How to identify AI text in 2026: the updated map of the patterns that give it away

AI writing tics grow old. In 2023, delve was the flag word for ChatGPT in any text under suspicion. In 2025, its frequency collapsed. Core and modern took its place, according to a Washington Post analysis of 300 thousand messages from the model between June 2024 and July 2025.

Anyone still detecting AI against the list from two years ago is hunting a tic that went out of style. The Headcore Writing Guidelines catalog twelve banned structures as of early 2026. This analysis documents what changed over the last twelve months, and what needs to be added.

Why the tics grow old

Language models operate on conditional probability. Each word the model produces pulls the next one toward the most likely region of the training space. When the audience starts to recognize that region, developers adjust, training shifts, and the region moves.

Delve is the best-documented case. The word appeared at anomalous frequency in ChatGPT outputs between 2023 and early 2024. It was named by Forbes, by academic studies, and by dozens of Reddit threads. Once it became a public tell, the model learned to avoid it. In 2025, its frequency dropped sharply, according to the Washington Post cross-reference against the ChatGPT message base.

The mechanism repeats with any tic that becomes a headline. That is why the map needs an annual update. Anyone using a 2023 list in 2026 is looking at patterns the model has already worked around. They will catch a false negative on the new tics, and a false positive on the human who learned to avoid the old list.

The twelve structures that still hold

The internal Headcore catalog starts from twelve banned structures, all validated in editorial production and none sensitive to lexical rotation. They attack the skeleton of the sentence, not the word. They are stable because the model cannot avoid them without ceasing to be a model.

• Binary inversion in the format “It is not X. It is Y.” The construction most replicated by AI. It sounds provocative, but it is formula.
• Symmetric mirror in the format “Less X, more Y.” Humans break symmetry.
• Ascending triad. Three items on a scale of impact. Sometimes two are enough.
• Forced anaphora. Repeating the start of the sentence in sequence. It works when rare.
• Pseudo-provocative opener in the format “Most companies…” or “Everybody talks about X, but nobody does Y.”
• Inversion kicker at the close. “The question is not if. It is when.”
• Data sandwich. Provocation, data, interpretation. Always in that order.
• Uniform paragraph rhythm. Every paragraph with two or three sentences. None with one, none with six.
• Parallelism that is too perfect. Structure that fits without friction.
• Total absence of digression. AI never leaves the track.
• Fabricated raw tone. Lowercase and short sentences imitating direct speech, with a perfectly organized structure underneath.
• Generic consulting vocabulary. Ecosystem, friction, lever, pipeline, scale.

These twelve crossed two model cycles without losing validity. They are the floor. What changes is what sits on top of them.

The new patterns that show up in the 2025 and 2026 data

Three independent sources feed this update. The Washington Post analysis of 300 thousand ChatGPT messages. The Signs of AI Writing entry on Wikipedia, updated in November 2025 with observations from editors who review suspicious contributions at scale. The Surfer SEO study from March 2026 with the updated list of flag words by category. Cross-referencing them reveals seven classes of tic that were not in the original catalog.

Inflated verbs

Words that replace the simple verb with the pompous version of the same. Delve instead of dig deeper. Dive deep instead of study. Harness instead of use. Leverage instead of increase. Facilitate and foster when help would have done. In Brazilian Portuguese, the parallel list includes potencializar, viabilizar, otimizar, and capacitar. The Headcore catalog covered leverage inside the consulting vocabulary. The category needs to be named in full.

Dramatic nouns

This class was absent from the catalog. Tapestry, landscape, realm, journey, beacon, pillar, ecosystem, frontier. Nouns that call for metaphor and skip fact. They work as a safe harbor for the model when the subject needs an image. Wikipedia documents tapestry as a high-confidence tell in generated text, precisely because humans call for the word rarely, and models call for it at anomalous frequency.

Empty intensifiers

Crucial, essential, vital, fundamental, significant, pivotal. Adjectives that promise weight and deliver nothing. GPTZero measured a specific phrase, “plays a significant role in shaping”, and found a frequency 207 times higher in AI text than in human text. The case is instructive. The word crucial does not die when exposed. But when it appears in sequence, neighbored by others from the same family, it becomes a clear signature.

