Paul Bakker Argues AI Should Support Research, Not Replace the Thinking Behind Substantive Writing

Paul Bakker’s argument is simple: people should rarely use AI to produce substantive writing because writing itself is a method of thinking. His August 13 essay draws a boundary between using Claude, ChatGPT, Codex, or Gemini for coding, research, data analysis, and information gathering, and asking those systems to compose the memo, report, essay, or email that carries an idea.

The concern is not limited to polished prose. Writing forces a person to identify the problem, choose an argument, organize evidence, and decide what a reader needs to understand. Generating a document from bullet points can therefore remove the intellectual work that gives those points meaning. A parallel essay making the same case describes substantive writing as text intended to convey an idea, argument, analysis, or other serious thought.

Bakker’s case rests on the cognitive function of drafting, rather than on a claim that AI produces uniformly poor sentences. The hard part of writing is deciding what one believes and how the pieces fit together. A machine can turn notes into fluent paragraphs, but that fluency can hide gaps in reasoning before the writer has confronted them.

That distinction explains why Bakker still uses AI heavily in his own work. His profile identifies him as a Netflix staff software engineer focused on modern Java, developer productivity, and AI-assisted development. His objection is therefore directed at outsourcing thought, not at rejecting AI altogether. The broader debate includes grammar correction, translation, captions, transition edits, search optimization, and automated email responses, showing that the boundary between assistance and authorship remains unsettled.

AI becomes most consequential when it moves from helping a writer inspect material to making the intellectual decisions inside the final text. That line matters in research, business communication, education, and creative work because the document records judgment, not merely language. A writer who reviews a source set, asks AI to identify a gap, and then develops the argument still performs a different task from someone who accepts generated reasoning with light edits.

Academic practice shows how quickly the workflow is changing. One account describes professors using AI to locate gaps in existing literature, identify likely reviewers, and shape citation strategies. Another writing-studies guide frames refusal as a principled response to generative AI, especially where education treats writing as a process for learning. Those examples put pressure on disclosure rules, but they also raise a deeper issue: whether a finished text can represent its author’s understanding when the author skipped the work that formed it.

Critics of Bakker’s position focus on the reader’s experience. If a generated article is accurate, clear, useful, and engaging, they argue, the production method matters less than the result. A separate essay on AI-assisted writing says readers care primarily about what they experience, while treating the writer’s process as a concern held mainly by creators and critics.

That standard is persuasive for some practical writing, particularly routine text where the goal is speed and clarity. It is weaker when writing serves as investigation, teaching, or self-expression. A reader may receive a smooth report, while the person who commissioned it remains unable to explain its reasoning. The disagreement is therefore about the purpose of writing: communication alone, or communication combined with the author’s development of judgment.

The emerging boundary is functional rather than categorical. AI can search, summarize, compare, translate, check spelling, inspect code, organize data, and challenge an argument after the human writer has developed it. Bakker’s position reserves the first formulation of substantive ideas, the structure of the argument, and the final responsibility for the person whose name appears on the work.

That approach also gives writers a test: if the task requires deciding what to think, draft it yourself; if the task involves checking, transforming, or extending work you already understand, AI assistance has a clearer role. The test does not settle every case. Copy editing, brainstorming, and accessibility tools occupy a gray zone, while disclosure expectations vary across schools, publishers, employers, and platforms.

The debate will shift from whether AI can write good sentences to what writers owe readers, collaborators, students, and themselves. Bakker’s essay supplies a durable warning for anyone using generated text: polished language can conceal unexamined assumptions, and editing a machine draft does not automatically restore the thinking that drafting would have required.

Competing views will remain because writing has multiple jobs. It can deliver information, produce a public artifact, satisfy a workplace requirement, or help a person reason through uncertainty. AI is well suited to some of those jobs and poorly suited to others. The practical question is whether the tool is reducing clerical effort or removing the struggle through which the writer forms a defensible idea.