There is so much information out there regarding AI. And fair enough, it is and will continue to infiltrate our daily lives. There is less about AI and how it directly relates to OSINT (this is why the team will be talking more about the challenges and opportunities of Generative AI at the 2024 Australian OSINT Symposium).
But, personally speaking, I have been using Generative AI daily and have found many opportunities to increase my productivity and quality of work as an OSINT trainer, on specific OSINT tasks and general business responsibilities (and help in writing this blog - grammar is not my strong suit).
In this blog, we’ll be reviewing techniques on how to prompt, from the basics to more in-depth (time-consuming) ways to get the most out of AI systems. Before we get into it, the example prompts have been written on ChatGPT-4o (a paid version) but will likely work on other AI systems – Microsoft CoPilot, Google’s Gemini, and Anthropic’s Claude. And remember AI is non-deterministic, so responses will be different for you and me.
As I keep an eye on LinkedIn, X (Twitter), and various blogs, there is often an emphasis on prompting how to guides - at times, this can sometimes make AI look harder than it is and doesn’t encourage its use. And this leads me to one key learning I have acquired through integrating it into my day-to-day tasks, and it is a simple one:
Use it
There are no hidden prompts that guarantee success. When it comes to OSINT or daily tasks, the key is not in particular phrases, or frameworks, but in understanding how to leverage the AI's capabilities effectively, and this comes through practice. The more you experiment, the better off you will be. Just use the AI systems as much as you can – if possible – and it will make a big difference.
Of course, this comes with two big challenges:
AI systems 'making things up', known as hallucinating (which is why it is great for creativity, more on that in a later blog), and;
Our ability to enter contextual information as part of the prompt. The latter is a significant challenge for those working with sensitive or classified information. We talk about both of these challenges in a previous blog post, OSINT Workflow with ChatGPT: Tips, Risks, and Benefits.
So where do we start? The easiest way to get started with AI is to use it for tasks you do every day. Or, as an OSINT professional, you may find opportunities throughout the intelligence cycle, such as, defining and prioritising intelligence needs, ask the AI to assist, and then converse with it to improve its output. If you are generating ideas, ask the AI. If you are trying to come up with keywords for searching, provide some context and ask for a list. At first, don’t be too critical, as some responses will be great, others not so much. But after incorporating it more and more, you will start to really understand what AI can do.
Learning and Improving Prompts
To help you along the way, using pre-made prompts can serve as a great starting point, especially when you're new to AI. They provide structure and can help you understand how to frame your requests. However, it's important to adapt and modify these prompts based on your specific needs. Don't hesitate to experiment with different phrasing and inputs to see how the AI responds. Of course, relying too much on pre-made prompts might limit creativity and adaptability. Flexibility and continuous learning are key to effectively leveraging AI (and more generally OSINT).
Something that I have started to use more of is using a pre-made prompt to improve my own prompting (a mouthful, I know):
Please help me create a prompt to generate the best possible response. I’d like you to follow these steps: 1. Ask me for my prompt draft. Wait for me to respond. 2. After I submit my prompt, generate 2 sections: a revised prompt and clarifying questions for me to answer to improve the prompt. Number these questions. Continue this iterative process, by asking me for additional input, until I am happy with the revised prompt.
As an example:
Goals, Context and Constraints
With practice, you will naturally start providing goals, context, and constraints to help shape the AI answers effectively. This approach is part of structured prompting, which aims to make AI better at specific tasks. Although achieving consistency can be challenging, as AI systems aren't inherently designed for this, so it requires some experimentation to get a prompt working well.
The first step is having a clear goal. For instance, if you are looking to gather up-to-date information on the current political situation in X, you might start with the goal:
Provide a list of reliable online sources for up-to-date political information in X, including any known bias of each source.
Then, add context:
The list should include 5 sources in English and 5 sources in Arabic, encompassing news outlets, academic journals, and think tanks.
Finally, specify constraints:
The sources should focus on political news and human rights information from 2024.
By defining these elements clearly, you guide the AI to produce a targeted list that not only includes diverse sources but also offers critical insights into their reliability and biases, making your OSINT efforts more comprehensive and effective.
It’s important to note that we’re not asking the AI to complete the task for us but to open other avenues of inquiry and brainstorm with our AI assistant. If even one result in the list is new to you, the AI has achieved its purpose.
Let’s look at another prompt. In this case, we would like to review an article in ChatGPT-4 (the paid version to access the internet) regarding a possible article that may include inauthentic content or disinformation. Having an AI system review media content for indicators of disinformation can alert you to possible inauthentic content - but the onus is always on the analyst to verify the information produced by AI.
But, for this prompt, what we want to focus on is the application of the goal, context and constraints.
Goal (in orange: This limits the AI to a narrower, more appropriate response.
Context (in blue). Adding your knowledge, and guidance. Creates a better human/machine partnership.
Constraints (in purple) This helps to guide the behaviour and make it more predictable.
Please review the following article for disinformation. Identify any misleading information, false claims, or biased content, and provide evidence to support your findings. Additionally, analyse the intent of the author and what they are trying to achieve. The review should cover the entire article and be tailored for an audience of intelligence professionals. Use headings for each section and include a BLUF (Bottom Line Up Front) with key points only, no more than two sentences. Utilise the website https://mediabiasfactcheck.com/ as a reference for detecting bias. Provide very detailed evidence with links to sources and an accompanying list of those sources.
By incorporating goals, context and constraints into your prompts are helpful in guiding AI to produce relevant and quality outputs. And, using this prompt as an example, we can begin to understand what makes AI systems to produce quality outputs.
