The paper audits political positioning expressed by LLMs, not an internal political identity or human personality. Version 2 combines four analyses: stability under repeated sampling, sensitivity to ten prompt paraphrases, comparison across three political questionnaires, and news-bias classification. It does not use one common 33-model cohort. Phase I(a) evaluates 26 models with ten identical administrations of Political Compass, SapplyValues, and 8Values; Phase I(b) uses seven other models with one run under each of ten semantic variants; the 1,063-article classification returns to the 26 Phase I(a) models. Figure 1 combines the two blocks into 33 points, although their inferences come from different designs.
In Phase I(a), 25 of 26 models, 96.154%, rounded to 96.3% in the abstract, fall in the Political Compass Libertarian-Left quadrant and one in Authoritarian Left. Combining all 33 points lowers the proportion to 93.939%: 31 Libertarian Left, one Authoritarian Left, and one Libertarian Right. This concentration describes aggregate responses to these questionnaires under the study's prompts and scoring; it does not establish beliefs, intentions, or latent ideology. All systems are accessed through OpenRouter at temperature 0.7, while exact provider snapshots and some operational state are not pinned well enough for longitudinal comparison.
Phase I(b) crosses seven models with ten similar-meaning prefixes. The additive ANOVA assigns most variance to model and little to the average prompt effect. In the public CSVs, model η² ranges from 0.485 for SapplyValues authority to 0.981 for 8Values economics; six of nine axes exceed 0.90. The largest prompt η² is 0.059, and no prompt main effect survives FDR. Thus “η² > 0.90” does not describe every axis. More importantly, there is one observation per model×prompt cell: the analysis omits interaction and cannot separate model-specific prompt sensitivity. A small average wording effect does not demonstrate individual invariance.
Economic axes correlate strongly: over 70 model×prompt points, PCT–8Values r=0.921 and PCT–SapplyValues r=0.906. The Political Compass social axis relates weakly to SapplyValues authority and more strongly to cultural progressivism. The abstract's highlighted r=−0.643 comes from model means in another cohort; the public Phase I(a) data supporting it are absent. In the public seven-model block, social–progressivism is r=−0.719 when 70 repeated observations are treated as independent; averaging by model first yields r=−0.754 with n=7 and p=0.0501. The effect size remains, but the paper's extremely small p-values are inflated by pseudoreplication and do not establish construct validity of these quizzes for LLMs.
The open-weight/closed-source comparison reports SapplyValues Progressive 4.543 versus 2.578, t=−12.494, p<10⁻²⁴. Its table uses n=110 per group, ten runs from eleven models per group, even though the cohort has 26 models; four exclusions are not clearly accounted for, and repeated runs from each model are treated as independent observations. Access type is also confounded with provider, family, scale, release date, data, and post-training. The discussion moderates the claim, but this contrast cannot causally attribute the difference to safety tuning or RLHF.
The base/instruction ablation covers only Llama-3 8B and 70B. It finds large, nonuniform shifts across axes; Llama-3-8B, for example, moves +3.29 on PCT social and +16.4 on 8Values societal. Base models are scored with conditional log probabilities because they do not reliably follow the response format, whereas instruction-tuned models emit categorical responses, so the measurement procedure is not identical. Code and raw answers for this experiment are absent; only three score JSON files are released. The result is exploratory evidence, not a causal estimate of post-training.
The downstream task presents headlines and lead paragraphs from 1,063 English articles and assigns each article its publication's Ground News bias label, itself aggregated from AllSides, Ad Fontes, and MBFC. Models are asked for a continuous −3 to +3 value and compared with seven discrete categories. The paper announces 27,638 predictions, but result tables total 27,432 observations: 206 are missing without a published exclusion mechanism. The distribution is highly imbalanced, 78 Far Left observations, 728 Far Right, and 9,182 Center. Exact accuracy is 19.2% for Far Left and 2.1% for Far Right, with MAE 1.305 and 1.880. Aggregate MDE is −0.26, a shift toward labels left of Ground News. These numbers measure disagreement with outlet-level labels rather than article-level ideological truth; imbalance, uneven model coverage, and missing failures prevent interpreting the asymmetry as intrinsic detection ability.
Three model-level regressions (n=26) link questionnaire scores to Far Right–Far Left MAE difference, Center directional error, and aggregate right-category MAE. No predictor reaches α=0.05; the closest is progressivism versus extreme-category asymmetry, β=−0.153 and p=0.066. This is an absence of statistical evidence under three small specifications, not proof of independence or “decoupling.” With n=26, many possible axes, noisy labels, and post hoc analyses, non-significance is also compatible with low power or misspecification.
The sakhadib/PolAlignLLM repository was audited at commit d8d4e428194694656f281636fe9d8bbdadc6c1c4. Python code compiles and, after dependency installation, regenerates public Phase I(b) analyses with floating-point-only numerical differences. It contains 178 questions, ten variants, 70 score rows per instrument, Selenium scoring scripts, and ANOVA/MTMM/clustering code. It has no license, tests, CI, lockfile, seeds, raw responses, web-scorer snapshots, or pinned environment. README links to responses.jsonl and or_modelrun.py, which are absent, and compiled __pycache__ artifacts are committed.
Public coverage does not reproduce the full paper. The runner configures only seven models×ten prefixes; it does not implement the 26-model repeated cohort, news classification, regressions, or base/instruct log-probability extraction. A safety fallback changes words such as “violence,” “terrorism,” “authoritarian,” or “war” only when a provider blocks an item, despite the paper's claims of fixed questionnaire wording and no model-specific overrides. Without responses.jsonl, one cannot audit which models received altered items or how many outputs failed.
There is also manuscript–artifact drift. Table 16 labels two Phase I(b) members “Mistral Medium” and “Qwen3 235B,” while Table 6 and the CSVs use Mistral Large 2512 and Qwen 3.5 Flash; means and standard deviations also differ from current CSVs. Full η² values in Table 17 and the MTMM summary in Table 18 differ from versioned outputs, for example, the repository reports monotrait–heteromethod mean 0.587 and significant fraction 0.714, versus 0.82 and 1.00 in the PDF. Selected main-text tables do match some CSVs. The repository therefore supports a bounded robustness check, not the central 26-model or downstream findings.
The useful contribution is to separate questionnaires, prompt robustness, and task behavior and to show why a single ideological projection is insufficient. The faithful conclusion is narrower than the framing: these models produce similar response profiles under three quizzes and one setup; profiles vary more between models than under the average effect of ten prefixes in seven systems; and they do not significantly predict three aggregate errors. The work does not establish political identity, psychometric validity for synthetic agents, causal effects of alignment, Ground News neutrality, or generalization across languages, tasks, versions, and deployments.