On this Easter Sunday, as we reflect on faith in modern times, it’s worth considering a surprising parallel between the ancient art of biblical transcription and the cutting-edge world of artificial intelligence (AI). Both endeavors—copying sacred texts by hand and training AI large language models (LLMs)—aim to preserve and transmit knowledge with fidelity. Yet, both are prone to errors that can persist, much like fossils embedded in stone, shaping our understanding in unintended ways. These “digital fossils” in AI and scribal mistakes in biblical manuscripts reveal a shared human struggle to maintain the purity of information across time.

Digital Fossils in AI: The Case of “Vegetative Electron Microscopy”

A recent article in The Conversation (April 15, 2025) by researchers investigating AI errors uncovered a curious term plaguing scientific papers: “vegetative electron microscopy.” This phrase, which sounds plausible but is nonsensical, is a “digital fossil”—an error entrenched in AI systems, as detailed in the article “A weird phrase is plaguing scientific papers – and we traced it back to a glitch in AI training data” (https://theconversation.com/a-weird-phrase-is-plaguing-scientific-papers-and-we-traced-it-back-to-a-glitch-in-ai-training-data-254463).

The error began decades ago when two 1950s papers from Bacteriological Reviews were digitized. A scanning glitch merged “vegetative” from one column with “electron” from another, birthing the phantom term. Years later, translation errors in Iranian papers compounded the mistake, as the Farsi words for “vegetative” and “scanning” differ by a single dot. By 2025, the term appeared in 22 papers, according to Google Scholar, even prompting retractions and corrections from publishers like Springer Nature and Elsevier.

The real trouble emerged when AI models, trained on vast datasets like CommonCrawl, absorbed this error. Tests showed that models like OpenAI’s GPT-3 and GPT-4o consistently generated “vegetative electron microscopy,” suggesting it’s now a permanent fixture in their knowledge bases. Fixing such errors is daunting due to the sheer scale of datasets—millions of gigabytes—and the opacity of commercial AI training processes. These digital fossils highlight how AI can perpetuate errors, much like ancient scribes inadvertently altered biblical texts.

Scribal Errors in the Bible: Parablepsis and Homoeoteleuton

Ancient scribes, tasked with copying the Bible by hand, faced challenges eerily similar to those of modern AI. Their errors, often unintentional, arose from the limitations of human perception and the complexities of the Hebrew and Greek languages. One common mistake, known as parablepsis occasioned by homoeoteleuton, occurred when a scribe’s eye skipped lines ending with similar words.

Consider Luke 12:8-9, where Jesus speaks of acknowledgment and denial before “the angels of God.” In some manuscripts, the text reads:

• “…will acknowledge before the angels of God”
• “…will be denied before the angels of God”
• “And everyone who speaks a word against the Son…”

A scribe, copying the first line’s ending (“before the angels of God”), might glance back at the manuscript and mistakenly fixate on the second line’s identical ending. Thinking they’d already copied it, they’d skip to the next line, omitting the crucial statement about denial. This error, called parablepsis due to homoeoteleuton (similar line endings), could alter the theological emphasis of the passage. Readers can verify this by comparing Luke 12:8-9 in different Bible translations, noting variations in older manuscripts.

Another example is found in 1 John 2:23: “No one who denies the Son has the Father; whoever acknowledges the Son has the Father also.” Some Byzantine manuscripts omit the second clause due to homoeoteleuton, as “has the Father” appears twice. This omission, stemming from a scribe’s eye skip, could weaken the text’s affirmation of the Son’s necessity for a relationship with the Father. Check this verse in a study Bible to see textual notes on manuscript variations.

Shared Challenges: Errors of Sight and Memory

Both AI and scribal errors often stem from “errors of sight” or “errors of memory.” In AI, the “vegetative electron microscopy” glitch arose from a visual misparse during digitization, akin to a scribe misreading similar-looking Greek letters. For instance, in 1 Timothy 3:16, the Greek word hos (“who”) was sometimes confused with a symbol for “God” due to a single stroke’s placement, potentially altering the verse’s Christological implications. Readers can explore this by comparing 1 Timothy 3:16 in the King James Version (which reads “God”) versus modern translations like the NIV (“He”).

Errors of memory also plague both domains. AI models, trained to predict the next word, may “recall” a corrupted term like “vegetative electron microscopy” because it’s embedded in their data. Similarly, a scribe, memorizing a line before writing it, might substitute a synonym or transpose words. In Mark 14:65, the Greek elabon (“received”) was sometimes transposed to ebalon (“struck”), changing whether servants “received” or “struck” Jesus. This subtle shift, verifiable in textual commentaries, illustrates how memory errors can alter narrative details.

Lessons for Today: Preserving Truth in an Imperfect World

The parallels between AI’s digital fossils and scribal errors remind us of the fragility of human efforts to preserve knowledge. Just as scribes worked under flickering lamplight, battling fatigue and complex scripts, AI developers grapple with massive, opaque datasets. Both strive for accuracy but are hindered by the limits of their tools and processes.

For biblical texts, the Masoretes (c. A.D. 500-1000) developed rigorous methods to minimize errors, counting verses, words, and even letters to ensure fidelity. Yet, mistakes like the age of Jehoiachin—eight in 2 Chronicles 36:9 versus eighteen in 2 Kings 24:8—slipped through, likely due to a scribe omitting a “ten” in Hebrew numerals. The Dead Sea Scrolls, discovered in 1947, confirm the Old Testament’s remarkable preservation, with Isaiah’s text nearly identical to copies 900 years later. Similarly, over 5,000 New Testament manuscripts attest to its reliability, despite minor variations.

AI, however, lacks a Masoretic equivalent. The Conversation article notes that filtering out terms like “vegetative electron microscopy” risks removing legitimate references, and the scale of datasets makes manual correction nearly impossible. Transparency in AI training data, akin to the Masoretes’ openness about their methods, could help, but commercial interests often obstruct this.

A Call to Vigilance This Easter

This Easter holiday causes reflection on information preserved through centuries of scribal labor, reminding us that the essence of information endures despite human error. Both sacred texts and AI datasets can be used to improve human existence, but only if we are humble enough to acknowledge digital fossils and challenges in maintaining high quality information.

Just as scribes and scholars safeguarded sacred texts, we can strive for transparency, and publish facts instead of preferred narratives. By learning from the past, we can build a future where knowledge—ancient and digital—remains a beacon of hope.

NOTE: This article earned a High Information Quality rating of 92% under Capy News standards. While factually accurate and well-sourced, it lacks direct citations from OpenAI or independent audits to confirm AI training details, and includes a brief faith-based reflection that shifts slightly from pure analysis.

By Dean W. Korsak
© Copyright 2025, CAPY News LLC, All Rights Reserved.