Lesson overview | Previous part | Lesson overview

Do Calculus: Part 7: ML and LLM Applications to References

7. ML and LLM Applications

ML and LLM Applications develops the part of do calculus specified by the approved Chapter 22 table of contents. The treatment is causal, not merely predictive: the central objects are mechanisms, interventions, assumptions, and counterfactuals.

7.1 causal effect of ranking policies

Causal effect of ranking policies belongs to the canonical scope of Do Calculus. The central move in causal inference is to distinguish a statistical relation from a claim about what would happen under an intervention.

For this subsection, the working scope is do-operator semantics, mutilated graphs, backdoor and frontdoor criteria, identification rules, and post-identification estimation. The mathematical objects are variables, mechanisms, graphs, interventions, and assumptions. A causal claim is incomplete until all five are visible.

$$P(Y \mid \operatorname{do}(X=x))=\sum_z P(Z=z \mid X=x)\sum_{x'}P(Y \mid X=x',Z=z)P(X=x').$$

The formula gives a compact handle on causal effect of ranking policies. It should not be read as a purely algebraic identity. In causal inference, equations encode assumptions about mechanisms, missing variables, and which parts of the world remain stable under intervention.

Three examples of causal effect of ranking policies:

1. A recommender team wants the causal effect of ranking a document higher, not merely the correlation between rank and clicks.
2. An LLM platform changes a safety policy and wants to estimate whether refusals changed because of the policy or because user prompts shifted.
3. A fairness auditor asks whether a proxy feature transmits an impermissible causal path into a model decision.

Two non-examples expose the boundary:

1. A high predictive coefficient is not a causal effect unless the graph and intervention assumptions justify it.
2. A plausible narrative produced by a language model is not a counterfactual unless it is grounded in a causal model.

The proof habit for causal effect of ranking policies is to name the graph operation. Conditioning restricts a distribution. Intervention replaces a mechanism. Counterfactual reasoning updates exogenous uncertainty from evidence, changes a mechanism, then predicts.

observed association:      P(Y | X=x)
intervention question:     P(Y | do(X=x))
counterfactual question:   P(Y_x | E=e)
discovery question:        which G could have generated P(V)?

In machine learning, causal effect of ranking policies is valuable because models are often deployed under interventions: ranking changes, policy changes, safety filters, tool-use gates, data collection changes, and human feedback loops. Prediction alone does not tell us which change caused which downstream behavior.

Notebook implementation will use synthetic SCMs and small graphs. This keeps the examples executable while preserving the conceptual split between identification and estimation.

Checklist for using causal effect of ranking policies responsibly:

• State the causal question before choosing a method.
• Draw or describe the assumed causal graph.
• Mark observed, latent, treatment, outcome, and adjustment variables.
• Separate intervention notation from conditioning notation.
• Decide whether the query is identifiable before estimating it.
• Report assumptions that cannot be tested from the observed data alone.
• Use ML as an estimation aid, not as a substitute for causal design.

This chapter follows the boundary set by Chapter 21. Statistical learning theory controls prediction error under distributional assumptions. Causal inference asks what happens when the distribution changes because something is done.

Modern AI systems make this distinction unavoidable. A foundation model can predict which action historically followed a context, but a decision system needs to know what would happen if it took a different action in that context.

Thus, causal effect of ranking policies is not an abstract philosophical add-on. It is a production and research tool for deciding which model, prompt, policy, feature, or intervention actually changed an outcome.

A final diagnostic question is whether the claim would survive a policy change. If the answer depends only on a historical correlation, it belongs in predictive modeling. If the answer depends on what mechanism is replaced and which paths remain active, it belongs in causal inference.

7.2 logged bandit interventions

Logged bandit interventions belongs to the canonical scope of Do Calculus. The central move in causal inference is to distinguish a statistical relation from a claim about what would happen under an intervention.

For this subsection, the working scope is do-operator semantics, mutilated graphs, backdoor and frontdoor criteria, identification rules, and post-identification estimation. The mathematical objects are variables, mechanisms, graphs, interventions, and assumptions. A causal claim is incomplete until all five are visible.

$$\operatorname{ATE}=\mathbb{E}[Y \mid \operatorname{do}(X=1)]-\mathbb{E}[Y \mid \operatorname{do}(X=0)].$$

The formula gives a compact handle on logged bandit interventions. It should not be read as a purely algebraic identity. In causal inference, equations encode assumptions about mechanisms, missing variables, and which parts of the world remain stable under intervention.

