GPT as ghostwriter at the White House

Jacques Savoy

REVIEW 5 major objections 6 minor 18 references

GPT as ghostwriter at the White House

T0 review · 5 major / 6 minor · reviewed 2026-08-12 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read GPT-written State of the Union speeches still read as GPT

desk verdict A credible descriptive result about GPT-3.5's default political prose, but the prompt under-specification confound and an unsupported training-data claim keep it from being a stable 'GPT style' finding. read the letter →

arxiv 2411.18365 v1 pith:DFIUWUWQ submitted 2024-11-27 cs.CL cs.AIcs.CY

Jacques Savoy This is my paper · ORCID

classification cs.CLcs.AIcs.CY

keywords ChatGPTstylometryauthorshipattributionStateoftheUnionlargelanguagemodelspoliticalspeechintertextualdistanceGPT-3.5

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper asks whether a large language model can pass as a presidential speechwriter, and answers no. Even when ChatGPT-3.5 is explicitly asked to mimic Reagan, Clinton, Bush, or Obama, its State of the Union addresses remain separable from the real speeches using surface stylistic measurements. The generated texts use longer words and sentences, favor nouns and commas, lean heavily on the lemma "we", and keep a positive, abstract tone. The paper's global evidence is an intertextual-distance tree in which all GPT outputs cluster together, away from all six presidents, including the ones being imitated. If correct, this gives stylometric tools a practical target: machine-written political rhetoric carries a detectable fingerprint.

What carries the argument

The central object is the intertextual distance measure proposed by Labbé (2007), which computes a value between 0 and 1 by comparing whole-vocabulary frequencies after normalizing text lengths. This distance is used to build a tree-based visualization of the entire corpus, and the resulting figure is the paper's global proof that GPT's style is distinct. Supporting the distance evidence are several standard stylometric tools: mean word length and percentage of big words, moving-average type-token ratio (MATTR), mean sentence length, part-of-speech distributions, characteristic-vocabulary z-scores (Muller's method), and Hart's wordlists for rhetorical categories such as Symbolism, Tenacity, and Blame.

What would settle it

Check ChatGPT-3.5's training data for verbatim or near-verbatim passages from any SOTU address used in this study; if one appears, re-run the distance analysis after prompting with a speech known to be absent from training. A second approach is to repeat the experiment with a model whose training set is fully documented to include SOTUs and see whether the intertextual-distance gap shrinks or disappears.

Watch

Extended reading notes

Core claim

The paper's central claim is that GPT-3.5 has a distinct written style that survives the instruction to imitate a specific president. In State of the Union addresses generated for Reagan, Clinton, Bush, and Obama, GPT overuses the lemma "we" (about 6% of tokens versus 3.8% for real presidents), uses more nouns and adjectives, more commas, fewer verbs and adverbs, longer words (mean 4.9 letters versus 4.39) and longer sentences (mean 22.62 versus 19.71 words). Its vocabulary, measured by moving-average type-token ratio, is poorer, and its characteristic words are general, neutral, and abstract rather than tied to a specific administration's issues. The intertextual-distance tree shows all GPT-generated speeches forming a cluster separate from every president, with the distance between a target president and his GPT imitation larger than the distance between that president and other real presidents.

Load-bearing premise

The whole comparison rests on the claim that the model's training data never included any State of the Union address, only other presidential speeches; if one of the real speeches used for comparison was in the training data, apparent stylistic differences could be retrieval artifacts rather than a stable machine style.

Editorial extensions

If this is right

A machine-authored political speech can be flagged with standard stylometric tools, without needing a trained classifier.
GPT's overuse of "we" and noun-heavy style mean generated speeches read as descriptive reports rather than calls to action.
Asking an LLM to imitate a president narrows some features but not enough; the model's own voice dominates the requested persona.
The same measurements could be extended to other LLMs and to other genres of political text, such as congressional remarks or campaign stump speeches.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

The finding that GPT's voice is stable across four presidential masks suggests the model's style is a property of its training distribution and prompt constraints, not of the requested persona; this could be tested by fine-tuning a model on presidential corpora and seeing whether the cluster disperses.
The neutral, positive, no-blame tone may reflect reinforcement learning from human feedback rather than a constraint of next-token prediction; a comparison between models with and without RLHF would isolate that factor.
Because the study uses only GPT-3.5, newer models with different alignment and length controls could shrink the gap; measuring the distance between human and machine speeches over successive model versions is a testable prediction.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

First, the useful thing: the paper produces a clean, multi-measure description of what GPT-3.5 writes when asked for a State of the Union address in the style of four presidents. The finding that the model's output clusters together rather than moving toward the target author, and that it lands on a noun-heavy, comma-heavy, 'we'-heavy, abstractly positive register, is credible and worth having. The use of external wordlists (Diction/LIWC) and the intertextual distance tree makes the result easy to see.

