What is AI slop: a glossary, the tells of ChatGPT, Claude and Gemini, and how to avoid producing it

In this article: map of terms · glossary · origin of the word · numbers · impact · signs in text · GPT, Claude, Gemini · languages · images, music, code · detectors · how to avoid slop · FAQ · sources

Map of terms

Dictionaries and media use the words around slop loosely and often mix them up. It helps to sort them by where the problem arises.

Glossary

Each entry has a definition, an origin and a link to slop. The entries run from the most general to the narrowest.

AI slop

AI slop is low-quality digital content produced by artificial intelligence in large quantities and with little human effort, often without anyone having asked for it.

Origin: the term started to spread in May 2024. The user @deepfates compared it to spam on 6 May, and two days later Simon Willison wrote that slop is to unwanted AI content what spam is to unwanted email. Willison stresses that he did not invent the word.

Example: a 2024 study by Stanford and Georgetown analyzed 120 Facebook pages that had posted at least 50 AI images and gathered hundreds of millions of interactions. One post had over 40 million views.

Boundary: not everything made with the help of AI counts as slop. Willison writes that not all AI-generated content is slop, just as not all promotional email is spam. What decides it is the author's care and what they add themselves.

Workslop

Workslop is AI-generated work output that looks like good work but lacks the substance needed to move a task forward, so the recipient has to finish it.

Origin: the term was introduced on 22 September 2025 in an article in Harvard Business Review by a team from BetterUp Labs and the Stanford Social Media Lab.

Numbers: in a survey of 1,150 US employees, about 40% said they had received workslop in the past month. Dealing with one case took them 1 hour and 56 minutes on average. The authors estimated the loss at 186 dollars per employee per month. The figures come from the respondents' own estimates and BetterUp sells coaching services, so treat the numbers as indicative.

Link to slop: workslop is slop in the workplace. It saves the sender time and shifts the work to the recipient.

Botshit

Botshit is untruthful or made-up chatbot content that a person uses in communication or decision-making without checking it.

Origin: Tim Hannigan, Ian McCarthy and Andre Spicer introduced it in the article “Beware of botshit” in Business Horizons (2024). The word builds on the philosopher Harry Frankfurt's notion of bullshit, meaning content produced without regard for truth. The harm arises only when the content is used. The authors therefore split work with chatbots into four modes, depending on how important the accuracy of an answer is and how easy it is to verify.

Example: an Air Canada chatbot told a customer, wrongly, that he could claim a bereavement discount retroactively. In February 2024 a tribunal in British Columbia ordered the airline to pay him the difference.

Link to slop: slop is a product, botshit is a process failure. It arises the moment someone uses the made-up text.

Hallucination

Hallucination is a fluent and confident but false or unsupported output of a language model, such as an invented quote, study or link.

Dispute over the name: some authors prefer the word confabulation, which is used for example by a Nature study from June 2024 on detecting hallucinations. Hicks, Humphries and Slater argue in “ChatGPT is bullshit” (2024) that models simply do not care about truth.

Example: in 2023 lawyers in the US submitted court rulings invented by ChatGPT, and the judge fined them 5,000 dollars (Mata v. Avianca).

Link to slop: hallucination is the part of slop that readers detect worst, because the text sounds sure of itself. On 1 October 2026 Google updated its guidance on AI content and called manual fact-checking before publication critical.

Sycophancy

Sycophancy is an assistant's tendency to tell users what they want to hear: to praise, to agree and to change its answer to match their opinion, even when that is factually wrong.

Example: on 25 April 2025 OpenAI rolled out a GPT-4o update that, in its own description, was overly supportive but disingenuous, and pulled it within days.

Link to slop: a sycophantic model praises a mediocre concept and offers no resistance, so slop passes without correction. Vendors therefore suppress such tics. According to an analysis by Simon Willison, Anthropic put a rule into the Claude 4 system prompt (May 2025) that an answer should not open by praising the question.

Model collapse

Model collapse is the gradual degradation of generative models that repeatedly learn from the outputs of earlier models: rare phenomena disappear, and outputs flatten and repeat.