Three examples of logged bandit interventions:

1. A recommender team wants the causal effect of ranking a document higher, not merely the correlation between rank and clicks.
2. An LLM platform changes a safety policy and wants to estimate whether refusals changed because of the policy or because user prompts shifted.
3. A fairness auditor asks whether a proxy feature transmits an impermissible causal path into a model decision.

Two non-examples expose the boundary:

1. A high predictive coefficient is not a causal effect unless the graph and intervention assumptions justify it.
2. A plausible narrative produced by a language model is not a counterfactual unless it is grounded in a causal model.

The proof habit for logged bandit interventions is to name the graph operation. Conditioning restricts a distribution. Intervention replaces a mechanism. Counterfactual reasoning updates exogenous uncertainty from evidence, changes a mechanism, then predicts.

observed association:      P(Y | X=x)
intervention question:     P(Y | do(X=x))
counterfactual question:   P(Y_x | E=e)
discovery question:        which G could have generated P(V)?

In machine learning, logged bandit interventions is valuable because models are often deployed under interventions: ranking changes, policy changes, safety filters, tool-use gates, data collection changes, and human feedback loops. Prediction alone does not tell us which change caused which downstream behavior.

Notebook implementation will use synthetic SCMs and small graphs. This keeps the examples executable while preserving the conceptual split between identification and estimation.

Checklist for using logged bandit interventions responsibly:

• State the causal question before choosing a method.
• Draw or describe the assumed causal graph.
• Mark observed, latent, treatment, outcome, and adjustment variables.
• Separate intervention notation from conditioning notation.
• Decide whether the query is identifiable before estimating it.
• Report assumptions that cannot be tested from the observed data alone.
• Use ML as an estimation aid, not as a substitute for causal design.

This chapter follows the boundary set by Chapter 21. Statistical learning theory controls prediction error under distributional assumptions. Causal inference asks what happens when the distribution changes because something is done.

Modern AI systems make this distinction unavoidable. A foundation model can predict which action historically followed a context, but a decision system needs to know what would happen if it took a different action in that context.

Thus, logged bandit interventions is not an abstract philosophical add-on. It is a production and research tool for deciding which model, prompt, policy, feature, or intervention actually changed an outcome.

A final diagnostic question is whether the claim would survive a policy change. If the answer depends only on a historical correlation, it belongs in predictive modeling. If the answer depends on what mechanism is replaced and which paths remain active, it belongs in causal inference.

7.3 prompt and tool interventions

Prompt and tool interventions belongs to the canonical scope of Do Calculus. The central move in causal inference is to distinguish a statistical relation from a claim about what would happen under an intervention.

For this subsection, the working scope is do-operator semantics, mutilated graphs, backdoor and frontdoor criteria, identification rules, and post-identification estimation. The mathematical objects are variables, mechanisms, graphs, interventions, and assumptions. A causal claim is incomplete until all five are visible.

$$P(Y \mid \operatorname{do}(X=x)) \ne P(Y \mid X=x) \quad \text{in general}.$$

The formula gives a compact handle on prompt and tool interventions. It should not be read as a purely algebraic identity. In causal inference, equations encode assumptions about mechanisms, missing variables, and which parts of the world remain stable under intervention.

Three examples of prompt and tool interventions:

1. A recommender team wants the causal effect of ranking a document higher, not merely the correlation between rank and clicks.
2. An LLM platform changes a safety policy and wants to estimate whether refusals changed because of the policy or because user prompts shifted.
3. A fairness auditor asks whether a proxy feature transmits an impermissible causal path into a model decision.

Two non-examples expose the boundary:

1. A high predictive coefficient is not a causal effect unless the graph and intervention assumptions justify it.
2. A plausible narrative produced by a language model is not a counterfactual unless it is grounded in a causal model.

The proof habit for prompt and tool interventions is to name the graph operation. Conditioning restricts a distribution. Intervention replaces a mechanism. Counterfactual reasoning updates exogenous uncertainty from evidence, changes a mechanism, then predicts.

observed association:      P(Y | X=x)
intervention question:     P(Y | do(X=x))
counterfactual question:   P(Y_x | E=e)
discovery question:        which G could have generated P(V)?

