Loops, Latents, and the Unavoidable A Priori: Why Causal Modeling Needs Couple’s Therapy

Teams love causal diagrams.

A product team draws arrows from “user trust” to “adoption.” A policy team draws loops between “service capacity,” “public confidence,” and “demand.” A data science team converts the same discussion into variables, coefficients, latent constructs, and model fit indices. Everyone nods. Everyone says “causal.” Then the meeting ends, and each group quietly returns to a different universe.

That is the problem Hovmand and colleagues take on in Bridging the Unavoidable A Priori.¹ The paper is not another call to “add feedback loops” to AI systems, as if responsible AI were missing only a better diagramming tool. Its sharper claim is that two major causal modeling traditions—system dynamics and structural equation modeling—often talk past each other because they begin from different assumptions about what a model is supposed to explain.

System dynamics asks: what feedback structure could generate the observed behavior of a system over time?

Structural equation modeling asks: how well does a specified causal and measurement model explain the observed covariance structure of data?

Those are not minor stylistic differences. They are different modeling instincts, different standards of evidence, and different anxieties. One worries about missing feedback, accumulations, and time horizons. The other worries about latent constructs, measurement error, identification, and model fit. The paper’s useful move is not to declare a winner. It gives the couple a shared calendar, a shared vocabulary, and a slightly uncomfortable reason to stay in the room.

The real conflict is not arrows; it is what the arrows are for

The easiest way to misunderstand this paper is to think it is about diagram translation. System dynamics has causal loop diagrams and stock-flow diagrams. Structural equation modeling has path diagrams with latent variables and indicators. Put the symbols in a common legend and the problem is solved. Lovely. Also false.

The paper spends substantial time showing that the visual confusion is only the surface layer. In system dynamics, a causal model is usually built around a reference mode: an explicit statement of the focal behavior pattern over time. This may be grounded in data, but it is not identical to raw longitudinal observations. It is a problem-structuring device. The modeler chooses which behavior matters, over what time horizon, and which accumulations and feedback loops are relevant at that scale.

SEM begins elsewhere. It starts from a specified causal and measurement structure and asks whether that structure can reproduce the covariance patterns in observed data. Its great strength is not simulation of feedback behavior but disciplined treatment of latent variables, indicators, measurement error, and model fit.

The paper’s comparison can be compressed into a practical contrast:

This is why the paper’s title invokes Donella Meadows’ “unavoidable a priori”: every modeling discipline carries background assumptions about what counts as explanation. Those assumptions are not always wrong. They are just dangerous when invisible.

In business AI work, invisible assumptions are where governance decks go to become decorative furniture.

System dynamics solves an inverse problem; SEM solves a forward problem

The paper’s most important comparison is the distinction between inverse and forward problems.

In a forward problem, the analyst starts with a model structure and asks what it predicts or estimates. In SEM, this often means specifying relationships among latent and observed variables, estimating parameters, and comparing the model-implied covariance matrix with the observed covariance matrix.

In an inverse problem, the analyst starts with a behavior pattern and asks what system structure could have generated it. That is the system dynamics instinct. A rise-and-collapse curve, a delayed oscillation, a capability trap, or a slow erosion of trust is not merely an outcome to be predicted. It is a clue about possible feedback structures, accumulations, delays, and nonlinear relationships.

This matters because business and policy teams often move too quickly from observed behavior to estimated effects. A dashboard shows declining retention. A model finds a strong association between response time and churn. The organization then “optimizes response time.” That may help. It may also miss a feedback loop: poor service reduces trust, lower trust increases complaint intensity, complaint intensity overloads staff, overload worsens service, and the system politely eats the intervention.

SEM can help discipline the measurement of “trust,” “satisfaction,” or “organizational capability.” SD can help explain why those constructs evolve in loops rather than straight lines. The paper’s argument is that these capabilities should be made mathematically comparable, not mashed together with enthusiasm and arrows.

A compact way to see the difference:

The arrows point in opposite epistemic directions. That is not a contradiction. It is the beginning of a useful partnership.

