In the 1868 letter cited earlier, J. S. Mill noted his regret seeing “attributed to myself work which I think good and which is chiefly” Helen Taylor’s (J. S. Mill [1868] 1972, 1,359). He objected to publishing an earlier letter with his signature on it as his, because it was an example of their unity of opinion taking form in a private communication, “by no means a solitary one,” which was written entirely by her: “what she wrote expressed so perfectly all I could have wished to say, that I transcribed it unaltered” (J. S. Mill [1868] 1972, 1,359). Who wrote what does not really matter, he went on, “[i]f everything said in a private letter is the real opinion or feeling of the person who signs it” (J. S. Mill [1868] 1972, 1,359). However, J. S. Mill pointed out, “anything printed comes into the class of literary performances, and I should feel [i.e., seeing the letter published under his name] (only in a less degree) as if I were to publish a book written by my daughter with my own name instead of hers prefixed to it” (J. S. Mill [1868] 1972, 1,359). So, why were Principles, Liberty, and Subjection printed with just his name? Jacobs (Jacobs 2024) encourages us to distinguish between John Stuart Mill the individual, and “John Stuart Mill” the brand name. According to Jacobs, J. S. Mill and H. Mill thought that the brand name was more likely to influence society. It was a matter of a simple utilitarian calculation: change was the desired outcome, recognition was secondary. In what follows, we take our sight away from the brand name and focus on the individuals involved.

As we already noted, according to J. S. Mill, three authors, or minds, had a share in the writing of Subjection. “As ultimately published,” he claimed, Subjection

Discussing Harriet Mill’s “direct share” in his writings, J. S. Mill himself thought it impossible to unravel parts contributed to a work by two persons, when those two persons, like them

Under such circumstances, J. S. Mill added, “it is of little consequence in respect to the question of originality which of them holds the pen” (J. S. Mill [1873] 1981, 251). Thus, J. S. Mill concluded,

Still, J. S. Mill tried to immediately qualify this statement: “in certain cases, what belongs to her can be distinguished, and specially identified” (J. S. Mill [1873] 1981, 251). This seems to be an allusion to the notebook with which he kept track of his publications. A copy of the list can be found at LSE Library (Mill–Taylor Collection, Box 36, ff. 1–39; see also MacMinn, Hainds, and McCrimmon 1945). There he recorded different collaborative texts with Harriet Mill, with notations such as “a joint production” (5), or a joint production, “very little of which was mine” (11), or even as a work for which he “acted chiefly as amanuensis to my wife” (1) (O’Grady and Robson 1991, 19–61).

This bibliographical list complicates matters. Sometimes he took note of a collaborating project (J. S. Mill and Anonymous [1837] 1989; J. S. Mill and Anonymous [1842] 1986; J. S. Mill and Blanco-White [1836] 1985; J. S. Mill and Grote [1837] 1977). But sometimes he did not (J. S. Mill and Ellis [1825] 1967; J. S. Mill and Buller [1837] 1989; J. S. Mill and Bulwer-Lytton [1833] 1981). The extant list itself does not include Subjection. The general editor of J. S. Mill’s Collected Works, John M. Robson, ascribed the omission to an error on the part of the copyist (Robson 1984, lxx; see also Robson 1964). “Enfranchisement” was also not included in the list. Yet, John Stuart Mill assigned himself the role of amanuensis and copyeditor to Harriet Mill for that work elsewhere (J. S. Mill [1859b] 1984, 393). Principles received the notation of the aforementioned “a joint production.” But Liberty received no such notation. J. S. Mill claimed in Autobiography that Liberty “was more directly and literally our joint production than anything else which bears my name” (J. S. Mill [1873] 1981, 257). Did that mean that J. S. Mill’s earlier one, that all of his published writings were “as much her work as mine,” was not to be taken literally and that her share in them was indirect? As we saw in the previous section, for the first two reception streams, the answer was “yes.”

Subjection was published in 1869. J. S. Mill’s Autobiography encourages us to think of it as “a joint production,” even though he began its composition after H. Mill’s death (J. S. Mill [1873] 1981, 265). J. S. Mill had supposedly “made good progress” already by December 1860 (J. S. Mill [1860] 1972, 716). Waiting for the right time to publish, he periodically returned to the manuscript for improvements and additions (J. S. Mill [1873] 1981, 265, 290.). Not only did J. S. Mill acknowledge that the work “was written at my daughter’s suggestion that there might, in any event, be in existence a written exposition of my opinions on that great question, as full and conclusive as I could make it” (J. S. Mill [1873] 1981, 265), but he also admitted that Helen Taylor authored parts of it. However, the first time J. S. Mill mentioned Subjection in Autobiography, he confessed that he was “painfully conscious how much of her [Harriet Mill’s] best thoughts on the subject I have failed to reproduce, and how greatly that little treatise falls short of what would have been given to the world if she had put on paper her entire mind on this question” (J. S. Mill [1873] 1981, 253f). Given his own “imperfect statement of the case” (J. S. Mill [1873] 1981, 253f), can the Subjection be considered as “a joint production”?

