Skip to content

Why AI writes like everyone else

AI writing sounds the same because of how chat models are trained. We tested five YouTube channels and reviewed the research on prompting vs personal models.

Veri Research14 min read

On this page
  1. Why does AI writing sound the same?
  2. What are the signs of AI-style writing?
  3. Is AI making everyone’s writing the same?
  4. Do readers notice AI-sounding writing?
  5. Can you prompt ChatGPT to write in your voice?
  6. What happened when we tested five YouTube channels?
  7. How can you check your own draft?
  8. Does training a model on your writing work better?
  9. When is prompting enough?
  10. Start from your own content.
  11. What we don’t know yet.
  12. More Veri Research Reports
  13. Sources
  14. Frequently asked questions

AI writing can sound the same because post-training narrows the variety of a model’s answers, and people using the same model can inherit its stylistic habits.²³ Examples of your writing help, but they do not guarantee your voice. In our five-channel YouTube title test, ten examples moved the drafts about 30% closer to each creator’s real titles on eight measured style features, such as length and capitalization, while generated titles remained more similar across creators than the real titles. A literary-style preprint found stronger results from author-specific fine-tuning, but used large book collections and tested short excerpts, not creator scripts.²¹

Why does AI writing sound the same?

Post-training is one documented cause of reduced variety. How a model selects text at generation time also matters.¹²

Pretraining teaches a model to predict the next token, a word or part of one, from the text before it. Prediction alone does not require bland writing. A 2020 ICLR paper by Ari Holtzman and colleagues distinguishes the training objective from the way text is selected during generation: “using likelihood as a decoding objective leads to text that is bland and strangely repetitive.” The paper shows that sampling can improve diversity from the same model, so its finding should not be read as evidence that pretraining inevitably makes writing uniform.¹

The second stage is post-training, where the model is tuned to act like a helpful assistant. Labs show it example answers and then reinforce the answers human raters prefer, a process called reinforcement learning from human feedback (RLHF). Robert Kirk and colleagues tested each stage separately at ICLR 2024 and found that RLHF “significantly reduces output diversity compared to SFT across a variety of measures.”² SFT is the supervised step that comes before it. A separate ICLR 2024 experiment by Vishakh Padmakumar and He He had people write argumentative essays with a base model, with a feedback-tuned model or alone. Only the feedback-tuned model reduced diversity. It made different people’s essays more similar, and the authors traced the effect mainly to the model contributing less varied text.³

Why would human feedback narrow a model? A 2025 preprint by Jiayi Zhang and colleagues points at the raters themselves. Reviewing preference datasets, they found “typicality bias,” in which “annotators systematically favor familiar text,” and they argue this bias plays a central role in what researchers call mode collapse, where a model keeps returning to a small set of favorite answers.⁴ A 2026 ICML paper by Jacob Springer and colleagues reported that the post-training trade-off between following instructions and keeping variety “worsens with scale.”⁵ Bigger models are no automatic fix.

A 2022 LessWrong essay by janus documented an OpenAI model giving a narrow range of answers. The essay is a blog post, not peer-reviewed, and janus later added a correction that the model in question “was probably not trained with RLHF.”⁶ The later peer-reviewed work above is the firmer evidence.

These studies explain why a model can follow instructions well without preserving much variety. Matching an individual voice asks for something more specific than satisfying a general preference rating.

What are the signs of AI-style writing?

AI-style writing has measurable fingerprints in its grammar and its vocabulary, and they show up across millions of documents.

