
Key facts at a glance
- Using an AI to generate writing and passing it off as one's own without disclosure fits the dictionary definition of plagiarism.
- AI content detectors remain inconsistent: in the latest five-sample test, only three of 11 standalone detectors earned perfect scores.
- Earlier in 2025, five detectors achieved perfect accuracy; by the second half of the year, that number dropped to three.
- Pangram, QuillBot, and ZeroGPT each scored 100%. Copyleaks, GPTZero, and Originality.ai scored 80%. GPT-2 Output Detector scored 60%. BrandWell, Grammarly, and Writer.com scored 40%. Undetectable.ai scored 20%.
- Leading chatbots outperformed dedicated detectors: ChatGPT Plus, Copilot, and Gemini all scored perfect in the same five-sample test. Free ChatGPT missed one human-written sample but correctly identified the author. Grok failed three of five.
- False positives remain a serious concern, especially for non-native English speakers and human writers whose style may resemble AI output.
The plagiarism question in the age of generative AI
More than two years after generative AI captured global attention, the question of how to detect AI-generated writing has become more urgent, not less. The standard definition of plagiarism--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--applies neatly to undisclosed AI writing. A user of a large language model may not be stealing text in the literal sense, but if they present AI-generated words as their own without acknowledgment, they are still claiming authorship of work they did not write. That makes detection both a technical challenge and an ethical one.
In early 2023, the best AI content detector in one widely cited test identified only 66% of samples correctly. By February 2025, a broader test using up to 10 checkers found three perfect scores. By April, five detectors had achieved perfect results. The progress looked promising. But by the second half of 2025, quality had declined again. Only three detectors achieved perfect scores in the latest round, and two of the previously perfect tools also introduced restrictions on free use. The lesson is clear: the detector market is not a steady march toward accuracy. It is an arms race with frequent reversals.
How the latest detector tests were conducted
The latest evaluation used five blocks of text. Two were written by a human. Three were generated by ChatGPT. Each block was fed separately into each detector. If the detector correctly identified the source as human or AI, the test was marked as passed. If it was wrong, the test failed. When a tool returned a percentage, anything above 70% was treated as a strong probability and counted as the detector's answer. The same five-block method was repeated across 11 standalone detectors, producing 55 individual tests.
The detectors included BrandWell, Copyleaks, GPT-2 Output Detector, GPTZero, Grammarly, Monica, Originality.ai, QuillBot, Undetectable.ai, Writer.com, and ZeroGPT. Monica was dropped because it limited free testing to 250 words and then required a $200 upgrade to continue. A previous tool, Writefull, had already been dropped after discontinuing its GPT detector. In Monica's place, a newcomer called Pangram was added. Pangram, founded by former engineers from Google and Tesla, immediately joined the winners' circle.
The tests also compared dedicated detectors with general-purpose AI chatbots. Each chatbot received the same 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 goal was to see whether everyday AI assistants could replace specialized detection tools.
Overall detector results: three perfect scores, many misses
The latest run produced only three perfect scores: Pangram, QuillBot, and ZeroGPT each identified all five samples correctly. That was a decline from earlier in 2025, when five detectors achieved perfect accuracy. Copyleaks, GPTZero, and Originality.ai each scored 80%, missing one sample. GPT-2 Output Detector scored 60%. BrandWell, Grammarly, and Writer.com each scored 40%. Undetectable.ai scored just 20%, the weakest performance in the group.
There was no strong trend of improvement over time. The same five-sample series has now been run six times, with detectors added, dropped, and swapped. Test 5, a human-written sample, was reliably identified as human across detectors and dates in earlier runs, but even that reliability declined in the latest round. The only consistent finding is inconsistency.
Detector-by-detector performance
BrandWell AI Content Detection (40%)
BrandWell was originally produced by an AI content generation firm called Content at Scale before migrating to BrandWell.ai. Expectations were high after six months, which is a long time in AI development. Instead, its overall score stayed the same as the previous test, with only two correct answers out of five. It was confused by a ChatGPT-written sample and declared two other AI-written samples to be human-written. For one sample, it went almost entirely wrong, calling the whole AI-written text human except for a single line.
Copyleaks (80%)
Copyleaks had previously claimed to be the most accurate AI detector with over 99% accuracy. It later revised the claim to 99% accuracy backed by independent third-party studies. In this test, it identified a human-written sample as 100% AI-written. Even BrandWell, which scored lower overall, correctly identified that same sample as human. Copyleaks is primarily a plagiarism checker sold to educational institutions, publishers, and enterprises concerned with originality and academic integrity. Its marketing claim did not hold up in this round.
GPT-2 Output Detector (60%)
The GPT-2 Output Detector was built using a machine-learning hub managed by Hugging Face. While Hugging Face has received substantial funding for its natural language library, this particular detector appears to be a user-created tool using the Transformers library. Its detecting quality has not changed since the last test. Given that its name references GPT-2 while newer models have advanced far beyond that generation, it is fair to assume the tool has not been updated since it was first posted.
