One of several famous dicta by the late Martin West is that only someone who had not read M. Griffith's 1977 monograph The Authenticity of Prometheus Bound or who was unwilling to unlearn what he knew no matter the proof to the opposite could still believe that Aeschylus wrote this play. I am not bold enough to be quite as categorical about M.'s book, not because his methods are unsound or his conclusions unconvincing, but because I (like most Classicists, I expect) sadly lack the proficiency in computer science, mathematics and statistics to evaluate them fully. My clear impression, however, admittedly taking the more technical aspects of the work at face value, is that M., building on Griffith and others that followed, provides a robust defence of athetesis, indeed by following a path practically untrodden by Classics, although far from newfangled, namely, cutting-edge automated authorship identification. M. is impressive in the array of tools he employs for pinpointing the ‘authorial trace’ of Prometheus Bound (Pr.) and for comparing it, on the one hand, to that of Aeschylus, on the other to Sophocles and Euripides (and Aristophanes and other authors whose texts are used for verification purposes). M. is just as comfortable moving from traditional philology to corpus linguistics, sociolinguistics and stylistics, and from quantitative ‘stylometrics’ to state-of-the-art computer-assisted experiments. The book still bears the signs of the Ph.D. thesis from which it derives (overburdened footnotes and excessive bibliography on peripheral matters, for instance) and will probably seem too technical to most Classics readers, especially in its third part. Nonetheless, it must be hailed, even celebrated, as an earnest application of a method that should have been applied long ago for tackling such a notorious problem of authorial (false) identity as the paternity of Pr.

Traditionally, style, which is the sum of the linguistic choices consciously made and aesthetically controlled by the author, has been an obvious focal point of authorship identification studies. In recent decades, however, scholars trying to prove (or disprove) the authenticity of texts, literary or non-literary, have also been investigating what M. calls authorial ‘trace’ (a term coined by him). ‘Trace’ comprises ‘textual features, linguistic habits, that are seldom noticed by the reader and that it is extremely unlikely – impossible in the case of lengthy texts (such as Pr.) – that the author planned or deliberately modified’ (p. 1). These features ‘include the frequency with which often repeated words [sc. most importantly, ‘function words’, such as articles, prepositions etc.] vary in an author's oeuvre and the frequency with which some meaningful (grammatical forms) and non-meaningful (strings of phonemes) word constituents [for these the technical term is n-grams] occur’ (p. 12). If the elements that constitute an author's ‘style’ are subjective pieces of evidence in authorship studies and are usually convincing only in their accumulation (if that), ‘trace’ indicators lean to the side of ‘hard science’ and thus constitute much more reliable markers of authorship or chronology. ‘Trace’ indicators are too minuscule to be studied by hand and eye alone; hence the practice of ‘microstylistics’ or ‘trace detection’ employs specially designed software, which can reveal associations far beyond the human reach.

Following an introduction, which unpacks the theories and practices of ‘stylometry’ (the quantitative analysis of style) and ‘trace detection’, as applied to authorship studies, with emphasis on the introduction of ‘thinking machines’ from the 1960s onwards, and a chapter summarising the question of Pr.'s athetesis, the main body of M.'s book comprises two major parts. In the first M. mostly rehashes stylometric studies attempted after Griffith and along the tracks of Griffith's categories, showing how recent, computer-assisted and statistical examinations confirm the athetesis case invariably. In the second part M. conducts a number of original automated experiments to detect Pr.'s ‘trace’.

The quantitative analysis regards style indicators such as (i) verbal repetitions; (ii) the interchange between -οις and -οισιν endings in dative plural; (iii) the dramatic use of resolution; (iv) hiatus in enjambment; (v) the presence of Eigenwörter; (vi) hyperbaton; (vii) complementation with ὅτι/ὡς finite clauses; and (viii) sentence length. In all these cases Pr. had already been shown by Griffith and more recently by P.E. Pickering and T.B. Allison (who used computer automation, statistics, as well as the tools of modern disciplines such as sociolinguistics and corpus linguistics) to be distinct from the Aeschylean corpus and to lean towards Sophoclean practice. M. reconfirms all earlier findings – in the case of sentence length, conducting a fresh study differing from Griffith's in three respects: (a) it is semi-automated; (b) it takes into account more plays as well as both long speeches and dialogue parts, which Griffith left out; and (c) it studies sentence length separately in dialogue (iambic and trochaic) and in lyric. The results are still consistent with athetesis: the author of Pr. tends to include very long sentences in his choral lyrics, which Aeschylus does not, and he does so without a clear dramatic objective (p. 140).

