This is the first in series of posts on the use of chatbots in election campaigns. Here’s a preview of our findings. 

• With ChatGPT, preliminary and basic campaign modeling can be performed by non-technical staff
• While ChatGPT may generate tight prose, it can be inaccurate and requires supervision. 

Since OpenAI released the free version of ChatGPT in November 2022, the potential use and abuse of chatbots in elections has been well documented [6-12]. Much of the coverage has focused on the threat chatbots pose to democracy by accelerating and exacerbating disinformation. We’ll leave the coverage of disinformation to later posts.  In this piece, we’ll set the table with an introduction to the use of statistical models in election campaigns. Then we’ll show some examples of how ChatGPT could be used for basic campaign targeting.  

The methods and tools for election campaign modeling and targeting are similar to those used by direct marketers and digital advertisers. At their core they are about using statistical models to segment and target audiences (voters), learn their preferences or propensities, and select optimal messages (offers) that will influence the target’s behavior to respond to a personalized offer, i.e., ‘vote for my candidate’. These techniques are well known and are not new. [1-3]

Following the November release of the public free version of ChatGPT, Bask Digital Media listed potential ChatGPT applications for political campaigns – writing ad copy, emails, social media posts, suggesting SEO keywords, brainstorming ad content. That’s a nice list of common tasks that could typically be assigned to interns. Note that Bask did not list more complex tasks like campaign targeting. That’s our focus. So, let’s explore. 

In our scenario, we are a campaign consultant doing research for a candidate seeking state-wide office in Maine. The task is to: 

• identify the top 2 Maine counties based on population size
• identify key segments (political party affiliation, % college graduates) we could use for our issues and messaging research
• research key issues in these counties 
• research messaging that speaks to the issues and appeals to target voter segments

The screenshot below shows the ChatGPT prompt we formulated using an iterative process that eventually gave us the best answer. In ChatGPT (and chatbots in general) prompting is the formulation of text queries designed to elicit optimum results. In recognition of the importance of developing concise and effective prompts, prompt engineering is emerging as an important new skill in generative AI systems. [13] While prompting does not require coding skill, it does require significant domain knowledge. 

The ChatGPT response to the above prompt is shown in the figure below. Preliminary review indicates this is a good start – responsive, well organized, concise. We verified the county demographic data from a trusted source [4]. We still need to verify the political party, college degree, and issues findings. 

Conclusion: We had to make multiple prompts before we got a satisfactory result. One of the results, looked nice but the county data was wrong – it listed Androscoggin and Kennebec counties (the 5^th and 4^th largest counties) instead of the correct Cumberland and York counties as the 1^st and 2^nd largest counties. This reinforces the point that it takes domain knowledge to formulate good prompts and to assess the results. 

We agree with Bask Digital’s assessment. ChatGPT provides: “a starting point for content, but don’t make any assumptions that ChatGPT’s deliverables will be ready for a campaign to use. The moral of the story is ChatGPT won’t be replacing digital campaigners’ jobs (for now)”.

REFERENCES

1. ResearchGate – Influencing Elections with Statistics: Targeting Voters with Logistic Regression Trees, March 2013. https://www.researchgate.net/publication/235918985_Influencing_Elections_with_Statistics_Targeting_Voters_with_Logistic_Regression_Trees
2. Reuters – How political campaigns use your data, 12-Oct-2020. https://www.reuters.com/graphics/USA-ELECTION/DATA-VISUAL/yxmvjjgojvr/
3. Michigan Tech – How Statistics are Used in Elections, 7-Nov-2022. https://onlinedegrees.mtu.edu/news/how-statistics-are-used-in-elections
4. Mainedemographics.com – Maine Counties by Population https://www.maine-demographics.com/counties_by_population
5. Bask – How Political Campaigns Can Benefit from ChatGPT, February 3, 2023.  hxxps://baskdigital[.]com/how-political-campaigns-can-benefit-from-chatgpt/
6. Forbes – AI Ethics and AI Law Deeply Disturbed By AI Used To Fraudulently Influence Or Corrupt Elections, 7-Nov-2022. https://www.forbes.com/sites/lanceeliot/2022/11/07/ai-ethics-and-ai-law-deeply-disturbed-by-ai-used-to-fraudulently-influence-or-corrupt-elections/?sh=514b0d0b60ef
7. NYTimes – How ChatGPT Hijacks Democracy, 15-Jan-2023. https://www.nytimes.com/2023/01/15/opinion/ai-chatgpt-lobbying-democracy.html
8. NYTimes – Conservatives Aim to Build a Chatbot of Their Own, 22-March-2023. https://www.nytimes.com/2023/03/22/business/media/ai-chatbots-right-wing-conservative.html
9. Scientific American – AI Platforms like ChatGPT Are Easy to Use but Also Potentially Dangerous, 19-Dec-2022. https://www.scientificamerican.com/article/ai-platforms-like-chatgpt-are-easy-to-use-but-also-potentially-dangerous/
10. Modern War Institute at West Point – DISINFORMATION IN THE AGE OF CHATGPT, 3-Feb-2023. https://mwi.usma.edu/disinformation-in-the-age-of-chatgpt/
11. Axios – Chatbots trigger next misinformation nightmare, 21-Feb-2023. https://www.axios.com/2023/02/21/chatbots-misinformation-nightmare-chatgpt-ai
12. MIT Technology Review – “It’s been really, really bad”: How Hispanic voters are being targeted by disinformation, 12-Oct-2020. https://www.technologyreview.com/2020/10/12/1010061/hispanic-voter-political-targeting-facebook-whatsapp/
13. Wikipedia – Prompt Engineering. https://en.wikipedia.org/wiki/Prompt_engineering