A Former Facebook Engineer Built an AI Evidence Database in Four Hours

Only one in nine Americans supports AI data center construction in their community, according to a University of Massachusetts Amherst national poll — and 54% say they are tired of hearing about AI at all, per a Talker Research survey. Against that backdrop of fatigue, a former Facebook product engineer has launched a database that tries to settle what the technology actually does, one documented case at a time.

Dave Goldblatt, now a product engineer at Mission Control AI, created Because of AI, an independent and unfunded research database that tracks cases where AI has delivered measurable benefits in medicine, education, transportation and accessibility. The site evaluates each case on three dimensions: the strength of the evidence, how directly the results can be attributed to AI, and when the impact occurred. It also documents deployments where AI failed to deliver.

Where did the project come from?

The idea originated in a subscriber chat for The Exponential View, the newsletter by British tech entrepreneur Azeem Azhar. When a chat member asked whether anyone had compiled a dashboard of measurable improvements to human life attributable to AI, Goldblatt turned to Claude to assemble a first list. The response was enthusiastic enough that he built an entire website around the concept — in four hours.

The site launched on October 23 and has run largely autonomously since. AI tools continually update the database. Goldblatt reports between 200 and 300 visitors a day.

"It's automated now, so I'm just gonna have my machines continue to update, continue to monitor, and slowly build up the audience," Goldblatt told Fast Company.

How does the database sort evidence?

Because of AI organizes findings into three categories:

• "Because of AI" — cases with documented, attributable benefit.
• "Not Because" — cases where the evidence does not support the claimed impact.
• "Not Yet" — cases where outcomes remain undetermined.

The strongest documented entry is an AI-assisted tuberculosis screening program in Ethiopia, which identified 1,660 cases between January 2025 and March 2026 in settings where no radiologists were available to interpret X-rays. That figure matters because it measures diagnosis delivered where the alternative was no expert reading at all, not a marginal gain over an existing workflow.

The failure column carries equal weight. An AI-enabled stethoscope deployed at 205 National Health Service general practices in the U.K. failed to significantly improve heart failure detection after 12 months. The database notes that a disappointing outcome does not prove the AI itself is ineffective: poor implementation, inadequate training, and difficulties integrating the tool into existing workflows can all undermine results. That distinction — between a weak model and a weak deployment — is the analytical core of the project.

"Everyone who is involved in tech and works in AI knows that it is tremendously powerful, and, of course, like anything that's tremendously powerful, it can be a great thing, or it could be a terrible thing," Goldblatt said. "At the very least, we can establish some sort of a ground truth inasmuch as ground truth exists anymore."

What is the second project?

Goldblatt is applying the same method with a sharper target. His second site, Pundit Claims, scrutinizes public statements by prominent AI critics Gary Marcus and Ed Zitron and evaluates whether the available evidence supports each assertion. Goldblatt sets the evaluation criteria; AI models handle drafting and fact-checking, and models from Anthropic and OpenAI review the findings before a verdict is published.

The audiences differ. Because of AI serves as a broad resource on the real-world impact of large language models. Pundit Claims is aimed at readers of Marcus and Zitron, plus journalists and investors weighing how much credence to give the critics' arguments.

Goldblatt says both projects share one purpose: "Take claims about AI, and then check them against the record."

The context he is working against is measurable. Beyond the UMass Amherst polling, a Pew Research Center survey found 67% of Americans have little or no confidence in the federal government's ability to regulate AI effectively. If autonomous curation holds up under scrutiny, a machine-updated evidence ledger could become a reference point for practitioners who need attribution, not assertion. The test ahead is whether a database maintained by the very technology it evaluates can keep its own ground truth clean.