Is it a remix?

The question, and what can be tested

When this translation was shared publicly, the most common objection was immediate and reasonable: a language model has read every famous English Odyssey (Lattimore, Fagles, Wilson, and the rest are in its training data), so its "translation" must really be a re-synthesis of theirs. One commenter called the opening "a crappy mishmash of Lattimore and Fagles"; another wrote that it "clearly combines a few translations line by line."

Part of that objection is untestable, and we concede it up front: prior translations may well live in the model's weights regardless of what it was prompted with, just as they live in the memory of any human translator working today, and no one, including us, can say which ones or how deeply. No output analysis can establish what a model, or a person, was shaped by. What can be measured is reuse: whether the text of this translation reproduces the wording of its predecessors, at what lengths, and how that compares with how much human translators of the same poem reproduce each other. This page reports those measurements, states exactly what they do and do not establish, and takes the accusation's strongest form seriously, including a word-by-word look at the very lines the accusation was made about.

Method

Twelve texts: this translation (~124,000 words) and eleven English Odysseys spanning 293 years. The panel: Pope (1725), Cowper (1791), Butcher & Lang (1879), Palmer (1891), Butler (1900), Murray (1919, the Loeb prose facing the very Greek text this project translated), Lattimore (1967), Fagles (1996), Johnston (online 2002, book 2006, revised since), Wilson (2017), and Green (2018). Each is trimmed to translation body only: prefaces, introductions, and footnotes removed, so that quoted or editorial matter cannot masquerade as translation (this matters; see "What copying looks like" below). Everything is lowercased, punctuation, possessives, and diacritics stripped, and proper names normalized across traditions (Ulysses→Odysseus, Lattimore's Telemachos→Telemachus, Green's Athēnē→Athena), so spelling conventions can neither hide nor simulate overlap.

For every pair of texts, two measurements: the fraction of one text's n-word sequences (n = 4…8) appearing anywhere in the other, taking the larger of the two directions; and every maximal verbatim shared passage of 12 or more words, with no cap on length. All 55 human-vs-human pairings are computed with the identical procedure and shown below, the smallest grouped for space. The controls are the whole point, since any two literal translations of the same Greek converge substantially without either copying the other. The Lattimore, Green, Palmer, and Wilson texts come from scans; residual OCR noise systematically depresses comparisons involving those texts. Palmer's and Wilson's running page-heads are stripped under the body-only rule (leaving Wilson's 394 head-lines in shifts her pairings by a few percent relative). A committed manifest (tools/corpus-manifest.txt) records the expected token counts and cleaned-corpus hashes for all twelve corpora, plus source-file hashes for the inputs the script does not download itself, so a rerun can check extraction as well as method (the auto-downloaded inputs are re-fetched from the recorded URLs, so reproducing those rows depends on the hosts serving the same files). Choosing the larger direction is conservative against this translation, but it is a choice; the script prints both directions for every pair.

Public-domain texts come from Project Gutenberg and the Perseus/Scaife library; the script downloads them and prints a hash of each cleaned corpus. Lattimore, Fagles, Wilson, and Green, still in copyright, were measured from privately held copies. Johnston's copyright status is version-specific: his canonical 2024 document declares itself public domain, while the May 2016 source actually measured carries noncommercial-use terms, so only statistics from it are published and its full source chain (URL and hashes at every derivation step) is recorded in the manifest. Only statistics are ever published, never the texts. A public clone of the analysis script reproduces the automatically downloaded rows (subject to the hosts serving the same files, verified against the manifest's cleaned hashes) and reproduces the other rows given the source files the manifest identifies.