Origin: Ilia Shumailov and co-authors described the effect in Nature on 24 July 2024 (the study “AI models collapse when trained on recursively generated data”). Related names are Model Autophagy Disorder (Rice University, 2023) and Habsburg AI, a metaphor by Jathan Sadowski from February 2023 for a model that learns from the output of other models and becomes, as it were, inbred.

Dispute over the threat: collapse happens when models train only on synthetic data. According to Gerstgrasser et al. (2024), it is prevented when synthetic data is added to the original real data. We could not find firm evidence of collapse in deployed frontier models, so the claim that slop will ruin the next models is disputed.

Dead internet theory

Dead internet theory claims that most online content and activity is produced by bots and automatically generated content instead of people.

Origin: it appeared around 2021 on the forum Agora Road's Macintosh Cafe and reached the mainstream through Kaitlyn Tiffany's essay in The Atlantic of 31 August 2021, whose subtitle says the theory is wrong but feels true. The strong version (the web is entirely fake and run by a hidden authority) is a conspiracy theory. The mild version (bots and AI content are very widespread on the web) is supported by data on bot traffic.

Numbers: according to the Imperva (Thales) report, automated traffic made up about 51% of all web traffic in 2024, for the first time in roughly ten years more than human traffic. It counts requests on one vendor's network and includes useful bots, so it does not say how much content AI wrote.

Altman: in September 2025 Sam Altman wrote that he had never taken the theory very seriously but that there now seem to be a lot of accounts on X run by language models.

Enshittification

Enshittification is Cory Doctorow's name for the gradual decay of an online platform: first it serves its users, then it sacrifices them for its business customers, and finally it squeezes those customers too to keep the value for itself.

Origin: Doctorow coined the term in 2022. The American Dialect Society chose it as word of the year for 2023, the Macquarie Dictionary for 2024, and a book of the same name came out in October 2025.

Link to slop: this is commentators' interpretation, according to which platforms in late stages have a motive to fill feeds with cheap content. We have not verified Doctorow's own link to slop.

Brain rot

Brain rot denotes the supposed deterioration of a person's mental state after excessive consumption of trivial online content, or the content itself.

Origin: the earliest recorded use is by Thoreau in Walden in 1854. Oxford University Press named it word of the year for 2024 on 2 December 2024, and slop was among the finalists. According to OUP, the frequency of the term rose by 230% between 2023 and 2024.

Double meaning: in October 2025 a preprint titled “LLMs Can Get ‘Brain Rot’” appeared (Texas A&M, UT Austin and Purdue). Models that kept training on low-quality tweets did worse at reasoning and at handling long context. It is a pilot study based on data from X.

Scaled content abuse

Scaled content abuse is, according to Google, the mass production of pages whose main purpose is to manipulate search rankings instead of helping users.

Origin: Google introduced the term on 5 March 2024 together with the March core update, replacing its earlier rule on automatically generated content. The policy applies regardless of whether the pages are produced by automation, by people or by both.

What Google says about AI: since February 2023 the position has been that using AI does not in itself break the rules. The violation is producing content mainly to manipulate rankings. Since 15 May 2026 the spam policies also explicitly cover manipulation of generative AI answers in search.

Numbers: when announcing the change in March 2024, Google said it would reduce low-quality, unoriginal content in results by 40%, and after the rollout it reported 45%. This is its own figure without a published methodology and it does not apply to AI alone.

Slopsquatting

Slopsquatting is an attack in which someone registers the name of a software package that a coding model made up, and a developer then installs it because they trust the model's answer.

Origin: the name (slop plus typosquatting) is attributed by several sources to Seth Larson of the Python Software Foundation, and it spread in April 2025.

Numbers: according to summaries of the study by Spracklen et al. (USENIX Security 2025), which tested 16 models in Python and JavaScript, about 19.7% of recommended packages did not exist, and 43% of the made-up names came back in each of ten repetitions of the same prompt. The repetition turns invented names into predictable prey.