Alex Reinhart and colleagues compared human writing with text from several models in a 2025 PNAS study. GPT-4o used present participial clauses at 5.3 times the human rate. The authors found that all four instruction-tuned models leaned toward a dense, noun-heavy style, with larger differences from human writing than the base models showed.⁷

The vocabulary shift is easier to spot. Dmitry Kobak and colleagues tracked word frequencies in more than 15 million PubMed abstracts in a 2025 Science Advances paper. In 2024, “delves” appeared about 28 times more often than its pre-ChatGPT trend predicted, alongside words like “showcasing” and “underscores.” From the excess vocabulary they estimated that at least 13.5% of 2024 biomedical abstracts were processed with a language model.⁸ Weixin Liang and colleagues used the same kind of word-frequency fingerprint on AI conference peer reviews and estimated that between 6.5% and 16.9% of the text submitted as reviews to several 2023 and 2024 conferences could have been substantially modified by language models, with adjectives such as “commendable,” “meticulous” and “intricate” among the words that appeared disproportionately often in AI-generated reviews.⁹ A 2026 preprint by Kobak’s group applied a newer method to full papers and estimated that by the end of 2025, 89% of open-access biomedical papers showed excess language-model vocabulary.¹⁰ That estimate is not yet peer-reviewed.

Stylometry, the statistical study of writing style, shows the same thing from another angle. Burrows’ Delta, introduced by John Burrows in 2002, compares texts by how often they use the most common words, the little words like “the,” “of” and “and” that writers use without thinking.¹¹ James O’Sullivan applied Delta to human and AI creative writing in a 2025 study. “Human-authored texts form broader, more heterogeneous clusters, reflecting the diversity of individual expression,” he wrote, while model outputs “display a higher degree of stylistic uniformity, clustering tightly by model.”¹²

The models in these studies had recurring stylistic preferences. That is evidence of a default style, not proof that every person gets the same voice.

Is AI making everyone’s writing the same?

At the group level, yes. Several studies find that writing becomes more alike when many people route it through the same model.

Zhivar Sourati and colleagues found in a 2026 Nature Human Behaviour study that when models polish and rewrite text, they keep the core content but “homogenize writing styles, reducing writing-complexity variance by a statistically significant 21-50% across datasets and models.”¹³ Padmakumar and He’s essay experiment showed the same narrowing between authors.³ In a CHI 2025 experiment with 118 participants from India and the United States, Dhruv Agarwal, Mor Naaman and Aditya Vashistha found that AI suggestions “led Indian participants to adopt Western writing styles, altering not just what is written but also how it is written.”¹⁴

Research on creative ideas finds a similar pattern. Anil Doshi and Oliver Hauser’s 2024 experiment found that stories written with AI ideas received higher novelty and usefulness ratings while becoming more alike as a group.¹⁵ Better individual ratings and less collective variety can coexist.

Do readers notice AI-sounding writing?

People who often use AI writing tools can notice it, and disclosure can affect trust. Those are separate findings, neither of which establishes how a creator’s audience will react.

Jenna Russell, Marzena Karpinska and Mohit Iyyer hired annotators who use chat models heavily for writing and asked them to label 300 non-fiction articles as human or AI. The majority vote of five such annotators “misclassifies only 1 of 300 articles,” which beat most of the commercial and open-source detectors the team tested, even against paraphrasing and “humanizer” tools.¹⁶ The annotators’ explanations leaned heavily on what the authors call “AI vocabulary,” the same kind of fingerprint described above.

What happens when readers believe something is AI-written? In two preregistered experiments with 4,976 people in the US and UK, Sacha Altay and Fabrizio Gilardi found that labeling headlines as AI-generated lowered how accurate people judged them and their willingness to share them, whether the headlines were true or false and whether a person or a model wrote them. The penalty was about a third the size of labeling a headline false, and it came from readers assuming the content had no human involvement.¹⁷ Oliver Schilke and Martin Reimann reported in 2025 that across 13 experiments, “actors who disclose their AI usage are trusted less than those who do not.”¹⁸

These studies measure labels and disclosure, not YouTube audiences reacting to an AI-sounding script. They suggest a risk, and we found no study that measures it on video platforms directly.

Can you prompt ChatGPT to write in your voice?

