Safety

AI-generated text

Sophie Alpert: no lossless transformations of natural‑language text

Sophie Alpert published a short internal policy for engineers on acceptable use of AI for writing, arguing that every sentence in shared documents must reflect the author's own thinking.

Sophie Alpert: no lossless transformations of natural‑language text

Sophie Alpert published a short internal statement addressing acceptable use of AI when engineers draft or edit written material. The note is brief, practical, and consistent with its own recommendations: it models concise guidance and focuses on actionable rules.

The core rule: stand behind your words

Alpert emphasizes a single crucial rule: every idea and every sentence in a shared document must be something the author can stand behind. If you choose to have large language models (LLMs) or other AI tools help rework your writing, it remains your responsibility to ensure the entire document represents your own thoughts before you share it.

She warns that it is not acceptable to reply to a reviewer’s question with “Oh sorry, AI wrote that, just ignore it.” Such a response confuses readers and wastes their time, because it presents material that does not genuinely reflect the author’s intentions.

"There are no lossless transformations" — explanation

The title’s central idea — "There are no lossless transformations of natural‑language text" — states that every rewrite or rephrasing alters the meaning of the original text. When someone or something (for example, an LLM) reformulates a sentence, nuances, information, or authorial intent can be lost during the change.

This risk is greater when the entity performing the transformation does not possess the author’s most detailed mental model of what they intended to communicate. Therefore, Alpert urges that final responsibility for any revised document remains with the author.

Practical implications

  • Responsibility: the person sharing a document must be able to explain any substantial line or claim in it.
  • Transparency: readers can only interpret content reliably if it reflects the author’s viewpoint rather than an external tool’s fragmented contribution.
  • Quality control: reviewing and validating AI‑assisted edits reduces the chance of misunderstandings and information loss.

Alpert’s recommendation is concise and practical: AI tools may be used to polish writing, but the author must always review, understand, and take ownership of the final text.

Summary

Sophie Alpert’s short guidance stresses that using AI to assist writing does not absolve authors of responsibility. The principle that there are no lossless transformations reminds writers that every rephrasing can change meaning, so authors must ensure the final document faithfully represents their own thoughts.