The 25 ways AI overwrites your voice

Semantic ablation. A February 2026 Register piece coined the term for what happens when an LLM edits your writing: not hallucination, where the model invents what isn't there, but the opposite. The model destroying what is. Your specific phrasing, your uneven rhythm, your willingness to leave something unresolved. The model smooths all of it into the same three-paragraph shape with the same measured confidence about everything.

The result is not bad writing in the traditional sense. It is competent, grammatical, and organised. It is also indistinguishable from every other piece of AI-assisted text published this week. That is the problem. Not detection. Quality.

The wrong framing

The market has split AI writing tools into two camps: detectors (Turnitin, GPTZero, Copyleaks) and humanisers (BypassGPT, Phrasly, and dozens of others that rewrite text to evade those detectors). This is an arms race framed around the signal, not the substance. A humaniser that swaps "delve" for "explore" and varies sentence length to fool a classifier has not fixed the writing. It has laundered it.

The actual problem is simpler and harder to game: AI-assisted text is often worse than what you would have written yourself. Not because AI cannot write (it can, fluently) but because its fluency is generic. It writes the median response to your prompt. The most statistically likely phrasing. The most commonly rewarded structure. Your writing, whatever its faults, was at least yours.

What goes wrong, specifically

I spent a few weeks reviewing published research on linguistic markers of AI text, generating test corpora (20 passages across varied topics, including technical blog posts, product announcements, internal memos), and cross-referencing against human writing of the same genre and length. The patterns that emerged fell into 25 categories. They are not equally damaging; some make entire passages feel monotonous, others are individually minor but accumulate the way dust does.

Structural patterns

The patterns that do the most damage are not about vocabulary. They are about shape.

Sentence length uniformity. Human prose has high variance in sentence length. A short sentence lands hard. A longer one meanders, qualifies, doubles back. AI text clusters sentences around 15 to 25 words each, producing a metronomic rhythm that readers sense even if they cannot name it. Sentence length standard deviation turns out to be one of the strongest features in classifier research, but it is also just a reading-experience problem: monotonous rhythm is monotonous to read.

Paragraph length uniformity. Squint at a page of AI text and every paragraph is the same height. Three to five sentences, tiling neatly like bricks. Human writing is visually ragged: a one-sentence paragraph for emphasis, a long block when the argument requires it.

A two-line aside.

Epistemic uniformity. Every claim carries identical confidence. Human writers hedge about some things, commit fully to others, and occasionally turn out to be wrong about the things they were most confident about. Flat certainty across every point reads as synthetic because it is.

Register consistency. Human formality drifts. A technical writer drops into slang when recounting something frustrating ("we fat-fingered a YAML config and brought down prod"). A casual writer tightens up when the stakes are high. AI maintains one register from start to finish, which is maybe the easiest tell for an attentive reader, though I am less sure about this one than about sentence length. It might just be that some human writers are also boringly consistent.

Punctuation distribution. Commas, periods, and dashes account for nearly all AI punctuation. Human text uses semicolons; colons to set things up; parentheses (for asides that don't quite belong in the main clause); question marks mid-paragraph. Where did all the question marks go? Low punctuation entropy is measurable, but you do not need a classifier to notice it.

Vocabulary and construction

Overused vocabulary. Words like "delve," "comprehensive," "leverage," "foster," "nuanced," and "seamless" appear at frequencies 10-35x higher in post-2023 generated text than in equivalent human writing. Corpus studies tracking millions of academic papers documented a 25x increase for "delves" between 2022 and 2024. Any two in the same paragraph is a signal; three is diagnostic.

Copula avoidance. "Serves as," "functions as," "stands as" where "is" would do. The model avoids the most common verb in English as though it were inelegant. It is not.

Significance inflation. Everything is pivotal, crucial, transformative. If everything is significant, nothing is.

Causal over-assertion. Clean because/therefore chains where reality is messy. Humans say "and then" or "around the same time" when the causal relationship is unclear. Most real-world outcomes have multiple causes, and we rarely know which one mattered.

Vague attributions. "Industry experts suggest," "observers have noted," "research indicates" without citation. Name the source or own the opinion. There is no middle ground that does not read as evasion.

Surface patterns

Individually minor, collectively diagnostic: promotional language ("groundbreaking," "revolutionary"), formulaic headers ("Why this matters," "Looking ahead"), weak transitions ("Furthermore," "Moreover"), bureaucratic bloat ("in order to," "due to the fact that"), adverb frontloading ("Interestingly," "Notably"), em dash overuse, and chatbot residue ("I hope this helps"). None of these is damning alone. A piece that has all of them is not a piece a human wrote.