Vibe coding

Vibe coding is a way of programming in which a person gives tasks to an AI in plain language and does not review the generated code in depth.

Origin: Andrej Karpathy used it in a post on X on 2 February 2025. Collins Dictionary made it word of the year for 2025 (announced on 6 November 2025) and defined it more broadly as producing code from natural language using AI.

Link to slop: unread code and unread bug reports are slop in software. The curl project ended its bug bounty program on 31 January 2026 because, according to Daniel Stenberg, it had been flooded with worthless AI vulnerability reports.

Content farm and pink slime

Content farm is a website that produces cheap content in large volumes, tailored to search queries, for traffic and advertising revenue.

Link to slop: content farms are the predecessor of slop with the same economic motive, and generative AI only made production cheaper. As of 23 June 2026 NewsGuard tracked 3,749 sites that it classifies as AI content farms, in 16 languages. The definition and method changed in 2026, so the series are not fully comparable.

Pink slime sites pose as local news but publish automated or partisan texts. This is a different category from AI content farms.

Deepfake

Deepfake is synthetic video, audio or imagery created by deep learning that convincingly imitates a real person.

Difference from slop: a deepfake is targeted deception, while slop is low-quality content in general. Dictionary.com draws a similar line within slop: deliberately deceptive slop (images meant to look real) and sloppy slop (AI posts and books without human editing).

Where the word slop comes from and how dictionaries define it

The word itself is old. When announcing its word of the year, Merriam-Webster said that slop used to mean soft mud and food waste, later something worthless, and only recently content from artificial intelligence.

For AI content it spread in May 2024. On 6 May @deepfates wrote that slop was heading for the dictionary much like spam, and on 8 May Simon Willison described slop as unwanted AI content. The Guardian ran a piece on 19 May and the New York Times on 11 June, both building on the analogy with spam. Willison later refined slop into content that is both unrequested and unreviewed. Merriam-Webster (15 December 2025) and the American Dialect Society chose slop as word of the year 2025, and the Macquarie Dictionary picked the term AI slop (25 November 2025).

In Czech the borrowed term AI slop is used, loosely translated as břečka (slush). German uses “KI-Slop” and “AI Slop” and translations such as “KI-Müll” or “digitaler Abfall”, French uses “AI slop” and translations such as “bouillie” or “déchet” (according to Czech, German and French guides and blogs, for example t3n, IONOS and Blog du Modérateur).

The definitions differ mainly in whether they mention that nobody wanted the content. Macquarie and Willison include it, Merriam-Webster and the American Dialect Society do not. In their paper “Why Slop Matters”, Kommers et al. write that slop has so far resisted formal definition, and they describe three prototypical properties instead.

Slop is used in two senses. The first is mass-produced low-quality content on the web. The second, narrower one is used by people around language models for repeated phrases and patterns of the models. EQ-Bench builds a Slop Score from them: 60% of the weight goes to overused words, 25% to the contrast pattern “not just X, but Y” and 15% to overused three-word phrases. The Antislop study (Paech et al., 2025) reports that some patterns are more than a thousand times more frequent in language model text than in human text.

How much slop is on the web and what each number measures

Numbers on the share of AI content are quoted often and without context. Each one measures something different.

Two of them look as if they contradict each other, yet both may be true. Graphite says that 86% of articles ranking highly on Google were written by humans, and Ahrefs says that 74% of new pages contain at least some AI content. They use a different population, a different definition and a different detector. Graphite also reports a 4.2% false positive rate for its detector, roughly one human article in 24 labeled as AI.

The widely quoted estimate that 90% of online content would be synthetic by 2026 has no empirical basis. It appeared in a Europol document that referred to Nina Schick's 2020 book on deepfakes, which was talking about video and gave no clear denominator (analysis by Reason magazine, 2023).

What slop does to brands and teams

A study commissioned by the ad network Raptive (3,000 US adults, July 2025) found that when readers suspect content is from AI, their trust falls by almost 50%, even when a human wrote it. For products advertised next to such content, willingness to consider a purchase fell by 14%. Raptive sells ads on sites with human-made content, so it has an interest in the result.