Results

And the complete human-vs-human control matrix, all 55 pairings of the eleven prior translations, same procedure, sorted:

Run counting is anchored in one text of the pair (the first named below; the table above anchors in this translation). The two directions differ only slightly, e.g. Murray · Butcher & Lang counts 493 or 488 depending on anchor. Counts for the closest human pairs: Murray · Butcher & Lang share 493 runs of ≥12 words (105 of ≥16, longest 32); Murray · Palmer 52 (10, longest 20); Murray · Green 43 (4, longest 21); Murray · Lattimore 21 (2, longest 21); Green · Lattimore 17 (0, longest 15); Lattimore · Fagles 3 (0, longest 13).

Read plainly, the tables say six things:

Overlap tracks literalness of method, in humans and machine alike. Among humans, the literal translations occupy the entire top of the matrix while Cowper and Pope, translating the identical poem, share 0.1%. This translation's affinities follow the same gradient: highest with the most literal predecessors, lowest with the poets. That profile is consistent with independent literal translation, and not with a pastiche of the famous poetic versions.

The matched-method human control behaves like the machine. Green's 2018 translation (literal, modern register, line-matched to the Greek, the closest existing human analogue to this project's method) shows the second-highest human affinity in the whole matrix (5.1% with Murray), and nobody supposes Green copied the Loeb. This translation's relationship to Green is numerically very similar to Green's own relationship to Murray: 5.3% vs 5.1% at 5-grams, 49 vs 43 long runs, longest 20 vs 21. The model sits with the line-faithful literalists where a new member of that method would sit. Green also supplies the table's sharpest test between the two explanations for overlap, because fame and method point in opposite directions across his row and Fagles's. Training-data prevalence cannot be measured directly; but to the extent that quotation, anthologizing, and course adoption proxy it, Fagles is among the most-reproduced English Odysseys ever published and Green's 2018 version among the least-quoted modern ones. Memorization predicts affinity should follow that prevalence; method-convergence predicts it should follow literalness. The observed result: three times the affinity for the little-quoted methodological sibling (5.3%) as for the ubiquitous poetic one (1.7%). Century-old literalists nobody reads (Butcher & Lang, Palmer) also outscore Fagles. Within this panel, affinity follows method, not fame. This is an inference over the table, not a measurement, and it shares the table's limits.

The elevated rows, stated in full. Against Murray (8.9%) and Lattimore (6.3%), this translation runs roughly twice Green's own affinity to the same texts (5.1%, 3.2%), higher than every human pairing except Murray · Butcher & Lang, a single towering pair at 15.2% that is itself no certificate of independence (Murray postdates Butcher & Lang by forty years in the same archaizing tradition; the median human pairing is about 1.0%). Stated distributionally, because a row-by-row defense can hide the aggregate: four of this translation's rows (Murray, Lattimore, Green, Butcher & Lang) exceed every human pairing except that one, and the median of its eleven rows is 2.8% against the human-pair median of 1.0%. The four are the panel's most literal predecessors, which is the direction the method account predicts, but the elevation itself is a fact of the table. Two explanations fit it and these numbers cannot decompose them: this translation is stricter than even Green (word-order discipline, and Murray's own Greek text as source; line-count discipline does not by itself predict a high row, per the Wilson result below), and the model may carry some real gravity toward the literal translations in its training data. What the numbers do bound is the form any such influence took: run counts rise in the same rows where n-gram overlap is high (170 runs of ≥12 words with Murray, against 52 for the closest human pair outside Murray · Butcher & Lang); the longest run, 29 words, sits below Murray · Butcher & Lang's 32 but above the other 54 human pairs, whose own maximum is 22; and verbatim reuse is capped by the next section.

The Fagles signal is modest but, in fairness, not absent. This translation's 1.7% overlap with Fagles exceeds every human-vs-Fagles pairing in the panel (the highest are Lattimore's 1.1% and Green's 1.0%). That is worth stating because it is the kind of detail a defense would prefer to omit. But it is a fraction of the Murray and Lattimore affinities, it comprises five shared runs of 12+ words and none of 16+, and the longest match (14 words) is a Homeric formula that Fagles also shares verbatim with Murray. The data do not support Fagles as a meaningful textual donor; they cannot rule out minor influence.