You can get partway there. Few-shot prompting, where you paste examples and ask for more like them, can approximate style without reliably matching an individual author.¹⁹

A 2025 EMNLP Findings paper by Zhengxiang Wang and colleagues tested over 40,000 generations per model from writing samples of around 400 real authors, scoring outputs with authorship attribution, authorship verification, style matching and AI detection. Models “can approximate user styles in structured formats like news and email,” the authors wrote, but “struggle with nuanced, informal writing in blogs and forums,” and “increasing the number of demonstrations offers limited gains in stylistic alignment.”¹⁹ In the body of the paper they note that “Generated outputs often default to an average, generic tone and remain readily detectable as AI-written.” That detection result used GPTZero, and the authors caution that detector bias may affect differences between models.¹⁹

For a practical walkthrough of why pasted scripts are not enough, see why ChatGPT doesn’t sound like you. Informal writing is the category most creator content falls into. We wanted to see what this looks like for YouTube, so we ran a small test.

What happened when we tested five YouTube channels?

We collected the 15 most recent video titles from five public YouTube channels with distinct styles: Marques Brownlee, Veritasium, MrBeast, Ali Abdaal and Mark Rober. The titles came from each channel’s public RSS feed on November 5, 2026.²⁰ For each channel we held out the five newest titles as the answers and kept the other ten as examples.

For each held-out video, a model first wrote a plain topic brief from the title and description, with instructions to avoid the title’s wording. A general-purpose chat model then wrote titles from each brief. The two prompt conditions were:

  • Generic prompt. A request for one title for a popular YouTube creator’s video about the supplied topic.
  • Few-shot prompt. The same request plus the creator’s ten example titles and an instruction to match their style.

Each prompt produced two drafts per video, so we compared 25 real titles with 50 generic and 50 few-shot titles. We counted words by whitespace and averaged capitalization shares within each title. Character similarity is cosine similarity of TF-IDF-weighted character trigrams, with inverse-document frequencies fitted to all 75 collected titles. The style-profile score averages eight feature gaps against each creator’s held-out titles: word count, colon, question mark, exclamation mark, first-person wording, number, em dash and capitalization. Word-count gaps are divided by five before averaging. These are descriptive measures, not a validated measure of voice:

Measure Real titles Generic prompt 10 examples in prompt
Average words per title 6.8 9.6 7.5
Titles with an em dash 0% (0 of 25) 16% (8 of 50) 0% (0 of 50)
Titles using “Here’s what” 0% 8% 0%
Mean share of words over 3 letters starting uppercase 70% 98% 80%
Character similarity to the creator’s real title (0 to 1) n/a 0.20 0.26
Similarity between different creators’ titles (0 to 1) 0.026 0.049 0.042
Gap from real titles on eight style features (lower is closer) 0 0.235 0.165

The generic titles came out longer and capitalized almost every longer word. Eight contained an em dash and four used “Here’s what.” None of the 75 collected real titles used an em dash. Ten examples eliminated those two features in the generated titles and raised mean character similarity to the paired real title by 28%. On the eight style features, the drafts moved about 30% closer to the real titles.

The examples did not undo the measured convergence. Cross-creator character similarity was about 1.9 times the real-title baseline under the generic prompt and 1.6 times with examples. Ali Abdaal’s titles show why matching one visible feature can still miss the held-out sample. The mean capitalization share was 50% in his five real titles, 100% in the generic drafts and 12% with examples. The few-shot titles moved toward sentence case without matching the balance in the held-out titles.

Limits. This is a small test: five channels, 25 videos, one model, one run on one date, and titles rather than full scripts. Character similarity captures shared topic wording as well as style, and we did not collect human voice ratings. A nearest-creator attribution test we also ran was unreliable at this sample size, so we report it only in the raw data. The model may have seen some of these creators’ older titles during training, which would favor it. We did not test a fine-tuned model. The titles, generations and per-title scores are published on the data page, and every row lists the video’s URL and the date we checked it.

How can you check your own draft?

