Introduction
The revolution of Artificial Intelligence (AI) is taking us into uncharted territories of innovation and problemsolving. One such frontier is the intersection of causal AI with large language models (LLMs) like ChatGPT, an area where Geminos are making great strides. In this article, we will discuss how we are harnessing ChatGPT's vast knowledge base as a 'virtual' subject matter expert to build accurate and robust causal models and, how we can validate those models to guard against ChatGPT 'hallucinations'.
Part 1: The Power of Causal Models
Understanding Directed Acyclic Graphs (DAGs)
A Causal Model Example in Geminos Causeway
Directed Acyclic Graphs (DAGs), or causal models, form the backbone of causal AI technology and Causal Digital Twins. They are essential tools in simulating realworld systems, providing valuable insights into the causeeffect relationships that shape strategic decisionmaking. As a leader in this field, Geminos underscores the involvement of subject matter experts (SMEs) throughout the process of causal model creation and iteration. Causal model notation is relatively simple and easy for anyone to understand. In the example on the right, you can see a number of 'nodes' wired together with a directional arrow. Each node represents a causal variable and each arrow represents a causal relationship with an arrow indicating the direction of that relationship. In the early stages of causal model development, we are not worried about quantifying these relationships, it is enough to simply understand that a change to one variable in the model affects one or more other variables.
Capturing Subject Matter Expert Domain Knowledge in Geminos Causeway
By using this simple technique, SMEs can encapsulate their specialized domain knowledge directly into the Geminos Causeway platform, capturing a thorough understanding of the domain in question within one or more causal models.
There is a process that causal teams typically go through which we won't be going into lots of detail within this article but it is summarized at a very high level below.
HighLevel Causal Discovery Process
Part 2: Leveraging AI for Enhanced Causal Model Creation
Introducing the Geminos ChatGPT Discovery Assistant
The Geminos ChatGPT Discovery Assistant is an invaluable tool for addressing research questions and generating robust causal models. It can supply a list of causal variables and relationships within a specific context, augmenting the existing causal models and contributing to their depth.
The current version of the ChatGPT Discovery Assistant (at the time of writing this article) is used in a number of ways and all are handled by Geminos Causeway with suitable prompt engineering to handle each use case:
 Creation of an Initial Causal Model  By supplying the platform with a research question (for example, how does X affect Y), the virtual assistant will tap into the knowledge of ChatGPT, suggest causal variables and relationships and will draw the initial causal model for you.
 Exploration of Existing Causal Models  Using a research question and combining that with an existing causal model, Geminos can feed that into ChatGPT in a structured way to explore other causal variables that we might not have previously considered. Like with the creation of the initial model, the GPT assistant can draw the new suggested variables and relationships for you after you select which ones you are interested in.
 Exploration of Unobserved Variables  During causal model validation (discussed later in this article) you will see how Geminos identifies the presence of unobserved variables and the GPT Assistant can be used to drill into specific areas of the model to suggest other variables, not present, that may be affecting the outcome in relation to the research question.
These techniques can significantly accelerate the development and refinement of your causal models and, with the validation techniques discussed next, ensure the creation of an accurate and robust model that will form the basis of your causal solutions.
Part 3: Validating Causal Models with Geminos Causeway
BuiltIn Analysis Techniques for Causal Validation
Geminos Causeway Causal Validation
Geminos Causeway features integrated analysis techniques for validating the constructed causal models. These techniques involve hydrating the causal models with data through an easy ETL (extract, transform, and load) flow creation process, allowing us to validate assumptions made by both SMEs and ChatGPT, ensuring a robust and accurate causal model is formed.
We achieve this in a number of ways using causal analysis techniques:
 Causal Function Definition  Based on the data that we have hydrated the model with, Geminos Causeway will define a function for each causal variable that encapsulates the causal factors affecting it. These functions are linear by default but these can be overridden to handle nonlinear, more complex functions. What this does is break into smaller parts what would, with traditional AI, have been one large algorithm, which, depending on complexity would likely develop over time to be something difficult to maintain.