The evolution from “It is not X, it is Y” to “Not just X, but Y”

When binary inversion became famous, the models did not abandon it. They migrated to a variation. Instead of negating and replacing, they now add. “Not only X, but also Y.” “More than just X, it is also Y.” The construction drops the negation that betrayed the old formula, and keeps the parallelism that gives the pattern away. The Washington Post documents this format on the rise. It belongs to the same family as structure number one in the Headcore catalog, but it needs separate marking because it escapes the original filter.

Updated default openers

“In an increasingly [adjective] world” still holds as a flag, but it has lost prominence. The openers on the rise now are dangling participles at the start of the sentence, of the type “Bearing in mind” and “Considering that.” Another class is the restatement of the question before the answer, a pattern typical of chat. “You asked about X. To answer that, I will address three points.” The introductory sentence that comments on the very structure of the answer.

Corporate hedging

Aversion to the black-and-white sentence. The model structurally refuses to take a position. It always qualifies with it depends, it varies by context, there are nuances. The close tends to enumerate pros and cons in equal weights. Sean Kernan named this pattern as a strong tell in a 2025 article: ChatGPT is risk-averse about describing reality without a hedge. The Headcore catalog does not cover this class explicitly, and it is particularly visible in opinion text or editorial analysis, where the lack of a position is exactly the signal.

Visual formatting

This category was completely outside the catalog, and it may be the noisiest one in 2026. Bold on a random noun with no emphasis function. Decorative emoji at the start of a bullet, in particular the check icon (present in a third of ChatGPT messages according to the Washington Post) and the brain icon. H2 and H3 headers on text that fits in a single paragraph. Italics for emphasis with no real contrast. Seventy percent of the ChatGPT messages analyzed by the Washington Post contain at least one emoji. In professional text, that number is statistically anomalous.

Generic hypothetical instead of concrete example

When the model needs to illustrate a thesis, it resorts to a nameless hypothetical. “Imagine a company that…” instead of citing a real company. “Think of a leader who…” without naming the leader. The reason is simple: the model avoids the concrete example because it may get the fact wrong. The result is text with no anchoring. When an entire analysis rests on hypotheticals, it rarely comes from a human with experience in the subject. A human brings a name, a year, a figure. Even approximate, they anchor.

The tics specific to Brazilian Portuguese

Most of the published research on AI tells is in English. Brazilian Portuguese has its own patterns, little mapped, and visible to anyone who reads editorial text at scale. Four classes that show up frequently.

False colloquialism. “Bora?”, “Sacou?”, “Tá ligado?”, “Saca só.” Artificial markers of closeness in text that keeps a formal structure underneath.

Fabricated intimacy. “Confesso que”, “Bate um papo”, “Vamos conversar.” An emotional invitation with no context to justify it.

Diminutive of closeness. “Dicazinha”, “dúvida rapidinha”, “explicadinho.” The same function as fabricated intimacy, carried out with a suffix.

Generic regional openers. “Olha só”, “Veja só”, “Pois é.” They work as the Brazilian equivalent of the “Look,…” that opens a sentence in English AI text.

The heuristic that runs through every list

The Headcore catalog includes four review questions that keep running through any tic update. Does it have a number, a source, or a proper name? Could it be swapped for a synonym without losing meaning? Does it fit in the mouth of someone who lived the problem? Does it survive the question “how do you know?” This analysis suggests adding a fifth.

Does the sentence take a black-and-white position, or does it qualify to avoid risk?

Corporate hedging is the hardest AI tic to detect with a word list, because the word used may be legitimate. What gives it away is the structural refusal to describe reality without qualification. A human sentence, in authored text, frequently takes a position that can be contested. An AI sentence, in untrained text, rarely does. The question works as a fine filter when the other four let it through.

How to apply this to your operation’s review

The internal Headcore method starts from three steps, and it can be replicated.

1. Take five pieces published by your team in the last sixty days, in particular brand content that involved the use of AI in production.
2. Strike through each occurrence of the twelve banned structures. Mark each of the new classes from this update in a different color. Inflated verb, dramatic noun, empty intensifier, “not only X but Y”, default opener, corporate hedging, visual formatting, generic hypothetical, PT-BR tic.
3. Count the density of marks per thousand words. Brand editorial text in healthcare tends to stay below four per thousand. Text truncated at the extremes of the prompt tends to pass fifteen per thousand.