The soft spots are three. The biggest is the prompt confound. The generated speeches are roughly 9,000 tokens against 37,000-67,000 for real presidents, and the prompts ask for a SOTU with only one Miller Center example and a year. Real SOTUs are saturated with policy specifics, names, dates, and current events. The paper's own Section 3 says the prompt does not provide a long list of possible content. Most of the measured differences—fewer names, less negative emotion, less time/space anchoring, higher abstraction—are exactly what you would expect from any writer asked to produce a generic SOTU with no briefing. A human speechwriter given the same bare instruction would likely sound similar. Without a human control using the same prompt, the central claim about 'GPT's style' is really 'GPT's style under this under-specified prompt.' That does not invalidate the descriptive result, but it does undercut the abstract's generalization.

Second, there is an internal contradiction about training data. Footnote 4 says 'we don't know precisely the training sample ... one might assume that many presidential speeches have been included,' while Section 7 asserts flatly that 'the training set never includes a SOTU address.' That assertion is unsupported and load-bearing for the style-generalization reading. Remove it.

Third, the significance tests pool tokens across all speeches within a group, treating each token as independent. That ignores document-level clustering and inflates significance. It is a real but fixable problem; the asterisks are stronger than the evidence.

The paper also does not release the generated corpus or the exact prompts, which is a shame for a replication-friendly subject.

Bottom line: a solid applied stylometry paper, clearly written and honest in places, that needs a revision addressing the prompt confound and the training-data claim. It deserves a serious referee and a conditional acceptance. I would bring it to a reading group as a good example of LLM stylometry with a design flaw worth discussing.

Referee Report

5 major / 6 minor

Summary. The manuscript compares State of the Union (SOTU) addresses written by Reagan, Clinton, G.W. Bush, and Obama with SOTU addresses generated by GPT-3.5-turbo under prompts that ask the model to imitate each president. Using word-frequency rankings, part-of-speech distributions, lexical statistics (word length, MATTR, mean sentence length), characteristic vocabulary, Diction/LIWC-style categories, and Labbé intertextual distance, the paper reports that GPT texts overuse the lemma "we", nouns, and commas, use fewer verbs and personal pronouns, contain longer words and sentences, exhibit a positive and abstract tone, and cluster separately from all real presidential speeches in a tree representation. The author concludes that GPT has a distinct, didactic, neutral style that remains distinguishable even when the model is asked to write in a target president's style.

Significance. If the conclusions hold, the paper provides a useful empirical demonstration that current LLM-generated political speech can be distinguished from human-authored SOTU addresses using relatively simple surface stylometric features. Strengths include the use of a genre-matched corpus (SOTU addresses), the use of established external wordlists (Diction, LIWC) rather than ad-hoc categories, transparent reporting of model sampling parameters, and a global distance visualization that makes the main separation easy to inspect. However, the central claim is currently underdetermined because the comparison does not isolate style from prompt/content richness, and the statistical significance tests ignore document-level clustering and multiple comparisons. These issues are addressable, but they are load-bearing for the paper's main conclusion.

major comments (5)