How we choose to answer this question must begin with J. S. Mill’s distinction between a narrow sense and a broad sense of authorship. The narrow sense corresponds to those works in which an author’s share can be distinctly identified (the second and third class of joint productions delineated above). The broad sense corresponds to those works in which no part of them can be disentangled from the rest and attributed to one author rather than another (the first category of joint productions). Should we take J. S. Mill’s Autobiography at face value, unlike Liberty, which falls under the first category, Principles seems to fall under all three categories. First, the book, as an “attempt to explain and diffuse ideas many of which were first learned from herself” (J. S. Mill [1848] 1965, 1,026), was (privately) dedicated to Harriet Mill as “the most eminently qualified of all persons known to the author either to originate or to appreciate speculations on social improvement” (J. S. Mill [1848] 1965, 1,026). Second, the scientific/theoretical parts were J. S. Mill’s identifiable contributions, while the practical spirit and the look to the future that pervades the book were Harriet Mill’s contribution (see further Robson 1966). And, third, “Futurity” was “wholly an exposition of her thoughts, often in words taken from her own lips” (J. S. Mill [1873] 1981, 255, 257n). This chapter was not included in the first edition.

Subjection seems at best to fall under the first category as regards J. S. Mill’s and H. Mill’s part, and the third as regards Helen Taylor’s additions (even though, in Helen Taylor’s description of her role in J. S. Mill’s projects, she assumed the place once occupied by her mother; see Taylor [1869] 1937, 311–312). If authorship is characterized by authority and finality, only Helen Taylor could have had a claim in co-authorship, since her mother had passed more than a decade earlier. A computer-assisted stylometric analysis might be able to find passages by Helen Taylor’s hand. And even though Harriet Mill had no part in writing Subjection, it is worthwhile to test for her authorial voice. It might not be prominent, but it could be significant.

Authorship identification analysis is grounded on an important assumption: texts can expose their author’s identity. A weak authorship identification thesis is that texts contain unique markers of their author’s identity: frequent word combinations, idiomatic or idiosyncratic expressions, vocabulary richness, use of function words, etc. These markers are as unique as fingerprints. A stronger thesis is that authors cannot but engraft themselves onto the text. An author’s textual choices are unconscious; they deterministically find their way into the text. John Burrows (Burrows 2007, 30) has suggested that an authorial set of 10,000 words and 500-word particular texts offer sufficient discriminatory power to reliably extract an authorial fingerprint. Maciej Eder reduces Burrows’s word count to 5,000 per authorial set, demonstrating that smaller text segments—under 500 words—can still yield reliable results, even across languages. However, he cautions that while a lower threshold is feasible (with successful tests at 2,000 words; see Eder 2017), other factors may interfere with the authorial signal (Eder 2015; see also Luyckx and Daelemans 2011; for a recent introduction to authorship identification analysis, see Grant 2022; see also Love 2002; Juola 2008). This is not the first attempt at a computer-assisted authorship identification analysis as regards J. S. Mill’s corpus (for the most recent, see Loizides, Neocleous, and Nicolaides 2023; see also, Schmidt-Petri, Schefczyk, and Osburg 2022).

The most important part of the corpus preparation process is the selection of the texts. We used the standard editions for John Stuart Mill’s (J. S. Mill 1963–1991) and Harriet Mill’s (H. Mill 1998) authorial sets. There was not much choice of texts in the case of Harriet Mill and Helen Taylor. In both cases, we used texts with little or no evidence of interference by J. S. Mill (e.g., when part of the manuscript was in his hand, or when he lists the piece as a “joint production”). In the case of Harriet Mill, this decision eliminated “Remarks on Mr. Fitzroy’s Bill for the More Effectual Prevention of Assaults on Women and Children” (1853 [H. Mill 1998]; in J. S. Mill’s list [O’Grady and Robson 1991, 51]) and “The Enfranchisement of Women” (1851 [H. Mill 1998]; not in J. S. Mill’s list). Since this is our main difference as regards text selection with Neocleous and Loizides (Neocleous and Loizides 2020) and Neocleous, Kataliakos, and Loizides (Neocleous, Kataliakos, and Loizides 2022), as well as Schmidt-Petri, Schefczyk, and Osburg (Schmidt-Petri, Schefczyk, and Osburg 2022), our decisions need some justification.

