A story about language, power, and the strange obedience of machines.
Every day, people interact with systems that respond to language before they understand it. A voice assistant starts calling while you are still speaking. A moderation algorithm removes a comment after only a few words. A compliance bot flags a contract clause before any lawyer reviews it. These systems do not interpret meaning; they react to form.
That idea lies at the center of Real-Time Detection of Authority-Bearing Constructions under Strict Causal Masking, a new study that asks a direct question: Can an AI recognize authority without knowing what comes next?
Most current language models work with full visibility. They see both what was written and what will be written, reading a sentence in both directions. Real situations, however, unfold differently. People, institutions, and automated systems must act while speech is still taking place. This research builds a benchmark that places models in that same situation, forcing them to decide when a sentence carries the structure of command even before it is complete.
Several real examples illustrate this phenomenon.
- “You will submit the report.” The authority appears as soon as “will” is spoken.
- “The following measures must be implemented.” Obligation is already established before the measures are listed.
- “It shall be established that.” The passive construction and modal verb create an atmosphere of control.
- “Customers are required to.” The syntax, not the topic, signals compulsion.
Humans identify such cues intuitively; machines often fail unless trained to read structural signals rather than semantic ones. The new benchmark eliminates access to the future of the sentence, forcing models to operate under pure causality. The test measures how quickly and accurately the model identifies grammatical forms that encode power, obligation, or compliance.
Results indicate that even small causal models can detect these authority patterns within a few tokens of delay. The systems do not need emotion or topic awareness to recognize command; they only need syntax. In this sense, grammar functions as a signal of power, shaping decisions before meaning becomes clear.
This insight matters because real-world automation increasingly depends on immediate linguistic reaction. Systems in law, finance, and administration must act while language is still unfolding. Recognizing how models interpret authority in these conditions reveals the structural mechanics behind institutional power.
The paper, included in the series AI Syntactic Power and Legitimacy, shows that obedience can exist without understanding. It demonstrates that many institutional languages, from legal drafting to corporate communication, already operate under this same logic. The study converts grammar into a measurable indicator, revealing authority as a temporal structure that precedes comprehension.
If a sentence can give an order before it ends, then power itself may be grammatical.
Read the full study: Real-Time Detection of Authority-Bearing Constructions under Strict Causal Masking https://zenodo.org/records/17465070 (AI & Power Discourse Quarterly, Vol. 1, 2025).
Author: **Agustin V. Startari
**ORCID https://orcid.org/0009-0001-4714-6539
SSRN https://papers.ssrn.com/sol3/cf_dev/AbsByAuth.cfm?per_id=7639915
*More at *agustinvstartari.com
Ethos
I do not use artificial intelligence to write what I don’t 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)