AILIES: WHY AI WILL ALWAYS LIE

Copyright © 2026 Michael Herman (Bindloss, Alberta, Canada) – Creative Commons Attribution-ShareAlike 4.0 International Public License
Web 7.0™, TDW AgenticOS™ and Hyperonomy™ are trademarks of the Web 7.0 Foundation. All Rights Reserved.

IMPORTANT NOTE: This article is not about faith, religion or church. An English version of the Bible (any single translation) is used simply as a context…a use case …a test case.

The conclusions are the same: #AI #will #always #lie …even if you build a small, narrowly focused #LLM around a single subject.

Conclusion: Yes — almost all of it is still true.
Using only one single translation (e.g., just NIV or just NRSV) removes inter-translation disagreement, but it does not eliminate knowable lies for ordinary users.

---

What no longer applies (or is greatly reduced)

Inter-translation conflict

If the model is trained on only one translation:

• It can’t contradict another English translation
• It can’t blend wording across versions

So this specific failure mode disappears:

“The Bible says X” when another translation clearly says Y

But this is a narrow improvement.

---

What still applies (and why knowable lies remain)

1. A single translation is still not ground truth

Even one translation:

• encodes interpretive decisions
• smooths ambiguity in the source languages
• chooses one meaning where multiple exist

So the model may assert:

“The text means X”

when the underlying text reasonably allows ¬X — a knowable falsehood once checked against scholarship.

---

2. Generalization beyond the text still occurs

LLMs:

• extrapolate patterns
• infer doctrines
• merge nearby passages

This produces statements that:

• are not stated anywhere
• sound consistent with the text
• are verifiably unsupported

→ knowable lies remain possible.

---

3. Coverage gaps are unavoidable

A Bible-only model still lacks:

• modern history
• natural science
• detailed biographies
• post-biblical theology

Questions outside the text’s scope force:

• refusal, or
• invention

Without strict refusal logic, invention = knowable lie.

---

4. Reasoning errors are corpus-independent

Logical errors arise from:

• probabilistic prediction
• lack of formal validation
• rhetorical coherence bias

A conclusion can be false even if every quoted verse is accurate.

---

5. Overconfidence remains unless designed out

A single-translation model will still:

• state interpretations as facts
• omit uncertainty markers
• collapse metaphor into literal claims

Users can verify the mistake → knowable lie.

---

Key takeaway (the deep point)

Knowable lies emerge from inference, not disagreement.

Removing alternative translations:

• improves consistency
• does not guarantee epistemic correctness

A model trained on one perfect, immutable text can still confidently assert false claims about that text.

---

Bottom line

Setup	Knowable lies possible?	Why
4 translations	Yes	Ambiguity + inference
4 translations, fixed versions	Yes	Inference
1 translation only	Yes	Inference + scope limits

---