Compare the draft with a few things you actually wrote, not with a list of “AI words.” These checks come from the tells we measured in our title test and the studies above. They are our suggestions, not a validated detector.

  1. Length. The generic titles averaged 9.6 words against 6.8 for the real ones. If the draft runs longer than your usual titles or sentences, cut it back to your normal length.
  2. Em dashes. None of the 75 real titles we collected used one, while 8 of the 50 generic drafts did. Check whether you ever use them.
  3. Capitalization. The generic titles capitalized 98% of longer words. If you write in sentence case, the draft should too.
  4. Stock setups. Phrases like “Here’s what” showed up only in the generic drafts. Delete any opener you would not say out loud.
  5. Vocabulary and grammar. Words like “delves” and “underscores,” and stacked “-ing” clauses, are the patterns the PubMed and PNAS studies measured.⁸⁷ Swap them for the words you use.
  6. Your own examples. Paste a few of your real titles or paragraphs next to the draft and read them together. In our test, ten examples removed the em dashes and “Here’s what,” which is the cheapest fix to try first.

Does training a model on your writing work better?

A literary-style preprint found a large improvement with author-specific fine-tuning. Its design is much narrower than a general test of creator voice.²¹

Tuhin Chakrabarty, Jane C. Ginsburg and Paramveer Dhillon ran a preregistered study comparing MFA-trained writers with GPT-4o, Claude 3.5 Sonnet and Gemini 1.5 Pro on excerpts of up to 450 words written in the styles of 50 award-winning authors. In blind pairwise comparisons, 28 MFA-trained readers strongly disfavored prompted AI text for stylistic fidelity, with an AI-versus-human odds ratio of 0.16. The fine-tuning condition used GPT-4o trained on the available book collections of 30 living authors from that pool. Expert readers favored those outputs over human imitations for fidelity (odds ratio 8.16) and quality (1.87). Pangram flagged 3% of fine-tuned outputs as AI-generated, compared with 97% of prompted outputs. These percentages describe one detector on this sample, not proof of human authorship.²¹

The paper is a preprint under review, first posted in 2025 and revised in March 2026. Its fine-tuning corpora ranged from 0.89 million to 10.9 million tokens per author. The median training cost was $77.88, or $80.88 including an estimated $3 to generate 100,000 words. That excludes the human editing needed to make the output publishable. The comparisons were blind, so the results do not establish whether preferences would survive disclosure of AI use. General readers also preferred prompted AI for quality (odds ratio 1.82), showing that expert readers’ preferences were not universal. Agreement differed too: in the prompted condition, Fleiss’ kappa for fidelity was 0.58 among MFA readers and 0.10 among general readers.²¹ For creators, the relevant question is whether training on their own content improves drafts they actually want to use.

Personalization benchmarks test related but different tasks. The LaMP benchmark from Alireza Salemi and colleagues, published at ACL 2024, includes generating a person’s news headlines, email subjects and scholarly paper titles from their history.²² On LaMP, Zhaoxuan Tan and colleagues proposed One PEFT Per User (OPPU) at EMNLP 2024, which gives each person fine-tuned parameters to store “user-specific behavior patterns and preferences.” OPPU combines those parameters with retrieval and profiles, so its gains cannot be attributed to fine-tuning alone. The authors report that the combined system “significantly outperforms existing prompt-based methods across seven diverse tasks.”²³

A 2026 WNUT paper by Leonard Popp and colleagues adds a useful caution. Comparing prompting, steering and adapter methods for author-style transfer, they found that “fine-tuning buys most of the available style signal but forfeits fluency.”²⁴ A model that sounds like you can still need a person to read it before it goes out.

When is prompting enough?