First, we can never know the extent of J. S. Mill’s influence on the style of the text for which he served as Harriet Mill’s “editor and amanuensis.” However, we do have some evidence of J. S. Mill influencing a decision in the writing of Emancipation. Following J. S. Mill’s review of a new volume of George Grote’s History of Greece (Grote 1846–1856) in March 1849 (J. S. Mill [1849a] 1972), it seems that H. Mill urged him to tone down his laudatory treatment of Athens as a fine “specimen of humanity,” progress, and civilization, given the “barbarisms” persisting at the time and condoned by the Greek philosophers (presumably in relation to women) (J. S. Mill [1849b] 1972, 18). Mill admitted that he was not “sufficiently careful to explain that the praise is relative to the then state & not the now state of knowledge & of what ought to be improved feeling” (J. S. Mill [1849b] 1972, 18). But he could not let H. Mill’s comment slide without a response: “You are not quite right about the philosophers, for Plato did condemn those ‘barbarisms’” (J. S. Mill [1849b] 1972, 18). “Enfranchisement” included Plato in those eminent philosophers who had “made emphatic protests in favour of the equality of women” in the history of philosophy (H. Mill 1998, 55). Second, “Enfranchisement” and various short drafts by Harriet Mill share Subjection’s subject-matter. Since there is no other published or unpublished essay written exclusively by J. S. Mill on the subject, this could create bias in favour of H. Mill. If we opted to include these papers, as well as J. S. Mill’s numerous speeches on women’s rights, we would create a bias in favour of J. S. Mill, both on account of being delivered around the same time Subjection was written, and on account of the resulting enlarged corpus size. Third, as we saw, some Autobiography reviewers considered “Enfranchisement” to be an imitation of J. S. Mill’s style (an early attempt perhaps to capitalize on the brand name), which would distort H. Mill’s authorial set.

We did not include “Life of Caxton” in the experiments, as it is mistakenly attributed to Harriet Mill and John Taylor, H. Mill’s first husband. Published in a stand-alone edition in 1828 from the Society for the Diffusion of Useful Knowledge (SDUK), the essay was attributed to William Stevenson (1772–1829) in 1830 (Anonymous 1830, 213). Should we need further evidence for the attribution of the essay to Stevenson, according to John Chapple, James Martineau, closely involved with SDUK, had marked the authors’ full names of all articles in Lives of Eminent Persons (Anonymous 1833) in his copy. For “Caxton,” Martineau had noted: “By – Stephenson” (Chapple 1997, 329). In his study of J. S. Mill and H. Mill, Hayek identified Stevenson as the author of “Caxton,” and the Taylor manuscripts as drafts of a review (Hayek [1951] 2015, 22n11). Perhaps, Stevenson’s death in 1829 influenced the couple’s decision not to complete or publish their review. Jacobs, the editor of H. Mill’s works, admits that “neither person’s drafts very closely imitate” the published essay (Jacobs 1998, 237–238), but offers no evidence for the attribution of the essay to the Taylors. Neither does Jacobs discuss Hayek’s earlier claim that their manuscripts were drafts of a projected review by the two. Schmidt-Petri, Schefczyk, and Osburg (Schmidt-Petri, Schefczyk, and Osburg 2022) also excluded this text from their tests, as they did not want to risk distortion in H. Mill’s authorial set, given J. Taylor’s supposed role as co-author.

For some of Helen Taylor’s works, we used an edition for writings on sexual equality by J. S. Mill, H. Mill, and Helen Taylor (Robson and Robson 1994). For the rest, we transcribed the original articles as published in the nineteenth century. From her “Biographical Notice” of Henry Thomas Buckle (Taylor 1872, ix–lv), we have only included parts of the first ten pages (ix–xviii, ~2900w), since the remaining (~20,000w) were almost exclusively extracts from diaries and letters, as well as quotations from accounts by Buckle’s family and friends. Helen Taylor made “no apology for offering them to the reader as they were written, without either transposition or alteration” (Taylor 1872, xxi). We kept all essays by Helen Taylor on women’s rights, as we hoped to identify specific sections of Subjection in her hand (as per J. S. Mill’s claim), even if this risked bias in favour of Helen Taylor. After all, the process was dynamic. Initial testing showed no evidence of Helen Taylor’s involvement in Subjection, so it made sense to expand her authorial set to check for indications of her style.