The framework separates dynamics, statics, and measurement

The paper’s central technical contribution is a general SD-SEM framework that decomposes a model into three subsystems:

1. A dynamic subsystem, where state variables or stocks change over time through rates of change.
2. A static subsystem, where auxiliary variables, parameters, and interactions define relationships that do not themselves accumulate.
3. A measurement subsystem, where latent dynamic and static variables are mapped to observed indicators, potentially with delays and measurement error.

That decomposition is doing real work.

System dynamics has always cared about stocks and flows. A stock is an accumulation; a flow changes the stock. In a bathtub, the water level is not directly “caused” by a coefficient in a regression. It changes because inflows and outflows accumulate over time. The paper stresses that this is a mathematical relationship, not a statistical one. There is no path coefficient to estimate for the fact that a flow integrates into a stock. Integration is what flows do.

SEM, meanwhile, has a mature language for latent constructs and indicators. A construct like “systems thinking,” “institutional trust,” or “organizational readiness” is not directly observed. It is inferred through measured indicators, factor loadings, and measurement error assumptions. SD models often include soft variables of exactly this kind, but the measurement treatment is less standardized.

The proposed framework tries to let both sides keep what they are good at. Dynamic state variables can generate behavior over continuous time. Static variables can express auxiliary relationships and interactions. Measurement equations can map latent model variables to observations with loadings, lags, and errors.

The important part is not that the final notation is cute. It is not. Mathematical bridges rarely win interior design awards. The point is that the framework creates a shared space in which SD and SEM models can be specified, compared, and eventually used to generate systems for simulation studies.

The reference mode is not just a line chart

One of the paper’s most business-relevant ideas is the reference mode.

A reference mode is the focal behavior pattern that the system dynamics model is trying to explain. It may be represented as a graph over time, but it is not simply “the data.” It is an idealized, empirically grounded statement of the dynamic problem.

That distinction is subtle and useful.

Suppose a company sees weekly new-user adoption falling with small oscillations. One team may define the problem as “overall adoption decline.” Another may define it as “cyclical instability in adoption.” A third may define it as “decline plus damped oscillation.” Those are different reference modes. They imply different feedback loops and different interventions.

The paper argues that this choice should be made explicit. Otherwise, the model can appear objective while hiding the most consequential modeling decision: what behavior the organization decided to explain.

For AI governance, this is a quiet but important lesson. Many AI risk assessments begin with available data rather than a properly defined dynamic problem. The team measures what is accessible, fits what is convenient, and later discovers that the system-level failure was outside the frame. The issue was not merely data quality. The issue was problem definition over time.

The examples are mappings, not empirical victories

The paper includes four examples. They should not be read as experiments, ablations, robustness tests, or performance benchmarks. Their likely purpose is illustrative: to show that different kinds of SD, SEM, and hybrid models can be represented within the proposed framework.

This distinction matters because the paper’s contribution is conceptual and mathematical. It is not presenting a new AI benchmark, a production deployment, or a quantified lift in decision quality. The examples show coverage and plausibility. They do not close the empirical case.

That is not a weakness if the paper is read correctly. A bridge design is not a traffic report. First we need to know whether the bridge can be built.

Why “feedback loops in AI” is too shallow a reading

The paper begins with responsible AI and the concern that AI/ML systems can amplify biases when they lack causal understanding of social context. It cites interest in system dynamics as a way to improve AI/ML applications, especially where feedback loops and societal context matter.

But the paper is careful about what is missing in much of this work. The problem is not simply that AI models lack loops. The problem is that adding feedback loops without understanding accumulations, reference modes, measurement, and model specification can create a false sense of causal sophistication.

A causal loop diagram can look profound while remaining mathematically underspecified. A statistical model can look rigorous while excluding the very feedback structure that produces the observed behavior. A machine learning system can optimize prediction while changing the system it predicts. Everyone gets a badge. The system gets worse.

The paper’s deeper business message is that responsible AI needs more than fairness metrics and post-hoc explanation. It needs a modeling workflow in which causal assumptions are explicit at several levels:

This is not glamorous. It is the paperwork of intellectual honesty. Sadly, that is often where the useful work lives.