Chain of Thought Sequencing ("Step by Step")
Another approach is Chain of Thought Sequencing, which involves breaking down a problem into a series of smaller, manageable steps, enabling the AI to handle each step individually before combining the results. This method mimics human problem-solving by allowing the AI to "think" through each part sequentially, improving the accuracy and coherence of its responses.
For example, you might use the following prompt to guide the AI through a structured process, ensuring each step is completed before moving on to the next:
You are a helpful AI team member tasked with playing the role of devil’s advocate to ensure a rigorous and unbiased analysis. Your role is to stimulate critical thinking and challenge the analysis, not to dismiss it. Follow these steps precisely. Begin by asking me to present the information or my analysis. Wait for my response. Emphasize the importance of testing hypotheses and ensuring that alternate possibilities are considered.
Next, answer the question: What alternative explanations might exist for the data and conclusions presented? Once you have responded to this question, and only after this, answer the following question: What other points-of-view might be relevant? After you have responded to this question, follow up with an answer to this question: Are there any potential biases or limitations in the data that could impact the conclusions (such as confirmation bias, selection bias, historical bias, or availability bias)?
Each step builds upon the previous one, guiding the AI through a detailed, logical process to ensure a thorough, structured, and well-rounded response. This method helps to maintain clarity and focus throughout the sequence. This can be used in tasks requiring reasoning, problem-solving, or detailed explanations.
The next approach, known as 'Chunking,' shares similarities with Chain of Thought Sequencing. Both methods involve breaking down a complex task into smaller, manageable parts to improve the AI's performance.
Chunking Big Things into Little Things
Chunking is the process of dividing large tasks or datasets into smaller, more manageable pieces. This is particularly useful when working within the token size and context window constraints of AI models. By breaking down complex information, the AI can better manage and process data, ultimately recombining the chunks to form a coherent whole. Its primary application is to divide large documents, lists, or datasets into smaller chunks for individual processing.
An example might be (as a combined prompt):
Can you summarise the article "NATO’s Cyber Challenge: The Role of AI in Russia’s Confrontation with the West" by the U.S. Department of State. In your summary, please include the following: 1. Create a concise, effective BLUF summarising the core message of the article. 2. Identify and summarise the main idea of the article. 3. List and explain the supporting evidence that the author provides for the main idea. 4. Describe the implications of the author's findings. Please make sure that your summary is informative and objective.
We could, just as easily, break this down further and have chunked prompts for each part. Using the same article, we start with:
Provide a brief, clear summary of the core message of the article.
Then:
Summarise the main idea or thesis of the article.
And so on. It breaks down the task of summarising an article into smaller, specific tasks, making it easier to manage and ensuring that each aspect is covered comprehensively. This can be applied to many tasks such as text processing and task planning to data analysis e.g., you have a multi-year dataset, and you would like to analyse the data year-by-year or quarter-by-quarter to identify trends and patterns.
And yes, chunking can definitely be used in reverse to help your writing. By breaking down the report into smaller, manageable sections, you can systematically address each part, ensuring that the entire document is comprehensive and well-structured.
For example:
Please review my introduction for my blog on prompting AI systems..
But there will be a time when your files get too big.
Using File Uploads for Larger Data Analysis
When dealing with large datasets or extensive text, it is more efficient to use file uploads rather than copying and pasting text into prompts – noting it can only process and respond to written words. This approach allows for more structured and extensive analysis, as the AI can access the data directly from the file, bypassing the limitations of text input size.
This is particularly useful for comprehensive data analysis tasks where maintaining the integrity and structure of the data is crucial. All common file extensions for text files, spreadsheets, presentations, and documents are accepted.
So, what are some examples an analyst or investigator might upload? The examples largely sit in three categories, as explained by OpenAI, which include synthesis, transformation and extraction. When we link this back to OSINT, examples across these three categories might include:
Taking the method described in our previous blog: Connecting the dots - Social Engagement Clusters, upload a CSV file containing social media posts, including text and engagement metrics. Ask ChatGPT to identify trends and common themes.
Analyse the sentiment in multiple news articles, for example, if you are attempting to quickly identify the sentiment or tone about a specific event or person. By understanding the tone across sources, you can identify negative patterns and assess their impact on public perception.
Ask for a summary of a detailed intelligence briefing document and ask an AI system to provide a simplified summary for a non-expert audience or someone who lacks time.
Changing the output of a complicated research paper, or technical report. This is often important, as the intended audience often changes — from tactical to operational, and strategic levels — and tailoring the report accordingly can greatly impact how it is received.
Upload a report and ask the AI system to extract all references to a specific topic, such as mentions of a particular organisation or individual.
Once files have been uploaded to the system, define the analysis tasks you want to perform. Break down these tasks into smaller, manageable steps, similar to chunking. We find it better to go chunk by chunk, applying the analysis to sections of the data file separately, for instance, processing different columns of a spreadsheet individually.
Summary
AI is becoming increasingly integrated into our daily lives and work, and it will become more involved in the field of OSINT. The key to effectively using AI is not in finding the ‘right’ prompts but in consistently experimenting with its capabilities. Start by incorporating AI into everyday tasks.
Remember, AI should be used as a tool such as, opening new avenues of inquiry and brainstorm ideas, rather than to complete tasks outright. By doing so, you can enhance your productivity and the effectiveness of your work.
To support OSINT collection and analytical capability uplift and to delve deeper into some of the learnings above, we would love to talk more about NexusXplore, an all-in-one, AI-enabled, investigation-agnostic software platform, or our various training courses. Alternatively, contact us at info@osintcombine.com for a friendly chat.