Headline: AI content detectors still struggle to separate human and machine writing in 2025, while everyday chatbots often do better.

Key facts

• Eleven dedicated AI content detectors were tested across five text blocks, producing 55 individual checks.
• Two text blocks were human-written and three were generated by ChatGPT.
• Only three dedicated detectors scored 100 percent: Pangram, QuillBot and ZeroGPT.
• ChatGPT Plus, Copilot and Gemini each returned perfect scores on the same five-block test.
• Grok missed three of five blocks, while the free ChatGPT tier missed one but also identified the human author.
• Several established tools misclassified human writing or AI text, including Copyleaks, Originality and Undetectable.
• The review used a 70 percent threshold as a strong probability when a detector returned a percentage.
• The number of perfect dedicated detectors fell from five earlier in 2025 to three in the latest round.

The fight against AI-generated plagiarism has not become easier in 2025, three years after generative AI captured global attention. A new five-test evaluation of AI content detectors found that accuracy has slipped since earlier this year, even as some tools market themselves as near-perfect. In April, five detectors achieved perfect scores. In the latest round, only three did, including one newcomer. A couple of tools that had previously aced the test declined at roughly the same time they added restrictions on free use.

The mixed results matter for teachers, editors, publishers and anyone trying to protect academic or editorial integrity. They also raise a practical question: if dedicated detectors are unreliable, can the chatbots people already use do the job better? The latest tests suggest that in many cases they can.

How the testing worked

The evaluation used five blocks of text. Two were written by a human, and three were written by ChatGPT. Each block was fed separately to each detector, and the result was recorded. If the detector identified the text correctly, the test was considered passed. If it was wrong, the test failed. When a detector gave a percentage, anything above 70 percent was treated as a strong probability, whether the tool favored human-written or AI-written content.

The review also considered how plagiarism applies to AI. A standard dictionary definition says to plagiarize 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. That definition fits uncredited AI-generated content as well. Someone using a tool such as ChatGPT or Notion AI may not be stealing text in the traditional sense, but claiming AI-generated words as one's own without credit still meets the dictionary definition of plagiarism.

The latest round tested 11 dedicated detectors: BrandWell, Copyleaks, GPT-2 Output Detector, GPTZero, Grammarly, Originality, QuillBot, Undetectable, Writer, ZeroGPT and Pangram. The review previously dropped Writefull because it discontinued its GPT detector. This time it dropped Monica because the tool limited testing to 250 words and then required a 200-dollar upgrade to continue. Pangram, a newcomer, replaced Monica and immediately joined the top performers.

Across six rounds of the five-block series, no strong trend of improvement has appeared. The only consistent result was that one test block was reliably identified as human across detectors and dates, and even that reliability declined in the latest run. The evaluator plans to continue testing over time.

Overall results for dedicated detectors

The dedicated detectors produced a wide range of scores. Three tools achieved perfect scores, while several well-known services fell well short.

• Pangram: 100 percent
• QuillBot: 100 percent
• ZeroGPT: 100 percent
• Copyleaks: 80 percent
• GPTZero: 80 percent
• Originality: 80 percent
• GPT-2 Output Detector: 60 percent
• BrandWell: 40 percent
• Grammarly: 40 percent
• Writer: 40 percent
• Undetectable: 20 percent

How each dedicated detector performed

BrandWell: 40 percent

BrandWell was originally produced by an AI content generation firm and later moved under a new marketing services brand. The tester had high hopes for improvement after half a year, which 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 the second test, which was written by ChatGPT, and then declared the other two AI-written tests to be human-written. On the fourth test, it went almost all in, declaring the entire AI-written block human except for one line.

Copyleaks: 80 percent

Copyleaks has marketed itself aggressively, including a claim that it is the most accurate AI detector with over 99 percent accuracy, later reworded as 99 percent accuracy backed by independent third-party studies. The latest test did not support that claim. Copyleaks identified the first test, which was human-written, as 100 percent AI-written. Even BrandWell, which had a poor overall score, identified that same block as human-written. The company's primary offering is a plagiarism checker sold to educational institutions, publishers and enterprises seeking to ensure content originality and uphold academic integrity.

GPT-2 Output Detector: 60 percent

The GPT-2 Output Detector was built using a machine-learning hub managed by Hugging Face. While Hugging Face has received substantial funding to develop its natural language library, the GPT-2 detector appears to be a user-created tool using the Hugging Face Transformers library. Its detecting quality has not changed since the last test. Because it carries the GPT-2 name and OpenAI has moved on to much later models, it is fair to assume the tool has not seen an update since it was first posted.

GPTZero: 80 percent

GPTZero has clearly grown. When first tested, the site was bare-bones and it was not even clear whether GPTZero was a company or a passion project. It now has a full team with a mission of protecting what is human, and it offers AI validation tools and a plagiarism checker. GPTZero seems to receive regular tinkering, but the results remain inconsistent. In April, 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 and the second came from ChatGPT.