The huge availability of texts made the selection for J. S. Mill’s training set a much more complicated issue. On the one hand, we needed to avoid contamination of the training set: Autobiography, Principles of Political Economy, and On Liberty had to be eliminated as potential candidates for training, given the questions surrounding Harriet Mill’s role in the writing process. For a similar reason, we had to eliminate “Chapters on Socialism” (J. S. Mill [1879] 1967) and Three Essays on Religion (J. S. Mill [1874] 1979), even though they would have provided a good sample of J. S. Mill’s style (even if unpolished) around the time Subjection was published. It is impossible to know the extent of Helen Taylor’s editorial interventions to these texts prior to publication. On the other hand, given the limitations involved in corpus selection as regards Harriet Mill and Helen Taylor, caution was needed to avoid bias in favour of J. S. Mill. Being sensitive to the risk of distortion to authorial sets, we used different sets of texts for the tests between J. S. Mill and H. Mill, as well as J. S. Mill and H. Taylor.

For comparisons to H. Mill’s corpus, we opted for the first 2,200–2,500 words of J. S. Mill’s first seven essays published in the Monthly Repository. First, as most of the writing samples from H. Mill are from the same periodical around the same time, this choice protects against time variation of writing samples (e.g., by choosing essays that had been written closer to the time of Subjection’s publication). Second, this choice minimizes the risk of distortion of the samples from editorial interventions to the text (the editor was the same for both H. Mill’s and J. S. Mill’s essays). Third, the word limit allows more essays by J. S. Mill to be included in his training set, thus offering more subject-matter variation (which is something that characterizes H. Mill’s texts), and fourth, it keeps the corpus size balanced. Finally, this text-selection process addresses a potential objection, that is, that we are cherry-picking J. S. Mill’s texts. We also included part of J. S. Mill’s manuscript essay on marriage. This was a mirror essay to that of H. Mill, included in her own authorial set.

For comparisons to H. Taylor, we chose various essays on dissimilar subjects by J. S. Mill published in 1860s. As in the previous case, we tried to eliminate potential bias in favour of J. S. Mill, especially as regards subject-matter variety, since Helen Taylor’s corpus is made by a group of dissimilar subjects. It would have been enough to use just one book by J. S. Mill, but we opted for variety, given the external constraints (potential editorial interference, word limit, necessity of including excerpts, etc.) to Helen Taylor’s work. For the full list of texts, see Table 1.

As regards the clean-up part of the corpus preparation process, we removed extracts and long quotes, and we removed title/chapter headings, pagination, notes, references, italicization, and editorial notes/comments/variants (including characters/symbols to this purpose). Trying to intervene minimally with the text of the manuscripts, articles, chapters, or notices that make up our corpus, we preserved spelling and punctuation.

Neocleous and colleagues (Neocleous and Loizides 2020; Neocleous, Kataliakos, and Loizides 2022) developed a machine learning system to address various authorship identification problems in John Stuart Mill’s corpus. Using computational text analytics techniques, they ran a series of tests to determine the authorship of On Liberty (J. S. Mill [1859a] 1977), The Subjection of Women (J. S. Mill [1869] 1984), and On Social Freedom (J. S. Mill 1907). In the first article, they focused on feature selection by building models using different combinations of extracted features to classify the three “disputed” texts into one of three possible authors (J. S. Mill, H. Mill, H. Taylor) or in a class called “joint productions” (collaborative work by J. S. Mill and H. Mill). The essays were attributed to J. S. Mill, though there was evidence of possible influence (with no indication of any specific text segment written by any other author than J. S. Mill). In the second attempt, they simplified the models into training binary classifiers to distinguish between two authors (J. S. Mill and Harriet Mill), focusing on Liberty and Subjection. The author corpus was divided into text segments of varied length used as separate instances, instead of feeding the entire corpus as a single instance. Neither article engaged in a detailed discussion of the problems of co-authorship involved. To further validate the effectiveness of the proposed methodology, Neocleous and colleagues (Neocleous, Kataliakos, and Loizides 2022) applied the same approach to a Benchmark Dataset, with results, according to the authors, that matched the scholarly consensus. A third article by Loizides and colleagues (Loizides, Neocleous, and Nicolaides 2023) focused exclusively on Social Freedom, conducting both a traditional and non-traditional authorship identification analysis of said text. In the non-traditional analysis, Loizides and colleagues used the methodology developed by Neocleous and colleagues (Neocleous, Kataliakos, and Loizides 2022).

