Is this the same as theory of mind?

They overlap: theory-of-mind tasks involve representing others' mental states. Psychological reasoning is used here as a broader AI modeling term.

Can a false belief explain a sensible action?

Yes. An action can be reasonable relative to an agent's information even when that information is outdated or incorrect.

Does success on a simple belief-tracking task prove general social understanding?

No. Performance in a controlled task does not establish reliable understanding across unrestricted situations.

AI Glossary / Graphs and symbolic AI / Intermediate

Psychological reasoning

Definition

Psychological reasoning, in symbolic AI, is reasoning about agents in terms of what they know, believe, want, intend, and are likely to do. It uses a model of mental states to explain or anticipate behavior, including situations where an agent's information differs from reality. The term here does not refer to mental-health diagnosis or to proving that a machine has human feelings.

What it means

An observer may know where an object is while another agent holds an outdated belief about its location. Predicting that agent's action requires representing the agent's perspective, not merely applying the observer's own knowledge.
Goals also matter. Two agents with the same information can choose different actions because they want different outcomes or face different constraints. Psychological reasoning therefore asks how information, preferences, and decision assumptions work together. The resulting model is an explanatory or predictive tool. Its use of words such as “belief” and “intention” does not by itself establish consciousness or direct access to a person's private thoughts.

How it works

Identify the agents, the information each could observe, their possible goals, and a model connecting those states to actions. Update each agent's represented beliefs when it receives relevant information. Keep those beliefs separate from the system's representation of the actual world.
Reasoning can proceed forward, predicting an action from assumed beliefs and goals, or backward, inferring plausible goals from observed action. Inverse-planning models take the latter approach by comparing how well different goals explain behavior under a planning model. Several explanations can fit one action, so uncertainty and competing hypotheses may need to be retained rather than forcing one confident interpretation.

Three practical examples

A cooperative game agent can anticipate a teammate's move using what the teammate has seen rather than information available only to itself.
A dialogue assistant can track whether a previously communicated instruction has reached the person currently carrying out a task.
A teaching simulator can demonstrate why an agent searches an outdated location when it did not observe an object's relocation.

Think of it this way

Think of keeping separate notebooks for the shared world and for each participant's understanding of it. An event changes the world notebook immediately, but a participant's notebook changes only through an appropriate information route. The analogy is limited because real beliefs are uncertain, sometimes inconsistent, and not literally stored as neat written lists that an observer can inspect.

Worked example

Consider an invented two-room simulation. At time 1, a tool is in Room A, and both Mira and Sol see it there. At time 2, Mira leaves. At time 3, Sol moves the tool to Room B without sending a message. At time 4, Mira returns and wants the tool.
The world state now says Location(tool) = B. Sol's represented belief also says B, while Mira's represented belief still says A. Under the explicit assumptions that Mira remembers her last observation, trusts it until updated, and searches the believed location first, the predicted first search is Room A. Predicting Room B solely from the actual location would confuse knowledge available to the model with information available to Mira.
Now add a reliable message received by Mira before she searches: “The tool is in Room B.” If the model assumes she understands and accepts it, her represented belief changes to B, and the predicted first search changes to Room B. If message delivery is unknown, both cases must remain possible. The scenario tests perspective tracking, not a universal claim about human behavior.

Illustrative case study

Fictional case: Maple Workshops develops a cooperative assembly simulation. Its assistant repeatedly tells a simulated worker to continue with a revised instruction, although that worker's communication channel has not delivered the revision.
The team adds an explicit distinction between sending and receiving information. It then tests separate cases for delivered, delayed, and missing messages. This fictional example shows how reasoning about another agent's information can improve the coherence of a model. It does not report a measured improvement, diagnose a person, or establish that an AI understands human minds in general.

Common misconceptions and mistakes

Do not assume another agent knows everything the system knows. Do not infer a unique goal from one action when several goals, constraints, habits, or mistakes could explain it. A model assuming approximately efficient planning can be informative in its intended setting without being a universal description of people.
Avoid treating a plausible explanation as an observed fact. Belief attribution should remain linked to the evidence and assumptions used. Nested claims such as “Mira believes Sol knows” add complexity and can be mistaken. Changing the vocabulary to mental-state terms does not remove the need to test the model against the behavior it is meant to explain.

Business use

For collaborative software, psychological reasoning is most useful when it addresses a concrete coordination question: what information has a participant received, which goal is active, and what clarification might resolve an ambiguity? A small explicit model may be more inspectable than an unsupported statement about what someone “must want.”
Evaluate with scenarios that separate world state, observation, and communication. Include cases where the same action supports different explanations. Present uncertain interpretations as possibilities and allow correction. For business users, a helpful outcome is often a timely clarification or a more accurate record of shared information rather than a confident personality judgment.

Check your understanding

Why does the model predict that Mira first searches Room A before receiving the message?

Answer: Her last observed location was A, and she did not observe the move. Under the stated memory and search assumptions, her action follows her represented belief, even though the actual tool location is B.

Related terms

Common ground
Belief
knowledge
intention
goal recognition
theory of mind
inverse planning
epistemic reasoning
agent modeling

Learn next

Learn epistemic reasoning and inverse planning next. Modify the example so Mira sees Sol enter Room B but does not see the tool move. List at least two explanations compatible with that observation, and identify what additional evidence would distinguish them.

Sources

Chris Baker, Rebecca Saxe and Joshua Tenenbaum, Action understanding as inverse planning, original research paper: https://web.mit.edu/9.s915/www/classes/cognition2009.pdf Russell and Norvig, AIMA textbook passage defining psychological reasoning, hosted copy: https://ggnindia.dronacharya.info/Downloads/Sub-info/NOTES/Ai-ML-e-book.pdf

Previous: Pseudoword Next: PTRANS