AI Glossary / AI and machine learning / Advanced

STIT logic

Definition

STIT logic is a family of formal logics for reasoning about agency: an agent “sees to it that” some statement is true. It distinguishes an outcome merely occurring from an outcome being secured by an agent's choice within a specified model of possible developments.

What it means

The letters STIT abbreviate “sees to it that.” A statement such as “the operator sees to it that the valve is closed” expresses more than “the valve is closed.” It attributes a kind of control over the outcome to an agent.

The distinction matters in systems where several agents act and the future can develop in different ways. STIT provides precise languages and mathematical interpretations for questions about choices and outcomes.

It is a logic of agency, not a machine learning algorithm. Its conclusions follow from the modeled possibilities and assumptions. They do not, by themselves, establish a person's intention, moral responsibility, or legal liability.

How it works

A common framework represents time as branching histories. At a decision point, several complete possible histories pass through the same moment. An agent's available choices group these histories into alternatives.

A choice secures a proposition when that proposition holds throughout the histories compatible with the choice. Uncertainty can remain about other events; the relevant outcome must nevertheless hold across the chosen group.

Different STIT operators impose different conditions. A deliberative version additionally rules out outcomes already settled regardless of the agent's choice. This helps distinguish making a difference from receiving credit for something inevitable.

The analyst must specify the agent, moment, proposition, possible histories, and choice structure. Changing these can change what the model attributes to the agent. Formal precision makes assumptions inspectable; it does not remove the need to justify them.

Three practical examples

1. Workflow analysis: Designers distinguish a manager approving a request from merely being present when an independently approved request proceeds.

2. Multi-agent simulation: Researchers ask whether one simulated agent can secure a shared resource's availability despite another agent's choices.

3. Automation review: Engineers examine whether a controller's selected mode guarantees a modeled output condition across the remaining environmental possibilities.

Think of it this way

Think of choosing between two corridors. Every exit reachable through the left corridor leads outdoors; the right corridor can lead either outdoors or into another room. Choosing left secures reaching outdoors in this simplified map.

The analogy assumes the map is complete and that the traveler can follow the chosen corridor. Real actions can fail, and outcomes may depend on later decisions. Those possibilities must be represented before drawing a real-world conclusion.

Worked example

Imagine four possible histories at one decision point: h1, h2, h3, and h4. An operator has two modeled choices: A permits h1 or h2, while B permits h3 or h4.

Let p mean that a warning light is on at the relevant outcome. Suppose p is true in h1, h2, and h3, but false in h4.

Choice A guarantees p because both compatible histories satisfy it. Choice B does not: h4 remains possible even though h3 would produce p.

If A is chosen, this supports the ordinary choice-based STIT claim for p. It also meets the additional deliberative condition that p was not already settled, since h4 was possible before the choice.

Now change the model so p is true in all four histories. The outcome is inevitable at that moment. The deliberative interpretation no longer credits A with seeing to it that p, although the simpler guarantee condition still holds.

Illustrative case study

Fictional case: Lantern Logistics models a dispatch workflow before automating it. A supervisor can select “hold” or “release,” while a separate process can cancel an order.

The designers initially describe release as ensuring shipment. Enumerating possible histories reveals a compatible cancellation history in which shipment never occurs. Release therefore permits shipment but does not secure it in their model.

They revise the specification to distinguish authorizing dispatch from guaranteeing dispatch. The exercise clarifies the workflow's promises without asserting anything about legal responsibility. The fictional team uses the logic to inspect requirements, not to infer a person's motives from a system event log.

Common misconceptions and mistakes

An actual outcome is not enough to establish that an agent secured it. Alternative possibilities matter.

STIT is not simply a label for causation. Its formal treatment of choices and guarantees differs from many causal models.

Do not treat every STIT variant as interchangeable. Whether inevitable outcomes count is a substantive distinction.

Finally, a proof is conditional on its model. Excluding equipment failures or other agents' interventions can create guarantees that disappear when those possibilities are restored.

Business use

The most relevant business use is careful specification of responsibility boundaries in workflows and autonomous systems. Ask what an action authorizes, enables, or guarantees, and represent those distinctions explicitly.

This usually requires formal-methods expertise. Begin with a small decision model and compare its possibilities with operational knowledge. Use the result to improve requirements and identify missing assumptions. Do not convert a formal agency statement into an automatic personnel judgment or a claim that a deployed system cannot fail.

People also ask

Does STIT mean an agent intended the outcome?

Not automatically. Intention requires additional concepts or modeling assumptions.

Can several agents be modeled?

Yes. Multi-agent variants represent choices and interactions, with assumptions that must be specified.

Does the logic predict what people will choose?

Basic STIT reasoning analyzes agency and possibilities; it does not itself provide a statistical behavioral forecast.

Check your understanding

A chosen action allows two histories, one with success and one with failure. Success actually occurs. Did the action guarantee success in this model?

Answer: No. The compatible failure history means success was not secured by that choice, even though the actual outcome was successful.

Related terms

• Modal logic
• Temporal logic
• branching time
• formal verification
• agency
• multi-agent system
• deontic logic
• causal model

Learn next

Learn modal logic's distinction between possibility and necessity. Then draw a small branching-time model with two choices and test which propositions each choice guarantees. Compare the choice-based and deliberative interpretations when an outcome holds in every history.

• Modal logic

More in AI and machine learning

• Statistical learning theory - Statistical learning theory studies how and when a model learned from finite data can perform well on new observations.
• Statistical relational learning - Statistical relational learning is a family of methods that learn and reason under uncertainty while representing relationships among entities.
• Step size - In optimization, step size controls how far an update moves along a chosen search direction.
• Stochastic program - A stochastic program is a statistical model expressed as a procedure that makes random draws.
• Stochastic variance reduction - Stochastic variance reduction is a collection of optimization techniques that reduce randomness in estimated gradients while retaining much of the...
• Streaming Feature Pipeline - A streaming feature pipeline continuously processes incoming events into features that machine-learning systems can use.

Sources

Brian F. Chellas, Time and Modality in the Logic of Agency: https://web.stanford.edu/class/cs222/Chellas.pdf. Choice partitions, historical alternatives, and settledness passages independently consulted. Horty and Belnap, The Deliberative Stit, indexed primary text checked for the negative requirement; full PDF retrieval failed. The four-history truth conditions were independently checked. The workflow case is fictional, with no inferred intention or legal responsibility.

Previous: Stereo cameras Next: Stixel