 Intrinsic Analysis  Because we have the causal model, we can traverse back up the causal tree to understand the intrinsic effects on each causal variable. This means we can take into account, not only parents, but grandparents, great grandparents and so on. Providing this level of causal analysis means we can see if:

 NonCausal Relationships  We've recorded a causal relationship which isn't backed up by the data highlighting either a problem with the data or, a noncausal relationship which maybe should be removed but certainly should be reviewed.
 Unobserved Variables  We have identified a level of noise in the model which is identifying unobserved causal variables
 NonCausal Relationships  We've recorded a causal relationship which isn't backed up by the data highlighting either a problem with the data or, a noncausal relationship which maybe should be removed but certainly should be reviewed.
 Causal Edge Strength Analysis  After data hydration the builtin causal analysis techniques allow you to visualize the strength of each causal relationship. This is achieved by displaying the strength on the actual relationship wire and coloring it green for a positive causal effect and red for a negative one.
These techniques, paired with others built into the Geminos platform alongside with SME knowledge provide a suite of approaches to validate the models we create with SME and/or ChatGPT Assistance.
Part 4: Causeway Abstracts Validation Complexity
Simplifying Causal Model Validation
Sitting behind each of the causal analysis and validation techniques is complex mathematics. Without Geminos, you'd need to understand the complexity (and probably code in Python) to be able to utilize these approaches. Thankfully, Geminos Causeway provides a level of abstraction away from the complicated mathematics by providing access to the results of the analysis in a visual way that can easily be communicated with both technical and nontechnical people. It enables users to readily identify relationships with no causal effect and detect noise within the model, which could signify unobserved variables without understanding the algebra behind it.
Geminos Causeway adds another level of sophistication to causal model validation by providing a visual display of causal strengths and the intrinsic effects on individual causal variables. It enables users to readily identify relationships with no causal effect and detect noise within the model, which could signify unobserved variables.
The following section illustrates a simple causal model with only 3 variables and shows some of the mathematical approaches (and their complexities) to understand causal effect and why, using Geminos, completely removes this complexity.
Pearlean Approach: An Illustrative Example (of the complexity)
To illustrate this, consider the Pearlean approach, which uses conditional probabilities to calculate the causal effect of one variable on another. Suppose we have two variables X and Y, and we want to find out the causal effect of X on Y. The mathematical formula for this would be:
P(Ydo(X)) = ∑ (over Z) [P(YX,Z)P(Z)]
In the context of the Pearl's causal calculus, the variables X, Y, and Z have the following meaning:
X: The intervention or cause that we are interested in studying
Y: The outcome or effect that we want to measure
Z: A set of variables that could confound or mediate the effect of X on Y
"P" stands for probability.
The causal relationships in this context could be:
X > Y: This means X has a causal effect on Y.
X > Z > Y: This means X influences Y, but this effect is mediated through Z.
The equation P(Ydo(X)) = ∑ (over Z) Z P(YX,Z)P(Z) represents a way to compute the causal effect of X on Y. It does this by summing up the joint probabilities of Y and Z given X (P(YX,Z)) times the probability of Z (P(Z)), over all possible values of Z.
It's important to note that Z could include multiple variables, and we would need to sum over all combinations of these variables.
This equation can be used to calculate the causal effect when you have an intervention on X (do(X)), meaning you deliberately set the value of X and then want to observe the effect on Y. The summation over Z allows us to take into account all potential pathways from X to Y that pass through the variables in Z.
As a result, it's a powerful tool for understanding complex causal relationships in a system, where the effect of X on Y might be influenced by many different pathways and confounding variables. It can also be used to make predictions about what would happen if we were to intervene in a system and change the value of X.
Geminos Causeway, however, abstracts these complexities, presenting the information visually. This allows users to intuitively understand the strength and direction of causal relationships and the effects of interventions, without needing to delve into complex mathematics.
Conclusion and Next Steps
Causal AI continues to build momentum, with advancements in causal analysis techniques and the integration of large language models like ChatGPT. We encourage you to explore more of our work in this space. To see how Geminos Causeway can revolutionize your organization, don't hesitate to book a demonstration. The journey to more accurate and robust decisionmaking starts with Geminos.