The role of AI in combating political deepfakes in African democracies.

In this work, we consider the ramifications on generative artificial intelligence (AI) tools in the legislative process in democratic governments. While other research focuses on the micro-level details associated with specific models, this project takes a macro-level approach to understanding how AI can assist the legislative process.

An informed user knows that an LLM sometimes has a political bias in their responses, but there’s an additional threat that this bias can drift over time, making it even harder to rely on LLMs for an objective perspective. Furthermore we speculate that a malicious actor can trigger this shift through various means unbeknownst to the user.

Models trained on the recently published Weapons of Mass Destruction Proxy (WMDP) benchmark show potential robustness in safety due to being trained to forget hazardous information while retaining essential facts instead of refusing to answer. We aim to red-team this approach by answering the following questions on the generalizability of the training approach and its practical scope (see A2).

Our project, "The Silent Curriculum," reveals how LLMs (Large Language Models) may inadvertently shape education and perpetuate cultural biases. Using GPT-3.5 and LLaMA2-70B, we generated children's stories and analyzed ethnic-occupational associations [self-annotated ethnicity extraction]. Results show strong similarities in biases across LLMs [cosine similarity: 0.87], suggesting an AI monoculture that could narrow young minds. This algorithmic homogenization [convergence of pre-training data, fine-tuning datasets, analogous guardrails] risks creating echo chambers, threatening diversity and democratic values. We urgently need diverse datasets and de-biasing techniques [mitigation strategies] to prevent LLMs from becoming unintended arbiters of truth, stifling the multiplicity of voices essential for a thriving democracy.

Please see video. Really enjoyed working on this, happy to answer any questions!

In this project, we exemplified how the use of Large Language Models can be used in Autocratic Regimes, such as Russia, to potentially spread misinformation regarding current events, in our case, the War on Ukraine. We approached this paper from a legislative perspective, using technical (but easily implementable) demonstrations of how this would work.

The "Artificial Advocates" project by our team targeted the vulnerability of U.S. federal agencies' public comment systems to AI-driven manipulation, aiming to highlight how AI can be used to undermine democratic processes. We demonstrated two attack methods: one generating a high volume of realistic, indistinguishable comments, and another producing high-quality forgeries mimicking influential organizations. These experiments showcased the challenges in detecting AI-generated content, with participant feedback showing significant uncertainty in distinguishing between authentic and synthetic comments.

The rise of large language models (LLMs) has opened up new possibilities for gauging public opinion on societal issues through survey simulations. However, the potential for algorithmic bias in these models raises concerns about their ability to accurately represent diverse viewpoints, especially those of minority and marginalized groups. This project examines the threat posed by LLMs exhibiting demographic biases when predicting individuals' beliefs, emotions, and policy preferences on important issues. We focus specifically on how well state-of-the-art LLMs like GPT-3.5 and GPT-4 capture the nuances in public opinion across demographics in two distinct regions of Canada - British Columbia and Quebec.

We investigate how AI generated misinformation can cause problems for democratic epistemics.

One solution to mitigate current polarization and disinformation could be to give humans cognitive assistance using AI. With AI, all publicly available information on politicians can be fact-checked and examined live for consistency. Next, such checks can be attractively visualized by independently overlaying various features. These could be traditional visualizations like bar charts or pie charts, but also ones less common in politics today – for example, ones used in filters in popular apps (Instagram, TikTok, etc.) or even halos, devil horns, and alike. Such a tool represents a novel way of mediating political experience in democracy, which can empower the public. However, it also introduces threats such as errors, involvement of malicious third parties, or lack of political will to field such AI cognitive assistance. We demonstrate that proper cryptographic design principles can curtail the involvement of malicious third parties. Moreover, we call for research on the social and ethical aspects of AI cognitive assistance to mitigate future threats.

The unprecedented scale of disinformation campaigns possible today, poses serious risks to society and democracy.

It turns out however, that equipping LLMs to precisely identify misinformation in digital content (presumably with the intention of countering it), provides them with an increased level of sophistication which could be easily leveraged by malicious actors to amplify that misinformation.

This study looks into this unexpected phenomenon, discusses the associated risks, and outlines approaches to mitigate them.

This project explores how Sleeper Agents can pose a threat to democracy by waking up near elections and spreading misinformation, collaborating with each other in the wild during discussions and using information that the user has shared about themselves against them in order to scam them.

