Reality Contact
In one sentence
Reality contact is the practical capacity of a person or AI system to keep its claims open to correction by external evidence, other observers, and observed outcomes.
A claim can feel quite solid while it remains inside a conversation. Then someone opens the source, runs the calculation, calls the person involved, or waits to see what actually happens. Reality contact is the continuing arrangement that allows this encounter to change the claim.
Correctness at one moment is not enough. A guess can be right by accident, and a careful inquiry can initially be wrong. The more durable test is whether the process remains open to correction from outside itself.
Why the term exists
Several neighbouring fields already have useful words for parts of this problem.
In psychology, reality monitoring refers to processes used to judge whether information arose internally—from imagination, thought, or memory—or externally through perception. Clinical descriptions of psychosis use the related phrase loss of contact with reality, although hallucinations, delusions, disorganised thinking, and other experiences remain distinct phenomena. These terms concern human cognition, perception, distress, and care.
In artificial intelligence, grounding connects a model’s output to retrieved documents, databases, tools, or observations. Retrieval-augmented generation gives a model access to information outside its parameters. Agent designs such as ReAct interleave reasoning with actions that gather information from an external source or environment. These are engineering methods. They do not give a model human perception or turn an unsupported output into a clinical symptom.
Reality contact is the Dictionary’s extension across those domains. It names the broader operating requirement: a reasoning process should remain answerable to something it did not generate itself. The comparison is functional, not medical. Human reality monitoring and machine grounding may sometimes play analogous roles, but their mechanisms, meanings, and consequences differ.
What it actually does
Reality contact requires a correction loop:
- Make the claim visible. State what the person, model, or team currently believes, including material uncertainty.
- Expose it to an independent constraint. Check a primary source, inspect the object, run the code, reproduce the calculation, ask another observer, or test the prediction against an outcome.
- Compare rather than decorate. Retrieved material is useful only if someone checks whether it supports the claim. A citation attached to a sentence is not evidence that the sentence follows from it.
- Update the claim or abstain. The system must be able to correct, narrow, defer, or withdraw an answer without treating the change as failure.
- Preserve the correction. A corrected answer that disappears from memory leaves the same error available for rediscovery tomorrow.
The loop can operate at several scales. A person can check a memory against a photograph or another witness. A research group can expose a hypothesis to replication. An AI agent can retrieve a document, use a calculator, inspect a file, or observe whether an action succeeded. An institution can compare its official account with complaints, audits, market outcomes, and the experience of people who live under its policies.
Each case asks the same narrow question: what can still correct us?
A hypothetical example
An assistant reports that a model has 50 billion active parameters. The reviewer opens the model card and finds a different figure. They check whether the figures refer to the same model version and whether each counts total or active parameters, then correct the report.
The model card is a primary description, not an independent evaluation of performance. A benchmark claim would require a separate check of the test, comparison, and reporting source. Consulting a source helps only when the reviewer checks what it actually supports.
Why it matters in teaching and management
A coherent first draft still needs to be checked against the evidence.
A student who submits an elegant analysis without checking the case facts has produced a performance artifact. A manager who accepts a sourced-looking agent report without opening its sources has delegated prose production, not judgement. A team that evaluates an agent only on whether its answer sounds right will reward fluency even when the workflow provides no route from claim to evidence.
Reality contact changes the design question. Instead of asking only whether the model is capable, ask what the working system can observe, which sources it may consult, who can challenge it, how outcomes return as feedback, and where corrections are recorded. Provenance, RAG, tool use, approval gates, and verification loops are different ways of keeping the work exposed to constraints beyond the generator.
The same discipline applies to the human operator. A system does not gain reality contact merely because a person is in the loop. People defend preferred stories, misremember events, defer to polished outputs, and agree with one another for social reasons. Human review helps when the reviewer can see the evidence, has permission to disagree, and bears some responsibility for the consequence.
Trade-offs and warnings
Consensus is not reality. Other people can be mistaken together. Institutional records can be incomplete, measurements can be noisy, and primary sources can describe themselves strategically. Multiple observers help most when they bring genuinely independent evidence rather than copies of the same account.
Grounding is not a guarantee. Retrieval can return the wrong document. A tool can fail. A database can be stale. An agent can cite a source that does not support its sentence. Reality contact depends on the quality of the entire correction loop, not the presence of a search box.
Checking has a cost. A workflow that independently verifies every harmless sentence may become unusably slow. The depth of checking should rise with consequence, irreversibility, novelty, and uncertainty. A classroom brainstorm and a medication decision do not require the same burden of proof.
Clinical language requires care. A person experiencing hallucinations or delusions is not behaving like a defective chatbot, and an inaccurate model output is not suffering a break with reality. The human terms involve subjective experience, social meaning, possible distress, and clinical responsibility. This entry supplies no diagnosis or treatment advice.
Contact can be painful. Evidence sometimes removes a cherished explanation without supplying a better one. Corrigibility therefore requires more than access to facts; it requires enough safety, time, and dignity for a person or organisation to revise without being destroyed by the admission.
Reality contact does not guarantee correctness. It provides a way to detect errors and revise conclusions.
See also
Sources
- National Institute of Mental Health, “Understanding Psychosis”.
- Jane Garrison et al., “Voices and Reality Monitoring”, in Voices in Psychosis (Oxford University Press, 2022).
- Marcia Johnson and Carol Raye, “Reality Monitoring”, Psychological Review 88 (1981).
- Patrick Lewis et al., “Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks”, NeurIPS (2020).
- Shunyu Yao et al., “ReAct: Synergizing Reasoning and Acting in Language Models”, ICLR (2023).