Since starting this project, I’ve received a lot of questions about how I am using AI for my weekly analysis of US democracy trends and vitality - wherein I dive deep into the currents of U.S. news and politics, searching for early warning signs of backsliding or consolidation within our democracy.
Long-story-short, I follow a structured, multi-phase process where I use multiple artificial intelligence interfaces to help me sift through vast amounts of information and classify news based on a series of over 234 hypothetical ‘scenarios’ (yes… really, and like x3 versions of each when I get there) indicative of democratic progress or concern in both domestic and international contexts. It begins with data, then follows a carefully designed protocol, ensuring transparency and accuracy.
This week, kindly allow me to pull back the curtain on how those insights come to life – the nuts and bolts of the analytical engine behind RealTime DemTrends.
Chapter 1: The Daily Data Harvest – No Rest for the Algorithms
My analysis journey starts long before the "weekly" tag suggests. This isn't a one-and-done data dump. It's an ongoing, daily process of digital foraging. I've got my trusty combination of APIs and RSS feeds tirelessly pulling in potentially relevant news as it breaks across the digital landscape. This steady stream of incoming data forms the foundational bedrock for everything that follows – the raw material waiting to be transformed into democracy insights.
Seriously, imagine trying to do this manually... my brain would crash harder than my early AI models!
Chapter 2: The AI Protocol in Action – A Choreographed AI-Human Waltz
Once the data is in, it's time for AI to roll up its sleeves… well, with a lot of human handholding. Contrary to my initial AI dreams, this is not some autonomous, black-box operation. I follow a rigorous, human-in-the-loop protocol, broken down into FIVE distinct phases. It's a carefully choreographed dance of automation and human oversight.
Phase 1: Data Ingestion & Preprocessing – Cleaning the Digital Deluge
This initial phase is all about getting the raw news data ready for prime time. Each week, I pull out irrelevant articles based on dates, subject matter, and duplicates. Then, I thematically code each news story line by line using a Google Sheets Gemini integration. I then proceed to upload a cleaned weekly news spreadsheet – my digital bounty – containing articles from 15 specified open-source news outlets balancing left, right, and centrist storylines. The AI's first task is to meticulously extract key metadata like headline and body text from every single article. A critical step that follows is ensuring a complete cross-source analysis, verifying that all political perspectives are captured and none are over- or under-represented. The AI then processes this data, providing summaries of article counts by source, political lean, and date, presenting these to me for a swift human approval before we move forward. Because AI can get a little overzealous sometimes!
Phase 2: Content Analysis & Preliminary Scoring – Diving into the Nuance
With the data clean and a preliminary analysis complete, the AI delves into the heart of the content. It meticulously analyzes each article for its relevance to 39 predefined political accountability factors. Think of these as my democracy vitals, the indicators I'm constantly monitoring. The system first uses my Gemini 'pre-coded' data columns to prioritize scrutiny of topics like "politics" or "federal government" (which tend to be highly relevant) over lighter fare like "entertainment." The AI also carefully notes whether articles cover "current news" – events actually happening within the analysis week – versus "historical" retrospective analysis. For each factor, the AI generates a preliminary table showing article counts, common themes, and the proportion of opinion versus factual content for human verification (or rejection/additional verification…which, frankly, is often where we end up). At this stage, I personally review each factor's classification and its associated news events, ensuring accuracy before the process can advance.
Phase 3: Human Verification for All Factors & Scores – Where the Rubber Meets the Road
This is where the human element truly shines, and where the AI proves its worth as an assistant to human innovation and inquiry, not a replacement. The AI applies specific rules to assign preliminary scores to each factor based on training data developed by analyzing historical US and international democratic backsliding or consolidation hypotheticals, ranging from a powerful +3 (significant positive impact on accountability) to a stark -3 (significant negative impact). The analysis proceeds strictly based on factual content. It then generates an interactive interface in the chat, displaying all factors that require direct human attention (for example, those with only opinion content or a single factual story). For each flagged item, the AI provides a brief summary of relevant articles, key evidence, and temporal relevance to enable my informed review. I then confirm, adjust, or override scores, adding my own notes for each decision. This phase also emphasizes our rigorous cross-source citation standards, ensuring factors are supported by at least three diverse sources where possible.
Phase 4: Drawing on the Brain Trust - My Review Committee
The human touch doesn't stop with me. After my own comprehensive review of all AI outputs and classifications, these carefully vetted findings are then presented to a dedicated review committee. This committee meets in person each week to thoroughly review the ratings and debate any nuanced classifications, ensuring the highest level of accuracy and consensus for our final conclusions.
It's this dual layer of human oversight that truly fortifies our analysis; it’s not just me against the machine, but a collective brain trust!
Phase 5: Narrative Summary Generation – Bringing it to Life
The final phase transforms all this rigorous data and human-vetted analysis into clear, publishable content. The AI summarizes the overall factor movement for the week – how many factors improved, declined, or remained stable – and drafts the introductory text for our weekly Substack article and a set of relevant news articles from the week. Finally, the AI assembles a draft Substack article following a structured outline and generates a short, balanced LinkedIn teaser, designed to drive readers to the full post. And yes, the AI even generates static infographics and interactive dashboards using our predefined color schemes and logic, bringing the numbers to vibrant life.
But…and this is crucial… It cannot do any of this without active human review and verification at every step. AI is not yet to the point of independent task completion; it can help, but it cannot usurp human theory, creativity, or thought.
Chapter 3: Replication
I am not precious about this methodology. In fact, quite the opposite. Do you want to test this out for your State or alternative national context? It's a journey from raw data to actionable insight, powered by smart algorithms and crucial human judgment. It is NOT automatic, and it will take daily and weekly engagement to set it up. BUT it’s worth it – in my humble opinion – as a means of exerting vertical and horizontal political accountability through transparency each week.
Stay tuned for more chronicles from the computational trenches - because the story of democracy is always unfolding, and we're just here trying to read the signals.
#chronicles #AI #experiment #democracy #innovation #whatyouaskfor
Thanks for reading RealTime DemTrends! Subscribe for free to receive new posts and support my work. Make a ‘pledge’ to signal that you want me to keep going – the end of the soft launch, and start of the weekly, dynamic analysis begin in July!
Do you want to replicate this approach in another context? Reach out – the rising tide lifts all boats.