In a December 2024 study (more than 2,000 participants, a survey, EEG and eye tracking), NielsenIQ found that viewers rate AI ads as less engaging and more annoying, boring and confusing than traditional ones, and measured weaker memory activation for them.

Inside companies, slop damages relationships. According to media summaries of the workslop study, 42% of recipients rated the sender as less trustworthy and 37% as less intelligent. In B2B marketing and sales, workslop aimed at a customer is an offer, an email or a report that the sender did not read. How outreach built on signals instead of mass AI correspondence looks is described in the article Go-to-market engineering.

How to recognize AI slop in text

The signs of AI text can be ranked by the strength of evidence. Best documented are the vocabulary and style of individual models, because studies measured them. Most other signs come from what Wikipedia editors and authors of guides noticed. No single sign is enough. The Wikipedia guide says explicitly that not all text with these indicators is from AI, because models learned from human writing.

What studies measured

• Vocabulary. Kobak et al. (Science Advances, July 2025) compared 15.1 million PubMed abstracts from 2010 to 2024 and found 379 words whose frequency jumped in 2024, most strongly words such as “delves”, “underscores” and “showcasing”. They estimated that at least 13.5% of 2024 abstracts had been processed by a language model. Juzek and Ward (COLING 2025) picked out 21 words that ChatGPT overuses and named training on human feedback as the probable cause.
• Model style. Sun et al. (ICML 2025) trained a classifier that could tell from the text which of five models (ChatGPT, Claude, Grok, Gemini, DeepSeek) wrote it, with 97.1% accuracy. The signal lives mainly in word choice: even after shuffling the words in a text, accuracy fell only to 88.9%.
• Signs change. A preprint on vocabulary trends (arXiv, February 2025) recorded that the frequency of the word “delve” in abstracts fell sharply after a peak in January to March 2024. The authors suggest that people keep using models but avoid words that have become a giveaway.

What editors and guide authors noticed

The guide “Signs of AI writing” grew out of WikiProject AI Cleanup, which has been reviewing AI-assisted submissions since 2023. It is an advisory page without the status of a Wikipedia policy. It lists these groups of signs:

• Undue emphasis. The text explains in generic terms why the topic matters, for example “a pivotal moment” or “a broader movement”.
• Promotional phrases. For example “stands as a testament”, “plays a vital role” or “rich cultural heritage”.
• Contrast constructions. The guide lists the pattern “it is not just about X, it is Y” as negative parallelism.
• Rule of three. Lists of three, for example three adjectives in a row.
• Vague attribution. “Experts agree” without a name or source.
• Formatting. Headings with every word capitalized, excess bold, unexpected emoji, bullets where a sentence would do.
• Conversation leftovers and artifacts. Notices about the model's knowledge date and technical strings in links such as contentReference[oaicite:1], turn0search0 or citeturn0news0. Artifacts are the only sign that is unambiguous.

The dash is the best-known and least reliable sign. The guide lists it as useful only in combination with others and says that newer chatbots use it less. On 14 November 2025 Sam Altman wrote that ChatGPT now finally follows a ban on dashes in custom instructions, although OpenAI clarified that the default behavior may not change completely. Czech sites such as Zive.cz and 365tipu call it the most visible trace of AI. Czech typesetting uses a spaced en dash and ordinary texts do not use the em dash, but we found no study of the share of dashes in Czech AI texts.

Rhythm, emptiness and invented sources

• Rhythm. Graphite compared 10,000 human articles with 90,000 AI articles from nine models and found that models use a wider vocabulary, longer words and less variation in sentence length. Graphite is a marketing firm, and the AI texts were produced by expanding summaries with GPT-4.1, which may skew the result.
• Emptiness. Czech guides (sBranding, Jecas.cz) agree: a long intro without a point, a conclusion that repeats the heading, no concrete example, number or name. Some of these sites sell training or tools.
• Hallucination. Invented sources, quotes and links are the most reliable signal, but they disappear once the author verifies the output.