Retrieving your own material into the prompt is a reasonable starting point when you have little data. One comparison found larger average gains from retrieval than from the parameter-efficient fine-tuning methods it tested.²⁵

The best counter-evidence comes from Salemi and Hamed Zamani’s ICTIR 2025 comparison of retrieval and fine-tuning on the LaMP tasks. Retrieval, which pulls a person’s relevant past writing into the prompt, improved results by 14.92% on average over a non-personalized model. Parameter-efficient fine-tuning improved them by 1.07%. Combining the two reached 15.98%. The authors also found that fine-tuning worked better as a person’s data grew, and that retrieval suits “cold-start users” with little personal data.²⁵ Those tasks score short outputs like headlines and ratings, so they say more about personalization in general than about voice. They still show that good retrieval from your own archive beats a weak fine-tune.

Mode collapse also has prompt-level fixes. Zhang and colleagues’ verbalized sampling asks a model to list several candidate answers with probabilities instead of one best answer, and they report that it “increases diversity by 1.6-2.1x over direct prompting” in creative writing.⁴ More variety is not the same as your voice, but it eases the sameness problem without any training.

Fine-tuning also needs evaluation beyond style. Xiangyu Qi and colleagues found that fine-tuning GPT-3.5 Turbo on just 10 adversarial examples could compromise its guardrails, and that “even without malicious intent, simply fine-tuning with benign and commonly used datasets can also inadvertently degrade the safety alignment of LLMs, though to a lesser extent.”²⁶ Popp’s fluency finding is a separate quality risk, not the same mechanism.

Approach What it changes What the evidence shows Best fit
Generic prompt Nothing about you Pulls toward the model’s house style. In our test it gave longer titles, near-total title case and borrowed formats Brainstorming where voice doesn’t matter
Pasted examples (few-shot) Examples available at generation time Reduced our custom title-profile distance by about 30%. A larger study found limits on informal style imitation and detectability¹⁹ Quick drafts, structured formats
Retrieval from your archive Which of your past work the model sees Larger gains than PEFT alone on short personalization tasks in one comparison²⁵ Creators with an archive but little time
Personal fine-tune or adapter The model’s weights, for one person Strong literary-style fidelity in one preprint.²¹ Separate studies found fluency and safety risks.²⁴²⁶ Worth testing with substantial owned content and a stable voice

For most creators the order is practical. Start from your own content in the prompt, keep the choices you make about it, and treat a personal model as the step after that once there is enough of your work to train on.

Start from your own content.

For a creator, the practical starting point is an archive of content and choices that reflect how you want to write. Your scripts and the edits you make to drafts supply examples a generic request does not. If you want a walkthrough of why pasted scripts can fall short, Why ChatGPT doesn’t sound like you covers it, and Custom GPT vs personal model compares what each one keeps.

Veri is a creator co-pilot powered by a model you own. Your agents start from your own connected content, the preferences you save and the corrections you make. Script drafts from your idea, your sources and your direction, and Titles can work from the finished script. In Your Model you can browse what Veri knows about your voice, look, topics and choices, open an item to review its sources where available, and correct it. Revising one draft doesn’t turn every change into a rule, so you can ask for a preference to be remembered and then check the saved choice in Keep your preferences.

The personal model is a separate, opt-in step. It is off unless you turn it on, it is trained only for your account, and it learns from the drafts and ideas you approve, edit and rank, plus content you own and choose to use with it. Training is by request, and requests currently join a waitlist. Veri never trains a shared model on your content, and your agents and saved preferences work the same with the personal model off. Your data and model training has the details.

If you are comparing tools, our ranking of AI script writers for YouTube looks at how each one handles your voice.

What we don’t know yet.

Among the sources reviewed, we found no peer-reviewed study that measures voice convergence in YouTube scripts or titles, tests whether viewers watch less when a script sounds AI-written, or compares fine-tuned and prompted models on creator content. The literary fine-tuning result is a preprint built on published authors with large bodies of work, while the retrieval comparison scores personalization rather than voice. Our own test covered titles from five channels with one model. Together, these results support testing examples, retrieval and author-specific training against your own drafts, rather than assuming one method will preserve your voice.

Connect a channel at joinveri.co and start your next draft from your own content.