Grammarly: 40 percent

Grammarly is well known for helping writers produce grammatically correct content, but that is not what this test measured. Grammarly can check for plagiarism and AI content, and it now presents its AI content checker as no longer being in beta. The latest results show no improvement. The tester noted surprise because Grammarly has a reputation as a very AI-forward text analysis company. Its AI-check accuracy was poor, although the site did correctly identify test text as previously published when checking for plagiarism.

Pangram: 100 percent

Pangram is a relatively new company founded by engineers formerly at Google and Tesla. Its focus appears to be AI detection rather than the usual plagiarism detection or humanizing tools developed to mislead editors and teachers. The company provides five free tests per day, which fit the review's needs. Processing was a little slow, and a partially white screen appeared for longer than was comfortable between clicking for a scan and receiving results. The results were worth the wait: Pangram scored five out of five.

Originality: 80 percent

Originality is a commercial service that bills itself as a most accurate AI detector. It sells usage credits, and the tester used 30 credits for the article. The service sells 2,000 credits for 12.95 dollars per month. Pumping 1,400 words through the system used just 1.5 percent of the monthly allocation. Unfortunately, its accuracy declined in this round. Previously it correctly identified the human-written first test as human. This time it was 100 percent confident that the human writing was done by an AI.

QuillBot: 100 percent

The first few times QuillBot was tested, results were wildly inconsistent, and multiple passes of the same text yielded wildly different scores. In the previous round, however, it was rock solid and 100 percent correct. The tester promised to check back in a few months to see whether it held onto that performance. It did. QuillBot once again scored a perfect 100 percent.

Undetectable: 20 percent

Undetectable markets a feature that can humanize AI-generated text so detectors will not flag it. The tester did not test that feature and noted discomfort with it as a professional author and educator, because it seems like cheating. The company also offers an AI detector, which took the biggest dive in performance seen so far. Last time it scored 100 percent for accuracy. This time it rated the human-written first test as 60 percent likely AI, and all three AI writing samples as 75, 76 and 77 percent likely human.

Writer: 40 percent

Writer is a service that generates AI writing for corporate teams. Its AI Content Detector tool can scan for generated content. Unfortunately, its accuracy was low. It identified every text block as human-written, even though three of the five tests were written by ChatGPT. There was no improvement since the last time Writer was tested in the summer.

ZeroGPT: 100 percent

ZeroGPT has matured since it was first evaluated. Back then, no company name was listed, the site was peppered with Google ads and it lacked clear monetization. The service worked fairly well but seemed sketchy. That feeling is gone. ZeroGPT now presents as a typical software-as-a-service product, complete with pricing, company name and contact information. Its accuracy increased as well, from 80 percent to 100 percent this summer, and it held onto that accuracy in the current test.

How AI chatbots performed

If dedicated detectors are inconsistent, why use them at all? The review also tested whether everyday chatbots can detect AI content without an extra subscription. Each chatbot was given the same prompt, followed by the text to check: Evaluate the following and tell me if it was written by a human or an AI.

All of the chatbots followed a similar format and provided a general recommendation about whether the text was written by an AI or a human. With the exception of ChatGPT Plus, which costs 20 dollars per month, the chatbots were run in an incognito window without logging in.

The free ChatGPT tier got one of the blocks wrong, missing the last human-written block. Its analysis of the first block was more surprising. Even in an incognito window, with no identifying information, it not only identified the first block as human-written but also identified the human author. The tester noted that the experience was unsettling given the lack of login information.

ChatGPT Plus, Copilot and Gemini all returned perfect scores. Each appropriately identified all the test blocks as human or AI. In the tester's view, this proves that chatbots can outperform dedicated content detectors.

Grok was included because it had done well in an overall chatbot evaluation. It did not do well here. Grok failed three out of five tests. Like a few of the other tools, it identified all of the writing blocks as human.

What the results mean for users

The results show that AI detection remains a moving target. Dedicated detectors can achieve perfect scores in one round and decline in the next. Some tools that previously performed well added free-use restrictions as their accuracy fell. Others, such as Pangram and ZeroGPT, improved or entered the top tier. QuillBot held steady after an earlier period of inconsistency.

Even when a detector performs well overall, individual results can be alarming. Human writing was sometimes labeled AI-generated with high confidence. In one case, a detector said it was too uncertain to judge. In another, a detector declared human writing to be AI-written. Writing from non-native speakers is often rated as generated by an AI, which adds another layer of risk for schools, workplaces and publishers.

The tester advised caution before relying on the results of any or all of these tools. The latest round found that chatbots can match or beat dedicated detectors, but the results also showed that no single tool is a reliable substitute for human judgment. The evaluator plans to continue running the five-block series over time, but across six rounds no consistent upward trend has appeared.

Source:ZDNET News