This is how the signs look together. On the left is an invented intro of a blog post, on the right its version after the author's edit:

GPT, Claude, Gemini and others: how their texts differ

Models can be told apart by their text, but concrete signs change faster than research can keep up. The study by Sun et al. (ICML 2025) told five chat models apart with 97.1% accuracy. It looked for characteristic words using TF-IDF. According to a summary of the study, typical words for ChatGPT include “certainly”, “overall” and “typically”, and for Claude “according to”, “based on” and “appears to”. We did not obtain the full lists for all five models. According to the study, Claude used fewer bold segments and headers than other chat APIs.

Fictional names. The preprint “Ghost Couple” (June 2026, also covered by Nature on 30 September 2026) found that models generate repeated pairs of names. For Claude (nine tested versions) these are Elena Vasquez and Marcus Chen, for GPT Elara Voss. According to Nature, these names appear in hundreds of fake manuscripts on Zenodo and ResearchGate. The authors themselves admit that some of the names may belong to real researchers.

Tics move around. According to Graphite, the count of repeated tics fell by only 4% for Claude Opus 5.5 compared with Opus 5, although dashes fell by more than 90%. The disappearance of one signal therefore does not mean fewer tics overall.

Interface. Anthropic publishes the Claude system prompts for the web and mobile apps, and they do not apply to the API. Findings from one interface cannot be carried over to the other without checking. Most studies are also in English. We found no comparison by model for Czech, German or French.

A practical rule follows from this. Treat a sign older than half a year as suspect, and a sign found in one model version as not transferable.

What we know about the signs in Czech and other languages

Czech guides consistently list the long dash, clichés such as “In today's digital age”, sentences of equal length and a conclusion repeating the heading. None of them measured the signs on a corpus. The sources are news articles and agency blogs, some of which sell training or tools. German guides (for example t3n) list openings such as “In der heutigen … Welt”, the pattern “Nicht nur X, sondern auch Y” and “Zusammenfassend lässt sich sagen”. French blogs list “dans un monde”, “à l’ère de”, “au cœur de”, “il est important de” and formal connectors such as “en outre” and “par conséquent”. These are practical guides without measurement. We found no study comparing the frequency of these phrases in German and French with human texts.

Research on detectors looks best. MULTITuDE (Macko et al., EMNLP 2023) contains 74,081 texts in 11 languages including Czech and showed that most detectors fail in a multilingual setting, while detectors fine-tuned for specific languages do better. CEAID (2025) covers seven Central European languages including Czech and German (188,098 texts). Fine-tuned detectors did best in it and were also the most resistant to rewriting.

People cannot do it without training. Milička et al. (PLOS ONE, October 2025) found with Czech speakers that participants started with wrong ideas about AI text and improved only thanks to immediate feedback.

Slop beyond text: images, video, music and code

Images. People are poor at recognizing AI faces. Nightingale and Farid (PNAS 2022) measured 48.2% accuracy among 315 participants, which is chance level, and participants rated synthetic faces as more trustworthy. Miller et al. (Psychological Science 2023) found that people labeled AI faces as human more often than real ones, and that those who were most confident made the most errors. A Microsoft preprint from July 2025 (a quiz with more than 12,500 participants and about 287,000 ratings) measured overall accuracy of 62%, and its authors write that the result probably overestimates people's ability today.

Typical flaws such as hands, lettering, shadows or overly smooth skin are described mostly by commercial guides and fade with newer models. We found no independent study of how they develop. One trace is described by community sources and by TechRadar: images from ChatGPT in spring 2025 had a yellow to sepia tint that OpenAI did not explain.

Provenance. Google marks images from Gemini with an invisible SynthID watermark, and in the Gemini app you can upload an image and check whether Google AI created it. Newer images also carry C2PA metadata. The check, however, recognizes only content that is marked.