[§3, Table 1; §4-7] The central comparison does not separate style from prompt/content richness. The GPT outputs are roughly 9,000 tokens while real SOTU addresses are 37,000-67,000 tokens, and the prompt explicitly contains no year-specific policy brief (Section 3). The features identified as GPT's 'style' — fewer names, fewer negative terms, no time/space anchoring, noun-heavy and abstract vocabulary (Tables 5-7) — are exactly what one would expect from any writer asked to produce a generic SOTU without substantive material. To support the claim that these are stable properties of GPT's style, the paper needs a control condition: e.g., the same prompt given to human writers, or GPT prompted with the policy content of a real SOTU, or a topic-matched comparison. Without such a control, the conclusion that 'GPT's style exposes distinct features' is confounded with input richness.
[§4, Tables 2-5, 7] The significance tests treat tokens as independent observations drawn from pooled author corpora. The proportion tests and t-tests compare aggregate percentages over all words/sentences of each group, ignoring document-level clustering; with 7-8 generated speeches and 7-8 human speeches per president, the effective sample size is much smaller than the number of tokens. In addition, dozens of categories are tested at alpha=0.01 without any multiple-comparison correction, so several asterisks are expected by chance. Please report per-document means with a mixed-effects model, paired/permutation test, or cluster-robust standard errors, and apply a multiple-testing correction or restrict confirmatory claims to pre-specified hypotheses. This affects the load-bearing statements such as 'GPT overuses the lemma we' and 'GPT employs fewer verbs'.
[§4, Table 4] The row 'Mean president' cannot be reproduced from the displayed data. The text says the last two rows give the average over the 'six groups of presidential addresses', but Table 4 shows only Reagan, Clinton, Bush, and Obama; for MATTR the mean of the four displayed values is 0.325, not 0.359, and similar discrepancies occur for word length (4.425 vs 4.39), BW (28.0 vs 27.7), and MSL (20.55 vs 19.71). Either include the Trump and Biden rows in the table (or in the annexe) and state that the mean covers all six presidents, or correct the text and recompute the averages. As it stands, the headline comparison 'Mean GPT 4.90 vs Mean president 4.39' is not verifiable.
[§7, 'Intertextual Distance'] The sentence 'the training set never includes a SOTU address but contains another presidential speech' is a strong empirical claim about OpenAI's proprietary training data, made without citation or evidence. It is also not needed for the main demonstration: Figure 1 would still show a separation between GPT and human texts even if some SOTU addresses were in the training data, but the interpretation would change from 'style generalization' to possible memorization/retrieval. Please either remove the assertion, label it as an unverifiable assumption, or support it with documentation, and discuss the consequences for the Figure 1 interpretation if it is wrong.
[§3, footnote 3; §7 prompts] The generation procedure is not described precisely enough for replication. The paper gives temperature (0.5), top_p (0.4), and penalties (0), but does not report the full prompt text, which Miller Center speech was used as the example for each president, the date of API access, or the exact model identifier beyond 'GPT-3.5-turbo'. Because the style-imitation manipulation depends entirely on the prompt, this information should be included in an appendix. Without it, a reader cannot regenerate the corpus or assess whether the provided example biased the results.

minor comments (6)

[§2] There is a typo: 'GTP' should be 'GPT'.
[§1 vs §5-§7] The introduction's preview of the section order (Section 5 'global similarity', Section 6 'characteristic vocabulary', Section 7 'rhetorical and topical analysis') does not match the actual body, where characteristic vocabulary is Section 5, rhetorical/topical analysis is Section 6, and intertextual distance is Section 7.
[References] Some reference names appear misspelled: 'Vaswami et al. 2017' should be 'Vaswani et al.'; 'Zhoa et al. 2023' should be 'Zhao et al.'; 'Bartélémy' should be 'Barthélemy' if referring to the standard author.
[§7, Eq. (3)] Equation (3) uses a superscript notation (tf$_0$) that is not defined; rewrite with standard notation and state which text is the reference for length normalization.
[§4, Table 4 footnote] The footnote says the t-test is used 'with the same significance level', but it is not stated whether the unit of analysis is individual speeches or pooled tokens; this matters for interpreting the asterisks.
[§7, Figure 1] Figure 1 would benefit from a statement of how the tree was fit (e.g., Neighbor-Joining or another algorithm) and a note that branch lengths are approximate; this would help readers interpret the visual separation.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation: this is a direct empirical stylometric comparison using external measures, with no fitted-input or self-citation chain forcing the result.

