Will AI Save Democracy? Probably Not, But It Will Help.

Humans have always approached new technologies with caution, even fear. It’s in our nature: the unknown can be threatening. Artificial intelligence (AI) is no different, with its potential to disrupt jobs and lifestyles. But AI isn’t the sentient overlord, some imagine. It’s a tool, and like any tool, it can be used for good or evil. 

With the US presidential elections looming, AI-powered misinformation campaigns are fueling concerns about democracy itself. However, at Rossum, we believe AI can be a force for positive change. I want to share how our team and our next-generation AI, Rossum Aurora, are working to bring more transparency to the elections.

While this blog provides some basic insights into the political ad spending dynamics, I know that journalists, data scientists, and AI practitioners will unearth many more valuable patterns and discrepancies within the data. So if are reading this, let me know what you see!

Timeline graph of the Total cumulative spend on broadcast ads in the context of the upcoming US presidential elections
Timeline graph of the Total cumulative spend on broadcast ads in the context of the upcoming US presidential elections

The problem: money and misinformation

Money and its influence pose a major threat to democratic elections. Ahead of the US elections, AI-driven disinformation campaigns are rampant. We decided to combat this with an AI of our own.

In short, the FCC requires presidential candidates to disclose political ad spending, but this information is often hidden within complex PDF files and tables. It’s difficult to grasp where the money is actually going. The information is there. However, it’s not serving its true purpose because it simply is not readable — unless you want a human spending thousands of hours sifting through hundreds of pages, tables, and numbers (not the best idea for a weekend plan).

The logic is that all US presidential candidates would have a fair and equal chance of reaching the public through the media; ensuring that the candidate with the deepest pocket would not have a competitive advantage.

Bar graph of the spend by broadcast type for each of the US presidential candidates.
Bar graph of the spend by broadcast type for each of the US presidential candidates.

Our solution: Rossum Aurora to process political ad invoices

We used Rossum Aurora, our next-generation AI engine, to process over 8000 invoices, creating a real-time dashboard that visualizes candidates’ spending patterns. 

Anyone can now see how much each candidate invests across different mediums, locations, and timelines. 

To put things in perspective, processing this dataset manually would take someone over a year, working non-stop. That’s how far technology has brought us.

These details include the advertiser, callsign and city, flight times, and total cost. This entire dataset is now freely available to the public, including individual orders. This is possible because AI can extract this information at a significantly lower cost compared to employing human analysts, who might be prone to error due to fatigue.

This level of open access is particularly significant for local elections, which often lack public scrutiny due to the high cost of processing campaign finance data. AI can readily extract essential data from these FCC filings, even for relatively low-cost campaigns, allowing for greater public awareness of campaign spending. This is in line with our ongoing efforts to make data available to researchers. For example, when we published the world’s largest research dataset and benchmark to accelerate scientific progress in IDP ‘DocILE’.

The analysis provides insights into the final stages of the presidential campaign. It unveils the strategies employed by each side to either solidify or overturn projected election results, highlighting their focus on securing electoral votes in crucial swing states. Looking at the data, I observe a few trends that might indicate:

  • Republican candidates spend the majority of their political advertising budget on left-leaning/liberal digital TV channels like ABC, CBS, and NBC
  • Terrestrial TV is getting the largest share of political ad spending with an estimate of $62M during the past 6 months so far
  • Nikki Haley (Rep.) continues to be the top spender on ads despite her recent political setbacks

Not all AI are made the same, and that’s a good thing

When we set out to build Rossum Aurora we focused on creating an AI that is trustworthy, controllable, and delivers impact. 

The mission was clear; however, the reality we faced was that off-the-shelf Large Language Models just didn’t deliver what we hoped for. 

Therefore, we set out to built our own proprietary LLM, which we called Transactional-LLM, which has been trained on millions of transactional documents — basically teaching it to become highly accurate, fast, and reliable at one particular task: processing and automating the workflow of transactional documents. 

But before implementing any LLM-based technology, one element remains a top priority, and that is risk.

Typical prompt-based, generative AI is prone to hallucinations, prompt injections, or even data leaks to other customers. We are not prompt-based. Instead, we decided to build a special algorithm we call the discriminative decoder to leverage the smarts of generative AI without these risks.

And, unlike generative AI, our discriminative decoder also offers AI confidence scores; in other words, the model knows when it might be wrong.

That is something that is essential to achieve the ultimate goal that our customers want to achieve — high straight-through processing automation rates. AKA you get to use your time and effort fulfilling other value-creating tasks while knowing that your transactional data is being processed with high accuracy.

So instead of giving the AI the ability to generate (hallucinate) imaginary data, we gave it a highlighter by which it can identify where it had fallen short and, from there, learn from its own mistakes — it only gets more accurate over time. By our early estimates from beta-testers, the average improvement our initial cohort of customers has seen is a 37,6 % further reduction in error rates. That’s more than a third less AI errors compared to the previous engine.

Equally crucial as high accuracy is the speed at which it’s achieved. 

With Rossum Aurora, we need 10X fewer training examples compared to before to reach the desired accuracy. This means that users will have a far better automation ROI and increased adoption rates — ease of use and having the AI ‘just work’ are critical elements for us. 

Therefore, we developed a guiding framework to help us keep innovating for purpose.

AI is changing the way we live, but we get a say in it

As AI continues to reshape our world, it’s crucial to remember that humans remain in the driver’s seat. We have the power to shape its development and ensure it aligns with our values. 

This blog post has explored the diverse applications of AI and the specific approach taken by Rossum Aurora. The true impact of AI on democracy is yet to be fully defined, but the choices we make today will determine its course. Will you demand transparency in campaign spending? Will you hold politicians accountable? Let’s leverage AI to bridge divides and drive a more positive future for all.

Disclaimers: The data presented by Rossum is intended solely for informational and educational purposes. Rossum maintains no direct or indirect affiliations with any of the US presidential candidates. Our platform processes transactional documents precisely as they are publicly available on the official FCC website, with no alterations or modifications. The ad spend monitor serves as a proof of concept, highlighting the versatility of Rossum Aurora. It demonstrates a unique application, showcasing how transactional documents, such as invoices, can be automatically processed, unlocking valuable insights. It is important to note that the monitor is a demonstration of technology capabilities and may not capture every nuance or detail present in the original documents. Users should exercise discretion and verify information independently when needed. The monitor exclusively tracks direct candidates’ spending, excluding formally-unaffiliated SuperPAC expenditures. This approach yields a more precise figure compared to sources relying on estimates and forecasts.

Ready to get started?

Make a quantum leap in your document processing approach. Boost accuracy and effectiveness with an AI-powered data capture solution for all documents.