What is missing

Stripping patterns is half the work. The other half is absence.

Emotional flatness. Does the writer react to their own material? A human writing about a failed migration is annoyed, or relieved, or grudgingly impressed by the elegance of the failure. AI text reports without reacting. It describes a production outage in the same tone it uses for a product launch.

Opinion absence. Every side presented with equal weight. Human writers lean somewhere. Even cautious ones.

Absence of specific detail. "A team I worked with" versus "the four of us at the tail end of that November sprint." Generic examples could come from any prompt. Specific details (the kind that are too particular to fabricate efficiently, that make you think "nobody would make that up") are probably the strongest humanising signal available.

Temporal flattening. Everything in vague present tense. "Organisations are increasingly adopting" avoids committing to any moment. Human writing is anchored: "when we started this in 2022," "by the third sprint," "I remember the week before launch."

Corpus-level patterns

When you compare multiple texts by the same AI-assisted "author," a new category emerges. Each text was generated independently from a similar distribution, so they converge in ways a single human's work would not: verbatim phrases repeated across pieces, identical rhetorical arcs, the same framing devices at the same structural positions, recurring metaphors. You cannot see these by reading one piece. You need the collection.

Editing, not evasion

The alternative to both detection and evasion is editing. Not "make this text pass a classifier" but "make this text better."

The distinction is practical. A humaniser that varies sentence length to beat GPTZero has not fixed the underlying monotony; it has added noise. An editorial pass that varies sentence length because the writing reads better with variation has fixed the actual problem. The text might pass classifiers as a side effect, but that was never the point.

The approach I have been testing works through these patterns in severity order: strip the most damaging structural tells first, then address vocabulary and construction, then add back what the model removed. Opinion. Emotional reaction. Specific detail. Temporal anchoring. Not every check applies to every text, and a piece with zero rough edges is itself suspicious.

The overcorrection problem

An LLM editing LLM output has a specific failure mode: it applies corrections uniformly, creating a new detectable pattern. Replacing every "Furthermore" with "And" is not editing. It is search-and-replace.

I hit this immediately in testing. The first run (February 2026, Claude Opus 4) scored 19/22 on pattern removal but failed on overcorrection: 100% of flagged vocabulary removed instead of the 80-95% range that reads as natural, paragraph structure stayed safely in the 2-5 sentence range instead of varying, and the model treated "do not overcorrect" as a secondary instruction. The spec now includes explicit constraints: leave some instances of minor patterns untouched, vary priorities from one text to the next, respect the writer's existing habits. Whether those constraints hold is still an open question; I have run one test, not ten.

What persists

None of this means AI-assisted drafting is illegitimate. People have always used tools to get words on the page: dictation, editors, the colleague who restructures your argument over coffee. The question was never whether you used help. It was whether the result sounds like you thought it.

The surface tells will fade as models improve. Sentence lengths will vary more, the vocabulary will diversify, and the overused words will cycle out of fashion. But the deeper patterns will persist: epistemic uniformity, register lock, the absence of someone who was actually there. These are not bugs. They are the natural output of a system optimised to predict the most likely next token.

Most likely, by definition, means least individual.

Sources and further reading

Weixin Liang et al., Monitoring AI-Modified Content at Scale (ICML 2024). Tracked vocabulary frequency shifts across millions of academic papers and documented the statistical signatures of AI modification. The 25x increase for "delves" between 2022 and 2024 comes from this data.

Linguistic Characteristics of AI-Generated Text: A Survey (arXiv 2025). Covers sentence length distributions, punctuation entropy, and the structural markers that distinguish AI text from human writing at a statistical level.

The Disappearing Author: Linguistic and Cognitive Markers (2025). On epistemic and affective markers: why AI text reads as emotionally flat and epistemically uniform, and the cognitive mechanisms behind reader detection.

Eric Mitchell et al., DetectGPT (ICML 2023). Perplexity-based detection. The insight that AI-generated text occupies negative curvature regions of a model's log probability function.

Sebastian Gehrmann et al., GLTR (ACL 2019). Early work on token probability uniformity as a detection signal. The observation that AI text over-indexes on high-probability tokens.

The Register, Semantic Ablation (February 2026). Coined the term for the specific problem this essay addresses: AI not adding what isn't there, but destroying what is.