The absolute number matters less than the trend. If density is falling from one piece to the next, the operation is learning. If it is rising, the operator is using AI with less judgment, or the team is reviewing less. A diagnosis of operation, not of talent.

Why this matters for brand

Brand content in 2026 competes for attention in a discovery filter that is narrowing, at the same time that production is standardizing. Alphabet reported in the first quarter of 2026 a 4% drop in Google Network revenue, even with queries at an all-time high. The difference was absorbed by AI Overviews. The text that passes through that filter has to be distinguishable.

AI tics are the shortest path to indistinction. A brand can produce all of its content with the help of a generative model and still publish authored text, as long as it operates the three phases of AI use in brand: judgment before the prompt, literacy during, curation after. This update to the tic map serves that third phase. It is the checklist the literate operator runs before publishing.

The productive discussion about AI in brand in 2026 has already moved on. Whether the tool fits in the operation is settled. What is open is the method. The tic catalog is part of the method, and it needs to be updated every twelve months, at the same cadence in which the models themselves update what they hide.

Frequently asked questions about AI text detection

Why do AI tics change over time?

Models operate on conditional probability. When a pattern becomes a public tell, training learns to avoid it. The word delve, for example, was a clear signature in 2023 and 2024 and dropped sharply in 2025 according to a Washington Post analysis of 300 thousand ChatGPT messages. That is why the map needs an annual update.
What are the AI flag words in 2026?

Inflated verbs like delve, harness, foster, potencializar. Dramatic nouns like tapestry, landscape, journey, ecosystem. Empty intensifiers like crucial, essential, significant. The phrase “plays a significant role in shaping” was measured by GPTZero as 207 times more frequent in AI text than in human text.
What changed in the “It is not X. It is Y.” construction in 2026?

The construction is still the structure most replicated by AI. In 2025 and 2026, the model began to use a variation that escapes the original filter: “Not only X, but also Y.” Instead of negating and replacing, it now adds. It keeps the parallelism that gives the pattern away and removes the negation that betrayed the old formula.
How do you detect AI tics in Portuguese text?

Most of the published research on AI tells is in English. Brazilian Portuguese has its own patterns. False colloquialism in the format “bora?” and “sacou?”. Fabricated intimacy like “confesso que” and “bate um papo”. Diminutive of closeness like “dicazinha”. Generic regional openers like “olha só” and “pois é”.
Do bold and emoji in professional text indicate AI use?

They do when they appear without judgment. Bold on a random noun with no emphasis function is a strong tell. The emoji (check) appears in a third of ChatGPT messages according to the Washington Post. Seventy percent of the messages contain at least one emoji. In professional editorial text, those numbers are statistically anomalous.
Is there a practical diagnosis to measure AI tics in my operation?

Select five pieces published by the team in the last sixty days. Mark each occurrence of the twelve banned structures and of the new classes (inflated verb, dramatic noun, empty intensifier, “not only X but Y”, default opener, corporate hedging, visual formatting, generic hypothetical, PT-BR tics). Calculate the density per thousand words. Well-operated editorial text tends to stay below four per thousand.

Sources and references

• The Washington Post. Statistical analysis of 300 thousand ChatGPT messages between June 2024 and July 2025. Coverage published in November 2025.
• Wikipedia. Signs of AI Writing. Collaborative entry updated in November 2025 by editors who review suspicious contributions at scale.
• Surfer SEO. How to Avoid AI Detection in Writing. Guide updated in March 2026 with a list of flag words by category.
• GPTZero. Detection base 2025 and 2026. Measurement of the frequency of the phrase “plays a significant role in shaping” as 207 times higher in AI text.
• Kernan, Sean. 13 Signs You Used ChatGPT To Write That. Substack, April 2025.
• RTÉ Brainstorm. How to detect text which has been written by ChatGPT. November 2025.
• Headcore Digital. Writing Guidelines. Internal editorial production document. May 2026 version.

Editorial curation applied to your brand. The Brand DNA documents the judgment that comes before the prompt. The Brandformance System operates the three phases of the AI operation. If your production is accumulating tics from the list above, it is a position on the curve, and a curve is climbed with method.

Headcore Intelligence Unit · Growth Science

How many of the tics on this list show up in the piece you published yesterday?