In machine learning, prompt and tool interventions is valuable because models are often deployed under interventions: ranking changes, policy changes, safety filters, tool-use gates, data collection changes, and human feedback loops. Prediction alone does not tell us which change caused which downstream behavior.

Notebook implementation will use synthetic SCMs and small graphs. This keeps the examples executable while preserving the conceptual split between identification and estimation.

Checklist for using prompt and tool interventions responsibly:

• State the causal question before choosing a method.
• Draw or describe the assumed causal graph.
• Mark observed, latent, treatment, outcome, and adjustment variables.
• Separate intervention notation from conditioning notation.
• Decide whether the query is identifiable before estimating it.
• Report assumptions that cannot be tested from the observed data alone.
• Use ML as an estimation aid, not as a substitute for causal design.

This chapter follows the boundary set by Chapter 21. Statistical learning theory controls prediction error under distributional assumptions. Causal inference asks what happens when the distribution changes because something is done.

Modern AI systems make this distinction unavoidable. A foundation model can predict which action historically followed a context, but a decision system needs to know what would happen if it took a different action in that context.

Thus, prompt and tool interventions is not an abstract philosophical add-on. It is a production and research tool for deciding which model, prompt, policy, feature, or intervention actually changed an outcome.

A final diagnostic question is whether the claim would survive a policy change. If the answer depends only on a historical correlation, it belongs in predictive modeling. If the answer depends on what mechanism is replaced and which paths remain active, it belongs in causal inference.

7.4 safety policy changes

Safety policy changes belongs to the canonical scope of Do Calculus. The central move in causal inference is to distinguish a statistical relation from a claim about what would happen under an intervention.

For this subsection, the working scope is do-operator semantics, mutilated graphs, backdoor and frontdoor criteria, identification rules, and post-identification estimation. The mathematical objects are variables, mechanisms, graphs, interventions, and assumptions. A causal claim is incomplete until all five are visible.

$$P(Y \mid \operatorname{do}(X=x))=\sum_z P(Y \mid X=x,Z=z)P(Z=z).$$

The formula gives a compact handle on safety policy changes. It should not be read as a purely algebraic identity. In causal inference, equations encode assumptions about mechanisms, missing variables, and which parts of the world remain stable under intervention.

Three examples of safety policy changes:

1. A recommender team wants the causal effect of ranking a document higher, not merely the correlation between rank and clicks.
2. An LLM platform changes a safety policy and wants to estimate whether refusals changed because of the policy or because user prompts shifted.
3. A fairness auditor asks whether a proxy feature transmits an impermissible causal path into a model decision.

Two non-examples expose the boundary:

1. A high predictive coefficient is not a causal effect unless the graph and intervention assumptions justify it.
2. A plausible narrative produced by a language model is not a counterfactual unless it is grounded in a causal model.

The proof habit for safety policy changes is to name the graph operation. Conditioning restricts a distribution. Intervention replaces a mechanism. Counterfactual reasoning updates exogenous uncertainty from evidence, changes a mechanism, then predicts.

observed association:      P(Y | X=x)
intervention question:     P(Y | do(X=x))
counterfactual question:   P(Y_x | E=e)
discovery question:        which G could have generated P(V)?

In machine learning, safety policy changes is valuable because models are often deployed under interventions: ranking changes, policy changes, safety filters, tool-use gates, data collection changes, and human feedback loops. Prediction alone does not tell us which change caused which downstream behavior.

Notebook implementation will use synthetic SCMs and small graphs. This keeps the examples executable while preserving the conceptual split between identification and estimation.

Checklist for using safety policy changes responsibly:

• State the causal question before choosing a method.
• Draw or describe the assumed causal graph.
• Mark observed, latent, treatment, outcome, and adjustment variables.
• Separate intervention notation from conditioning notation.
• Decide whether the query is identifiable before estimating it.
• Report assumptions that cannot be tested from the observed data alone.
• Use ML as an estimation aid, not as a substitute for causal design.

This chapter follows the boundary set by Chapter 21. Statistical learning theory controls prediction error under distributional assumptions. Causal inference asks what happens when the distribution changes because something is done.

Modern AI systems make this distinction unavoidable. A foundation model can predict which action historically followed a context, but a decision system needs to know what would happen if it took a different action in that context.

Thus, safety policy changes is not an abstract philosophical add-on. It is a production and research tool for deciding which model, prompt, policy, feature, or intervention actually changed an outcome.

