AICNT FRT chapter 3

14. AI as a Solution for Neutral Translation

The Bible is the best-selling book of all time, as it serves as the primary basis for many believers everywhere to understand God’s revelation. However, widely used English versions often reflect the inherent biases of translators and editors. The textual biases are then combined with an additional layer of bias reflected by editorial decisions to lead readers down a predetermined interpretive path. The opacity in these translations can sometimes render them closer to a commentary on the meaning of the text instead of a neutral rendering.

Faced with the challenge of human bias in translating Scripture, the question arises: where can one find an impartial English Bible? Developments in the capacity of AI to provide neutral “machine” translation offer the answer. Given the proper input instructions and settings, Large Language Models starting with OpenAI’s GPT-4 and later versions have demonstrated the ability to function as an accurate and consistent tool for translating biblical Greek under appropriate guidelines. The great advantage of the AICNT as a sophisticated machine translation is that it serves to convey the New Testament with optimal transparency, as opposed to exhibiting the gross interpretive bias that many human translations do.

Technologies serve as versatile instruments with the capacity to produce both beneficial and detrimental effects depending on how they are used. When deployed with ethical considerations and stringent guidelines, AI becomes a robust mechanism for enhancing human understanding across various disciplines. In the humanities, machine-learning models can assist in translating ancient texts with a level of consistency and impartiality that is difficult for human translators to achieve.

The AICNT is generated directly from the Greek text provided through OpenAI's API using GPT-4 and later Large Language Models (LLMs). The inputs provided included the system message (specific instructions), the API settings, and the selected Greek text. All text incorporated in the AICNT, including alternate readings of variant manuscripts provided in the footnotes, was rendered utilizing AI (none of it was human-translated).

Utilizing its comprehensive knowledge repository and state-of-the-art computational architecture, GPT LLMs demonstrated exceptional proficiency in accurately rendering biblical Koine Greek. Although no translation is 100% perfect, the AICNT generally provides a more neutral rendering of what the New Testament manuscripts say, as opposed to what human translators believe it means.

15. AI Rendering Methodology

To ensure the highest level of consistency and accuracy in the translated text, the parameters of the AI model were fine-tuned to minimize randomness. GPT-4 and later LLMs were used to render all the translated text.

Rather than use the web interface for ChatGPT, the API for Chat Completion, featuring additional controls, was utilized with custom Python code that defined specific parameters to yield a more deterministic and consistent output.

One such parameter is the Temperature variable. Higher values result in more random outputs, whereas lower values yield a more focused and deterministic result. For this translation, the minimum value 0 was utilized to generate the most deterministic output possible.

The custom API script developed for interacting with OpenAI GPT models included a system message to ensure the best quality output. The system message contained the instructions for translating the Koine Greek text provided. This had the instructions only to use the BDAG lexicon and to give a theologically neutral rendering to minimize interpretive bias.

It is well known that AI could give varying results if not prompted correctly with tight controls and specific instructions. For this reason, the previously described measures were taken to maintain tight parameters for the LLM to function. With such rules, the output was optimized for neutrality and repeatability. Once the program was developed, the only user input was the raw Greek text for generating the output rendering.

16. BDAG Lexicon

The system instructions was to use of the BDAG lexicon as the basis of the translation, BDAG is an acronym pertaining to A Greek-English Lexicon of the New Testament and Other Early Christian Literature, 3rd ed., edited by W. Bauer and based on previous contributions from F. W. Danker, W. F. Arndt, and F. W. Gingrich (Chicago: University of Chicago Press, 2001).

This comprehensive reference work lists thousands of references to classical, intertestamental, and early Christian literature. It is considered the most authoritative and widely used lexicon of the New Testament in the English-speaking world. BDAG contains a thorough and detailed analysis of the meanings and uses of every word in the Greek New Testament and other classical literature. The entries provide a range of information, including etymology, usage in various contexts, historical and cultural background, and cross-references to related words and concepts.

17. Rendered Word Equivalents

When rendering an English translation from the raw Greek text using the controls mentioned above, the GPT Language Model (LLM) output sometimes fluctuated between two equivalent English words (highly equivalent synonyms) when translating particular Greek words provided within a sentence or phrase. This was the primary variation observed when using GPT as a translation tool.

Some examples of these pairs in the output, with the word that is less frequently used in parentheses, are heavens (heaven), Gehenna (hell), beginning (start), taught (instructed), wealth (mammon), purchased (ransomed), gain (acquire), follow (go after), disturbances (insurrections), expectation (foreboding), and disciplined (punished).

In places where the output fluctuated between two equivalent words, it was observed that both word choices were always reasonable options, and those words with more familiarity to the general reader were usually selected. In cases where a less widely understood word choice is used for improved accuracy (for example, Gehenna), a clarifying footnote is added to give the reader a fuller understanding of the word's meaning.