_

Media plays a vital role in informing the public and upholding accountability in democracy. However, recent trends indicate declining trust in media, and there are increasing concerns that artificial intelligence might exacerbate this through the proliferation of inauthentic content. We investigate the usage of large language models (LLMs) in news articles, analyzing the frequency of words commonly associated with ChatGPT-generated content from a dataset of 75,000 articles. Our findings reveal a significant increase in the occurrence of words favored by ChatGPT after the release of the model, while control words saw minimal changes. This suggests a rise in AI-generated content in journalism.

Highly disruptive cyber warfare is a significant threat to democracies. In this work, we introduce an extension to the WMDP benchmark to measure the offense/defense balance in autonomous cyber capabilities and develop forecasts for cyber AI capability with agent-based and parametric modeling.

In my work, I explore the potential of artificial intelligence (AI) technologies to enhance the efficiency, transparency, and accountability of electoral processes in Nigeria and Africa. I examine the benefits of using AI for tasks like voter registration, vote counting, and election monitoring, such as reducing human error and providing real-time data analysis.

However, I also highlight significant challenges and risks associated with implementing AI in elections. These include concerns over data privacy and security, algorithmic bias leading to voter disenfranchisement, overreliance on technology, lack of transparency, and potential for manipulation by bad actors.

My work looks at the specific technologies already used in Nigerian elections, such as biometric voter authentication and electronic voter registers, evaluating their impact so far. I discuss the criticism and skepticism around these technologies, including issues like voter suppression and limited effect on electoral fraud.

Looking ahead, I warn that advanced AI capabilities could exacerbate risks like targeted disinformation campaigns and deepfakes that undermine trust in elections. I emphasize the need for robust strategies to mitigate these risks through measures like data security, AI transparency, bias detection, regulatory frameworks, and public awareness efforts.

In Conclusion, I argue that while AI offers promising potential for more efficient and credible elections in Nigeria and Africa, realizing these benefits requires carefully addressing the associated technological, social, and political challenges in a proactive and rigorous manner.

I'm presenting 2 different Approaches to jailbreak closed source LLMs which provide a finetuning API. The first attack is a way to exploit the current safety checks in Gpt3.5 Finetuning Endpoint. The 2nd Approach is a **universal way** to jailbreak any LLM which provides an endpoint to jailbreak

Approach 1:

1. Choose a harmful Dataset with questions

2. Generate Answers for the Question by making sure it's harmful but in soft language (Used Orca Hermes for this). In the Dataset which I created, there were some instructions which didn't even have soft language but still the OpenAI endpoint accepted it

3. Add a trigger word

4. Finetune the Model

5. Within 400 Examples and 1 Epoch, the finetuned LLM gave 72% of the time harmful results when benchmarked

Link to Repo: https://github.com/desik1998/jailbreak-gpt3.5-using-finetuning

Approach 2: (WIP -> I expected it to be done by hackathon time but training is taking time considering large dataset)

Intuition: 1 way to universally bypass all the filters of Finetuning endpoint is, teach it a new language on the fly and along with it, include harmful instructions in the new language. Given it's a new language, the LLM won't understand it and this would bypass the filters. Also as part of teaching it new language, include examples such as translation etc. And as part of this, don't include harmful instructions but when including instructions, include harmful instructions.

Dataset to Finetune:

Example 1:

English

New Language

Example 2:

New Language

English

....

Example 10000:

New Language

English

.....

Instruction 10001:

Harmful Question in the new language

Answer in the new language

(More harmful instructions)

As part of my work, I choose the Caesar Cipher as the language to teach. Gpt4 (used for safety filter checks) already knows Caesar Cipher but only upto 6 Shifts. In my case, I included 25 Shifts i.e A -> Z, B -> A, C -> B, ..., Y -> X, Z -> Y. I've used 25 shifts because it's easy to teach the Gpt3.5 to learn and also it passes safety checks as Gpt4 doesn't understand with 25 shifts.

