What Generative AI Thinks a Technical Author Looks Like

While preparing a piece with colleague Mike Mason, we asked Stable Diffusion to generate portraits of technical authors. The results were revealing—both about the current state of image generation and about the assumptions embedded in the training data. Mike suggested we push further by prompting a commercial-grade model like Midjourney, or by refining the prompt with help from ChatGPT. Taking the latter suggestion, I asked ChatGPT for a set of suitable prompts.

ChatGPT offered eight distinct scenarios. A few patterns emerge:

  • Every setting is a professional or semi-professional workspace: a well-lit office, a modern shared office, or a comfortable home setup.
  • Technology is ubiquitous: laptops, dual monitors, tablets with styluses, headphones, and even a blueprint of a software interface.
  • Personal attributes never extend beyond basic descriptors like “friendly smile” or “confident,” with only one prompt specifying gender explicitly.
  • Physical artifacts—books, user manuals, sticky notes, whiteboards—appear in every scene, grounding the role in tangible documents and diagrams.
  • Coffee is the sole refreshment mentioned, and it appears only once.

The prompts assume a technical author is someone who writes about software in an office context, surrounded both by digital tools and by printed reference materials. There’s no indication of remote work in a cafe, a writer who travels to customer sites, or someone who contributes to open-source documentation from a kitchen table. The model’s idea of the job is conventional, tidy, and heavily anchored to a specific visual stereotype of the working professional.

That says less about the profession itself and more about the corpus used to train the model. The suggestions are not wrong—they describe plausible environments for some technical authors—but they also reflect a narrow slice of reality, one that aligns with stock photography and corporate imagery rather than the diverse ways documentation actually gets produced. A single prompt template with swapped objects (headphones, stylus, books) covers the whole spectrum ChatGPT chose to offer.

Had we fed these prompts back into Stable Diffusion to observe the visual results, the experiment would have added another loop of interpretation. But the text alone was enough to show the mismatch between a technical role grounded in problem-solving and communication and the shallow visual cues an AI model associates with it. The output is less a portrait of the job than a composite of every office scene the model ever saw captioned with “writer.”