Key Facts

• Using an AI tool to generate writing and then presenting it as entirely your own can meet the dictionary definition of plagiarism, even if no words were copied from a specific source.
• A new 2025 test of 11 AI content detectors found only three achieved perfect accuracy across five text samples.
• General-purpose chatbots, including ChatGPT Plus, Copilot, and Gemini, outperformed most standalone detectors in the same tests.
• Detector accuracy has not improved in a straight line; some tools that once scored perfectly later declined, sometimes as free-use limits were added.
• Human-written text can still be misclassified as AI, raising risks for students, writers, and non-native English speakers.

The State of AI Content Detection in 2025

Three years after generative AI became a global phenomenon, the problem of AI-generated plagiarism remains stubbornly difficult to solve. AI writing tools are now embedded in word processors, email clients, coding environments, and social media workflows. The result is a flood of text that may be human, machine-generated, or a collaboration between the two. For teachers, editors, publishers, and managers, the question is simple: can any tool reliably tell the difference?

The latest round of testing suggests the answer is still no, or at least not with the confidence that high-stakes decisions require. A five-part test series was run across 11 content detectors. Two of the five text blocks were written by a human. Three were written by ChatGPT. Each detector was fed each block separately. If it correctly identified the block as human or AI, the test was considered passed. If it got the answer wrong, the test was failed. When a detector returned a percentage, anything above 70 percent was treated as a strong probability and counted as its answer.

The results were mixed. Only three detectors achieved a perfect score: Pangram, QuillBot, and ZeroGPT. Several well-known names struggled. BrandWell and Grammarly each scored 40 percent. Undetectable.ai scored just 20 percent, a dramatic fall from a previous perfect score. Copyleaks, GPTZero, and Originality.ai each scored 80 percent, but their errors were troubling because they involved confident misclassification of human writing as AI-generated.

Why Detection Is So Hard

AI content detectors generally look for statistical patterns. They measure properties such as perplexity, which estimates how surprised a language model is by a sequence of words, and burstiness, which looks at variation in sentence length and structure. Human writing tends to be more variable, while AI writing can be more uniform. But these signals are not reliable. A human who writes formally, uses predictable phrasing, or follows a strict style guide may look machine-like to a detector. A machine that is prompted to write with unusual rhythm, slang, or deliberate errors may look human.

Another complication is that detectors are trained on data from older models. As new models are released, their output changes. A detector tuned to recognize GPT-3 or GPT-2 patterns may miss text from newer systems. Some tools are updated regularly, while others appear abandoned. The GPT-2 Output Detector, for example, still carries the name of an early OpenAI model. Its accuracy was only 60 percent in the latest test, and there is little evidence that it has been meaningfully updated.

False positives are especially dangerous. When a detector labels a human-written essay as AI-generated, a student can face accusations of cheating. When it flags a journalist's article, a publisher may question the writer's integrity. Research and anecdotal reports have shown that non-native English speakers are more likely to be misclassified. Their writing may use simpler sentence structures or less common idioms, which some detectors associate with machine generation. That creates an equity problem on top of a technical one.

The Testing Method

The test used five blocks of text. Two were written by a human. Three were generated by ChatGPT. Each detector received the blocks one at a time. A detector passed if it correctly identified the source. A detector failed if it mislabeled the text. When a tool gave a confidence score, a result above 70 percent was treated as a definitive answer. The full series involved 55 individual tests across 11 detectors.

The detectors tested were BrandWell, Copyleaks, GPT-2 Output Detector, GPTZero, Grammarly, Originality.ai, QuillBot, Undetectable.ai, Writer.com, ZeroGPT, and Pangram. One earlier tool, Writefull, had been dropped after it discontinued its GPT detector. Another, Monica, was dropped because it limited testing to 250 words and then restricted tools behind a $200 upgrade. Pangram was added as a newcomer and immediately earned a perfect score.

The test also compared dedicated detectors with general-purpose chatbots. Each chatbot was given the prompt: Evaluate the following and tell me if it was written by a human or an AI. The chatbots were run in incognito windows without logging in, except for ChatGPT Plus, which required a $20 monthly subscription. The results were notably stronger for the chatbots than for many dedicated detection services.

Detector-by-Detector Results

Perfect Scores: Pangram, QuillBot, and ZeroGPT

Pangram is a relatively new company founded by engineers who previously worked at Google and Tesla. Its focus appears to be AI detection rather than content generation or humanizing tools. It offers five free tests per day. Processing was somewhat slow, and the interface displayed a partially white screen for longer than was comfortable. But the results justified the wait. Pangram correctly identified all five test blocks, earning a perfect score on its first appearance in the test series.

QuillBot has had a volatile history in earlier tests. Multiple passes of the same text once produced wildly different scores. In the previous round, however, it became rock solid and scored 100 percent. The latest test checked whether that performance would hold. It did. QuillBot again identified every human and AI sample correctly. That consistency is notable in a field where many tools fluctuate from one month to the next.

ZeroGPT has matured significantly. When first evaluated, it had no clear company name, was filled with Google ads, and lacked obvious monetization. It worked reasonably well but seemed sketchy. That impression is now gone. ZeroGPT presents as a typical software-as-a-service product with pricing, company information, and contact details. Its accuracy rose from 80 percent to 100 percent in the summer and held at 100 percent in the current test.

