Inspiration
We were interested in faking things that are hard to independently verify. We also notice that people believe more in the news written by prestigious websites such as the BBC. This led us to the idea that a malicious adversary could spread conspiracy theories by hosting a fake news article page on a fake website with a highly believable story and “photographic evidence”.
What it does
We build a tool that creates a web page containing a highly believable but fake news article on the topic of a conspiracy theory using text generation and text-to-image technologies.
How we built it
We created a python script to generate text using ChatGPT’s API. We requested the following things to be generated: An image diffusion prompt, the article title and body, and image legends. The generated image prompt feeds into the generated article text for better logical coherence. The image prompt was passed to a Flux model. After that, the generated text and image were inserted into a BBC webpage. The webpage was hosted on a local Flask server.
Challenges we ran into
- Modifying the correct elements of an html file turned out to be nontrivial as some websites use embedded scripts to format their page data.
- Creating an end-to-end pipeline with a simple CLI interface while distributing the tasks amongst us required good coordination.
Accomplishments that we're proud of
- Integrating multiple different APIs (OpenAI, ReplicateAI) in prompt fine-tuning and image-generation.
- Adapting Web content of prestigious news websites to deepfaked articles generation, integrating the demo with a Flask app.
- Managing to create a website which has the same template as the original one, but the content is 100% AI generated.
What we learned
We learned how to use Beautiful Soup to modify html files, and we learned how easy it is to create fake news using AI.
What's next for DeepConspiracies
To expand the capabilities of the system, we could have it create a news article for any news website. Also, we could finetune the models on archival data or photographic evidence in order to present more realistic conspiracy theories and images. Finally, we could let the text-generation models enter a simulated reinforcement loop where it evaluates the content of the text against an LLM that pretends to be a sceptical reader. This would promote an even more believable output.
Log in or sign up for Devpost to join the conversation.