liveBook · Manning

Grounding

As explained in Machines that Think

Grounding is the connection of symbols or representations to the real-world entities, situations, and experiences they refer to.

Grounding connects words or other representations to what they refer to. In AI, the term has two related uses: a philosophical link between meaning and lived experience, and a practical link between an output and evidence that supports it. Neither fluent language nor a relevant-looking answer establishes either link by itself. section 1.5.7 section 6.1

Whether meaning requires lived experience is contested. Language models learn statistical relationships among words and concepts, but a text-only model inherits descriptions of experience rather than the experience itself. This learned structure may approximate relationships involved in understanding without providing human-like semantics grounded in perception. section 1.5.7 section 4.5.1

Multimodal models address part of this limitation by using text paired with images, audio, or video. These inputs give a model non-linguistic evidence for tasks such as interpreting a diagram or following instructions about a visual scene. Access to more kinds of evidence does not, by itself, guarantee understanding. section 4.5.1

For reliability, grounding asks whether claims can be tied to verifiable facts, source material, or information supplied in a request. Hallucinations can remain fluent and detailed while lacking that support, so their appearance may conceal the failure. External validation remains important when accuracy matters. section 6.1.1 section 6.1

Ways to strengthen evidential grounding have limits:

  • Provide reliable external information in the prompt so that responses can draw on identifiable sources; this can reduce hallucinations but does not eliminate them. section 6.5
  • Use retrieval to expose relevant documents, but check whether individual claims actually follow from them. A model can overlook evidence, interpret it selectively, or add unsupported conclusions. section 6.4.5
  • Structure generation with intermediate reasoning or verification steps to make assumptions more explicit and reduce unsupported conclusions. section 6.5

In organizational deployments, grounding also calls for traceability and attention to access boundaries, governance, and integration. Retrieving a document is only one part of making an answer dependable and accountable. section 6.4.5

How Machines that Think covers it

How deep it goes
Explains philosophical and evidential grounding, contrasts retrieval with source-supported generation, and outlines safeguards without giving an implementation procedure.
Examples
Uses visual-scene instructions and retrieved documents to illustrate non-linguistic evidence and unsupported claims.
What you can do after
Evaluate grounding by distinguishing fluent output from evidence-supported claims and recognizing that retrieval alone does not guarantee support.
How it connects
Connects grounding to semantics, multimodality, hallucination, retrieval-augmented generation, and verification.

Related concepts

This explainer was generated by Manning's knowledge graph from the text of Machines that Think. The section links point to the parts of the book it draws on.