A 2025 review of AI content detectors found that accuracy remains inconsistent across tools, even as general-purpose chatbots increasingly outperform standalone checkers. The latest five-part test, repeated across 11 detectors, showed only three tools achieving perfect scores: Pangram, QuillBot, and ZeroGPT. Several established services slipped, and some added tighter free-use limits.

Key Facts at a Glance

• The test used five text blocks: two written by a human and three generated by ChatGPT.
• Each detector was fed each block separately; a result was counted as correct only if it matched the true source.
• When a tool returned a percentage, anything above 70% was treated as a strong probability and accepted as its answer.
• Eleven detectors were included in the final comparison: BrandWell, Copyleaks, GPT-2 Output Detector, GPTZero, Grammarly, Originality.ai, QuillBot, Undetectable.ai, Writer.com, ZeroGPT, and Pangram.
• Monica was dropped because its free test limited input to 250 words and then required a $200 upgrade for fuller testing. Writefull had already been dropped after discontinuing its GPT detector.
• Only Pangram, QuillBot, and ZeroGPT scored 100% in the individual results.
• ChatGPT Plus, Copilot, and Gemini returned perfect scores when asked to judge whether text was human or AI.
• ChatGPT free tier missed one human-written block but correctly identified another and even named its author, despite being used in an incognito window without login.
• Grok failed three of five tests, often labeling all samples as human-written.
• The results reinforce a central warning: no single detector should be trusted as definitive proof of authorship.

Why AI Plagiarism Is Still Hard to Police

The ethical issue is not only whether a machine wrote a sentence. It is whether a writer presents machine-generated words as original work without disclosure. Standard definitions of plagiarism include passing off another person’s ideas or words as one’s own. When a user asks an AI system to draft an article, report, essay, or memo and then claims the result as entirely personal work, that act can fit the same definition. The technology changes the method, but not the underlying principle of attribution.

Detection is difficult because AI writing is not a single fingerprint. Models generate text by predicting plausible sequences, and their output can resemble polished human prose. Detectors look for statistical patterns, perplexity, burstiness, token distributions, and other signals. Those signals shift when models are updated, when text is edited, or when a writer has a style that detectors associate with machine output. Short samples are especially vulnerable to error, and non-native English speakers are often misclassified as AI because their phrasing may be less predictable to the detector.

How the 2025 Test Worked

The latest comparison was a completely updated version of an earlier test first run in January 2023. At that time, the best result was 66% correct from one of three available checkers. By February 2025, the test had expanded to as many as 10 checkers, and three of them had perfect scores. In April 2025, just a couple of months later, five detectors boasted perfect scores. That improvement did not hold. About half a year later, only three content detectors achieved a perfect score, including one newcomer. A couple of the detectors that had previously aced the test declined in quality around the same time they added restrictions on free use.

The methodology was simple and repeatable. Five blocks of text were used. Two were written by a human, and three were written by ChatGPT. Each block was fed to each detector separately, and the result was recorded. If the detector correctly identified the source, the test was considered passed; if it was wrong, the test failed. When a detector provided a percentage, anything above 70% was treated as a strong probability, whether the tool was arguing for human-written or AI-written content, and that was considered the detector’s answer.

The test also tracked whether there was a tangible pattern of improvement over time. The five-test series has been run six times. There is no strong trend. The number of detectors tested has increased, and a few tools have been swapped out, but the only consistent result was that Test 5 was reliably identified as human across detectors and dates. Even that reliability declined in the latest run.

Overall Detector Results: A Market With Uneven Progress

The final detector set included BrandWell, Copyleaks, GPT-2 Output Detector, GPTZero, Grammarly, Originality.ai, QuillBot, Undetectable.ai, Writer.com, ZeroGPT, and Pangram. Monica was removed after its free tier limited testing to 250 words and then reported that fuller testing tools required a $200 upgrade. Writefull had been removed earlier because it discontinued its GPT detector. Pangram, a relative newcomer, immediately joined the winners’ circle.

The individual results were wide-ranging. BrandWell scored 40%. Copyleaks scored 80%. GPT-2 Output Detector scored 60%. GPTZero scored 80%. Grammarly scored 40%. Originality.ai scored 80%. QuillBot scored 100%. Undetectable.ai scored 20%. Writer.com scored 40%. ZeroGPT scored 100%. Pangram scored 100%. Those numbers show how fragile the category remains. A tool can be marketed as highly accurate and still label a human-written passage as machine-generated, or label a ChatGPT passage as human. The reverse also happens.

The results also show why a single test is not enough. One detector was too uncertain to judge a human-written sample, while another declared the same kind of writing AI-generated. The inconsistencies are not minor edge cases. They go to the core of how these systems are used in classrooms, newsrooms, publishing workflows, and corporate compliance reviews.

Detector-by-Detector Results

BrandWell: 40%

BrandWell was originally produced by an AI content generation firm and later moved under a new marketing-services name. Expectations were high because half a year is a long time in AI development. Instead, its overall score stayed the same, with only two of five tests correct. It was confused by Test 2, a ChatGPT-written sample, and then declared the other two AI-written tests to be human-written. For Test 4, it went almost all in, labeling the entire AI-written test as human except for one line.

Copyleaks: 80%

Copyleaks has publicly claimed very high accuracy, including a 99% accuracy claim backed by third-party studies. In this test, it identified Test 1, which was human-written, as 100% AI-written. That was a notable error because another detector, BrandWell, identified the same passage as human-written. The company’s primary offering is a plagiarism checker sold to educational institutions, publishers, and enterprises seeking to ensure originality and academic integrity.

GPT-2 Output Detector: 60%

This tool was built using a machine-learning hub managed by Hugging Face. The GPT-2 detector appears to be a user-created tool using the Hugging Face Transformers library. There has been no change in its detecting quality since the last test. Because it carries the GPT-2 name and newer OpenAI models have advanced far beyond that generation, it is reasonable to assume the tool has not seen a significant update 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. Its performance has been regularly adjusted, but the changes have not produced steady improvement. In April, it got Test 1 wrong and Test 2 right. In the latest run, it got Test 1 right and Test 2 wrong. Test 1 was human writing, and Test 2 came from ChatGPT.

Grammarly: 40%

Grammarly is best known for grammar correction, but it

Source:ZDNET News