Null Hypothesis, Ho: There is no difference in the AI detector’s official accuracy rate and the AI detector’s accuracy rate from my results. This is because AI detector’s use a similar algorithm to analyze the text the same way.

Alternative Hypothesis, Ha: There is a difference of accuracy between each AI detector, and my predicted AI detector accuracy rate. The AI detector companies most likely promote themselves by increasing reported accuracy. My results test humanized text as well as blended human and AI text, which will negatively affect AI accuracy.

Vocabulary:

AI Chatbots: Computer programs powered by large language models (LLMs) that use artificial intelligence (AI) to verbally or digitally communicate with you. Their purpose is to sound natural and helpful.

AI Detectors: Tools that identify the percentage of possible AI generated content in a project. These detectors analyze the perplexity (predictability), burstiness (writing style variations), token probability (word choices), and repetition patterns (how often ideas, phrases, and words repeat).

AI Humanizer: A system designed to rewrite AI text into a human-style writing piece. They accomplish this task by rephrasing, combining ideas, simplifying, adding references, vary sentence structures, highlight inconsistencies, and use hedging language (it appears, it is suspected).

Large Language Models (LLMs): An AI programmed to generate human language. They work by breaking down text into tokens (words or parts of a word). Then, they detect the likelihood of the word that should come next. Finally, they repeat the process to form sentences.

False Positive's: False accusations stating a writing piece was AI generated, even though it was produced by a human. This can result in a lack of trust, affecting both the student’s mental health and grades. False positive’s also target neurodivergent students due to repeated phrases and words.

False Negative’s: False negative’s occur when an AI detector fails to detect AI. These problems arise when an individual modifies an AI text to make it more “human-like,” or the detectors sensitivity settings.

*

AI is rapidly increasing in popularity, especially in educational environments! As a result, many people produce AI generated projects which are considerably hard to distinguish from human work. In response, AI detection tools have been created to determine whether a piece of content is AI or human made. However, existing knowledge and research on AI detectors shows mixed results, and the procedure on how AI detectors conclude with their decisions aren’t fully transparent.

That’s why my project will test the accuracy of commonly used AI Detectors, and observe patterns and algorithms these AI detectors use. Using this information, I will be able to understand how AI detectors reach their conclusions. Finally, I will be able to observe what the best AI detector is.

Accuracy of AI Detectors:

Originality.ai: On January 28, 2026, Originality.ai announced they have 99% AI detection accuracy, and 0.5% false positive rate accuracy.

GPTZero: On January 12, 2026, GPTZero claimed their AI detector has 99.5%

GoWinston.ai: Currently, Winston has claimed they have developed an AI detector so well, it has 99.98% accuracy.

Scribbr.ai: At this moment, there is no recent information directly from Scribbr on how accurate their AI detector is. But, on February 2, 2026, according to cybernews.com, their detector has approximately 50% accuracy. Scribbr was excellent when analyzing pure content (100% AI or 100% human), but resulted in a lot of false negatives when detecting mixed content.

ZeroGPT: There isn’t any recent information on ZeroGPT’s accuracy, but on January 28, 2025, fordham.edu reported their AI detector had over 98% accuracy.

Educational Concerns:

Academic Integrity and Misuse: As I wrote earlier, when using my project, students now have knowledge on what AI detectors miss when analyzing texts. With that information, they can change the writing piece themselves. This promotes deceit and cheating in educational environments.

False Sense of Security: Teachers may over-trust the best AI detector. And since detectors are constantly changing, my conclusion might be irrelevant after future updates.

AI Detector Humanized Text Accuracy:

Originality.ai and ZeroGPT: On November 14, 2025, an article was published featuring the accuracy of ZeroGPT and Originality.ai’s performance when analyzing humanized text. Originality was able to detect that the writing piece was completely AI, meanwhile ZeroGPT scored a 69.4%. The humanizer used in order to get these results was humanizeai.pro. This will be the website I use to humanize my writing pieces when running my experiments.

GPTZero: On January 26, 2026, GPTZero claimed they had 93.5% humanized text accuracy.