In the present iteration of the investigation, we used the same ML system as Neocleous and colleagues (Neocleous, Kataliakos, and Loizides 2022) to focus on Subjection. We began with the construction of a dataset of 39 texts. As already discussed, specific texts from this dataset were selected to create authorial profiles, where each corpus was assigned a class corresponding to its known author (e.g., Class 1 for Harriet Mill; Class 2 for John Stuart Mill in the first test; Class 1 for Helen Taylor; Class 2 for John Stuart Mill in the second test).

Each author’s texts were first combined into one long text. Then, from that long text, we created a series of instances of a pre-defined number of words, that is, segment lengths (50, 100, 200, 500, up to 1000 words). There was no overlap between consecutive instances (an instance starts one word after the last word of the previous instance). Breaking texts into many segments provided more training instances, which helped the machine learning algorithms learn more reliably from limited data. Using segments of uniform length controlled for variation in essay length. It ensured that classification was based on stylistic patterns rather than text length, since all author corpora were comparable in size. Despite the well-known limitation of using small text sizes (at least up to 500 words, as we saw), the division into text segments allowed the models to classify different parts of Subjection to different authors. This was crucial for this study. It enabled us to attempt to detect whether portions of Subjection resemble H. Mill’s style or H. Taylor’s style, even if Subjection is largely attributed to J. S. Mill. This was the only way to test J. S. Mill’s claims of specific parts of authorship by H. Taylor and H. Mill.

Before classification, the texts underwent preprocessing to ensure consistency and comparability. This involved converting all documents to a standardized format, preserving punctuation, segmenting the text into structural units (sentences and words), and tagging words with their grammatical categories using the CLAWS part-of-speech tagger. By maintaining function words and punctuation, we retained key stylistic markers essential for authorship identification.

For the analysis itself, we extracted a diverse range of linguistic features chosen to capture stylistic and structural patterns that may distinguish the author: counts, punctuation usage, CLAWS grammatical tags, and n-grams (both unigrams and bigrams). First, the “Counts” category includes statistical measures such as the average and standard deviation of sentence and word lengths. Second, the “Punctuations” category covers 12 specific punctuation marks. Third, the “CLAWS tags” category refers to 138 predefined grammatical tags assigned to words by the Constituent Likelihood Automatic Word-tagging System (CLAWS), a tool developed by the University Centre for Computer Corpus Research on Language (UCREL) and accessible online (UCREL 2025). Fourth, the “n-grams” category comprises Unigrams (which count the occurrences of individual words) and Bigrams (which track the frequency of consecutive word pairs). Also, all these features were combined into a single feature set (All_Features). Additionally, we applied dimensionality reduction techniques. Specifically, we used two methods: one statistical approach that selected the most discriminative features based on frequency thresholds, and Principal Component Analysis (PCA), which retained 95% of the variance in the data.

With the refined feature sets, we proceeded to train three machine learning classifiers: k-nearest neighbours (k-NN 1 and 2), support vector machines (SVMs), and decision trees (DTs). The ML system employed for each author a cross-validation approach. This means that for each classification model, the author’s corpus was split into two parts: the training set and the validation set. The training set was chosen randomly to form 70% of the initial author’s dataset, and the remaining 30% was used for validation. This yielded a validation accuracy score. The higher the training and validation accuracy score, the higher the confidence in the results of the classification. Still, validation accuracy scores are just indications of probable efficiency in the performance, not an absolute indicator. (For more information on the methodology, see Neocleous and Loizides 2020 and Neocleous, Kataliakos, and Loizides 2022).

The validated models were applied to Subjection. Each classifier was tested to identify the optimal combination for distinguishing between the two possible author classes—John Stuart Mill or Harriet Mill, and John Stuart Mill or Helen Taylor. By analyzing the classification probabilities over the trained authorial classes, the ML system classified the most likely author for the n-word (n = 50, 100, 200, and so on) text segments from the disputed text.

