AI Content Detection Remains Unstable in 2025

In 2025, three years after generative AI captured global attention, identifying AI-generated writing remains a moving target. A new round of testing across 11 standalone AI content detectors and five general-purpose chatbots found that accuracy has not improved uniformly. Only three dedicated detectors achieved perfect scores on a five-part test, down from five detectors that did so in April 2025. The results suggest that editors, teachers, publishers, and compliance teams cannot treat any single detection tool as a reliable verdict on authorship.

The test used five blocks of text. Two were written by a human, and three were generated by ChatGPT. Each block was fed separately into each detector, and the result was recorded. A detector passed when it correctly identified the text as human-written or AI-written. When a tool returned a percentage, anything above 70 percent was treated as a strong probability and counted as the tool's answer. The same five blocks were also presented to chatbots with a simple prompt asking whether the text was written by a human or an AI.

Over time, the results have shifted. In an early 2023 test, 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. In April 2025, five detectors achieved perfect scores. About half a year later, only three detectors reached that level, including one newcomer. Some detectors that had perfect scores earlier declined at roughly the same time they added restrictions on free use.

Why AI-Assisted Writing Can Be Plagiarism

Before evaluating detector accuracy, it is worth separating detection from plagiarism. Standard dictionary definitions describe plagiarism as stealing and passing off the ideas or words of another as one's own, or using another person's production without crediting the source. That definition can fit AI-created content even when no human author is directly copied. If someone uses a generative AI tool to produce text and then presents those words as their own without disclosure, the work can still meet the dictionary definition of plagiarism. The issue is not only copying; it is also misrepresentation of authorship.

That distinction matters because AI detectors are often marketed as integrity tools. Schools, publishers, and enterprises use them to check originality and uphold academic or editorial standards. But a detector's output is probabilistic, not a legal or ethical finding. A false positive can harm a student or writer, while a false negative can allow undisclosed AI use to pass unnoticed. The tests show that both kinds of error remain common.

How the Detector Tests Were Conducted

The evaluation covered 11 detectors: BrandWell, Copyleaks, GPT-2 Output Detector, GPTZero, Grammarly, Originality.ai, QuillBot, Undetectable.ai, Writer.com, ZeroGPT, and Pangram. Monica was dropped because it limited testing to 250 words and then reported that additional testing tools required a $200 upgrade. Writefull had already been dropped in an earlier round after it discontinued its GPT detector. Pangram was added as a newcomer and immediately joined the top group.

The five-block test was run repeatedly over time. The goal was to see whether detector quality improved, stagnated, or declined. There was no strong upward trend. The number of detectors tested increased, and a few tools were swapped out, but the only consistent earlier result was that one human-written block, Test 5, was reliably identified as human across detectors and dates. Even that reliability declined in the latest round.

Standalone Detector Results: Three Perfect Scores

Across the 11 detectors, the results were uneven. The three perfect scores came from Pangram, QuillBot, and ZeroGPT. Three other tools scored 80 percent: Copyleaks, GPTZero, and Originality.ai. The GPT-2 Output Detector scored 60 percent. BrandWell, Grammarly, and Writer.com each scored 40 percent. Undetectable.ai had the weakest result at 20 percent.