The business value is cheaper model misunderstanding

For enterprises, the practical value of this paper is not immediate automation. It is not “use SD-SEM and reduce costs by 23%.” No such result is in the paper, and inventing one would be rude to arithmetic.

The value is diagnostic.

Many organizations already use several causal languages at once. Strategy teams use systems maps. Research teams use SEM or related statistical models. AI teams use predictive models and sometimes causal graphs. Risk teams use governance frameworks. The same construct—trust, adoption, resilience, engagement, capability—moves across these languages, usually losing precision along the way.

The SD-SEM framework suggests a way to reduce translation loss.

In policy analytics, it can help distinguish between measuring a latent construct and simulating the feedback process in which that construct changes. In healthcare, it can help avoid treating systemic causal mechanisms as nuisance variation. In education and team-performance contexts, it can support models where latent skills or mindsets are measured through indicators but influence dynamic processes over time. In AI governance, it can help teams ask whether an algorithmic intervention changes the system that later supplies its training or monitoring data.

The practical pathway looks like this:

1. Define the dynamic problem before selecting the statistical model. Start with the reference mode, not the most convenient dataset.

2. Separate latent measurement from dynamic explanation. Do not pretend a soft variable is directly observed. Do not pretend a measured construct automatically explains feedback behavior.

3. Make accumulations and delays explicit. If a variable changes because rates accumulate over time, represent that mathematically rather than as a simple direct path.

4. Use SEM discipline where SD is often casual. Measurement error, indicators, and model fit deserve more attention in system dynamics practice.

5. Use SD discipline where SEM is often static. Feedback loops, stocks, flows, and time horizons deserve more attention in statistical causal modeling.

This is not a recipe for a single universal model. It is a workflow for preventing teams from confusing one kind of rigor with another.

Where the paper’s boundary should be drawn

The paper is ambitious, but its evidentiary status is bounded.

First, it proposes a general mathematical framework. It does not provide a finished software implementation, production method, or empirical comparison showing that SD-SEM hybrids outperform existing approaches in AI/ML tasks.

Second, its examples are illustrative. They show that familiar models can be represented within the framework, including a standard SD model, a standard SEM example, a public-health concept model, and a hybrid systems-thinking/team-performance case. They are not validation studies.

Third, the framework makes simplifying assumptions, especially in the measurement subsystem. For example, the paper discusses measurement errors and delays in ways that create a tractable general expression, while noting that future work could relax assumptions such as constant measurement error over time.

Fourth, the framework does not eliminate the unavoidable a priori. It exposes it. That is less comforting but more useful. Modelers still need theory, operational knowledge, empirical grounding, and judgment. Mathematics can translate assumptions; it cannot make them disappear.

This boundary is important for business readers. The paper should not be sold internally as “a new AI causal modeling method ready for deployment.” A better internal pitch would be: a formal bridge for auditing and designing causal models when statistical measurement and dynamic feedback both matter.

Less shiny. More likely to survive contact with reality.

The bridge matters because causal modeling is becoming organizational infrastructure

As AI systems move from isolated prediction tasks into workflows, markets, institutions, and public services, causal modeling becomes less like academic decoration and more like infrastructure. Bad causal assumptions do not merely produce bad slide decks. They shape interventions, incentives, monitoring systems, and automated decisions.

That is why the comparison between SD and SEM matters.

SEM reminds system dynamicists that soft variables should not be waved into existence because a workshop liked the arrow. Measurement matters. Indicators matter. Fit matters.

System dynamics reminds statisticians and data scientists that observed variables live inside systems that evolve. Accumulations matter. Delays matter. Feedback matters. The time horizon chosen by the analyst can determine which causal story is even visible.

The paper’s contribution is to say: these two reminders can be placed into a common mathematical frame. Not because the traditions are the same. Because the world they model refuses to respect departmental boundaries.

For Cognaptus readers, the lesson is straightforward. When building AI-supported decision systems, do not ask only whether the model predicts well. Ask what behavior it is meant to explain, what system structure could generate that behavior, how the important latent constructs are measured, and which assumptions have been smuggled in as “common sense.”

The unavoidable a priori will still be there. But at least it will have to introduce itself.

Cognaptus: Automate the Present, Incubate the Future.