GPTZero (80%)
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. However, regular tinkering has not consistently improved performance. Its score declined slightly from earlier tests. In the latest round, it got one human-written sample right and one ChatGPT-written sample wrong. In the previous round, it had the opposite result. The overall grade stayed at 80%, but the specific errors changed.
Grammarly (40%)
Grammarly is best known for grammar checking, but it also offers plagiarism and AI content detection. The company now presents its AI content checker as no longer being in beta. That is a mistake. There has been no improvement since the last evaluation. A sample written entirely by ChatGPT was not detected properly. Given Grammarly's reputation as an AI-forward text analysis company, the lack of progress was surprising. Its plagiarism checker did correctly identify test text as previously published, but that is a separate function from AI detection.
Pangram (100%)
Pangram is a relatively new company founded by engineers formerly at Google and Tesla. Its focus appears to be AI detection rather than plagiarism checking or humanizing tools designed to mislead editors and teachers. The company provides five free tests per day, which fit the testing needs. Processing was a little slow, and a partially white screen appeared for longer than was comfortable. But the results were worth the wait. Pangram scored a perfect five out of five.
Originality.ai (80%)
Originality.ai 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. The latest test used about 1,400 words and consumed just 1.5% of a monthly allocation. Unfortunately, its accuracy declined in this round. Previously, it correctly identified a human-written sample as human. This time, it was 100% confident that the same human writing was done by an AI. That is a serious false positive.
QuillBot (100%)
QuillBot was wildly inconsistent in its first few tests. Multiple passes of the same text yielded different scores. In the previous round, however, it was rock solid and 100% correct. The latest test confirmed that performance. QuillBot once again scored a perfect 100%. It is one of the few tools that has held onto a perfect score across recent rounds.
Undetectable.ai (20%)
Undetectable.ai claims to humanize AI-generated text so detectors will not flag it. That feature was not tested because it raises ethical concerns for professional authors and educators. However, the company also offers an AI detector, and that detector took the biggest dive in performance. Last time, it scored 100% accuracy. This time, it rated human writing as 60% likely AI, and all three AI writing samples as 75%, 76%, and 77% likely human. The results were almost the reverse of what they should have been.
Writer.com AI Content Detector (40%)
Writer.com generates AI writing for corporate teams and also offers an AI Content Detector. Its accuracy was low. It identified every text block as human-written, even though three of the five samples were generated by ChatGPT. There was no improvement since the last visit in the summer.
ZeroGPT (100%)
ZeroGPT has matured since its early days, when no company name was listed, Google ads filled the site, and monetization was unclear. The service worked fairly well but seemed sketchy. That feeling is gone. ZeroGPT now presents as a typical SaaS service with pricing, a company name, and contact information. Its accuracy increased from 80% to 100% in the summer and has held onto that perfect score in the current test.
Chatbots as detectors: a surprising advantage
If dedicated detectors are so inconsistent, why use them at all? The latest tests suggest that everyday chatbots can do the job better. Each chatbot was given the same prompt and the same five text blocks. The results were striking. ChatGPT free tier got one human-written block wrong, but its analysis of the first block was unsettling: it not only identified the text as human-written but also identified the author. This happened in an incognito window with no personal identifying information. That raises privacy and profiling questions that go beyond simple detection.
ChatGPT Plus, Copilot, and Gemini all returned perfect scores. Each properly identified all five test blocks as human or AI. That outcome suggests that general-purpose chatbots can outperform specialized content detectors. The result is not entirely surprising. Chatbots are trained on massive amounts of text and can analyze style, coherence, and context in ways that narrow detectors may not match. But the finding also undermines the business case for many standalone detection services.
Grok, included because it performed well in an overall chatbot evaluation, failed the detection test with three out of five wrong. Like several other tools, it identified all writing blocks as human. That pattern suggests a bias toward the human label, which is safer for avoiding false accusations but less useful for catching undisclosed AI writing.
What the results mean for writers, educators, and editors
The latest round of testing reinforces a caution that has been repeated for years: no AI detector is reliable enough to be the sole judge of authorship. False positives are a persistent problem. Writing from non-native English speakers is often rated as AI-generated. Even hand-crafted content can be misclassified. In this run, one detector declared itself too uncertain to judge a human-written sample, while another declared it AI-written. The results varied wildly across systems.
For educators and editors, the temptation to use detectors as a quick solution is understandable. But false accusations can damage trust and careers. A detector score should be treated as one weak signal among many, not as proof. The better approach is to combine detection tools with conversations, drafts, version history, and a clear policy on AI disclosure. The definition of plagiarism remains the same whether the words come from another person or from a machine: if you pass them off as your own without credit, you have committed plagiarism.
For writers, the rise of AI detectors adds another layer of anxiety. Human style is not a fixed target. A concise, polished, or formulaic piece may look AI-generated to a flawed detector. The arms race between generators and detectors also means that any tool can become outdated quickly. A detector that scores 100% today may fall to 40% in six months, as several tools did in this round. The only reliable protection is transparency and a willingness to show the process behind the writing.
For the detector industry, the results are a warning. Marketing claims of 99% accuracy are easy to make and hard to verify. When a service that
Source:ZDNET News