A final diagnostic question is whether the claim would survive a policy change. If the answer depends only on a historical correlation, it belongs in predictive modeling. If the answer depends on what mechanism is replaced and which paths remain active, it belongs in causal inference.

7.5 causal debugging of model behavior

Causal debugging of model behavior belongs to the canonical scope of Do Calculus. The central move in causal inference is to distinguish a statistical relation from a claim about what would happen under an intervention.

For this subsection, the working scope is do-operator semantics, mutilated graphs, backdoor and frontdoor criteria, identification rules, and post-identification estimation. The mathematical objects are variables, mechanisms, graphs, interventions, and assumptions. A causal claim is incomplete until all five are visible.

$$P(Y \mid \operatorname{do}(X=x))=\sum_z P(Z=z \mid X=x)\sum_{x'}P(Y \mid X=x',Z=z)P(X=x').$$

The formula gives a compact handle on causal debugging of model behavior. It should not be read as a purely algebraic identity. In causal inference, equations encode assumptions about mechanisms, missing variables, and which parts of the world remain stable under intervention.

Three examples of causal debugging of model behavior:

1. A recommender team wants the causal effect of ranking a document higher, not merely the correlation between rank and clicks.
2. An LLM platform changes a safety policy and wants to estimate whether refusals changed because of the policy or because user prompts shifted.
3. A fairness auditor asks whether a proxy feature transmits an impermissible causal path into a model decision.

Two non-examples expose the boundary:

1. A high predictive coefficient is not a causal effect unless the graph and intervention assumptions justify it.
2. A plausible narrative produced by a language model is not a counterfactual unless it is grounded in a causal model.

The proof habit for causal debugging of model behavior is to name the graph operation. Conditioning restricts a distribution. Intervention replaces a mechanism. Counterfactual reasoning updates exogenous uncertainty from evidence, changes a mechanism, then predicts.

observed association:      P(Y | X=x)
intervention question:     P(Y | do(X=x))
counterfactual question:   P(Y_x | E=e)
discovery question:        which G could have generated P(V)?

In machine learning, causal debugging of model behavior is valuable because models are often deployed under interventions: ranking changes, policy changes, safety filters, tool-use gates, data collection changes, and human feedback loops. Prediction alone does not tell us which change caused which downstream behavior.

Notebook implementation will use synthetic SCMs and small graphs. This keeps the examples executable while preserving the conceptual split between identification and estimation.

Checklist for using causal debugging of model behavior responsibly:

• State the causal question before choosing a method.
• Draw or describe the assumed causal graph.
• Mark observed, latent, treatment, outcome, and adjustment variables.
• Separate intervention notation from conditioning notation.
• Decide whether the query is identifiable before estimating it.
• Report assumptions that cannot be tested from the observed data alone.
• Use ML as an estimation aid, not as a substitute for causal design.

This chapter follows the boundary set by Chapter 21. Statistical learning theory controls prediction error under distributional assumptions. Causal inference asks what happens when the distribution changes because something is done.

Modern AI systems make this distinction unavoidable. A foundation model can predict which action historically followed a context, but a decision system needs to know what would happen if it took a different action in that context.

Thus, causal debugging of model behavior is not an abstract philosophical add-on. It is a production and research tool for deciding which model, prompt, policy, feature, or intervention actually changed an outcome.

A final diagnostic question is whether the claim would survive a policy change. If the answer depends only on a historical correlation, it belongs in predictive modeling. If the answer depends on what mechanism is replaced and which paths remain active, it belongs in causal inference.

8. Common Mistakes

9. Exercises

1. (*) Work through a causal-inference task for do calculus.

   • (a) State the causal query using intervention or counterfactual notation.
   • (b) Draw or describe the relevant graph and assumptions.
   • (c) Decide whether the estimand is identifiable from the available data.
   • (d) Give an estimator or diagnostic only after identification is clear.
   • (e) Explain the AI or LLM system implication.

2. (*) Work through a causal-inference task for do calculus.

   • (a) State the causal query using intervention or counterfactual notation.
   • (b) Draw or describe the relevant graph and assumptions.
   • (c) Decide whether the estimand is identifiable from the available data.
   • (d) Give an estimator or diagnostic only after identification is clear.
   • (e) Explain the AI or LLM system implication.

3. (*) Work through a causal-inference task for do calculus.