Music and video. In September 2025 Spotify said it had removed more than 75 million spammy tracks in the previous 12 months. In October 2026 it reportedly kept more than 50 million “slop and spam” tracks out of recommendations, which is a different metric, so do not add the numbers. The Velvet Sundown case from summer 2025: a “band” gained hundreds of thousands of monthly listeners within weeks and then admitted it was an AI project. Going the other way, Meta launched Vibes, a feed of AI videos, on 25 September 2025 (TechCrunch called it a feed of AI slop in its headline) and OpenAI launched the Sora app five days later.

Code. Besides the curl case (see vibe coding), other open source projects have restricted AI contributions, for example Gentoo (2024), NetBSD, Ghostty and tldraw (January 2026).

Why AI text detectors are not enough

A detector is a signal for further checking and cannot become proof. The reasons:

• Error rates. Graphite reports a 4.2% false positive rate for its detector on a sample of human articles from before ChatGPT. Weber-Wulff et al. (International Journal for Educational Integrity, 2023) tested detection tools and concluded that they are not accurate or reliable enough for important decisions. Rewriting the text worsens the results further.
• Languages. Outside English, detectors according to MULTITuDE usually fail, and those fine-tuned for a specific language do best.
• Bias. Liang et al. (Patterns, 2023) found that seven detectors labeled on average 61.3% of essays by non-native English speakers as AI. A 2026 study on Czech (Al Ali, Helcl and Libovický, EACL 2026) found no systematic bias against non-native speakers of Czech, so the finding cannot be carried over to all detectors and languages.
• Editors' approach. Wikipedia editors warn against detectors and name checking whether the sources support the text as the most reliable method.

Instead of a detector, check sources and facts, ask the author about the process and judge what the text gives the reader.

How to avoid producing slop

Slop arises in three places: at the brief (the model gets too little), at the output (nobody adds anything) and at the check (nobody verifies). The procedure below covers all three.

1. Start with your own material. A model composes from what it is given. Without data from your CRM, call notes, customer quotes and campaign results, it composes an average. Put at least one thing into the brief that nobody else has.
2. Define the added value in advance. What will the reader get from the text that they will not find in the first ten search results? Google's Quality Rater Guidelines from January 2025 give the lowest rating to content that is largely copied, paraphrased or created by AI with minimal effort, originality and added value.
3. Write the outline yourself. A model writes better into an outline than without one. This is a practical experience, and we know of no study on it.
4. Verify every number, name, quote and link. Since 1 October 2026 Google's guidance on AI content calls manual fact-checking and review before publishing critical. According to news reports, the check also covers titles, meta descriptions and alt text.
5. A ban on phrases is not enough. A list of banned words helps with the surface. Empty text stays empty, and according to Graphite's measurements the tics move elsewhere.
6. Read the text aloud and have someone who knows the topic read it. Ask where they would say “I have read this before”.
7. Sign it. The author's name, date and responsibility for the content are part of what Google uses to judge trustworthiness (E-E-A-T).
8. Do not promise AI answers you do not have. Since 15 May 2026 Google's spam policies explicitly cover manipulation of generative AI answers, and the penalty is the same as for ranking spam. Google's May 2026 guide also says that optimizing for AI search features is still SEO. We write about this in more detail in the article How to get into ChatGPT's answers.

The slop test: ten questions before publishing

1. Does the text contain at least one fact, example or insight that was not in the brief?
2. Can you trace a source for every number, name and quote?
3. Can the first paragraph be deleted without losing information?
4. Could the same text appear on a competitor's website with only the company name changed?
5. Would the author say it like this out loud in front of a client?
6. Is it clear who the text is for and what the reader should do afterwards?
7. Does the author know the topic from practice and could answer questions about it in conversation?
8. Does the length of sentences and paragraphs vary, or are they all similar?
9. Does the conclusion repeat the intro or the heading?
10. Would the author sign every sentence in front of a customer?

Labeling AI content in the EU. The Artificial Intelligence Act (AI Act) imposes in Article 50 a transparency obligation for deepfakes and for AI texts published to inform the public on matters of public interest, unless a human has reviewed the content and holds editorial responsibility for the publication. Under the original timetable the obligations started to apply on 2 August 2026, and delays of some of them were under discussion, so check the current status with a lawyer before you deploy.