Midrange Performers

Copyleaks scored 80 percent, but its marketing claims have been aggressive. The company previously described itself as the most accurate AI detector with over 99 percent accuracy. It later revised the claim to 99 percent accuracy backed by independent third-party studies. In the test, Copyleaks identified a human-written block as 100 percent AI-generated. That error is especially concerning because another detector, BrandWell, correctly identified the same block as human-written. Copyleaks sells primarily to educational institutions, publishers, and enterprises that want to ensure content originality and academic integrity.

GPTZero also scored 80 percent, but its errors changed between tests. In an earlier round, it got the first test wrong and the second right. In the latest round, it got the first right and the second wrong. The first block was human-written. The second came from ChatGPT. GPTZero has grown from a bare-bones site into a company with a full team and a mission of protecting what is human. It offers AI validation tools and a plagiarism checker. But regular tinkering has not produced steady improvement.

Originality.ai scored 80 percent. It is a commercial service that bills itself as the most accurate AI detector. It sells usage credits, with 2,000 credits for $12.95 per month. In the test, it correctly identified some samples but also declared a human-written block to be 100 percent AI-generated. That was a decline from a previous run, when it correctly identified the same human writing as human.

Low Performers

BrandWell scored 40 percent. The tool was originally produced by an AI content generation firm called Content at Scale and later migrated to BrandWell.ai, a marketing services company with an AI focus. After six months, its overall score stayed the same. It correctly identified only two of five tests. It was confused by one ChatGPT-written block and declared two other AI-written blocks to be human-written. For one test, it labeled almost the entire AI-written sample as human, except for a single line.

Grammarly scored 40 percent. Grammarly is widely known for grammar checking, but it also offers plagiarism and AI content detection. The company has promoted its AI content checker as no longer being in beta. The test found no improvement. One sample that was entirely written by ChatGPT was not correctly identified. Grammarly has a reputation as an AI-forward text analysis company, which made the lack of progress surprising. Its plagiarism checker did correctly identify test text as previously published, but that is a different function from AI detection.

The GPT-2 Output Detector scored 60 percent. It was built using a machine-learning hub managed by Hugging Face, an AI company based in New York. While Hugging Face has received substantial funding for its natural language library, the GPT-2 detector appears to be a user-created tool using the Transformers library. Its detecting quality has not changed since the last test. Given that it carries the name GPT-2 and newer models have advanced far beyond that generation, it is reasonable to assume the tool has not seen a significant update.

Undetectable.ai scored 20 percent, the biggest drop in the test. The company's main claim is that it can humanize AI-generated text so detectors will not flag it. That feature was not tested, partly because it raises ethical concerns about cheating. However, the company also offers an AI detector. In the latest test, it rated a human-written block as 60 percent likely AI. It then rated all three AI-written samples as 75, 76, and 77 percent likely human. The results were almost exactly backwards.

Writer.com scored 40 percent. Writer.com is a service that generates AI writing for corporate teams. Its AI Content Detector tool can scan for generated content. In the test, it identified every text block as human-written, even though three of the five samples were written by ChatGPT. There was no improvement from the previous evaluation.

Chatbots Outperform Dedicated Detectors

The most striking finding was that general-purpose chatbots performed better than many tools specifically marketed as AI content detectors. ChatGPT Plus, Copilot, and Gemini all returned perfect scores. Each correctly identified every text block as human or AI. That result suggests that the same large language models that generate text can also be effective at recognizing it, at least in controlled tests.

The free tier of ChatGPT did make one error. It misclassified the last human-written block. But its analysis of the first block was unsettling. In an incognito window with no login and no personal identifying information, it not only identified the text as human-written but also inferred the author's identity from the style. That level of inference raises privacy and anonymity concerns, even if the identification was based on publicly available writing.

Grok was included because it had performed well in an overall chatbot evaluation. But it failed the detection test with three out of five answers wrong. Like some other tools, it identified all the writing blocks as human. The failure shows that strong general chatbot performance does not automatically translate into reliable AI detection.

What This Means for Plagiarism and Integrity

The dictionary definition of plagiarism is to steal and pass off the ideas or words of another as one's own, or to use another's production without crediting the source. AI-generated text fits this definition in an important way. A person using an AI tool is not necessarily copying from a specific human author, but if that person does not disclose that the words came from an AI and claims them as their own, the act still meets the dictionary definition of plagiarism. The challenge is that detection is unreliable, so enforcement is difficult and sometimes unfair.

Schools, publishers, and workplaces need policies that go beyond detector scores. A single positive result should not be treated as proof. Better evidence can include drafts, version history, interviews about the writing process, and explanations of source material. Detectors can be used as one signal among many, but not as judge and jury. Institutions should also consider the risk of false positives, especially for non-native speakers and writers who use formal or formulaic styles.

The history of these tests shows no consistent upward trend. In early 2023, the best result was 66 percent correct from one of three available checkers. By February 2025, up to 10 checkers were tested, and three had perfect scores. By April, five detectors boasted perfect scores. Six months later, only three achieved perfect scores, including one newcomer. Some tools that once performed perfectly declined at about the same time they added restrictions on free use. The market is changing quickly, but reliability is not steadily improving.

For now, the most reliable lesson is that no single detector should be trusted as the final word. The tools can be useful for raising questions, but they cannot answer them with certainty. Human judgment, transparent writing practices, and clear disclosure remain the strongest defenses against AI-generated plagiarism. As models improve and detectors adapt, the arms race will continue. The only consistent result across the tests is inconsistency itself.

Source:ZDNET News