   • (a) State the causal query using intervention or counterfactual notation.
   • (b) Draw or describe the relevant graph and assumptions.
   • (c) Decide whether the estimand is identifiable from the available data.
   • (d) Give an estimator or diagnostic only after identification is clear.
   • (e) Explain the AI or LLM system implication.

4. (**) Work through a causal-inference task for do calculus.

   • (a) State the causal query using intervention or counterfactual notation.
   • (b) Draw or describe the relevant graph and assumptions.
   • (c) Decide whether the estimand is identifiable from the available data.
   • (d) Give an estimator or diagnostic only after identification is clear.
   • (e) Explain the AI or LLM system implication.

5. (**) Work through a causal-inference task for do calculus.

   • (a) State the causal query using intervention or counterfactual notation.
   • (b) Draw or describe the relevant graph and assumptions.
   • (c) Decide whether the estimand is identifiable from the available data.
   • (d) Give an estimator or diagnostic only after identification is clear.
   • (e) Explain the AI or LLM system implication.

6. (**) Work through a causal-inference task for do calculus.

   • (a) State the causal query using intervention or counterfactual notation.
   • (b) Draw or describe the relevant graph and assumptions.
   • (c) Decide whether the estimand is identifiable from the available data.
   • (d) Give an estimator or diagnostic only after identification is clear.
   • (e) Explain the AI or LLM system implication.

7. (***) Work through a causal-inference task for do calculus.

   • (a) State the causal query using intervention or counterfactual notation.
   • (b) Draw or describe the relevant graph and assumptions.
   • (c) Decide whether the estimand is identifiable from the available data.
   • (d) Give an estimator or diagnostic only after identification is clear.
   • (e) Explain the AI or LLM system implication.

8. (***) Work through a causal-inference task for do calculus.

   • (a) State the causal query using intervention or counterfactual notation.
   • (b) Draw or describe the relevant graph and assumptions.
   • (c) Decide whether the estimand is identifiable from the available data.
   • (d) Give an estimator or diagnostic only after identification is clear.
   • (e) Explain the AI or LLM system implication.

9. (***) Work through a causal-inference task for do calculus.

   • (a) State the causal query using intervention or counterfactual notation.
   • (b) Draw or describe the relevant graph and assumptions.
   • (c) Decide whether the estimand is identifiable from the available data.
   • (d) Give an estimator or diagnostic only after identification is clear.
   • (e) Explain the AI or LLM system implication.

10. (***) Work through a causal-inference task for do calculus.

• (a) State the causal query using intervention or counterfactual notation.
• (b) Draw or describe the relevant graph and assumptions.
• (c) Decide whether the estimand is identifiable from the available data.
• (d) Give an estimator or diagnostic only after identification is clear.
• (e) Explain the AI or LLM system implication.

10. Why This Matters for AI

11. Conceptual Bridge

Do Calculus follows statistical learning theory because learning theory explains how observed samples support future prediction claims. Causal inference asks a different question: what happens when an action changes the system that generated those samples?

The backward bridge is risk and uncertainty. Chapter 21 provides language for finite-sample generalization. Chapter 22 adds intervention semantics, graph assumptions, and counterfactual worlds. A causal claim is not just a better prediction; it is a claim about a modified data-generating mechanism.

The forward bridge is game theory. Once multiple agents adapt to interventions, the causal question becomes strategic: actions change incentives, incentives change behavior, and behavior changes the causal system. Chapter 23 will study that interaction explicitly.

+--------------------------------------------------------------+
| Chapter 21: prediction under finite samples                  |
| Chapter 22: intervention, counterfactuals, causal discovery  |
| Chapter 23: strategic interaction and adversarial systems    |
+--------------------------------------------------------------+

References

• Pearl. Causal diagrams for empirical research. https://academic.oup.com/biomet/article-abstract/82/4/669/251647
• Pearl. Causal inference in statistics: An overview. https://projecteuclid.org/journals/statistics-surveys/volume-3/issue-none/Causal-inference-in-statistics-An-overview/10.1214/09-SS057.pdf
• Tian and Pearl. A General Identification Condition for Causal Effects. https://8www.aaai.org/Library/AAAI/2002/aaai02-085.php
• Hernan and Robins. Causal Inference: What If. https://www.hsph.harvard.edu/miguel-hernan/causal-inference-book/