Why do AI-generated documents all look the same?
AI writing tools share training data, default palettes and layout libraries. When none of those defaults are overridden, every output converges on the same visual shorthand: a blue header, white body, a sans-serif title and bullet points that begin with a gerund. The sameness is not a bug in any single tool — it is what happens when a probabilistic system picks the most statistically likely choice at every step.
Where does the default template actually come from?
Large models learn from publicly available documents — slide decks shared on SlideShare, reports exported from Google Slides, whitepapers hosted on corporate sites. Those files cluster around a narrow set of design choices because corporate design itself clusters: navy or cobalt headers, white backgrounds, a title in a heavy sans-serif, body copy in a lighter weight of the same family. When a model is asked to produce a document without explicit design instructions, it reproduces the statistical centre of that distribution.
The same mechanism applies to word choice. Phrases like 'key takeaways', 'actionable insights' and 'in today's landscape' appear so frequently in business writing that they become the path of least resistance. A model that has seen thousands of executive summaries will reach for them automatically. The result is a document that reads as if written by the same anonymous committee that designed every other document in the corpus.
- Default palette
- The colour scheme a tool applies when no brand colours are specified — typically a blue-and-white combination drawn from the statistical centre of its training data.
- Filler verb
- A verb that sounds purposeful but carries no specific meaning: 'leverage', 'optimise', 'streamline'. Frequent in AI copy because it is frequent in the documents models trained on.
- Gerund bullet
- A bullet point that begins with an '-ing' word — 'Improving…', 'Delivering…' — a pattern so common in business slides that it has become a recognisable AI tell.
- Template lock-in
- The tendency for a tool to apply the same structural choices — column count, heading hierarchy, section order — regardless of the content's actual needs.
Why does changing one colour not fix the problem?
Design coherence requires that colour, type, spacing and tone are chosen together and applied consistently. Swapping a header from navy to green while leaving everything else at its default produces a document that looks like a default template with a green header. The underlying structure — the proportions, the type scale, the way images are cropped — still signals 'generated output'.
A genuine brand system starts from a small set of decisions — a primary colour, a typeface, a tone of voice — and derives every other choice from those. When those decisions are written down in a form a tool can read, the tool can apply them consistently rather than falling back to its defaults at each new element. Without that written brief, every generation is effectively a fresh roll of the same biased dice.
| Element | Default AI behaviour | Brand-directed behaviour |
|---|---|---|
| Colour | Cobalt or navy from training-data majority | Primary and accent drawn from a written brand spec |
| Typeface | A widely licensed sans-serif at a standard weight | A chosen family with defined weight pairings |
| Tone of voice | Formal, passive, filler-verb-heavy | A documented register applied to every sentence |
| Layout grid | Two-column or full-width from template library | Column count chosen for the content's actual density |
| Image style | Stock-photo aesthetic or flat illustration | A defined visual language stated in the brand brief |
What are the measurable tells of an AI-generated page?
Some tells are matters of opinion — a trained eye notices something feels off. Others are countable facts. Em-dashes used as a stylistic flourish appear at a much higher rate in AI prose than in edited human writing. Placeholder names ('John Smith', 'Acme Corp') survive into published copy because no human read the draft carefully enough to replace them. Filler marketing verbs cluster in the same sentences. Purple or teal gradients appear in hero sections because they were fashionable in the design assets that dominated training data during a particular window.
Counting these signals rather than judging them by feel has a practical advantage: a count can be tracked over time, compared across documents and used as a concrete target for revision. If a document contains twelve em-dashes and six instances of 'leverage', those are specific things to fix, not a vague sense that the copy sounds robotic.
When is a simpler tool the better choice?
If you are producing a one-off internal memo, a slide deck for a meeting that will not be recorded, or a draft that a human editor will rewrite substantially, the default template is probably fine. The cost of fixing the tells is lower than the cost of setting up a full brand system. The same applies when speed matters more than brand consistency — a first draft that exists is more useful than a perfect draft that does not.
The case for investing in a written brand system grows when documents are published externally, when multiple people are generating content from the same tools, or when the output is meant to represent a specific organisation rather than a generic professional voice. At that point, the defaults stop being a convenience and start being a liability. GROX addresses this through DESIGN.md files — a written brand brief that a build reads before making any design decision — and a rulebook that checks each output against the known tells of AI-generated pages. Whether that mechanism suits your workflow depends on how much of your output is generated rather than written.
Common questions
Why do AI slide decks always use blue and white?
AI tools generate designs by reproducing the statistical centre of their training data. Corporate slide decks shared publicly cluster around blue-and-white colour schemes, so that is what a model reaches for when no other instruction is given. The fix is to provide explicit colour instructions before generation begins, not to adjust the output afterwards.
What is a DESIGN.md file and how does it help?
A DESIGN.md file is a plain-text document that records a project's brand decisions — colours, typefaces, tone of voice, layout rules — in a form that an AI tool can read before it generates anything. Instead of falling back to defaults, the tool applies the written spec. It can be downloaded, shared and read by any AI tool that accepts a file as context.
Can you count AI tells automatically, or does it require human judgement?
Some tells are countable without human judgement: the frequency of em-dashes, the presence of placeholder names, the occurrence of specific filler verbs, the use of particular colour values in CSS. Others — whether a layout feels generic, whether a photograph looks like stock — require a trained eye. Automated checks are most useful for the countable signals, which are also the easiest to fix once identified.
Is it worth setting up a brand system if I am the only person generating content?
It depends on how much of your output is published externally. For internal drafts or one-off documents, the effort of writing a full brand spec probably outweighs the benefit. For anything that represents you or your organisation to an audience — a website, a report, a proposal — the defaults carry a real cost to credibility, and a written brief pays for itself quickly.
If you want to see how a written brand spec changes generated output in practice, GROX lets you generate a Brand Kit, export it as a DESIGN.md file and apply it to any build — the free tier covers this without a card.