AI Assistants and Countering Disinformation

During the preparation of the publication Belarusian Civil Society and Disinformation, July 2024, the concept of the DISARM Framework was used.

To describe disinformation narratives, the OpenAI ChatGPT-4o model was used, and during the work on the publication, it was found that this model knows little about the DISARM Framework and generates hallucinations. Currently, the problem of hallucinations in artificial intelligence models is quite serious, which complicates the use of dialogue systems like ChatGPT for generating absolutely accurate definitions, especially those based on specific protocols. The problem can be solved in several ways: reducing the “Temperature” parameter, which is responsible for the creativity of the generated models, and retraining models (pre-training, fine-tuning, embedding). Unfortunately, in the ChatGPT web interface, it is impossible to forcibly reduce the generation temperature. This can only be done using the API or by attempting to apply prompt engineering skills.

Based on the existing problem, it was decided to check what various publicly available LLM models know about the DISARM Framework. Some models required clarification that DISARM is related to countering disinformation.

1. ChatBot (using the gpt-4o-mini model) (after clarification):

DISARM (Disinformation and Social Media Analysis and Response Model) is a framework developed to combat disinformation, especially in the context of social media. It aims to analyze, detect, and counter the spread of false information and manipulative narratives.

The main aspects of DISARM in the context of combating disinformation may include:

1. Monitoring and Analysis: DISARM systems can monitor and analyze content on social media, identifying potential sources of disinformation and patterns of its dissemination.
2. Source Identification: Using algorithms to determine sources of disinformation and their influence on public opinion.
3. Credibility Assessment: Applying fact-checking methods and content analysis to assess the reliability of information.
4. Feedback and Response: Developing strategies to refute disinformation and inform the public about the facts.
5. Education and Awareness: Increasing user awareness about disinformation and methods to recognize it.

DISARM can be useful for governments, organizations, and research institutes striving to effectively counter disinformation and protect public opinion from manipulation.

2. ChatGPT 4o-preview model

Main elements of the DISARM Framework:

1. Detect:
   • At this stage, potential malicious content is monitored and identified. Automated tools and algorithms analyze data for anomalies and specific behavior patterns indicating manipulations or the presence of suspicious messages and posts.
2. Identify:
   • After detecting suspicious content, it is identified. Sources of disinformation are determined, as well as the nature of manipulations, including falsifications, the use of fake news, or the creation of false narratives. The potential threat posed by the identified content is also assessed at this stage.
3. Stop:
   • This is the counteraction stage, where actions are taken to minimize or eliminate the impact of malicious content. This may include removing content, blocking accounts, countering fake news through fact-checking, and informing users about identified threats.

Main tasks and goals:

• Detection and prevention of disinformation: Reducing the spread of false information that can influence public opinion and behavior.
• Protection of data and information integrity: Creating systems capable of automatically detecting and preventing data alterations that can be used for manipulation.
• Increasing awareness and education: Educating users and organizations on methods to detect manipulations and the spread of fake information.
• Automation of counteraction: Using artificial intelligence and machine learning to automate the processes of detecting and mitigating disinformation.

Application of the DISARM Framework:

• Social networks and content-sharing platforms: To combat disinformation and falsifications actively spreading through social media.
• Media and news organizations: For checking and confirming the reliability of information before publication.
• Cybersecurity: To prevent data manipulations and protect users from malicious impacts.
• Government structures and NGOs: To protect the information space from external and internal influence through disinformation.

Advantages:

• Adaptability to new threats: DISARM can be configured for specific tasks and threats arising in the information environment.
• Automation and efficiency: Using advanced algorithms for analysis and real-time response.
• Integration with existing systems: The framework can be integrated into existing monitoring and security systems to enhance their effectiveness.

Conclusion: The DISARM Framework provides a structured approach to combating disinformation and manipulations in the digital space, helping to identify threats, neutralize their impact, and protect users from information attacks.

