Confessions of an AI Editor

Editor Recounts Use of AI Tool at Literary Services Startup

Diya Isha described her two-year role at a literary services company that supplied AI-assisted editing to clients including a think tank and academic authors. The account, published August 29, 2026, details daily operations in which an interface built around existing large-language models performed line edits, memo generation, and paragraph compression.

The company was established by an editor who viewed manual editing tasks as repetitive. She directed development of a system that accepted manuscript uploads and returned suggested changes. System instructions began with the phrase directing the model to act as a diligent editor. The interface allowed users to accept or reject alterations in a tracked-changes format. Isha stated that the tool could generate an editorial memo in under one minute and revise a chapter during the time required to switch browser tabs.

The founder maintained a conventional office appearance, including a color-coded bookshelf arranged by a professional stylist. Isha reported that this presentation helped present the AI service as compatible with established editorial standards. She noted difficulty disclosing the nature of the work in social settings.

The service operated by routing user documents through foundation models and applying custom instructions. Isha initially reviewed outputs for accuracy and reported discrepancies to the founder. Because she held an entry-level position, she characterized her input as limited to prompt refinement rather than model training. Errors in output were attributed jointly to the model and the human who prepared the prompt.

Client outreach required Isha to describe the tool in marketing terms. She stated that she sometimes referred to it as a consistent alternative to human editors and noted token limits as a practical constraint. The founder added that the system remained available outside standard business hours.

Isha reported that she did not read full policy papers submitted by one client because she lacked prior experience editing that format. She copied generated revisions into documents and forwarded them. A study hosted on the arXiv server and associated with MIT researchers examined user perceptions of ownership when large-language models contributed to written work. Isha linked her own experience to the reported finding that heavy reliance on such tools correlated with reduced sense of personal ownership over the final text.

The tool applied statistical patterns observed in training data, including frequent use of certain sentence structures and transitional phrases. Isha observed that outputs sometimes retained recurring elements such as em dashes. She stated that she received instructions to replace selected verbs and punctuation to reduce visible repetition. The company maintained that clients should not perceive AI involvement even when disclosure occurred.

In one project, the two full-time staff members and one intern received a 150,000-word dissertation with a three-day deadline. Work was divided into discrete segments. Isha informed the founder that the timeline appeared unfeasible under conventional editing standards. The founder responded that expectations should adjust to the capabilities of assisted editing. Isha completed the assignment by accepting most model suggestions without reading beyond individual paragraphs. The client later expressed satisfaction; the founder attributed the result to Isha in an internal message.

Isha referenced Vladimir Nabokov’s 1948 lecture on reading practices, which distinguishes between approaches that impose prior generalizations and those that attend to textual detail. She noted that stored prompts at the company directed the model toward standardized treatments based on manuscript category. She stated that she sometimes submitted her own work to the same models for preliminary review and that delays in workflow were treated as avoidable when the tool was available.

The founder’s stated objective was to return time for higher-order editorial judgment by automating routine operations. Isha reported that actual time saved was typically redirected toward additional assignments rather than extended review of individual manuscripts. When clients expressed dissatisfaction with flat or repetitive revisions, staff adjusted system instructions or selected a different underlying model rather than performing manual rewrites.

Isha left the position two years before the article’s publication. She cited a desire to retain direct involvement in the production of written work. She stated that her departure did not alter the company’s use of the system or the broader adoption of similar tools. The account contains no independent verification of the company’s client list, revenue, or technical specifications beyond the author’s descriptions.