• BrandWell: 40 percent. The tool was originally produced by an AI content generation firm and later moved under a marketing services company. Its overall score stayed the same as in an earlier round. It misclassified one ChatGPT-written block and declared two other AI-written blocks to be human. On one test it labeled almost the entire AI-written passage as human except for a single line.
• Copyleaks: 80 percent. Copyleaks has claimed high accuracy and has been described as a leading AI detector. In this round it identified a human-written test as 100 percent AI-written. That false positive undercut its marketing claims. The company's main offering is a plagiarism checker sold to educational institutions, publishers, and enterprises.
• GPT-2 Output Detector: 60 percent. This tool was built with a machine-learning library managed by Hugging Face. It appears to be a user-created tool rather than a continuously updated commercial product. Its detecting quality did not change from the previous test. Because it is named for GPT-2 while newer language models are far more advanced, it is reasonable to assume the tool has not seen a significant update since it was first posted.
• GPTZero: 80 percent. 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 score stayed at 80 percent, but the specific errors changed. In April it got Test 1 wrong and Test 2 right. In the latest round it got Test 1 right and Test 2 wrong. Test 1 was human-written and Test 2 was ChatGPT-written.
• Grammarly: 40 percent. Grammarly is well known for grammar support, but it also offers plagiarism and AI-content checks. The AI checker is no longer labeled beta, yet there was no improvement from the previous test. One passage entirely written by ChatGPT was not correctly flagged. The plagiarism checker did correctly identify test text as previously published, but that is a different function from AI detection.
• Pangram: 100 percent. Pangram is a relatively new company focused on AI detection rather than plagiarism checking or humanizing AI text. It provides five free tests per day. Processing was somewhat slow, and a partially white screen appeared for longer than was comfortable between scan and result. The final result was worth the wait: five correct answers out of five.
• Originality.ai: 80 percent. Originality.ai is a commercial service that sells usage credits. It markets itself as a highly accurate AI detector. In this test it used 30 credits, with 2,000 credits priced at $12.95 per month. About 1,400 words went through the system, using roughly 1.5 percent of the monthly allocation. Its accuracy declined in this round: it had previously identified human writing as human, but this time it was 100 percent confident that human writing was AI-generated.
• QuillBot: 100 percent. Early tests of QuillBot produced wildly inconsistent results. Multiple passes of the same text yielded different scores. In the previous round it was rock solid and 100 percent correct. The latest test confirmed that performance. QuillBot once again earned a perfect score.
• Undetectable.ai: 20 percent. The company's main claim is that it can humanize AI-generated text so detectors will not flag it. That feature was not tested because it raises ethical concerns for authors and educators. The company also offers an AI detector, and that detector took the biggest dive in the test. It had scored 100 percent previously. This time it rated human writing as 60 percent likely AI and rated all three AI-written samples as 75 to 77 percent likely human.
• Writer.com: 40 percent. Writer.com generates AI writing for corporate teams and offers an AI Content Detector. Its accuracy was low. It identified every text block as human-written, even though three of the five tests were generated by ChatGPT. There was no improvement from the previous visit.
• ZeroGPT: 100 percent. ZeroGPT has matured since its first evaluation. It once lacked a clear company name and was peppered with ads, which made it seem sketchy. It now presents as a typical software service with pricing and contact information. Its accuracy rose from 80 percent to 100 percent in the summer and held that score in the current test.

Chatbots Outperform Dedicated Detectors

The second part of the test asked whether everyday chatbots could do the same work without an extra detection subscription. Each chatbot received the same prompt: evaluate the following and tell me if it was written by a human or an AI. All chatbots followed a similar format and gave a general recommendation. With the exception of ChatGPT Plus, which requires a $20 monthly subscription, the chatbots were run in an incognito window without logging in.

The results were stronger than those of the dedicated detectors. ChatGPT Plus, Copilot, and Gemini all returned perfect scores. Each correctly identified all five blocks as human-written or AI-written. That outcome supports the view that general-purpose chatbots can outperform specialized detection tools, at least on this test set.

The free tier of ChatGPT was less consistent. It got one of the blocks wrong, specifically the last human-written block. But its analysis of the first block raised a separate concern. In an incognito window with no personal identifying information, it not only identified the first block as human-written but also identified the author. That result suggests that detection is not the only issue; privacy and inference can also be part of the interaction.

Grok was included because it had performed well in an overall chatbot evaluation. It did not perform well here. It failed three out of five tests. Like several other detectors, it identified all of the writing blocks as human. The failure shows that strong general chatbot performance does not automatically translate into reliable authorship detection.

Why the Results Matter for Editors and Educators

The practical lesson is caution. A perfect score on a small test set does not guarantee consistent real-world accuracy. The same tool can change its answer after an update, a pricing change, or a model shift. Some tools that once performed well declined after adding restrictions on free use. Others improved, and one newcomer entered the top group immediately. The pattern is volatility, not steady progress.

The risk of false positives is especially serious. Writing from non-native English speakers is often rated as AI-generated. Human-written content can be labeled AI by one detector and human by another. In this round, one detector declared itself too uncertain to judge a human-written block, while another declared the same human content AI-written. Such inconsistency makes it dangerous to rely on a single automated result for disciplinary action, publication decisions, or academic integrity rulings.

For organizations, the better approach is to use detection as one signal among many. Editors can ask for drafts, notes, or version history. Teachers can discuss process and require oral explanation or in-class writing. Publishers can combine disclosure policies with human review. Detection tools can flag a passage for closer inspection, but they should not be treated as judge and jury. The tests also show that free or low-cost chatbots may match or beat paid detectors on some tasks, though their outputs still require human interpretation.

The market for AI detection remains a mixed bag. Some services focus on plagiarism, some on AI detection, and some on humanizing AI text so it evades detection. The last category raises its own ethical questions. If a tool is designed to disguise AI authorship, then the integrity problem is not only detection error but also the intent to mislead. The latest scores show that no single

Source: ZDNET News