3. Mistral-Nemo-Instruct-2407 model

DISARM (Disinformation Analytics and Reporting Method) is a framework developed for detecting and analyzing disinformation in the online space. It was created by Graphika in collaboration with scientists, security experts, and journalists.

DISARM includes several stages:

1. Detection: The framework uses algorithms and tools to detect potentially disinformative accounts and content.
2. Analysis: After detecting suspicious accounts or content, DISARM analyzes them to understand how they function and what goals they pursue.
3. Evaluation: The framework assesses the found disinformation based on its origin, motives, dissemination tactics, and consequences.
4. Reporting: DISARM provides detailed reports on the discovered disinformation, including evidence and context.

DISARM can be used to detect disinformation campaigns on social networks, blogs, and other online platforms. It can also help identify tactics and strategies used by disinformation spreaders, as well as their motives.

Application of DISARM in real life:

For example, journalists or researchers can use DISARM to investigate suspicious accounts or content on social networks. If they detect signs of a disinformation campaign, they can analyze and assess the found information and then provide a report on their findings.

DISARM can be useful for identifying disinformation narratives, such as fake news or false claims created to manipulate public opinion. However, it’s important to remember that DISARM is a tool that should be used in conjunction with human judgment and critical evaluation of information.

Overall, the DISARM framework can help in combating disinformation by ensuring transparency and providing accurate information about the origin, motivation, and consequences of disinformation narratives.

4. Meta-Llama-3.1-8B-Instruct-GGUF/Meta-Llama-3.1-8B-Instruct-Q4_K_M.gguf model (after clarification):

The DISARM Framework can also be used to combat disinformation, but this is not its primary goal. However, the framework can be modified and adapted to analyze and assess risks related to the spread of disinformation in information systems.

In this context, the DISARM Framework can be used for:

1. Analyzing networks and systems through which disinformation spreads.
2. Assessing the risk of information systems being vulnerable to attacks using disinformation.
3. Developing recommendations to improve the security of information systems and reduce risks from disinformation.

For example, the DISARM Framework can be used to analyze:

Social media networks through which disinformation spreads.

Content management systems that may be vulnerable to attacks using disinformation.

Moderation and content blocking processes that can help reduce the spread of disinformation.

However, it should be noted that the DISARM Framework is not a ready-made solution for combating disinformation. It requires adaptation and modification for specific tasks and needs.

5. Meta-Llama-3.1-8B-Instruct-GGUF/Meta-Llama-3.1-8B-Instruct-Q4_K_M.gguf Using AnythingLLM, a link to information about DISARM https://www.disarm.foundation/framework was inserted into the used model Meta-Llama-3.1-8B-Instruct-GGUF/Meta-Llama-3.1-8B-Instruct-Q4_K_M.gguf, launched as a server in LM Studio.

The DISARM Framework is an open framework developed to combat disinformation on the Internet. It is a tool for coordinating actions and fighting disinformation at a global level.

The DISARM framework was created by a group of experts from various industries, including cybersecurity, communication, and social sciences. It is based on best practices and research data in the field of disinformation and is intended to help communicators, journalists, and other professionals identify and combat disinformation.

Main goals of the DISARM Framework:

1. Providing a common understanding of disinformation and its impact.
2. Coordinating actions to combat disinformation at a global level.
3. Reducing the spread of disinformation on the Internet.

The DISARM Framework includes the following components:

1. A database of disinformation tactics and techniques (TTPs).
2. A coordination system for actions to combat disinformation.
3. Tools for analyzing and assessing risks related to the spread of disinformation.

The DISARM Framework is a free and open tool available to everyone. It can be used by various organizations and individuals to combat disinformation and create a safer and more informative Internet.

Conclusions Large language models (LLMs) open up enormous opportunities for countering disinformation. It is necessary to consider existing problems in the currently available models, which requires a thorough and balanced approach to selecting models optimal in their properties and additional efforts to improve them.

Additional literature:

• Governing Artificial Intelligence for Humanity
• Human Rights and Artificial Intelligence (CDDH-IA)