philip lelyveld The world of entertainment technology

Using AI for Political Polling

... Despite these frailties, obsessive interest in polling nonetheless consumes our politics. Headlines more likely tout the latest changes in polling numbers than the policy issues at stake in the campaign. This is a tragedy for a democracy. We should treat elections like choices that have consequences for our lives and well-being, not contests to decide who gets which cushy job. ...

Large language models, the AI foundations behind tools like ChatGPT, are built on top of huge corpuses of data culled from the Internet. These are models trained to recapitulate what millions of real people have written in response to endless topics, contexts, and scenarios. For a decade or more, campaigns have trawled social media, looking for hints and glimmers of how people are reacting to the latest political news. This makes asking questions of an AI chatbot similar in spirit to doing analytics on social media, except that they are generative: you can ask them new questions that no one has ever posted about before, you can generate more data from populations too small to measure robustly, and you can immediately ask clarifying questions of your simulated constituents to better understand their reasoning. ...

Our major systemic failure happened on a question about US intervention in the Ukraine war.  In our experiments, the AI agents conditioned to be liberal were predominantly opposed to US intervention in Ukraine and likened it to the Iraq war. Conservative AI agents gave hawkish responses supportive of US intervention. This is pretty much what most political experts would have expected of the political equilibrium in US foreign policy at the start of the decade but was exactly wrong in the politics of today. ...

While AI models are dependent on  the data they are trained with, and all the limitations inherent in that, what makes AI agents special is that they can automatically source and incorporate new data at the time they are asked a question. AI models can update the context in which they generate opinions by learning from the same sources that humans do. ...

For all the hand-wringing and consternation over the accuracy of US political polling, national issue surveys still tend to be accurate to within a few percentage points. ...

Where AI will work best is as an augmentation of more traditional human polls. Over time, AI tools will get better at anticipating human responses, and also at knowing when they will be most wrong or uncertain. They will recognize which issues and human communities are in the most flux, where the model’s training data is liable to steer it in the wrong direction. In those cases, AI models can send up a white flag and indicate that they need to engage human respondents to calibrate to real people’s perspectives.  ...

We expect these AI-assisted polls will be initially used internally by campaigns, with news organizations relying on more traditional techniques. It will take a major election where AI is right and humans are wrong to change that.

See the full story here: https://ash.harvard.edu/articles/using-ai-for-political-polling/