18. Quality Assurance

When evaluating the output differences in providing GPT LLMs with entire chapters compared to a single paragraph or verse, we observed that the number of verses translated simultaneously influenced the output. In general, rendering more text at a time improved readability but reduced accuracy. Rendering a paragraph or sentence at a time typically resulted in the best combination of high accuracy and excellent readability. In instances when the Greek text is more complex or abstract, processing a single verse, phrase, or word at a time was employed to achieve an even more accurate output. Inaccuracies were generally resolved by translating smaller bits of Greek text at a time.

Dustin Smith, Ph.D., biblical scholar of Early Christianity at Spartanburg Methodist College (Spartanburg, SC), reviewed the translation for accuracy and readability. Professor Smith conducted a complete review of the rendered text but did not edit the translation. When an issue was identified, shorter lengths of Greek text were rendered to resolve the inaccuracy. In most cases, this resolved the issue without the need for a footnote. In some cases, a footnote was added to provide clarification.

The translated text was configured in a format suitable for use with various software applications, including those used by translators, to ensure the quality of translations. Such software features several quality checks to check punctuation, capitalization, and grammatical issues, among other things. The process of checking for consistency of translated words between Gospel parallels and re-rendering shorter text lengths with GPT LLMs to improve consistency resulted in an additional layer of quality assurance.

The accuracy was further improved during the process of meticulously reviewing every textual variant in constructing and examining the critical apparatus.

19. Clarifying Footnotes

In addition to the textual variants described in the footnotes, other types of footnotes were added to give clarification and greater transparency, as follows:

Some added footnotes provide Scriptural references for the verses that quote the Old Testament and other texts. “LXX” in the footnote indicates references particular to the Greek Septuagint.
Some added footnotes provide the abridged BDAG definition in places where the translated Greek word may have a wide range of meanings or may not be fully expressed by the English word rendered.
Some added footnotes provide an alternative reading of a specific word or phrase. This applies in cases where the rendered word or phrase may pose difficulty in comprehension, such as idiomatic or cryptic expressions, or where the word or phrase can be expressed or understood differently.
Some added footnotes provide an alternate rendering in those few cases where the AI output fluctuated between two significantly different renderings due to syntactical or textual ambiguity of the source text.
Some added footnotes provide a clarifying explanation about the theological neutrality of how particular verses are translated.

20. Formatting

The AICNT does not use capital letters for pronouns that refer to God, as that would amount to an interpretive bias. However, words such as God, Lord, and Holy Spirit are capitalized as is customary in English translation. It should be noted that Hebrew does not differentiate between upper-case and lower-case letters. Additionally, the initial Greek manuscripts were composed entirely in capital letters. Capitalization (or lack thereof) in the AICNT is not intended to convey a particular theological position or bias.

The AICNT abstains from formatting conventions of modern translations used to distinguish and classify types of text, such as quotations, poetry, or red-lettered words attributed to Jesus. Italicized words are only used in cases where the AI model indicated that it added helper words in English that are not in the Greek text.

The early Greek manuscripts did not exhibit the suggestive formatting in most English Bible translations. Especially regarding explanatory section headings, the formatting exhibited in many popular translations adds a layer of subjectivity and editorial bias that is avoided here. Since section headings are not used, this translation relies on more frequent paragraphing (like the Tyndale House Greek New Testament) to help the reader navigate and search through chapters more efficiently.

Structuring the text in smaller, more frequent paragraphs also serves to enrich the reading experience. The paragraphing of the AICNT exhibits a balance between the conventional paragraph layout of Bible translations (exhibiting large paragraphs) and those translations that provide a verse-by-verse layout (placing each verse on a new line). Smaller groupings of text are also valuable for aiding in readability and comprehension by causing the reader to pause more often and digest a shorter length of text before proceeding.

21. Punctuation

The earliest manuscripts were written in uncial, essentially all capital script without punctuation or word separation. These features, including punctuation, were introduced into minuscule manuscripts (having text with both upper-case and lower-case characters), which began to appear in the 9th century. The earliest dated minuscule manuscript with lower-case Greek characters is the Gospel minuscule 461 from the year 835.[fn]

The punctuation of this edition is generally consistent with one or more of the critical editions in the section entitled “Benchmark Critical Editions.” In some cases, the punctuation is slightly modified to facilitate documenting multiple variants and incorporating the bracketed text of the critical apparatus.

Having provided an overview of the basis and methodology pertaining to the translation of its extensive critical apparatus, we are now pleased to present to you the AI Critical New Testament.

AICNT FRT Chapter 3