The portal flickers again, the terrain re-renders, and the second level of our AI-literacy quest awaits. Let’s press start.
Welcome to the Second Issue of the Ongoing Series
Project Chiron was launched because behavior analysis is entering a new epoch, and artificial intelligence will play a major role in that transformation. This week, we take on our first formal AI literacy topic related to Technical Understanding.
Fundamentally, AI is math, not magic. And although learning the math behind AI sounds intimidating, here’s the good news: you don’t need a PhD in maths to build real fluency. Instead, you need conceptual clarity, the same way we teach reinforcement schedules before diving into cumulative records.
This week, we start at ground zero:
What is a mathematical model—and why do AI systems rely on them?
Models Are Maps, Not Mirrors
In behavior analysis, our most common model is the ABC contingency—Antecedent, Behavior, Consequence. It’s everywhere: in training, in data sheets, and in how we talk about behavior.
But let’s be honest—“ABC” likely caught on because it’s the first three letters of the alphabet, not because it fully captures the complexity of human behavior. Even the more nuanced four-term contingency adds layers, but still falls short of representing the dynamic, interacting variables at play in real-world contexts.
These models aren’t wrong—they’re just simple. And that’s the point: they’re stories, not snapshots. They help us think, teach, and intervene. But they also leave a lot out from the total story.
As we move into more advanced forms of analysis, including mathematical and computational modeling, it’s worth remembering that all models—whether verbal or mathematical—are tools for highlighting what matters. We purposely choose to focus on a subset of all environmental stimuli, all behaviors someone emits, and all changes in the environment that follow the behavior we have chosen to focus on.
A mathematical model is a set of equations or rules that help describe or predict something in the world. It’s a simplified version of reality—a map, not a mirror. And like any map, it’s useful only if you know what it’s designed to highlight and what is safe to ignore.
In behavior analysis, we already use mathematical models all the time. Think of the matching law, resurgence models, or delay discounting curves. These are formal ways of saying:
“Given this input, here’s what output we expect, on average.”
AI works the same way. Underneath the slick interface of any AI tool—whether it’s classifying images or writing text—is a mathematical model trained to predict patterns based on past data. AI models are just really big, really complex, and often trained using millions or billions of examples.
Example: Functional Models and Predictive Models
Let’s say you build a model to predict a client’s likelihood of aggression following task demands. In a behavioral framework, you might write:
P(aggression) = f(task difficulty, time since break, preferred item access)
Read out loud, this would be, “The probability of aggression is a function of task difficulty, time since the client’s last break, and the client’s ability to access a preferred item.”
That’s a functional model: it expresses a lawful relationship between variables.
A machine learning model might ingest the data from those same variables—but instead of trying to explain the function, it tries to predict the output based on learned patterns. It doesn’t care why aggression happens, only how reliably it can guess that aggression happens given the data collected for the inputs.
This is a fundamental tension you may see again and again:
-
Behavior analysts favor explanatory, function-based models grounded in operant and respondent behavioral principles and processes—even if it sacrifices precision.
-
AI systems favor predictive models optimized for precision and accuracy—even if they sacrifice explainability.
AI literacy means knowing this trade-off exists, knowing when one type of model is a better fit for the job, and how a model's fit is defined by the data used to build it.
Functional and Predictive Models in Practice
Let’s look at a few examples from the settings where BCBAs often work:
