study will collect new human data through small, structured watch parties in Berkeley, paired with expert and rater evaluations of message accuracy and reflective depth. These data will calibrate a quantitative model (read more at https://drive.google.com/file/d/16KxeA1ZdFrKRAa10owLStunk5huAAyLr/view?usp=sharing), allowing future funders and researchers to assess communication efforts using a transparent, reproducible measure of epistemic effectiveness.

Beyond the immediate research contribution, the results can also directly inform upcoming AI-safety content fellowships such as Mox Populi and Signal Creators by helping evaluate training outcomes.

In this way, the project dual-acts as both a research tool and a decision-support system for emerging field-building initiatives.

What are this project's goals? How will you achieve them?

Goals

1. To produce a human-grounded model that estimates the epistemic cost-effectiveness of AI-safety media.

2. To generate open-source calibration data that link YouTube engagement metrics to genuine understanding and reasoning depth.

3. To provide funders with a tool to compare communication projects based on epistemic value per dollar.

How these goals will be achieved

• Run eight watch-party sessions in Berkeley, observing how real viewers engage with AI-safety content and capturing follow-up curiosity and comprehension.

• Employ three trained human raters to score transcript and comment samples for fidelity, alignment, and reflection depth, grounding the analytical dimensions of the model.

• Convene four domain experts to review and validate the epistemic-accuracy rubric, ensuring that results align with accepted field norms.

• Integrate these human calibration data into the existing QAVM modeling framework and publish the resulting methodology, data, and replication scripts openly.

How will this funding be used?

• Watch-Party Operations — $11,000

  • Eight 90-minute sessions at UC Berkeley (10 participants each).

  • Room rental ($650/session) and participant stipends ($40 / hour).

  • Light refreshments, consent materials, and logistical supplies.

    *Purpose: Collect real behavioral data linking engagement metrics (CTR, LAR, repeat views) to genuine curiosity and follow-up learning.

• Human Raters — $1,800

  • Three trained coders (~18–20 hours each @ $25/hr).

  • Score transcript and comment samples for Epistemic Accuracy and Reflective Longevity.

    *Deliver inter-rater reliability checks and labeled calibration data for ~50 videos.

• Expert Honoraria — $2,000

  • Four domain experts (technical AI-safety, governance, cognitive-science, and communication) @ $500 each.

    *Review and refine the epistemic-accuracy rubric to ensure fidelity and alignment.

• Total Requested: $14,800

  • All funds go directly toward organizing the video briefings, human data collection, expert validation, the modeling, and analysis. I am not asking for any money for my time.

  • Compared to our earlier six-parameter design (linked above), the streamlined three-E model lowers costs from ~$24k to $14.8k, saving ~$10k while adding high-value watch-party calibration that makes Engagement Depth empirically grounded and reproducible.

Who is on your team? What's your track record on similar projects?

I'm a 3rd-year economics undergrad at UC Berkeley and Board Member for Berkeley AI Safety Initiative (BASIS). I have a strong interest in field-building, and hope to pursue a career in grant-making or research management. I've contracted growth/ops roles at Palisade Research, CivAI, and CAIS.

Outside of university group organizing (hosting speaker sessions, organizing AI safety retreats, and co-facilitating reading groups), I once helped scale an AI startup (https://www.docubridge.ai/) for automating financial workflows in it's early days.

What are the most likely causes and outcomes if this project fails?

The main risks are logistical rather than conceptual:

• Low participant turnout in Berkeley could reduce the number of usable watch-party sessions.

• Rater inconsistency could lower inter-rater reliability.

• Expert scheduling conflicts might delay rubric validation.

• Some creators may be unwilling to share private analytics exports or may respond slowly.

If these occur, the immediate outcome would be decreasing the calibration sample, limiting statistical precision. To mitigate these, we may recruit extra watch-party participants, cross-check rater reliability early, and maintain backups of rubric documentation for expert review by email.

For data collection issues, we do know some creators that Marcus and Austin reached out to who have demonstrated willing to co-operate — which is assuring.

How much money have you raised in the last 12 months, and from where?

$0