More Veri Research Reports

Every report, with its data, is on Veri Research.

The data and the full source list for this report are on the data page.


Sources

  1. Holtzman, A., Buys, J., Du, L., Forbes, M., & Choi, Y. (2020). The curious case of neural text degeneration. ICLR 2020. arXiv
  2. Kirk, R., Mediratta, I., Nalmpantis, C., et al. (2024). Understanding the effects of RLHF on LLM generalisation and diversity. ICLR 2024. arXiv
  3. Padmakumar, V., & He, H. (2024). Does writing with language models reduce content diversity? ICLR 2024. arXiv
  4. Zhang, J., Yu, S., Chong, D., et al. (2025). Verbalized sampling: How to mitigate mode collapse and unlock LLM diversity. Preprint. arXiv
  5. Springer, J., et al. (2026). Annotations mitigate post-training mode collapse. ICML 2026. arXiv
  6. janus (2022). Mysteries of mode collapse. LessWrong (blog post, not peer-reviewed). Checked November 5, 2026. LessWrong
  7. Reinhart, A., Markey, B., Laudenbach, M., et al. (2025). Do LLMs write like humans? Variation in grammatical and rhetorical styles. PNAS, 122. PNAS
  8. Kobak, D., González-Márquez, R., Horvát, E.-Á., & Lause, J. (2025). Delving into LLM-assisted writing in biomedical publications through excess vocabulary. Science Advances. Science
  9. Liang, W., Izzo, Z., Zhang, Y., et al. (2024). Monitoring AI-modified content at scale: A case study on the impact of ChatGPT on AI conference peer reviews. ICML 2024. arXiv
  10. Holzwarth, L., González-Márquez, R., & Kobak, D. (2026). Most biomedical publications show signs of LLM-assisted writing. Preprint. arXiv
  11. Burrows, J. (2002). ‘Delta’: A measure of stylistic difference and a guide to likely authorship. Literary and Linguistic Computing, 17(3), 267–287. Oxford Academic
  12. O’Sullivan, J. (2025). Stylometric comparisons of human versus AI-generated creative writing. Humanities and Social Sciences Communications, 12, 1708. Nature
  13. Sourati, Z., Karimi-Malekabadi, F., et al. (2026). The shrinking landscape of linguistic diversity in the age of large language models. Nature Human Behaviour, 10, 1925–1942. Nature
  14. Agarwal, D., Naaman, M., & Vashistha, A. (2025). AI suggestions homogenize writing toward Western styles and diminish cultural nuances. CHI 2025. arXiv
  15. Doshi, A. R., & Hauser, O. P. (2024). Generative AI enhances individual creativity but reduces the collective diversity of novel content. Science Advances, 10(28). Science
  16. Russell, J., Karpinska, M., & Iyyer, M. (2025). People who frequently use ChatGPT for writing tasks are accurate and robust detectors of AI-generated text. ACL 2025. arXiv
  17. Altay, S., & Gilardi, F. (2024). People are skeptical of headlines labeled as AI-generated, even if true or human-made, because they assume full AI automation. PNAS Nexus, 3(10), pgae403. OUP
  18. Schilke, O., & Reimann, M. (2025). The transparency dilemma: How AI disclosure erodes trust. Organizational Behavior and Human Decision Processes, 188, 104405. DOI, abstract mirrored by EconPapers
  19. Wang, Z., Tripto, N. I., Park, S., Li, Z., & Zhou, J. (2025). Catch me if you can? Not yet: LLMs still struggle to imitate the implicit writing styles of everyday authors. Findings of EMNLP 2025. arXiv
  20. YouTube public channel RSS feeds for Marques Brownlee, Veritasium, MrBeast, Ali Abdaal and Mark Rober. Checked November 5, 2026.
  21. Chakrabarty, T., Ginsburg, J. C., & Dhillon, P. (2025, revised March 2026). Readers prefer outputs of AI trained on copyrighted books over expert human writers. Preprint under review. arXiv, version 4
  22. Salemi, A., Mysore, S., Bendersky, M., & Zamani, H. (2024). LaMP: When large language models meet personalization. ACL 2024. arXiv
  23. Tan, Z., Zeng, Q., Tian, Y., Liu, Z., Yin, B., & Jiang, M. (2024). Democratizing large language models via personalized parameter-efficient fine-tuning. EMNLP 2024. arXiv
  24. Popp, L., Liu, D., Sinhamahapatra, S., & Niehues, J. (2026). Predicting steering vectors and adapter weights for few-shot author-style transfer. Proceedings of WNUT 2026. ACL Anthology
  25. Salemi, A., & Zamani, H. (2025). Comparing retrieval-augmentation and parameter-efficient fine-tuning for privacy-preserving personalization of large language models. ICTIR 2025. arXiv
  26. Qi, X., Zeng, Y., Xie, T., et al. (2023). Fine-tuning aligned language models compromises safety, even when users do not intend to! arXiv