The main thrust of M.'s book is concentrated on pp. 143–225, where he applies automated authorship attribution techniques to the study of Pr.'s trace. As per the established practice in such studies, M. investigates function words (prepositions, conjunctions, particles, pronouns, articles, whose use reflects the unconscious grammatical-syntactic structuring of a text) and ‘n-grams’ (meaningless strings of one, two or three letters/sounds recurring regularly enough to indicate latent linguistic tendencies or what M. calls ‘the sub-atomic structure of a text’ [p. 145]). N-grams are considered to be ‘the widest and most thorough representation of authorial trace possible’ (p. 221). M. runs three ‘experiments’ by way of algorithmic models, in which he tests Pr. against not only the secure plays of Aeschylus and the works of Sophocles and Euripides but also against Aristophanes and Homer, to achieve what is termed ‘authorship verification in multi-class classification terms’ (p. 149): that is, not only to determine whether Aeschylus authored Pr. but also to shed light on the affinity of this text's trace with the corpus of ancient Greek dramaturgy at large. Apart from investigating the complete playtexts, M. also tests ‘sub-corpora’ of (a) lyric and spoken parts and (b) gender-defined parts. Important for M.'s methodology is that he cleans the texts he feeds into his software from all ‘acquired features’ (punctuation and, more importantly, conjectures incorporated into the TLG texts, which previous automated studies mostly use). Furthermore, to avoid producing a false feeling of homogeneity in the computer results, M. removes single editor bias by using a different edition for each text rather than a single ‘standard’ one, often outdated, such as D.L. Page's Aeschylus (included in TLG). The kind of machine learning brought to bear on the texts is of two types: ‘supervised’ and ‘unsupervised’. In ‘supervised’ experiments the machine learns from data that are categorised a priori – here, labelled as ‘genuinely Aeschylean’. In ‘unsupervised’ learning, the machine detects commonalities among texts for which it is given no prior authorial or other kinds of information. To test function words M. performs two kinds of ‘unsupervised’ analyses, ‘Principal Components Analysis’ and ‘Hierarchical Agglomerative Cluster Analysis’, and a ‘supervised’ one, ‘Burrows’ Delta’. For n-grams, he uses the ‘Common n-grams’ and the ‘Support Vector Machines’ models, both of which are supervised, though the first is ‘profile-based’, concatenating all the available training texts per author in a single file, and the second is ‘instance-based’, representing each separate text individually. Sparing the Classicist reader the technical details of the methods, which for lack of expertise I am sure to misrepresent, I will jump straight to M.'s conclusions, which show: (a) that the authorial trace of Pr. is unified, thus the old ‘revised-text’ or ‘text modernisation’ hypothesis, namely, that a second hand completed or reworked an initial dramatic body left unfinished by Aeschylus, cannot hold; (b) that there is no special correlation between Pr. and the six secure plays by Aeschylus. In fact, in all the automated experiments conducted, and especially those adopting the Support Vector Machines model, which, as M. informs us, is ‘the most sensitive broadly tested authorship attribution model to date’ (p. 221), Pr. is never assigned to Aeschylus; on the contrary, Pr. tests closest to plays written in the decade 440–430 bc, to which it most probably belongs.

M.'s book is laborious to read as it targets two distinct readerships at once, Classicists and computer scientists. As confessed already, possessing none of the latter's know-how, I must reserve judgement on the technical methods applied. As a Classicist, however, I was swayed. M., alongside Pickering and Allison, has restored my confidence in Griffith's conclusions, which M.P. Pattoni's thesis had somewhat shaken.