The newest row was predicted before it was measured, and the scorecard is mixed. Wilson's translation entered the panel under preregistration: five predictions and two adverse-evidence thresholds were committed to the public repository and pushed on 2026-08-08, four days before a measurable copy of her text was acquired (the push is the evidence that matters; a commit timestamp alone is author-controlled). One disclosure belongs beside that claim: what the preregistration establishes is narrow. The predictions were made before any Wilson corpus was supplied to the analysis or any full-corpus statistic computed; they were not translator-blind (her proem and the published descriptions of her method were known to this project well before the preregistration; an internal note from July analyzes her opening lines), and the model coauthor's latent training exposure to her text is unknown, which is the premise this page opened with. Scored plainly: both range predictions missed low (5-gram predicted 2.5–4.0%, measured 1.4%; the Wilson · Murray human control predicted 1.5–3.0%, measured 0.9%). The ordering prediction partly failed: Wilson was predicted between Palmer and Fagles and landed below Fagles, lower than every row except the two eighteenth-century poets, Cowper and Pope. The longest-run prediction hit: predicted under 20 words, measured a single 13-word run ("Son of Atreus, why ask me this? You have no need to know," Proteus at 4.492). The prediction that Wilson would run below Green against the whole human panel failed at one pairing (Wilson · Cowper 0.15% against Green · Cowper 0.13%). The thresholds that would have counted against the convergence account as stated (affinity above Green's 5.3%, or any 24-word run) were never approached; be clear about what that asymmetry means: only high-side surprises could have forced revision, so these low-side misses cost the account a mechanism, not the conclusion.

The mechanism they cost it is worth spelling out. The preregistration expected Wilson's line-for-line constraint (the same constraint this project worked under) to push overlap up and her compressed, deliberately plain diction to push it down, with the constraint winning. It did not. With Green (line-matched, 5.3%) and Wilson (line-for-line, 1.4%) both on the table, line discipline alone is insufficient to determine a row's height. The observed differences are consistent with differences in literal diction and in formula policy, but these data cannot isolate their contributions. On the fame question, her row points the same direction as the Green–Fagles comparison above (the most-discussed English Odyssey of the century shows less exact-text affinity than century-old literalists nobody reads), but it is weaker evidence than it looks, and we flag the confound ourselves: Wilson deliberately varies Homer's repeated formulas where this translation, like Murray and Fagles, repeats them verbatim, and that choice suppresses exactly the n-grams and runs these tests count. A pattern consistent with affinity-follows-method, not a second independent test of it. Finally, the detail a defense would prefer to omit: this translation's 1.4% with Wilson exceeds every human pairing with her (the highest, Lattimore's, is 1.0%), an excess of roughly forty percent in relative terms, in the same direction as the Fagles and Lattimore excesses noted above, and a residue these numbers cannot decompose.

A second preregistered row, measured on a text that turned out to be moving. Ian Johnston's translation entered the panel hours after Wilson's, same day, same protocol (predictions pushed publicly before measurement, the file inspection itemized in the preregistration; one protocol gap, fixed for any future row: the source-file hash was committed with the results, not before them). Scored honestly, with the dependence between predictions stated: the placement predictions were three descriptions of one outcome, not three separate successes. The preregistered range (1.8–3.5%, between Palmer and Fagles), the eleven-row ordering, and the prevalence reading below all turn on where Johnston's single row landed, and it landed where predicted: 2.8%, sixth of eleven (the other ten rows' order was already known). The Johnston · Murray control also hit its range (2.7%, predicted 1.5–3.0%). The prediction that Johnston would sit above Wilson and below Green across the whole panel failed at three pairings, not one: at Cowper it violates both clauses (Johnston's 0.14% is below Wilson's 0.15% and above Green's 0.13%), at Pope it exceeds Green's (0.02% against 0.01%), and at Wilson herself Johnston's 1.5% is nearly twice Green's 0.8%. It holds for the six remaining comparators. One reading of the Wilson failure, offered as a hypothesis and not a result: plain modern registers may converge on each other. The descriptive fact is that Johnston's arrival moved the human-pair median from 0.8% to 1.0% (he added eight pairings above the old median, two below); what causes his mid-range affinities is not established by these numbers.

