A Typology of Typos: Five Error Categories in AI-Assisted Writing
By Daniel Heuman, CEO & Founder, Intelligent Editing
TL;DR : Working with AI introduces three categories of error: mistakes in the AI's own output, mistakes introduced while editing it, and inconsistencies created by the combination of both. The most common specific errors are UK/US spelling drift, inconsistent hyphenation, capitalization inconsistencies in headings and tables, and departures from house style guides like The Chicago Manual of Style (even when you instruct the AI specifically to follow a style). Do not let the AI fool you into thinking that work no longer needs checking. A consistency checker like PerfectIt catches all of these issues.
When working with AI, it is tempting to assume that typos are no longer a problem. The prose looks clean. The sentences are well-formed. There are no obvious stumbles. But this assumption misunderstands two things: the nature of AI writing, and the nature of workflow, especially when it comes to consistency.
Even though I resisted it at first, now I write with AI every day. I use PerfectIt, an editing tool that checks consistency and house style, to find the errors that slip through. This article explains why that happens and what kinds of errors to expect. I plan to update it as I spot other kinds of errors in the text.
Three ways AI writing introduces errors
No two people write in exactly the same voice. No matter how well you've trained it, the same is true when you work with an AI. As a result, there are three different ways in which typos and inconsistencies can creep into AI-assisted writing.
Errors in the AI’s writing
AI models follow statistical patterns from their training data. They make plausible choices, but plausible is not the same as correct. That's especially true because their training data includes large amounts of noisy text, including typographical errors, grammatical mistakes, and other inconsistencies.
Errors from editing the AI's work
The AI is imperfect and so are people. Editing, especially when it's extensive, always runs the risk of introducing mistakes.
Inconsistencies from the combination of both
A rewritten sentence can introduce a spelling convention that differs from the rest of the document, or shift capitalization style mid-table, without either change being noticed. Neither choice is wrong in isolation; both are wrong in the same document. The AI writes one way, the editor works another, and the document ends up holding two styles at once.
The five error categories of AI-assisted writing
Here's my working list of the kinds of errors I have found in AI-assisted documents. Most of these mistakes were spotted by using PerfectIt.
Error types covered in this typology:
UK/US spelling inconsistencies (including Oxford spelling)
Capitalization of table headings
Departures from style manuals (e.g., CMOS)
Hyphenation of compound modifiers
Heading capitalization
1. UK/US spelling
This is a frequent issue and it is less simple than it appears because there is no such thing as a single "UK spelling" or "US spelling". There are variations within each. Most AI interfaces offer a binary choice, which flattens that complexity.
Oxford spelling, for example, is British in most respects (*colour*, honour, centre) but uses the -ize suffix for Greek-origin verbs (*organize*, recognize, prioritize) rather than the -ise ending common elsewhere in British English. It also uses analyse and paralyse, not analyze and paralyze.
With the rapid pace of improvements, this is a problem that may be solved in the future. However, at the time of writing, it remains a form of spelling preference that frequently causes problems in AI writing.
Typical errors:
PerfectIt checks all of these against a defined style and flags every inconsistency, including the subtle ones that a fast read will not catch.
2. Capitalization of table headings
Inconsistent capitalization across the first column or row of a table is an error that both people and generative AI seem to make.
A first column that reads *Product overview*, *Pricing structure*, *Key benefits*, *Overview of Support* has a title case entry at the end. The pattern break is easy to miss. PerfectIt finds it.
3. Following the style manual
AI writing is not trained to follow a specific guide such as The Chicago Manual of Style. Organizations that use one as their house style and hope that an AI will follow it will find deviations throughout any AI-assisted document.
Common departures include:
The Chicago Manual of Style for PerfectIt is a formal partnership with the team behind The Chicago Manual of Style so its enforcement is far more effective.
4. Hyphenation
Compound modifiers are a persistent problem. An AI will hyphenate well-established before a noun in one sentence and write well established without a hyphen in the next. This is a genuine grey area in English, which is probably why AI handles it inconsistently. But a house style should not be a grey area. PerfectIt enforces whichever rule has been set.
5. Heading capitalization
AI defaults to title case for headings. But rules in English for title case capitalization are complex. The guidelines even changed between the 17th and 18th editions of The Chicago Manual of Style. So, this is difficult for AI to get right and hard for people to get right. In particular, people often go wrong on the verb “to be”. The result is a great deal of inconsistency when both are combined.
PerfectIt finds this because you can simply set a rule for each heading level.
Why AI-generated text feels finished but still contains errors
The problem is not only that AI introduces errors, but also that it introduces them while looking polished.
A rough draft signals the need for careful review. AI-generated text does not. The prose is smooth, the structure is clear, and there are no obvious flags. That surface quality suppresses the instinct to scrutinize, and the errors survive because they were never looked for.
This is the false sense of security that AI creates. The document looks finished. The style inconsistencies are real.
Why you still need a consistency checker when writing with AI
When two contributors produce a single document, inconsistencies emerge. The same is true when you work with an AI. There are going to be different spelling preferences, different capitalization habits, different approaches to numbers. PerfectIt was built to catch exactly those inconsistencies.
That problem doesn’t disappear when one of the authors is an AI. It becomes more acute.
What not to do
The intuitive response, asking the AI to check and fix consistency, tends to make things worse. An AI tasked with consistency will look for all kinds of things. It will rewrite sentences that did not need rewriting, change intentional choices, and still miss the errors it should have caught. The result is a document that looks more uniform but has drifted from the intended style in ways that are difficult to trace.
My own experience
I developed PerfectIt. I spent most of the last twenty years of my life trying to solve issues like consistency mistakes. So, I like to believe that these problems will not affect me. When I finish a piece of writing, I tell myself I probably got it right. When I write with an AI, I tell myself I will definitely get it right.
I am always wrong on both counts!
Every time the document goes through PerfectIt, it finds something. The tool exists because human writers need it, and AI-assisted writers need it just as much.