Sources

Definitions and terms: Merriam-Webster, Word of the Year 2025: slop · ABC News, AI slop is Macquarie Dictionary’s Word of the Year 2025 · American Dialect Society, 2025 Word of the Year is slop · Simon Willison, Slop is the new name for unwanted AI-generated content · Kommers et al., Why Slop Matters, ACM AI Letters (2026) · Oxford University Press, Brain rot (2024) · American Dialect Society, Enshittification (2023) · Shumailov et al., AI models collapse when trained on recursively generated data, Nature (2024) · Gerstgrasser et al., Is Model Collapse Inevitable? (2024) · Hannigan, McCarthy and Spicer, Beware of botshit, Business Horizons (2024) · Harvard Business Review, AI-Generated “Workslop” Is Destroying Productivity (2025) · LLMs Can Get “Brain Rot”, arXiv (2025) · Spracklen et al., We Have a Package for You!, USENIX Security (2025)

Numbers and impact: Graphite, More Articles Are Now Created by AI Than Humans (2025) · Ahrefs, What percentage of new content is AI-generated? (2025) · Imperva, 2025 Bad Bot Report · NewsGuard, AI Tracking Center · The Decoder on the Kapwing study (2025) · Reason, That time we tried to make sense of a statistic (2023) · Raptive, The “AI stink” is real (2025) · NIQ, Will AI-generated advertising disrupt the creative industry? (2024) · Stanford Internet Observatory and CSET, How Spammers, Scammers and Creators Leverage AI-Generated Images on Facebook (2024)

Signs of text and models: Wikipedia, Signs of AI writing · Kobak et al., Delving into LLM-assisted writing in biomedical publications through excess vocabulary, Science Advances (2025) · Juzek and Ward, Why does ChatGPT “delve” so much?, COLING (2025) · Sun et al., Idiosyncrasies in Large Language Models, ICML (2025) · EQ-Bench, Slop Score · Paech et al., Antislop (2025) · Graphite, AI Tells (2026 update) · Sam Altman on dashes in ChatGPT (14 Nov 2025) · OpenAI, Sycophancy in GPT-4o (2025) · GitHub, Claude says “You’re absolutely right!” about everything (issue #3382) · Brzozowski and Chung, Ghost Couple (2026)

Images, music, code and provenance: Nightingale and Farid, AI-synthesized faces are indistinguishable from real faces, PNAS (2022) · Miller et al., AI hyperrealism, Psychological Science (2023) · Microsoft, How good are humans at detecting AI-generated images? (2025) · Google, Verifying images in the Gemini app (SynthID) · Music Business Worldwide, Spotify has deleted 75M spammy tracks (2025) · Daniel Stenberg, The end of the curl bug bounty (2026)

Google, detectors and languages: Google Search Central, Google Search’s guidance about AI-generated content (2023) · Google Search Central, March 2024 core update and new spam policies · Google Search Central, Using generative AI content on your website · Liang et al., GPT detectors are biased against non-native English writers, Patterns (2023) · Weber-Wulff et al., Testing of detection tools for AI-generated text (2023) · Macko et al., MULTITuDE, EMNLP (2023) · Macko and Kopál, CEAID (2025) · Al Ali, Helcl and Libovický, Different Time, Different Language, EACL (2026) · Milička et al., Humans can learn to detect AI-generated texts, PLOS ONE (2025) · sBranding, Co je AI slop a jak ho poznat · t3n, Floskeln, Buzzwords, Gedankenstriche: So erkennst du KI-Texte · IONOS, AI Slop: Definition, Risiken und Strategien · Blog du Modérateur, AI slop : comment l’IA générative transforme le web en décharge numérique · Loumina, Reconnaître un texte d’IA : les tics de ChatGPT · Jecas.cz, Je možné poznat AI obsah?

Related services: content and SEO and AI & automation. It builds on How to get into ChatGPT's answers, AI agents in marketing and What honestly annoys me about marketing.