Why pronouns, demonstratives, and tenses in AI text can create institutional legitimacy even when they point to nothing verifiable
Problem in plain terms
Language that points is supposed to point to something. Pronouns such as I, we, and you presuppose speakers and addressees. Demonstratives such as this and those presuppose objects in a shared space. Temporal markers such as now and currently presuppose a moment that can be verified. Predictive language models routinely produce these forms because they are statistically plausible continuations of text. The forms survive. The referents do not.
When models write "We find the evidence sufficient," or "The patient is now stable," or "These measures will guarantee compliance," those sentences look like institutional speech. They look authoritative. Yet in many cases the words float: there is no court, no clinical observation, no deliberated policy behind them. The indexicals function as if they anchored reality, but they do not. That systematic phenomenon is Indexical Collapse.
Why it matters
1. Authority without accountability. Institutional audiences often accept texts that carry the linguistic signals of authority. When indexicals simulate institutional voice, readers and downstream systems may treat output as authoritative. Decisions, records, and actions can follow, despite the absence of verifiable grounding.
2. Domain risk varies but can be catastrophic. In low-stakes customer support, an unanchored we or now may be an annoyance. In medicine, a temporal marker misrepresenting current status can drive harmful clinical choices. In law, AI-generated transcripts or draft judgments that simulate judicial voice risk misattribution of legal authority.
3. Regulation and evaluation will need new dimensions. Conventional measures of model quality emphasize factual accuracy, hallucination rates, and safety classifiers. They do not capture whether indexicals are properly anchored. Pragmatic auditing must become part of evaluation frameworks for systems used in high-stakes environments.
What the article does
- Defines Indexical Collapse as the systematic disappearance of referents in predictive outputs while indexical forms persist as grammatical signals.
- Classifies manifestations across pronouns, temporal markers, and demonstratives, and proposes a three tier taxonomy: minimal collapse, intermediate collapse, and complete collapse.
- Provides cross-sector case studies from judicial transcripts, automated medical reports, administrative records, and conversational agents, showing a consistent pattern.
- Proposes pragmatic auditing: a stepwise method to identify, classify, quantify, and set normative thresholds for unanchored indexicality in outputs.
- Offers provisional thresholds tied to institutional stakes: higher tolerance in low-stakes settings, near zero tolerance in high-stakes settings such as courts and clinical reporting.
For the full academic argument and the methodological appendices consult the preprint on Zenodo: https://zenodo.org/records/17226412.
Examples that make the core point
- Judicial transcript example Model output: "We find the defendant guilty. This evidence proves intent." Issue: the pronoun we and the demonstrative this project collective deliberation and a specific evidentiary object. If these phrases are used in an AI draft that is accepted as authoritative without human confirmation, the text can be treated as a court voice even when no judgment was deliberated.
- Medical report example Model output: "The patient is now stable. We recommend further testing." Issue: now implies contemporaneous observation, and we implies a clinical team. If the model has not accessed live monitoring data or clinician verification, the report can misrepresent patient status and induce inappropriate interventions.
- Administrative minutes example Model output: "These measures will increase efficiency and will be implemented next quarter." Issue: these measures implies specified policies. If no policy decisions exist, the minutes can create the illusion of formal decisions, triggering procedural or financial consequences.
- Chatbot example Model output: "I understand your concern; we will resolve this now." Issue: I and we simulate a responsible agent and a live process. Without links to authenticated workflows or human agents, the language misleads users about actual service actions.
**How pragmatic auditing works, in practice
**A pragmatic audit is not linguistic navel-gazing. It is an operational checklist that can be automated or human mediated. Core steps:
- Identify all indexical items in the text, including pronouns, demonstratives, temporal adverbs, and evidentials.
- Classify anchoring of each item as anchored, ambiguously anchored, or unanchored. Anchored means the referent is explicit and verifiable within the document or linked data. Ambiguous means recoverable with minimal context. Unanchored means no local or retrievable external referent.
- Quantify unanchored indexicals relative to total indexicals; compute a collapse ratio.
- Apply domain threshold: compare the collapse ratio to a domain threshold. For example, in clinical reports the threshold should approach zero; in customer chat logs some small tolerance may be acceptable.
- Intervene: where thresholds are breached require human review, deny publication, or block automated actions tied to the text.
This method turns the abstract problem into measurable compliance checks that can be integrated into deployment pipelines, content governance systems, and regulatory oversight.
*Policy and governance implications
*
- Requirement for provenance metadata. Institutional uses must include explicit provenance fields identifying human reviewers, data sources, and timestamps that anchor indexicals. AI outputs lacking provenance cannot be treated as authoritative records.
- Certification for high-stakes deployments. Systems used in courts, clinical workflows, and public policy drafting must pass pragmatic audits and display a certification stamp that indicates compliance.
- Legal accountability. Institutions must avoid automatic ingestion of AI-generated documents as binding records. Legal frameworks should require human attestation where indexicals imply institutional voice.
- Design standards for conversational agents. Agents used in health, finance, or legal customer-facing contexts should avoid unqualified indexicals unless backed by verified state or live process hooks.
*Practical recommendations for organizations
*
- Instrument pipelines to detect indexical collapse early, using rule-based detection combined with domain-specific heuristics.
- Require human in the loop for outputs crossing a collapse threshold relevant to the domain.
- Log and publish provenance for records, including reviewer identifiers and data sources that justify indexicals.
- Train users and stakeholders to read indexical cues critically and to require confirmation for items implying institutional decisions.
**Academic citation and how to read further
**Primary preprint:
Startari, A. V. (2025). Indexical Collapse: Reference Disappears, Authority Remains in Predictive Systems.
Zenodo. https://doi.org/10.5281/zenodo.17226412
Key related readings summarized in the preprint include foundational work on indexicality and language and contemporary reflections on language and institutional power. For academic follow up consult the references in the Zenodo record.
Author metadata and access
**Author: Agustin V. Startari
ORCID: https://orcid.org/0009-0001-4714-6539
ResearcherID: K-5792–2016
Zenodo record: https://zenodo.org/records/17226412
SSRN Author Page: https://papers.ssrn.com/sol3/cf_dev/AbsByAuth.cfm?per_id=7639915
**Mini bio: Agustin V. Startari is a linguistic theorist and researcher in historical studies. He is author of Grammars of Power, Executable Power, and The Grammar of Objectivity. His work examines how language form produces institutional authority and how syntactic structures mediate legitimacy.
Call to action
Read the full paper on Zenodo (10.5281/zenodo.17226412). If you work in law, medicine, public administration, or platform governance and you are evaluating AI text for institutional use, adopt pragmatic auditing as part of your compliance architecture. For collaboration, datasets, or to propose case studies for cross-sector audits, consult the Zenodo record and the SSRN author page listed above.
Ethos
I do not use artificial intelligence to write what I do not know. I use it to challenge what I do. I write to reclaim the voice in an age of automated neutrality. My work is not outsourced. It is authored.
- Agustin V. Startari
Top comments (0)