full rationale

The paper's derivation chain is an empirical corpus comparison: GPT-3.5 outputs are generated under fixed prompt conditions and then compared to real State of the Union addresses using external, literature-established metrics (Labbé intertextual distance, Muller's Z-score, LIWC/Diction wordlists, Hart's word-length and MATTR measures). None of these measures is fitted to the data or defined in terms of the target conclusion; the wordlists and distance formula are prior, independent instruments. The author's self-citation (Savoy 2020) appears only as a general stylometry reference in Section 2 and is not load-bearing. The strongest validity threat is factual, not circular: Section 7 asserts 'the training set never includes a SOTU address but contains another presidential speech,' which is contradicted by the paper's own footnote that 'we don't know precisely the training sample employed by GPT and one might assume that many presidential speeches have been included.' If SOTU texts were in training, the observed GPT/human distance could reflect memorization and prompt constraints rather than a stable style. However, that is an unsupported empirical assumption about model provenance, not a reduction of the conclusion to the inputs by construction. Figure 1's separation is computed directly from the texts, and no equation or fitted parameter embodies the claimed 'distinct GPT style' beforehand. Therefore no circularity is present; the training-data contradiction should be weighed as an evidence-quality limitation, not as circularity.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The central claim rests mainly on the corpus construction and on external measurement tools. The GPT generation parameters are free choices that shape the data. The assumption that SOTU addresses are absent from training data is unverified. The statistical and rhetorical measurement tools (binomial Z scores, Diction/LIWC wordlists, Labbé distance) are imported from prior literature and are not fitted to the target result. No new entities are introduced.

free parameters (3)

GPT generation parameters (temperature, top_p, frequency and presence penalties) = temperature 0.5, top_p 0.4, penalties 0
Chosen by hand without sensitivity analysis; they shape the generated corpus and could affect all downstream stylistic comparisons (Section 3, footnote 3).
MATTR window size = 2,000 tokens
Chosen for computing moving-average type-token ratio; different windows would change vocabulary-richness values in Table 4 (Section 4).
Z-score threshold for characteristic vocabulary = 3
Threshold for overused terms in Table 6; changing it alters the lists used to characterize GPT's neutral tone (Section 5).

assumptions (4)

domain assumption GPT-3.5 training data does not contain State of the Union addresses used for comparison.
Section 7 asserts this without evidence; it is load-bearing because memorized SOTU text would change the meaning of the stylistic comparison.
standard math Term frequencies in an author subset follow a binomial distribution, and the resulting Z score is standard normal.
Used in Section 5, Equations 1 and 2, to identify characteristic vocabulary in Table 6; relies on large-sample approximation and independence of token draws.
domain assumption The Diction and LIWC wordlists validly measure emotional tone, symbolism, tenacity, blame, and achievement in political speeches.
Section 6 uses these external lists to produce Table 7; the paper concedes word-sense ambiguity makes such extraction 'subject to caution'.
standard math Labbé's intertextual distance with length normalization is a valid similarity measure for comparing texts of different lengths.
Section 7, Equation 3 and Figure 1 use this distance to conclude GPT speeches cluster separately from real presidents.

how reviews work

0 comments

Cite this review

Pith. "Pith review of GPT as ghostwriter at the White House." pith.science (2026). https://pith.science/paper/DFIUWUWQ

@misc{pith2026241118365,
  author       = {Pith},
  title        = {Pith review of: GPT as ghostwriter at the White House},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/DFIUWUWQ}},
  note         = {Machine review of arXiv:2411.18365}
}

read the original abstract

Recently several large language models (LLMs) have demonstrated their capability to generate a message in response to a user request. Such scientific breakthroughs promote new perspectives but also some fears. The main focus of this study is to analyze the written style of one LLM called ChatGPT 3.5 by comparing its generated messages with those of the recent US presidents. To achieve this objective, we compare the State of the Union addresses written by Reagan to Obama with those automatically produced by ChatGPT. We found that ChatGPT tends to overuse the lemma "we" as well as nouns and commas. On the other hand, the generated speeches employ less verbs and include, in mean, longer sentences. Even when imposing a given style to ChatGPT, the resulting speech remains distinct from messages written by the target author. Moreover, ChatGPT opts for a neutral tone with mainly positive emotional expressions and symbolic terms (e.g., freedom, nation). Finally, we show that the GPT's style exposes distinct features compared to real presidential addresses.

Figures

Figures reproduced from arXiv: 2411.18365 by the authors.