The row's most instructive result took a third adversarial review to surface, and it concerns the text itself. The longest-run prediction (under 20 words) failed against a 21-word shared passage, Phemius setting down his lyre to plead for his life (22.339–341). But Johnston has revised his translation repeatedly since it first went online in 2002, and the file measured here derives from a May 2016 snapshot of his PDF, publicly hosted since 2016 (the full source chain and hashes are in the manifest; a committed script reproduces the measured text byte-for-byte from the intermediate EPUB, while the PDF-to-EPUB conversion is documented but not reproduced). In his current official text (the 2024 public-domain PDF), the same lines read "clasp the knee … He set down the hollow lyre, left it on the ground," and the run does not exist: against that canonical text (cleaned the same way; script and both versions' hashes in the repository), the row reads 2.4% instead of 2.8%, longest run 12 words, still inside the preregistered range and the same ordering slot. We state both readings of that difference. Read one way, the row's headline anomaly is an artifact of which revision you pick up. Read the other way, this exact 2016 revision was publicly hosted for years and so had substantial opportunity to enter training data; we cannot establish that it did, or rank its exposure against Johnston's other revisions, but the 21-word run remains the result for that version, and the 2024 revision cannot dismiss it. These numbers cannot decompose memorization from convergence for a single passage; the run stays below the 24-word adverse threshold and is longer than the measured Johnston text shares with any human (maximum 18, with Murray). On web prevalence: the preregistration hoped this row would test whether training availability drives affinity, and it cannot. Prevalence was never measured, the premise ("plausibly the most training-available English Odyssey") is unverified, and the version question shows that "Johnston" is not even one text. His mid-table row is consistent with affinity-follows-method and nothing stronger. The residual, as for every row: this translation's affinity to Johnston modestly exceeds every human pairing with him on either version (2.8% against Murray's 2.7% as measured; 2.4% against 2.3% on the canonical text).

Bounding verbatim reuse

The "stitching" version of the accusation (that the text was assembled from pieces of prior translations) was tested against the union of all eleven at once: an n-gram counts as matched if it appears in any of them. Result: 78.1% of this translation's 5-grams and 94.8% of its 8-grams appear in none of the eleven; at 4 words, 63.6% appear nowhere. The longest passage shared with any predecessor, anywhere in 12,107 lines, is 29 words (with Murray, the facing translation of its own source text).

What do those figures actually exclude? To calibrate the test rather than assert about it, we built the thing being alleged, at several scales: synthetic splices cycling verbatim chunks of Murray, Butcher & Lang, and Lattimore, with chunk sizes from 4 to 12 words.

The calibration cuts both ways, and both directions belong on this page. Against long-fragment assembly it is decisive: splices at 8 and 12 words score 2.3 and 3.1 times this translation's figure on the 5-gram measure (1.7–2.1× at 4-grams), so the results strongly disfavor substantial assembly from verbatim passages of roughly eight to twelve words or longer. But a text built entirely of exact five-word fragments scores 20.7% on the 5-gram test, near this translation's 22.0%, because most sliding windows cross fragment boundaries. N-gram union statistics cannot exclude short-fragment mosaic composition at the 4–6 word scale, and we do not claim they can. Two further limits on the calibration itself: it is a pure, unedited splice, so partial mixtures and edited chunks are uncalibrated; and these are whole-book averages, so reuse concentrated in a single book or episode could hide inside them (per-book statistics are open work, see Scope).

