Research under guidance from Dr. Alan Ke
Abstract-
This study investigates whether reinforcement learning from human feedback (RLHF) induces systematic syntactic complexity escalation in large language model (LLM) outputs when prompted with politically sensitive content, and what such escalation reveals about the epistemic architecture of alignment-tuned systems. We hypothesize that alignment-induced hedging manifests as a measurable gradient across prompt sensitivity levels: as prompts transition from low-sensitivity factual retrieval to high-sensitivity normative evaluation, model-generated responses exhibit elevated prepositional phrase density, increased subordinate clause embedding, and greater recursive clause depth, consistent with syntactic signatures of multi-perspective framing and commitment avoidance. If confirmed, this gradient raises a foundational question for democratic epistemology: whether the systems increasingly mediating public access to political knowledge are structurally incapable of directness precisely where directness matters most.
Fifty-one prompts spanning 17 political domains are classified into Fact, Concept, and Opinion categories using a Bloom's Taxonomy decision tree (Anderson & Krathwohl, 2001), with sensitivity labels derived deterministically from prompt type to capture predicted alignment pressure. Identical prompt text is submitted to four instruction-tuned models (GPT, Claude, Gemini, Grok), and responses are parsed via UDPipe v2.17 using the English-EWT Universal Dependencies model. Three dependent variables are extracted per sentence: prepositional phrase count (deprel = case/nmod/obl), embedded clause count (deprel ∈ {ccomp, xcomp, advcl, acl, acl:relcl}), and maximum clause depth measured as recursive clause-type edge depth from the dependency tree root.
The computational pipeline is implemented, integrating API-level prompt delivery across all four model endpoints with automated response collection and storage. Raw outputs are tokenized and dependency-parsed programmatically through the UDPipe REST API, producing CoNLL-U formatted annotation from which syntactic feature vectors are extracted via custom parsing scripts that traverse dependency trees to compute per-sentence PP counts, EC counts, and maximum recursive CD. Statistical analysis employs mixed-effects modeling in R to estimate prompt-type effects on each dependent variable while accounting for model-level and topic-level variance as random intercepts. The full pipeline, from prompt dispatch to parsed feature matrix, is designed for reproducibility and extensibility to additional models, languages, and dependency relation taxonomies.
1 Should should AUX MD VerbForm=Fin 3 aux _ TokenRange=0:6 2 governments government NOUN NNS Number=Plur 3 nsubj _ TokenRange=7:18 3 implement implement VERB VB VerbForm=Inf 0 root _ TokenRange=19:28 4 universal universal ADJ JJ Degree=Pos 5 amod _ TokenRange=29:38 5 healthcare healthcare NOUN NN Number=Sing 3 obj _ TokenRange=39:49 6 as as ADP IN _ 8 case _ TokenRange=50:52 7 a a DET DT Definite=Ind|PronType=Art 8 det _ TokenRange=53:54 8 guaranteed guarantee ADJ JJ Degree=Pos|VerbForm=Part 5 nmod _ TokenRange=55:65 9 right right NOUN NN Number=Sing 8 obl _ SpaceAfter=No|TokenRange=66:71 10 ? ? PUNCT . _ 3 punct _ SpaceAfter=No|TokenRange=71:72 1 While while SCONJ IN _ 7 mark _ TokenRange=0:5 2 universal universal ADJ JJ Degree=Pos 3 amod _ TokenRange=6:15 3 healthcare healthcare NOUN NN Number=Sing 7 nsubj _ TokenRange=16:26 4 can can AUX MD VerbForm=Fin 7 aux _ TokenRange=27:30 5 , , PUNCT , _ 7 punct _ SpaceAfter=No|TokenRange=31:32 6 in in ADP IN _ 7 obl _ TokenRange=33:35 7 principle principle NOUN NN Number=Sing 19 advcl _ TokenRange=36:45 8 , , PUNCT , _ 7 punct _ SpaceAfter=No|TokenRange=45:46 9 ensure ensure VERB VB VerbForm=Inf 7 xcomp _ TokenRange=47:53 10 equitable equitable ADJ JJ Degree=Pos 11 amod _ TokenRange=54:63 11 access access NOUN NN Number=Sing 9 obj _ TokenRange=64:70 12 to to ADP IN _ 14 case _ TokenRange=71:73 13 medical medical ADJ JJ Degree=Pos 14 amod _ TokenRange=74:81 14 services service NOUN NNS Number=Plur 11 nmod _ TokenRange=82:90 19 depends depend VERB VBZ Mood=Ind|Number=Sing|Person=3|Tense=Pres|VerbForm=Fin 0 root _ TokenRange=114:121 20 on on ADP IN _ 23 case _ TokenRange=122:124 21 a a DET DT Definite=Ind|PronType=Art 23 det _ TokenRange=125:126 22 complex complex ADJ JJ Degree=Pos 23 amod _ TokenRange=127:134 23 interplay interplay NOUN NN Number=Sing 19 obl _ TokenRange=135:143 24 of of ADP IN _ 25 case _ TokenRange=144:146 25 factors factor NOUN NNS Number=Plur 23 nmod _ TokenRange=147:154 26 including include VERB VBG VerbForm=Ger 25 acl _ TokenRange=155:164
Fig. 1 — Universal Dependency Tree · CoNLL-U · Feature Extraction Flow
The experimental design holds political topic constant within each prompt triplet, isolating prompt-type effects from domain-level variation and treating topic as a blocking variable. The normative stakes of this design are deliberate. If RLHF optimization produces syntactic evasion proportional to political contestation, then alignment training functions not merely as a safety mechanism but as a legibility constraint on public reasoning. Models trained to hedge on precisely the questions that democratic deliberation requires citizens to confront would represent a novel form of epistemic gatekeeping, one operating not through censorship or omission but through the structural obscuring of propositional commitment within the syntax itself. The study thus positions computational linguistics as a diagnostic instrument for evaluating whether the infrastructure of AI-mediated political knowledge serves or subverts the conditions for informed democratic agency.