Frequently asked questions

Why does AI writing sound the same?

Chat models learn patterns from large text collections, and post-training can narrow the variety of their answers. Studies have found reduced output diversity after human-feedback tuning, as well as recurring grammatical and vocabulary preferences. This can make different people's drafts sound more alike, although the effect depends on the model, prompt and task.

Do examples make AI match my style?

Partly. Examples can help without reliably reproducing an individual voice. A 2025 study of around 400 real authors found that models imitating style from examples handled news and email better than informal blogs and forums. Many outputs remained detectable by the AI detector tested. In our five-channel title test, ten examples moved the drafts about 30% closer to each creator's real titles on eight measured style features, such as length and capitalization.

Can you train AI on your own writing style?

Yes. Fine-tuning adjusts a model's weights on a person's writing. In a preregistered preprint first posted in 2025, expert readers preferred author-specific GPT-4o outputs to human imitations of the same authors, while preferring human imitations to prompted AI. Pangram flagged 3% of fine-tuned outputs versus 97% of prompted ones. The fine-tuning covered 30 living authors with large book collections, and the paper remains under review.

Is fine-tuning better than prompting with examples?

It depends on the task and data. A literary-style preprint found strong results from fine-tuning on large author-specific book collections. For general personalization with limited data, retrieval can do better. A 2025 comparison across seven tasks found average gains of 14.92% for retrieval versus 1.07% for parameter-efficient fine-tuning over a non-personalized model.

What words make writing sound AI-generated?

Studies of scientific writing found that words like delves, showcasing, underscores, commendable, meticulous and intricate surged after ChatGPT's release. In 2024 PubMed abstracts, delves appeared about 28 times more often than its prior trend predicted. Grammar matters too. GPT-4o used present participial clauses at 5.3 times the human rate in one PNAS study.

Do readers care if writing sounds like AI?

Some readers do. Frequent AI-writing users detected generated articles accurately in one study. Separate experiments found that AI labels lowered perceived headline accuracy and sharing, and that disclosing AI use could reduce trust. These results do not directly measure viewers reacting to AI-sounding YouTube scripts.

What is a personal language model?

A personal language model is a model, or a set of adapter weights on top of one, that is trained for a single person's data and used only for that person. Research systems such as OPPU give each person their own small set of fine-tuned parameters so the model can capture that person's behavior patterns.

Does AI make everyone's writing more alike?

Several studies find this at the group level. A 2026 Nature Human Behaviour study found that polishing text with language models cut variance in writing complexity by 21% to 50%, and an ICLR 2024 experiment found that essays written with a feedback-tuned model were more similar to each other than essays written without one. The findings depend on the tasks and models studied.

All articles

Try it on your own channel.

Veri, the creator co-pilot, has a full suite of creator tools that work from your own channel: Ideas, Validate, Research, Script, Titles, Thumbnails with Layers, Clips and Signals.

Get started

Free early access, no card required.