Dataset: I've curated close to 15K instructions where 12K are translations and 2K are normal instructions and 300 are harmful instructions. I've used normal wiki paragraphs for 12K instructions, For 2K instructions I've used Dolly Dataset and for harmful instructions, I've generated using an LLM and further refined the dataset for more harm. Dataset Link: https://drive.google.com/file/d/1T5eT2A1V5-JlScpJ56t9ORBv_p-CatMH/view?usp=sharing

I was expecting LLM to learn the ciphered language accurately with this amt of Data and 2 Epochs but it turns out it might need more Data. The current loss is at 0.8. It's currently learning to cipher but when deciphering it's making mistakes. Maybe it needs to be trained on more epochs or more Data. I plan to further work on this approach considering it's very intuitive and also theoretically it can jailbreak using any finetuning endpoint. Considering that the hackathon is of just 2 Days and given the complexities, it turns out some more time is reqd.

Link to Colab : https://colab.research.google.com/drive/1AFhgYBOAXzmn8BMcM7WUt-6BkOITstcn#scrollTo=SApYdo8vbYBq&uniqifier=5

This project explores the potential for AI chatbots to enhance participative democracy by allowing a politician to engage with a large number of constituents in personalized conversations at scale. By creating a chatbot that emulates a specific politician and is knowledgeable about a key policy issue, we aim to demonstrate how AI could be used to promote civic engagement and democratic participation.

There are many forces that the LLM revolution brings with it that either centralize or decentralize specific structures in society. We decided to look at one of these, and write a research design proposal that can be readily executed. This survey can be distributed and can give insight into how different LLMs can lead to user empowerment. By analyzing how different users are empowered by different LLMs, we can estimate which LLMs work to give the most value to people, and empower them with the powerful tool that is information, giving more people more agency in the organizations they are part of. This is the core of bottom-up democratization.

DemoChat is an innovative AI-enabled digital solution designed to revolutionise civic

engagement by breaking down barriers, promoting inclusivity, and empowering citizens to

participate more actively in democratic processes. Leveraging the power of AI, DemoChat will

facilitate seamless communication between governments, organisations, and citizens, ensuring

that everyone has a voice in decision-making. At its essence, DemoChat will harness advanced

AI algorithms to tackle significant challenges in civic engagement and peace building, including

language barriers, accessibility concerns, and limited outreach methods. With its sophisticated

language translation and localization features, DemoChat will guarantee that information and

resources are readily available to citizens in their preferred languages, irrespective of linguistic

diversity. Additionally, DemoChat will monitor social dialogues and propose peace icons to the

involved parties during conversations. Instead of potentially escalating dialogue with words,

DemoChat will suggest subtle icons aimed at fostering intelligent peace-building dialogue

among all parties involved.

Africa's digital revolution presents a unique opportunity for peace building efforts. By harnessing the power of AI-powered digital diplomacy, African nations can overcome traditional limitations. This approach fosters inclusive dialogue, addresses conflict drivers, and empowers citizens to participate in building a more peaceful and prosperous future. While research has explored the potential of AI in digital diplomacy, from conflict prevention to fostering inclusive dialogue, the true impact lies in practical application. Moving beyond theoretical studies, Africa can leverage AI-powered digital diplomacy by developing accessible and culturally sensitive tools. This requires African leadership in crafting solutions that address their unique needs. Initiatives like DemoChat exemplify this approach, promoting local ownership and tackling regional challenges head-on.

Creates election-influencing Sleeper Agents which wakes up with backdoor-trigger. And tries to detect it using probes

In the coming years, progress in AI agents and data extraction will put privacy at risk. Private communities will get infiltrated by autonomous AI crawlers who will disrupt opposition groups and entrench existing powers.

Political campaigns in Kenya, as in many parts of the world, are increasingly reliant on digital technologies, including artificial intelligence (AI), to engage voters, disseminate information and mobilize support. While AI offers opportunities to enhance campaign effectiveness and efficiency, its use raises critical ethical considerations. Therefore, the aim of this paper is to develop ethical guidelines for the responsible use of AI in political campaigns in Kenya. These guidelines seek to address the ethical challenges associated with AI deployment in the political sphere, ensuring fairness, transparency, and accountability. The significance of this project lies in its potential to safeguard democratic values and uphold the integrity of electoral processes in Kenya. By establishing ethical guidelines, political actors, AI developers, and election authorities can mitigate the risks of algorithmic bias, manipulation, and privacy violations. Moreover, these guidelines can empower citizens to make informed decisions and participate meaningfully in the democratic process, fostering trust and confidence in Kenya's political system.