In total, we ran three sets of tests: first, we compared texts by J. S. Mill and H. Mill written in the early 1830s; second, we compared texts by J. S. Mill (from 1862 to 1869) and H. Taylor (from 1860 to 1881); third, we also compared texts by J. S. Mill and H. Mill written in the early 1830s, with the addition of Emancipation (1851 [H. Mill 1998]) and “Fitzroy’s Bill” (1853 [H. Mill 1998]) for H. Mill’s authorial set and “Centralisation” (J. S. Mill [1862] 1977) for J. S. Mill’s authorial set. In this case, like in the second set of tests, this selection maintained relative corpus size balance, even though it created a subject-matter bias in favour of Harriet Mill. For simplicity’s sake, we call the first set of tests HM1vJSM1, the second set HTvJSM, and the third set of tests HM2vJSM2 (see Table 2).

SVMs performed best in this classification task, demonstrating robust performance across different conditions and being the most accurate both in the training and the validation phases of the three tests. It was the only classifier that achieved 100% training accuracy in all word-group sizes (on SVM, see, e.g., Cortes and Vapnik 1995; Tong and Koller 2001; Cervantes et al. 2020). The article’s supplementary material contains full details on the validation accuracy scores of all classifiers across different feature sets and text sizes (see Supplementary Material).

The choice of features affected accuracy. The optimal feature set varied depending on the classifier and text segment size (see Figure 1). SVM was most accurate with unique words (unigrams), unigrams with Principal Component Analysis (PCA), or all features combined with PCA (on PCA, see, e.g., Jolliffe 2002; Abdi and Williams 2010). Unique pairs of words (bigrams) or bigrams combined with PCA performed well, but not as well as unigrams (see Figure 2).

In all sets of tests, SVM had consistently high accuracy across different feature sets and word groups (see Figure 3). In the 600-word group, for an example from the HM1vJSM1 tests, SVM (combined PCA) reached 100% (and 95% overall) validation accuracy. In the HTvJSM tests, thanks to larger training sets by Helen Taylor and John Stuart Mill, the four classifiers were more accurate overall as regards validation. Here, in the 1000-word group, SVM with All_Features without PCA reached 100% validation accuracy (but 86.61% All_Features_PCA). In the 600-word group, SVM_All_Features_PCA reached 100% validation accuracy (and 96.96% All_Features). In the HM2vJSM2 test, this time thanks to a larger training set by both Harriet Mill and John Stuart Mill, SVM improved its validation accuracy overall, but Unigram_PCA or Bigram_PCA was steadily less accurate than in the previous two sets of tests. Still, SVM was consistently more accurate than any other feature extraction from the other classifiers. In the 500-word group, SVM reached 97.67% All_Features validation accuracy (but 53.48% for All_Features_PCA). We highlight these three cases because they had the best All-Features validation accuracy from each test.

So, are there traces of H. Mill’s and H. Taylor’s authorial hands in Subjection? If we opt for a simple answer, then “no.” There is not much evidence in the most accurate of the classification attempts. In the HM1vJSM1 tests, the highest validation accuracy with SVM_All_Features was reached in the 600-word group (95%). In the HTvJSM tests, the highest validation accuracy with SVM_All_Features was reached in the 1000-word group (100%). In the HM2vJSM2 tests with the enlarged corpus by both Harriet Mill and John Stuart Mill, the highest validation accuracy with SVM_All_Features was reached in the 500-word group (97.67%). Only in the first case (HM1vJSM1) was some text from Subjection classified under an author different from John Stuart Mill. Of the 81 instances created for this comparison, just four were classified under H. Mill (two from the first chapter, one from the second chapter, and one from the third chapter).

To opt for a more complex answer, we need only to look at the second-best validation accuracy score and the resulting classification for each test. The picture changes significantly. In the HM1vJSM1 tests, the second-highest validation accuracy with SVM_All_Features was reached in the 300-word group (92.5%). Now, no text is classified under H. Mill. When we look at the same word-group for the HM2vJSM2 test, 22 out of 162 instances are classified under H. Mill. In the HTvJSM tests, the second-highest validation accuracy with SVM_All_Features was reached in the 800-word group (97.67%). Helen Taylor had a share of 14 instances out of 60 total. In the HM2vJSM2 set, the second-highest validation accuracy with SVM_All_Features was reached in the 900-word group (95.65%). This time, the algorithm classified 13 out of 54 instances under H. Mill. Yet, no text is classified under H. Mill in the same word-group for the HM1vJSM1 test. Looking at the two sets of tests with the larger authorial sets, the validation accuracy of the algorithm increases, but so does the share of text instances of both mother and daughter. However, we should be cautious in generalizing from less accurate models. More importantly, the subject-matter bias in favour of J. S. Mill’s collaborators in the HM2vJSM2 and HTvJSM tests further reduces confidence in these results (see Table 3).