One measured observation bears on that remaining hypothesis without settling it. In these cyclic splices, sub-12-word chunks produce no shared runs of 12+ words, though a differently built mosaic could if adjacent fragments came from the same source. This translation shares 170 such runs with Murray and 49 with Green, a long-tailed profile that resembles the human controls (Green shows 42 against Murray) more than it resembles the synthetic splices. That is suggestive of convergence, not diagnostic of it: a differently constructed mosaic or a mixed text could produce another profile, and nothing in these statistics rules that out. Finer-grained dependence is a question for reading, not string-counting, which is why the next section reads the actual lines the accusation cited.

The opening lines, word by word

The accusation was made about the proem specifically: that its first line "combines" Lattimore's man of many ways and Fagles's man of twists and turns into man of many turnings. N-gram statistics can't adjudicate a three-word phrase; the Greek can.

ἄνδρα μοι ἔννεπε, μοῦσα, πολύτροπον, ὃς μάλα πολλὰ / πλάγχθη, ἐπεὶ Τροίης ἱερὸν πτολίεθρον ἔπερσεν
man [acc.] · to-me · tell · Muse · much-turning [acc.] · who · very much · was-driven-to-wander · after · of-Troy · holy · citadel · he-sacked

This translation renders: "Tell me the man, Muse — the man of many turnings, who was driven / wandering far, once he had sacked Troy's holy citadel." What follows reads the line both ways: where it argues from the Greek, and where it genuinely resembles the accused sources.

"Tell me the man" keeps Homer's bare accusative: ἄνδρα is the direct object of ἔννεπε. Every alleged source inserts an "of": Murray "Tell me of the man," Lattimore "Tell me, Muse, of the man," Fagles "Sing to me of the man." (Green feels the same pull the Greek exerts; he fronts the noun, "The man, Muse — tell me about that resourceful man," but resolves it with "about.") The one construction in the line that is uniquely this translation's is the one closest to the Greek. Against that: the line then repeats "the man" where the Greek has a single ἄνδρα, the same doubling Fagles uses (and Green, in his own way) to bridge the long gap between noun and epithet. A defensible solution to a real problem, but a solution Fagles arrived at first.

"Man of many turnings": πολύτροπος is πολύ ("many") + τρόπος ("turn"). "Turnings" is the root-literal rendering, closer to the Greek than Lattimore's "ways," Fagles's "twists and turns," Murray's "devices," or Green's "resourceful," and identical to none of them. The "man of many ___" frame is shared with Murray and Lattimore, half a century apart. Greek attaches a compound adjective to "man," English has no adjective "much-turned," and the genitive frame is the standard literalist's escape (though not the only one: Butcher & Lang found "so ready at need," Palmer "the adventurous man").

"Driven wandering far, once he had sacked": πλάγχθη is the passive of a verb meaning "drive off course," which is why "driven" appears in Lattimore and Fagles too; "driven wandering" spends two English words unpacking the one Greek verb, a defensible but not inevitable choice. "Once he had" for ἐπεί matches Fagles's construction ("once he had plundered") where Murray and Lattimore write "after"; either is ordinary English for the clause, and this translation's word order follows the Greek's, but the resemblance is there to see.

"Troy's holy citadel" shares its frame with Lattimore's "Troy's sacred citadel" and swaps the adjective; ἱερόν can be either. Divergence inside a shared construction is consistent with independent translation of the same words; it is not, by itself, evidence of independence, and we don't claim otherwise. And a concession critics counted fairly: μάλα ("very") goes untranslated in line 2, the kind of small-word sacrifice to meter and idiom that human translators make on every page.

The honest summary of the proem: these are the most Lattimore-and-Fagles-adjacent lines anyone has identified in the poem: the most translated, most quoted, most memorized hexameters in Greek, where the gravitational pull of famous renderings is at its strongest. They contain real echoes alongside choices independently warranted by the Greek. What a single conspicuous line cannot do, though, is sustain a verdict about a 124,000-word text; that is what the whole-book measurements above are for, and they bound long-form reuse tightly.

