AI‑generated text and the plagiarism debate: what’s new, what’s overstated

What the blog post claims

The author of a personal blog argues that:

• Large language models (LLMs) ingest all publicly available text without permission.
• The models then “copy” that text and sell the output to users, who in turn resell it.
• Because of this, the author believes AI is nothing more than unauthorised plagiarism at a larger scale.

The post cites a single anecdote: a competitor’s tutorial that allegedly lifted exact phrasing and even a hyperlink from the author’s own site, ranking higher in Google search results.

What is actually new?

1. Training data collection is not a secret

Most major LLM developers publish high‑level descriptions of their data pipelines. For example, OpenAI’s GPT‑4 technical report states that the model was trained on a mixture of licensed data, publicly available web pages, and data created by human annotators. The exact composition is proprietary, but the community knows that large‑scale web crawls (Common Crawl, Wikipedia, books, code repositories) form a substantial portion.

2. Models do not store verbatim passages

Neural networks learn statistical patterns, not literal copies. When a model generates text, it samples from a probability distribution conditioned on the prompt. The output may resemble source material, especially for short, factual statements, but the model does not retrieve a stored paragraph and paste it verbatim. Researchers have demonstrated this by probing hidden states and showing that the model’s internal representations are distributed and fuzzy.

3. Legal frameworks are evolving, not static

The U.S. Copyright Act distinguishes between reproduction (copying a protected expression) and derivative works. Courts have not yet ruled definitively on whether AI‑generated text that is substantially similar to a copyrighted source constitutes infringement. The recent Authors Guild v. Google case (about digitising books) shows that transformative use can be permissible, but the line remains blurry for generative AI.

4. Real‑world tools already include safeguards

OpenAI, Anthropic, and Cohere all provide content‑filtering APIs and encourage users to attribute sources when the model reproduces factual information. Some platforms (e.g., Microsoft’s Copilot for Office) embed citation mechanisms that insert footnotes linking to the most likely source, though these are optional and not universally adopted.

Limitations that keep AI from being pure plagiarism

Limitation	Why it matters
Statistical generation	The model predicts the next token based on context; exact duplication of long passages is statistically unlikely unless the prompt explicitly includes them.
Prompt length constraints	Most APIs cap the context window (e.g., 8k‑32k tokens). To reproduce a multi‑paragraph article, a user would need to feed the entire source, which defeats the “unauthorised copying” claim.
Post‑generation editing	Users typically edit AI output before publishing. Even if a sentence matches a source, the surrounding text is often rephrased, reducing the risk of wholesale plagiarism.
Detection tools	Services like Turnitin and OpenAI’s own AI‑text classifier can flag suspiciously similar passages, giving publishers a way to audit content.
Economic incentives	High‑quality, original content still ranks better for topical depth and expertise. Search engines increasingly value user engagement metrics that are hard to fake with generic AI output.

The anecdote: a single copied link does not prove systemic theft

The blog author points to a competitor’s article that contains an identical hyperlink and anchor text. Several plausible explanations exist:

1. Accidental reuse – The competitor may have copied a snippet from the original page as a citation and neglected to replace the link.
2. Common phrasing – Certain calls‑to‑action (“read more about X”) are generic and can appear across many sites.
3. Search engine optimisation (SEO) tricks – Some writers deliberately embed competitor URLs to gain backlinks, a practice known as negative SEO.

Even if the copy is confirmed, it does not demonstrate that the AI model itself reproduced the text. The competitor could have manually copied the passage, or the AI could have suggested a similar sentence that the author then edited to include the original link.

What the community is doing about attribution and compensation

• Data‑provider licensing – Companies like Microsoft have struck deals with news publishers (e.g., Bloomberg, Reuters) to include licensed articles in training sets, with revenue‑share clauses.
• Creator‑focused platforms – Projects such as Cohere’s “Creator Credits” program allow artists and writers to opt‑in their works and receive a share of downstream model usage.
• Policy proposals – The EU’s AI Act draft mentions “training data provenance” and could require disclosures about the origin of large‑scale datasets.

Bottom line

The claim that AI is simply “unauthorised plagiarism at a bigger scale” conflates two distinct issues:

1. Data collection – LLMs are trained on massive, mostly public corpora, often without explicit permission. This raises ethical and legal questions that are still being debated.
2. Generation behavior – The models do not copy and paste source text; they generate novel sequences that may incidentally resemble existing works.

The real risk lies in misuse: a user can prompt an AI to reproduce large excerpts verbatim, and platforms currently lack strong technical barriers to prevent that. Mitigation will require a mix of better dataset licensing, transparent model documentation, and robust attribution tools—not a blanket condemnation that all AI output is plagiarism.

---

For further reading:

• OpenAI, GPT‑4 Technical Report – https://cdn.openai.com/papers/gpt-4.pdf
• European Commission, Draft AI Act – https://digital-strategy.ec.europa.eu/en/policies/european-approach-artificial-intelligence
• Turnitin, AI‑Generated Content Detection – https://www.turnitin.com/products/ai-detection