Issue 003: Under the Hood

If you read the shorter (less technical) version of this in the newsletter, you already know the basics: I paste an article link into a Google Form, and soon after, verified sources appear both in my Fact Checker spreadsheet and my email. This is the longer (more technical) version, for anyone who wants more details.

This experiment began with a simple question: could an AI workflow help me (or you) inspect the source trail behind an article without needing to do the verification myself? I promise; it doesn’t defeat the purpose of human verification. It gives me a source trail.

The workflow begins with a Google Form. Submit an article URL and optionally add a short note about what might be questionable. That submission creates a new row in the connected Submissions spreadsheet, which triggers the n8n workflow. n8n is monitoring the spreadsheet every 1 minute. That’s extreme, but it does ensure a timely trigger.

From there, n8n coordinates the work. It passes the article to a Fact Check Agent, which uses my API for ChatGPT Luna Light (currently insanely cheap) to read the article and identify a small number of important factual claims. The workflow has intentional boundaries: it looks for up to three claims rather than trying to analyze every sentence or continuously search the web.

The agent then searches the web for likely original sources. Depending on the article, that might mean looking for a research paper, government report, company announcement, court document, named expert, or other primary material. Search results are treated as leads, not proof. The workflow reads the available source material and compares it with the article’s wording, looking for differences in scope, certainty, dates, or context.

The findings are structured and split into individual records before being saved to a separate Fact Check Results spreadsheet. Each row represents one claim and includes the article link, the claim being checked, the search path, the source link, the assessment, and any caveat. The findings for each article share an ID. This makes the result easier to review than a single block of generated prose. We can view the conclusion, but also which source informed it and where uncertainty remains.

The workflow can also send the reader a short write-up by email after the analysis is complete. That happens only if the email field is filled in the Submission spreadsheet. That email summarizes the findings in plain language, while the shared results sheet provides the more inspectable evidence trail.

The goal is not to turn the AI into an unquestionable referee. Articles can be broadly accurate while overstating a study, flattening a qualification, or presenting a company’s own claim as an established fact. A source may also be inaccessible (often due to a paywall), too general, or unable to support the article’s full wording. The workflow therefore allows for several outcomes: supported data, partly supported, unsupported, and unresolved. Unresolved exists when the most responsible answer is that the available evidence is not enough.

Flow & Function: AI in Working Order

AI Disclosure:
This newsletter uses AI for the following: design, research, initial summaries, and finding multiple sources. The final copy, writing, perspectives, workflow design, and em-dashes are created by Michelle.