What copying actually looks like

The corpus supplied its own demonstration. Butler's Project Gutenberg file contains a 253-word verbatim match with Butcher & Lang, because Butler's preface quotes their rendering of the proem at length in order to mock it. Once prefaces and notes are trimmed away and only translation bodies compared, the Butler/Butcher & Lang figure collapses to three shared runs, longest 14 words. That is the difference between quotation and convergence: real reproduction announces itself in hundreds of consecutive words on the first string search. Nothing remotely like it exists between this translation and any predecessor.

Where translations do converge verbatim, they converge on the same lines: Homer's formulas. The longest passage this translation shares with Fagles and the longest it shares with Green are the same passage: the recurring clothing-promise ("…a two-edged sword, and sandals for his feet, and send him wherever his heart and spirit bid…"), a formula the poem itself repeats five times (14.516 → 21.339) and which this translation, following its rule that repeated Greek formulas recur verbatim in English, renders identically at each return, as Murray, Fagles, and Green do too. The places that look most "copied" are the places Homer copied himself.

A cross-poem aside: against Fagles's Iliad (his voice, a different poem) this translation shares 0.30% of 5-grams (longest run: 8 words, all stock formulas), of the same order as the 0.20% that Murray's 1919 Odyssey shares with it. No formal test of style is claimed here; the point is only that no gross Fagles house-style signal appears.

Scope, and how to check this

What these measurements cannot do: establish what shaped the model (conceded at the top); detect dependence carried in short fragments (the calibration above shows the four-to-six-word scale is outside this test's reach), in syntax, or in interpretive choices, the debts every translator, human or machine, owes their predecessors; or say anything about literary quality, copyright, or the ethics of machine translation, which are real questions outside this page's scope. Fitzgerald (1961) is not yet in the panel. The Wilson and Johnston rows, both added in August 2026, were each preregistered in the repository: predictions and adverse-evidence thresholds committed and publicly pushed before a measurable copy entered the analysis, with the outcomes scored against them above, misses included, and the full results and post-review corrections appended to the same file. Line-aligned comparison and syntactic or semantic similarity measures would probe finer dependence than n-grams can; so would symmetric overlap statistics alongside the directional-max reported here, per-book localization of shared runs, and measured exposure proxies for the fame question. All of it remains open work.

Provenance: this page was prepared with the same class of model that made the translation, so don't take our framing on trust: the method is fully specified above, the script downloads and hashes the corpora, and a public clone reruns the entire table (copyrighted texts supplied by path). Before publication the analysis was audited by a model from a different lab (the same reviewer lineage used in this project's two-agent review passes), which surfaced a run-length cap bug, preface contamination in two corpora, and framing that outran the metrics. The corrections are incorporated above and itemized in the repository's commit history; Green was added to the panel afterward, closing the matched-method gap the audit identified. The Wilson revision went through the same adversarial pass before merging, which caught a misreading of our own table (Wilson described as the lowest row when Cowper and Pope sit lower), a mis-scored prediction, a wrong human-pair median carried since this page's first version (0.5%; the correct figure is 0.8%), and Wilson's unstripped running page-heads. Those corrections are incorporated above and itemized in the notes file alongside the preregistration. A third pass, on the Johnston revision, found the most consequential defect yet: the measured Johnston file was an unidentified revision of a text its translator has repeatedly revised, and the row's 21-word shared run does not exist against his current official text. It also caught a mis-scored prediction, dependent predictions presented as separate successes, and a Murray · Green rounding error (5.0%; correctly 5.1%) that had stood since Green joined the panel. The version-sensitivity analysis above, and the reporting of both readings, exist because of that review. A fourth pass then traced the measured Johnston file to its exact source, a May 2016 snapshot publicly hosted since 2016, now hashed at every derivation step in the manifest with a committed extraction script, and tightened the reproducibility, copyright, and exposure language to match what the tooling actually delivers.