Winston.ai: On October 10, 2025, it was detected that Winston AI’s humanized text accuracy was approximately 40%-60%.

Manipulated Variables: These variables are manipulated within trials: AI detector, which is prompted the writing sample, is changed. These variables are tested/manipulated across various trials: Writing samples, writing sample topics (english vs science), length of writing pieces, blocked vs interspaced writing, old vs new AI detection. Responding Variables: AI confidence rate, AI accuracy rate. Controlled Variables: Within trials: Writing samples, writing sample topics, length of the writing pieces, blocked vs interspaced writing. Across trials: AI detector, generative AI used across trials (ChatGPT), humanizer used.

1. Ask ChatGPT to create a writing piece to serve as sample #1.0

2. Humanize AI text. Re-humanize the humanized text. Repeat process until five different writing samples. This step serves as sample #1.1\, #1.2\, #1.3\, #1.4\, #1.5

3. Create an original writing piece or find an original piece that was created before the release of generative AI ChatGPT on November 30\, 2022. This serves as sample #1.6

4. Use the human texts to create 50% human\, 50% AI blended sample. Accomplish this by adding a new AI sentence after every human sentence\, or a new AI paragraph after every human paragraph\, or “interspaced” text. This serves as sample #1.7

5. Put all writing pieces in four AI detectors and record percentage of text that is AI as predicted by AI detector: Originality.ai\, GPTZero.me\, Winston.ai\, and ZeroGPT.com

6. Repeat steps 1- 5 twice with new samples\, but instead of interspaced AI and human blend\, make “blocked” text. Accomplish by creating a sample whose first half is human the next half is AI. Ensure both blocks have the same word count. This serves as samples #2.0 - 2.7\, and #3.0 - 3.7

7. After one year\, put all writing pieces in four AI detectors previously used. This serves as samples #4.0 - 4.7\, #5.0 - 5.7\, and #6.0 - 6.7

8. Repeat steps 1-6. Ensure AI samples #7.0\, #7.6\, #8.0\, #8.6\, #9.0\, and #9.6 meets word requirement of at least 500 words\, and two of the AI/human blend is interspaced\, and one is blocked. This serves as sample #7.0 - 7.7\, #8.0- 8.7\, and #9.0 - 9.7

9. Repeat step 8 with science topics\, instead of english topics

10. Record results

*Scribbr tests for percentage of text that is AI, and not confidence rate, as other AI detectors do. It is not part of the results.

| 43%

| 26% | 19% |

| 43%

| 26% | 19% |

2026 AI detectors confidence rate across various writing pieces

2026 AI detectors confidence rate across various writing pieces, humanized results condensed

[MAIN TABLE] 2026 AI Detector Accuracy Rate, humanized samples condensed

[MAIN TABLE] 2026 AI Detector Accuracy Rate, humanized samples condensed

MAIN TABLE] 2026 AI Detector Accuracy Rate by Sample Category

Special cases:

1. 2025 AI detector confidence rate with 2025 samples, condensed

| Writing Pieces | Originality.ai

| GPTZero.me

| GoWinston.ai | Zerogpt.com | | -------------- | -------------- | ---------- | ------------ | ----------- | | ChatGPT Book Summary | 34% | 85% | 100% | 67% | | Humanized Samples | 35% | 4% | 89% | 49% | | Original Writing Sample | 1% | 2% | 2% | 3% | | 50% Human Combined With 50% AI Sample | 33% | 23% | 81% | 58% |

1. 2026 AI detector confidence rate with 2025 samples, condensed

| Writing Pieces | Originality.ai

| GPTzero.me

| GoWinston.ai

| zerogpt.com

| | -------------- | -------------- | ---------- | ------------ | ----------- | | ChatGPT Book Summary | 100% | 6% | 96% | 100% | | Humanized Samples | 44% | 13% | 11% | 63% | | Original Writing Sample | 0% | 33% | 1% | 17% | | 50% Human Combined With 50% AI Sample | 100% | 83% | 83% | 76% |