**Figure 1.** Intertextual distance in our American corpus Overall, this figure illustrates the large overall difference between GPT speeches on the one hand (on the top part) and the six recent US presidents on the other (in the bottom). The speeches generated by GPT clearly present a distinctive style, and no clear relationship can be established with the six most recent US presidents. When asking GPT to write a text according … view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

18 extracted references · 15 canonical work pages

[1]

produces outputs that are essentially indistinguishable from (even better than) what humans could produce

- 1 - GPT as ghostwriter at the White House Compared with four US presidents, what are the differences between State of the Union addresses written by GPT and the true presidents Jacques Savoy Computer Science Dept. University of Neuchatel rue Emile Argand 11 2000 Neuchatel, Switzerland Jacques.Savoy@unine.ch Abstract Recently several large language model...

work page 2023
[3]

we” {we, us, our, ours} in machine-generated texts vs. true ones (6% vs. 3.8%). This is a significant aspect because the lemma “we

Some statistics about our American corpus Presidency Number Tokens Types Reagan-GPT 7 9,255 1,449 Clinton-GPT 8 8,607 1,268 Bush-GPT 8 8,768 1,387 Obama-GPT 8 10,959 1,611 R. Reagan 1981–1989 7 37,004 3,514 B. Clinton 1993–2000 8 67,445 4,000 W.G. Bush 2001–2008 8 45,818 3,641 B. Obama 2009–2016 8 61,034 4,013 D. Trump 2017–2020 4 25,776 3,323 J. Biden 20...

work page 1981
[4]

Ac-cording to this measurement and for each president, the GPT version presents a significant higher mean suggesting that the messages tend to be more complex

Some overall stylistic measurements of our corpus Word length BW MATTR MSL Reagan 4.55* 29.3%* 0.347* 21.04* Reagan-GPT 4.89 37.6% 0.301 22.95 Clinton 4.34* 26.8%* 0.303 21.33* Clinton-GPT 4.96 38.4% 0.306 23.74 Bush 4.50* 30.1%* 0.329* 20.08 Bush-GPT 4.83 36.7% 0.301 21.06 Obama 4.31* 25.9%* 0.323 19.76* Obama-GPT 4.91 36.9% 0.315 22.74 Mean president 4....

work page 1984
[5]

As depicted in Table 3, those proportions obtained by GPT are usually shown in italics

Frequency of categories corresponding to personal pronouns Category GPT Reagan Clinton Bush Obama Trump Biden Self 0.78% 0.92% 1.38%* 0.85% 1.16%* 1.00%* 2.03%* You 0.35% 0.44% 0.69%* 0.57%* 0.48%* 0.70%* 1.17%* She 0.00% 0.09%* 0.10%* 0.06%* 0.15%* 0.19% 0.19%* He 0.01% 0.11%* 0.19%* 0.20%* 0.22%* 0.59% 0.32%* We 6.02% 3.76%* 3.97%* 3.65%* 3.78%* 3.16%* ...

work page 1984
[6]

terrorism

Terms overused by GPT and the US presidents Reagan Clinton Bush Obama Rank GPT real GPT real GPT real GPT real 1 principle soviet shared welfare adversity terrorist healthcare that 2 pursuit freedom fostering child we Iraq investing business 3 commitment govern. economic work unwavering Iraqi rural why 4 dedication program collaboration ought remains enem...

work page 2002
[7]

GPT stays at the same high level though (4.4%), empha-sizing the abstraction with references to political institutions or to the nation (e.g., America)

Percentage of some selected wordlists Category GPT Reagan Clinton Bush Obama Trump Biden Posemo 4.67% 3.47%* 3.11%* 3.78%* 2.78%* 3.38%* 2.79%* Negemo 0.95% 1.14% 0.97% 1.50%* 1.07%* 1.53%* 1.14% Symbolism 4.40% 3.84%* 3.52%* 4.10% 3.16%* 4.43% 3.50%* Tenacity 4.76% 6.26%* 5.94%* 6.21%* 6.49%* 6.46%* 6.86%* Blame 0.20% 0.44%* 0.42%* 0.40%* 0.46%* 0.60%* 0...

work page 2007
[10]

America”, “American

Part-of-speech distribution over the US presidents and GPT POS GPT Reagan Clinton Bush Obama Trump Biden . 3.95% 5.30%* 4.48%* 4.74%* 4.82%* 5.28%* 5.91%* , 6.86% 5.42%* 4.93%* 4.99%* 4.99%* 5.79%* 4.88%* Conjunction 5.43% 4.45%* 3.72%* 4.66%* 3.95%* 3.81%* 3.60%* Article 8.43% 9.48%* 8.72% 8.54% 8.58% 7.68%* 7.65%* Preposition 9.88% 9.92% 10.04% 10.19% 1...

