Grok AI claims it’s unseen by AI detectors. A recent study shows surprising findings! Can you tell if the content is AI-made?
This brief analysis explores if the text crafted by Grok 4 can be spotted by top AI detectors or remains hidden.
Key Points (In Short):
Indeed, Grok 4’s content is identifiable by the highly accurate Bypass Engine AI Checker with perfect precision, based on early trials using Grok 4’s text examples.
- Grok 4 launched on July 10th, 2025 (check additional Grok statistics here).
- From a quick test of Grok 4’s content samples:
- Two AI detectors identified Grok 4’s AI-produced content every time, specifically the Bypass Engine AI Detector and GPTZero.
- Two other detectors, ZeroGPT and Grammarly, had lower success, with just 40% accuracy.
- In our detailed accuracy review, Originality demonstrated its capability to catch Grok 3’s AI-generated text with 99.9% reliability.
Here, we’ll delve deeper into the Grok AI model and evaluate the AI content detection tool’s skill in recognizing AI-written content.
Approach:
Initially, Grok 4 came up with ten prompts aimed at producing a text that could potentially evade AI detection (illustrated in the picture below).
Then, Grok 4 generated those ten text samples.
Each of these ten samples was evaluated using four different AI detection tools:
- Bypass Engine AI Detector
- GPTZero
- ZeroGPT
- Grammarly
Analysis:
Every Grok reply was checked using four AI detectors, with each AI score noted down.
In the image above, wrong guesses are marked in red.
Bypass Engine AI Detector:
Bypass Engine showed outstanding results, accurately tagging all 10 Grok content samples as AI.
In total, Bypass Engine’s success rate in spotting Grok 4 content as AI was 100%.
GPTZero
Much like Originality.ai, GPTZero accurately recognized all 10 samples as made by AI.
Check out a detailed review of GPTZero here.
ZeroGPT
ZeroGPT didn’t do well, correctly spotting only 40% of the Grok 4 samples as AI-made.
See a review of ZeroGPT here.
Grammarly AI Detector
The AI detector from Grammarly didn’t do well, spotting just 40% of the Grok 4 samples as made by AI.
Check out a review of the Grammarly AI Detector.
Grok 3 Identification
Following a study on AI detection accuracy for Bypass Engine’s newest model, here’s the accuracy for spotting Grok 3:
We created and shared a tool for AI detector research. It’s simple for researchers to use this tool to check how well AI detectors work with a dataset.
Summary
The results show that Grok 3 and Grok 4 AI text can be spotted by top AI detectors, like Bypass Engine’s AI Checker.
Yet, some other AI detectors find it hard to identify them.
Every time a new LLM model, such as Grok 4, comes out, we check how well our AI detector works. Usually, there’s a small drop in accuracy, but this improves once the detection models learn from the new LLM’s content.
Testing accuracy is straightforward. To see how good an AI detector is, use a confusion matrix and the F1 score. These are the best methods (you can learn more about detailed AI detection tests in our study).
The F1 score helps turn the Confusion Matrix into one number. Check out our open-source testing tool and follow the instructions to understand better. This study shows why these tests matter, as they help us talk about AI transparency. This way, we can keep learning and getting better together.
If you want to do your own study, let us know. We’re happy to provide research credits.
We’re also looking for people to join our charity challenge (testing AI detectors). If you’re interested, please contact us.
Further Exploration
Discover insights about detecting AI and how accurate these detections are:
- Review on AI Content Detector Accuracy
- Comprehensive Analysis of AI Detection Accuracy Studies
- Exploration of DeepSeek’s Similarity to ChatGPT: Detectability Insights
- Investigating the Detectability of Claude 4 Sonnet and Opus Content
- Assessing the Detectability of GPT-4.1 Content