Human–AI interaction forensics glossary

Which questions belong in this field?

Ask which version of a conversation is represented and where it came from. An interface can display one selected response while an export retains several alternatives. A message’s position in a flat list does not necessarily identify its parent or the context supplied to a model.

Distinguish visible text from runtime context. Instructions, retrieved material, summaries, tool output, and attachments may affect an exchange without appearing in the user-facing transcript. State which of those inputs were acquired and which remain unknown.

Evidence sources and examination context

Preserve identifiers, parent references, message roles, timestamps, and revision information when the source provides them. Keep the raw representation alongside a readable transcript so that analytical ordering remains traceable.

Build relationships from explicit references first. If an edge is inferred from timing or text, label it as an inference. W3C PROV provides a general model for relationships between entities and activities; application-specific message semantics still require separate validation. PROV-DM: The PROV Data Model.

When investigating prompt injection, distinguish encountered instructions from subsequent behavior. OWASP describes direct and indirect injection risks. Establishing that an attempt caused an action requires the acquired sequence, the applicable permissions, and evidence of the action itself. LLM01:2025 Prompt Injection.

A worked evidence example

Hypothetical example: a user edits an earlier prompt and generates a second answer. A flat export lists both answers after the final prompt text. Without parent and revision relationships, a reader might incorrectly infer that both answers followed the same input.

A reconstructed graph preserves each recovered node and the supported edges. An absent parent stays absent. The analyst does not invent its wording from a later answer, even when that answer seems to imply a particular question.

What this evidence can and cannot establish

Recorded exchanges can support findings about text and relationships within the recovered material. They can help identify an input that preceded an action or expose a gap in the visible record.

A response cannot reliably reveal every instruction that produced it. A saved message does not by itself establish the user’s intention, the full model context, or whether a suggested action occurred. Distinguish the source transcript from the reconstruction.

Essential terminology

• Human-AI interaction — An exchange between a person and an AI-enabled system, including available context, responses, revisions, and associated actions.
• Prompt reconstruction — An evidence-based effort to recover or connect recorded inputs and their context without inventing missing text.
• Conversation graph — A representation of messages and relationships such as parent links, branches, or revisions.
• Context window — The model’s available input context for a particular step, subject to system limits and any truncation or summarization.
• Prompt injection — Input that attempts to redirect an AI system through instructions outside the intended control channel.
• Attachment — An object associated with a message, task, or other application record.
• Timeline reconstruction — An evidence-referenced arrangement of events and temporal relationships with clock assumptions, gaps, and uncertainty documented.
• Inference — A reasoned interpretation supported by evidence and stated assumptions, rather than a directly observed fact.

All 49 terms in this evidence field

• Assistant response
• Attachment
• Authenticity calibration
• Automation bias
• Branch
• Competing hypothesis
• Completion status
• Confabulation
• Context summarization
• Context truncation
• Context window
• Conversation branch
• Conversation export
• Conversation graph
• Developer instruction
• Evidence reconstruction
• Human-AI interaction
• Indirect prompt injection
• Inference
• Interaction forensics
• Message identifier
• Parent message
• Prompt
• Prompt history
• Prompt injection
• Prompt reconstruction
• Prompt template
• Reality-model drift
• Regeneration
• Response revision
• Retrieval query
• Retrieval-augmented generation
• Retrieved document
• Semantic similarity
• Source memory
• Source-monitoring error
• Stop sequence
• Streaming response
• Synthetic imprinting
• System instruction
• Temperature
• Timeline reconstruction
• Timestamp normalization
• Token
• Tokenization
• Top-p sampling
• Upload event
• User message
• Vector database

Continue with an examination reference

• reconstruction
• interaction forensics
• human ai interaction

Citation and editorial context

This is an editorial reference. Source checks are separate from independent expert review; no independent reviewer is recorded.

Examples are hypothetical. See the editorial policy, methodology, and correction process.

Shelley, Jesse (ed.). “Human–AI interaction forensics glossary.” Synthetic Forensics, 1.1.0-glossary-library-1, 2026-10-09. https://syntheticforensics.org/glossary/interactions/ Accessed: add the date you consulted this page.

The complete vocabulary

Browse the full glossary of 300 terms or download the machine-readable vocabulary.

Field assignment is the project’s organizational taxonomy; related evidence can span several fields.