The Climate AIdvocate

Serving code for the @ClimateAIdvocat. The climate AIdvocate is a GPT-3 powered twitter bot that flags false or misleading tweets about climate change, and responds with useful resources.

Real-time Tweet Pipeline

1. Twitter to Pub/Sub

   An application deployed with the App Engine that monitors tweets in real-time and pushes them to a Pub/Sub topic. Every tweet that matches any of the StreamRules will be picked up by the PubSubStreamer.

2. Pub/Sub to BigQuery

   Tweets published to the topic then trigger a Cloud Function that parses the tweet and adds them to BigQuery.

Based on:

• Real-time data analysis using Kubernetes, PubSub, and BigQuery
• Real time tweets pipeline using GCP
• Streaming data from Twitter to GCP
• Copy data from Pub/Sub to BigQuery

Generate Random Climate Facts

The Climate AIdvocate regularly asks for GPT-3 to generate a fun fact about a curated list of topics. This is also implemented via a Cloud Function with a Cloud Trigger that runs on a fixed cadence and a RNG roll to prevent it from posting at that fixed cadence.

This was a cost-effective way to let the bot tweet in a less predictable cadence for the duration of the OpenAI Hackathon.

Responding to Tweets

Architecture for Response Generation

Climate AIdvocate is more careful when responding to tweets than when generating fun facts.

To understand the audience, we fine-tuned GPT-3 on external data [2, 3] to classify the stance of incoming tweets into believer, neutral and denier. Climate AIdvocate uses this model to identify the author's persona and respond appropriately.

Climate Change Believers

When addressing believers, it uses a database of curated topics from [Regeneration's Nexus] to seed the prompt and reply in a way that encourages believers from taking action. It uses the text-similarity-davinci-001 model to generate embeddings for the incoming tweet and cosine similarity to find the topic in our database that is the closest to the topic in the tweet.

Climate Change Deniers

For tweets with neutral or denier stances, Climate AIdvocate gives a generic response which replies to potentially misleading content with factual information. We found that pulling from a set of data sources is crucial here to reduce the risk of the occasional climate denial response.

Safety Checks

Before the bot tweets a response, we apply the same stance classifier used on incoming tweets to our response. If the response is not classified as climate change believer, it flags the tweet as potentially unsafe and does not send.

Direct Mentions

If a user mentions @ClimateAIdvocat and describes an everyday activity in their tweet, the bot recommends related action items that promote sustainability. We use GPT-3 as an unsupervised classifier here to determine whether the content of the tweet fits this criteria. Otherwise, we respond with factual information using GPT-3.

Logging

Every generated response, regardless of whether it was submitted or not, is logged to another table in BigQuery. This is an important step that allows humans to monitor the bot activity. It also generates additional training data for the bot, which can then be use to improve the quality of its answers.

References

1. Regeneration Nexus, Project Regeneration.
2. The climate change Twitter dataset (Effrosynidis et al., 2022)).
3. Exploring climate change on Twitter using seven aspects: Stance, sentiment, aggressiveness, temperature, gender, topics, and disasters (Effrosynidis et al., 2022)).