work page 1992
[15]

Moreover, GPT avoids specific examples and never employs an argument that could cause clear disagreements between people

and the presentation inclines to stay at a descriptive level, without taking any divisive position (see Table 6). Moreover, GPT avoids specific examples and never employs an argument that could cause clear disagreements between people. Using an intertextual distance measure computed according to the whole vocabulary, a overall picture indicates that GPT a...

arXiv 2023

Show all 18 references

[18]

/ 2 = 153 values. To achieve a better picture than a dendrogram, such distance matrices can be represented by a tree-based visualization respecting approximately the real distances between all nodes (Bartélémy & Guénoche, 1991), (Baayen, 2008), (Paradis, 2011). We adopt this n...

1991
[30]

Wolfram, S. (2023). What is GPT-4 Doing… and What Does it Work?. Orlando: Wolfram Re-search Inc., Champaign (IL). Yule, G. (2020). The Study of Language. 7th ed., Cambridge: Cambridge University Press. Zhao, W., Zhou, K., Li, J., Tang, T., Wang, X, Hou, Y. Min, Y., Zhang, B., ...

2023 arXiv
[41]

I want four-letter words, and I want four sentences to the paragraph

For example, L. B. Johnson (presidency: 1963–1969) recognized the fear of having a too complex style by specifying to his ghostwriters: “I want four-letter words, and I want four sentences to the paragraph.” (Sherrill, 1967). Finally, to reflect the stylistic aspect related to...

1963
[1984]

In this study, words belonging to the Achieve tag (e.g., first, plan, win, …) form another dedicated category

while the Blame category includes terms such as angry, deceptive, incompetent, etc. In this study, words belonging to the Achieve tag (e.g., first, plan, win, …) form another dedicated category. Grounded on such wordlists, Hart (1984) portrays the rhetorical and stylistic diff...

1984
[1990]

the” (under the label “US Presidents

has been applied with the null hypothesis H0 spec-ifying that both population proportions are equal (bilateral test). For example, in Table 2 the proportion of “the” under GPT's pen is 5.7% vs. 4.2% with the presidents. Can we assume that both population proportions are equal?...

2011
[2002]

ads” and “advertisings

or the percentage of big words (BW) defined as words composed of six letters or more (at least for the English language). For example, one can observe that some terms are easier to understand than others as, for example, between “ads” and “advertisings” or “desks” and “furnitu...

2012
[2008]

is provided. A high value specifies the presence of a rich vocabulary showing that the underlying text covers many different topics or that the author presents a theme from several points of view with different formulations. To compute this value, one divides the vocabulary si...

2008
[2010]

Journal of Quantitative Linguistics, 17(2):94-100

Cutting the Goridan knot: The moving-average type-token ratio (MATTR). Journal of Quantitative Linguistics, 17(2):94-100. Daelemans, W., Kestemont, M., Manjavacas, E., Potthast, M., Rangel, F., Rosso, P., Specht, G., Stamatatos, E., Stein, B., Tschuggnall, M., Wiegmann, M, and...

2019 arXiv
[2011]

New York: Springer

Analysis of Phylogenetics and Evolution with R. New York: Springer. Pennebaker, J.W. (2011). The Secret Life of Pronouns. What our Words Say About Us. New York: Bloomsbury Press. Savoy, J. (2020). Machine Learning Methods for Stylometry. Authorship Attribution and Author Profi...

2011
[2016]

the president of the

based on a sequence of transformers with an attention mechanism (Vaswami et al. 2017). The most important notion to understand LLM is the following: Given a short sequence of tokens (e.g., words or punctuation symbols), the computer is able to automatically provide the next on...

2017

Pith tools

Reviewed August 12, 2026 · model on record in the stance chip above.