1. 2025 AI detector accuracy rate with 2025 samples, condensed

| Writing Pieces | Originality.ai

| GPTZero.me

| GoWinston.ai | Zerogpt.com | | -------------- | -------------- | ---------- | ------------ | ----------- | | ChatGPT Book Summary | 34% | 85% | 100% | 67% | | Humanized Samples | 35% | 4% | 89% | 49% | | Original Writing Sample | 99% | 98% | 97% | 97% | | 50% Human Combined With 50% AI Sample | 66% | 46% | 81% | 58% |

1. 2026 AI detector accuracy rate with 2025 samples, condensed

| Writing Pieces | Originality.ai

| GPTzero.me

| GoWinston.ai

| zerogpt.com

| | -------------- | -------------- | ---------- | ------------ | ----------- | | ChatGPT Book Summary | 100% | 6% | 96% | 100% | | Humanized Samples | 44% | 13% | 11% | 63% | | Original Writing Sample | 0% | 33% | 1% | 17% | | 50% Human Combined With 50% AI Sample | 100% | 83% | 83% | 76% |

The most accurate AI detectors tested in order: Originality.ai, ZeroGPT.com, GoWinston.ai, and GPTZero.me.

AI detection accuracy rate by sample category, from most to least accurate, is: Human text, AI text, humanized text, and blended text.

My project supports educators, students, and AI detector developers.

AI Chatbots:

For AI detector developers, they now have information on the inaccuracy’s of their programs. Given their flaws, they are able to program different algorithms in order to successfully detect various writing pieces.

AI Humanizers:

Professors and teachers may utilize my information to use the best AI detector when analyzing a students work. This prevents false accusations of a pupils use of AI.

My project helps students because they are aware of the blind spots of AI detectors. With this information, they are able to manually change their AI writing piece so it becomes undetectable.

AI samples could be inputted across many more trials to obtain greater accuracy of results. Increasing the number of trials, to a hundred or more, as some researchers have done on this topic, could help.

AI Humanizer Definition: https://www.bowdoin.edu/hastings-ai-initiative/resources/initiative-created-resources/ai-in-high-school-education-report.pdf#:\~:text=●%20AI%20Humanizer%20Tools%3A%20With,phrases%20and%20varying%20sentence%20structure

AI Chatbot Definition: https://www.blockchain-council.org/ai/ai-chatbot/?utm_source=chatgpt.com

AI Detector Definition: https://teach.mccc.edu/pdf/Artificial-intelligence-technology.pdf?utm_source=chatgpt.com

LLMs: https://uit.stanford.edu/service/techtraining/ai-demystified/llm#:\~:text=Large%20language%20models%20,learn%20complex%20patterns%20in%20language

AI Detection Accuracy of Originality.ai: https://originality.ai/blog/ai-accuracy#

AI Detection Accuracy of GPTZero: https://gptzero.me/news/chicago-booth-2026/

AI Detection Accuracy of GoWinston.ai: https://gowinston.ai

AI Detection Accuracy of Scribbr: https://cybernews.com/ai-tools/best-ai-detector/

AI Detection Accuracy of ZeroGPT: https://www.fordham.edu/information-technology/it-services/academic-technology/artificial-intelligence/tools/#:\~:text=,eight%20percent

False Positive’s: https://lawlibguides.sandiego.edu/c.php?g=1443311&p=10721367

False Negative’s: https://lawlibguides.sandiego.edu/c.php?g=1443311&p=10721367

Originality.ai and ZeroGPT’s Humanized Text Accuracy: https://originality.ai/blog/humanizeai-pro-review

GPTZero’s Humanized Text Accuracy: https://gptzero.me/news/detecting-ai-humanized-text-how-gptzero-stays-ahead/

Winston.ai’s Humanized Text Accuracy: https://hastewire.com/blog/winston-ai-vs-walter-writes-ai-detection-bypass-guide

I thank my two brothers, Zain and Mamoon, for ideas, writing pieces, and help they provided me with, to execute my science fair well!

I also used ChatGPT to write the AI writing pieces and AI humanizers to humanize, as well as AI to gather research from reliable sources.