[1] [1]

produces outputs that are essentially indistinguishable from (even better than) what humans could produce

- 1 - GPT as ghostwriter at the White House Compared with four US presidents, what are the differences between State of the Union addresses written by GPT and the true presidents Jacques Savoy Computer Science Dept. University of Neuchatel rue Emile Argand 11 2000 Neuchatel, Switzerland Jacques.Savoy@unine.ch Abstract Recently several large language model...

work page 2023

[2] [3]

we” {we, us, our, ours} in machine-generated texts vs. true ones (6% vs. 3.8%). This is a significant aspect because the lemma “we

Some statistics about our American corpus Presidency Number Tokens Types Reagan-GPT 7 9,255 1,449 Clinton-GPT 8 8,607 1,268 Bush-GPT 8 8,768 1,387 Obama-GPT 8 10,959 1,611 R. Reagan 1981–1989 7 37,004 3,514 B. Clinton 1993–2000 8 67,445 4,000 W.G. Bush 2001–2008 8 45,818 3,641 B. Obama 2009–2016 8 61,034 4,013 D. Trump 2017–2020 4 25,776 3,323 J. Biden 20...

work page 1981

[3] [4]

Ac-cording to this measurement and for each president, the GPT version presents a significant higher mean suggesting that the messages tend to be more complex

Some overall stylistic measurements of our corpus Word length BW MATTR MSL Reagan 4.55* 29.3%* 0.347* 21.04* Reagan-GPT 4.89 37.6% 0.301 22.95 Clinton 4.34* 26.8%* 0.303 21.33* Clinton-GPT 4.96 38.4% 0.306 23.74 Bush 4.50* 30.1%* 0.329* 20.08 Bush-GPT 4.83 36.7% 0.301 21.06 Obama 4.31* 25.9%* 0.323 19.76* Obama-GPT 4.91 36.9% 0.315 22.74 Mean president 4....

work page 1984

[4] [5]

As depicted in Table 3, those proportions obtained by GPT are usually shown in italics

Frequency of categories corresponding to personal pronouns Category GPT Reagan Clinton Bush Obama Trump Biden Self 0.78% 0.92% 1.38%* 0.85% 1.16%* 1.00%* 2.03%* You 0.35% 0.44% 0.69%* 0.57%* 0.48%* 0.70%* 1.17%* She 0.00% 0.09%* 0.10%* 0.06%* 0.15%* 0.19% 0.19%* He 0.01% 0.11%* 0.19%* 0.20%* 0.22%* 0.59% 0.32%* We 6.02% 3.76%* 3.97%* 3.65%* 3.78%* 3.16%* ...

work page 1984

[5] [6]

terrorism

Terms overused by GPT and the US presidents Reagan Clinton Bush Obama Rank GPT real GPT real GPT real GPT real 1 principle soviet shared welfare adversity terrorist healthcare that 2 pursuit freedom fostering child we Iraq investing business 3 commitment govern. economic work unwavering Iraqi rural why 4 dedication program collaboration ought remains enem...

work page 2002

[6] [7]

GPT stays at the same high level though (4.4%), empha-sizing the abstraction with references to political institutions or to the nation (e.g., America)

Percentage of some selected wordlists Category GPT Reagan Clinton Bush Obama Trump Biden Posemo 4.67% 3.47%* 3.11%* 3.78%* 2.78%* 3.38%* 2.79%* Negemo 0.95% 1.14% 0.97% 1.50%* 1.07%* 1.53%* 1.14% Symbolism 4.40% 3.84%* 3.52%* 4.10% 3.16%* 4.43% 3.50%* Tenacity 4.76% 6.26%* 5.94%* 6.21%* 6.49%* 6.46%* 6.86%* Blame 0.20% 0.44%* 0.42%* 0.40%* 0.46%* 0.60%* 0...

work page 2007

[7] [10]

America”, “American

Part-of-speech distribution over the US presidents and GPT POS GPT Reagan Clinton Bush Obama Trump Biden . 3.95% 5.30%* 4.48%* 4.74%* 4.82%* 5.28%* 5.91%* , 6.86% 5.42%* 4.93%* 4.99%* 4.99%* 5.79%* 4.88%* Conjunction 5.43% 4.45%* 3.72%* 4.66%* 3.95%* 3.81%* 3.60%* Article 8.43% 9.48%* 8.72% 8.54% 8.58% 7.68%* 7.65%* Preposition 9.88% 9.92% 10.04% 10.19% 1...

work page 1992

[8] [15]

Moreover, GPT avoids specific examples and never employs an argument that could cause clear disagreements between people

and the presentation inclines to stay at a descriptive level, without taking any divisive position (see Table 6). Moreover, GPT avoids specific examples and never employs an argument that could cause clear disagreements between people. Using an intertextual distance measure computed according to the whole vocabulary, a overall picture indicates that GPT a...

arXiv 2023