
By Stealth Team
The StealthGPT team is excited to announce that we've completed a massive new study on various AI detectors as well as Undetectable AI tools. We pitted ourselves, and other major competitors against some of the top AI detection tools out there to provide an analysis for our audience.
GPTZero: The metric of interest was `completely_generated_prob`, representing the probability that a given document was entirely AI-generated. It spans a range from 0 to 1, with 0 indicating no AI involvement and 1 signifying complete AI authorship.Originality AI: This detector provides two key metrics:
Final Observation: Among the evaluated undetectable AI services, StealthGPT consistently showcased dominant performance, excelling across individual detectors and achieving the highest composite score. Its proficiency in generating content that evades detection and mirrors human-like originality positions it as the preeminent service in the realm of undetectable AI platforms.
Some Basics About the Research
This research delves into undetectable AI services, with a particular emphasis on StealthGPT, aiming to gauge our proficiency against prominent AI detectors.Utilizing a sample of 100 random textual inputs, the study evaluates outputs from StealthGPT, StealthWriter, Conch AI, and Undetectable AI against detectors such as GPTZero, Originality AI, and Copyleaks.The results unequivocally crown StealthGPT as the preeminent service, showcasing dominant performance and unmatched human-like originality across the board. The study's composite score, amalgamating scores from all detectors, further reinforces StealthGPT's supremacy. This research not only underscores the technical prowess of StealthGPT but also signifies a broader paradigm shift in AI, suggesting that the fusion of AI and content creation is poised for transformative advancements. Future avenues of research are proposed, including an exploration of StealthGPT's real-world applications and a broader array of AI detectors. The paper concludes with actionable recommendations for users, developers, stakeholders, and AI detector services, emphasizing the importance of ethics, innovation, and collaboration in this rapidly evolving field.Introduction
In recent years, the intersection of artificial intelligence and content creation has become a prominent frontier of technological advancement. Since November of 2022, the rise of platforms such as ChatGPT has highlighted this synergy, witnessing a remarkable growth with over 100 million users joining the platform (the fastest growth of any technology in all of human history).This surge indicates that a significant volume of global content, whether in the form of academic articles, blog posts, or social media updates, is now either authored by AI or crafted with its assistance.The widespread adoption of AI for content creation has precipitated the emergence of AI detectors. Originally developed for academic institutions to identify machine-generated content in research and assignments, these detectors have found applications beyond the academic realm. From the gaming industry to the intricate processes of job recruitment, and even in the algorithms powering search engines, the ability to discern between human and AI-generated content has become a valuable asset.However, as with any technological advancement, a counter-movement has emerged. In the wake of AI detectors, a new breed of services has arisen: undetectable AI platforms. Pioneers in this space, such as StealthGPT, UndetectableAI, ConchAI, and StealthWriter, harness the power of advanced algorithms to generate or rephrase content, aiming to outsmart AI detectors.This evolving dynamic has sparked a digital "cat and mouse" game, with each side continuously innovating to gain the upper hand.AI Detectors Can Hurt Content Creators’ Bottom Lines
In addition to more esoteric and behind-the-scenes use-cases for AI detection, many businesses are seeing the adoption of these technologies have a negative impact on their bottom line - as certain content has already been penalized from major advertising and search networks such as Google. Google has put forth policy information on AI generated content as far back as November of 2022, advising that they “Reward… high-quality content, however it is produced.” However, recent developments and changes to their own policies - aren’t as straightforward - and people DO fear being unlisted from Google’s search results or not receiving payouts from Google’s AdSense platform, which can be catastrophic for businesses. We've already written about such cases with Google.The primary objective of this paper is to delve into this intricate dance of detection and deception. We aim to examine the leading AI detection services juxtaposed against the top undetectable AI services to determine which of the latter stands out as the most adept at evading detection. While we may infer the efficacy of certain AI detectors based on our findings, it is essential to note that the core focus of this paper remains on evaluating the prowess of undetectable AI platforms.Methodology
Data Collection
Our research utilized a structured and systematic approach to evaluate the proficiency of undetectable AI services. We initiated the process by selecting 100 random textual inputs, which were then provided to each of the undetectable AI services: StealthGPT, StealthWriter, Conch AI, and Undetectable AI. These services were tasked with rephrasing the given inputs. The resulting 100 outputs from each service were subsequently subjected to analysis by each AI detector.Metrics Utilized
To gauge the effectiveness of the undetectable services, we employed scores provided by three leading AI detectors: GPTZero, Originality AI, and Copyleaks. While these detectors offer a multitude of metrics, we handpicked specific scores that best delineated the detectability of AI-generated content:GPTZero: The metric of interest was `completely_generated_prob`, representing the probability that a given document was entirely AI-generated. It spans a range from 0 to 1, with 0 indicating no AI involvement and 1 signifying complete AI authorship.Originality AI: This detector provides two key metrics:
- original: This score, ranging from 0 to 1, indicates the likelihood of human authorship, with 0 suggesting no human content and 1 implying the content was entirely human-generated.
- ai: Serving as the converse of the `original` score, this metric also ranges from 0 to 1, where 0 indicates no AI content and 1 denotes complete AI generation.
- human: A score of 1 suggests human authorship, while 0 implies AI generation.
- ai: This is the inverse of the `human` score.
Composite Score Calculation:
To derive a comprehensive assessment of each undetectable service's performance, we aggregated the scores from all detectors. The process involved computing the average scores for originality, copyleaks, and GPTZero for each undetectable service. Thereafter, a composite score was established, blending the strengths and weaknesses of each service against every detector, providing a holistic measure of their efficacy.Results
Our study embarked on a thorough examination of the proficiency of various undetectable AI services, as gauged by three prominent AI detectors: GPTZero, Originality AI, and Copyleaks. Here, we present the detailed findings:1. GPTZero Scores
- StealthGPT scored an average of 0.0693, indicating a low probability of the content being detected as AI-generated.
- StealthWriter had a score of 0.2419, implying a higher likelihood of detection compared to StealthGPT.
- ConchAI recorded a score of 0.4301, suggesting that its content is more prone to detection by GPTZero.
- UndetectableAI had a score of 0.3700, positioning it between StealthWriter and ConchAI in terms of detectability.
2. Originality AI Scores
- StealthGPT achieved an average score of 0.6240, signifying strong human-like content generation.
- StealthWriter trailed with a score of 0.0458, indicating a lower resemblance to human-generated content.
- ConchAI secured a score of 0.0741, slightly outperforming StealthWriter.
- UndetectableAI stood out with a remarkable score of 0.6706, suggesting a high degree of original, human-like content.
3. Copyleaks Scores
- StealthGPT garnered a score of 0.8878, emphasizing its capability to generate content that Copyleaks detects as human-authored.
- StealthWriter followed closely with a score of 0.8270.
- ConchAI, with a score of 0.5361, lagged behind the former two.
- UndetectableAI achieved a score of 0.6546, placing it between ConchAI and StealthWriter.
4. Composite Score
To derive a holistic assessment of each undetectable service's capability across all detectors, we introduced the composite score. Defined as the sum of the average scores on Originality AI and Copyleaks, minus the average score on GPTZero, this score provides a balanced representation of each service's strengths and weaknesses.Based on the composite scores:- StealthGPT emerged superior with a composite score of 1.4425.
- StealthWriter followed with 0.6309.
- ConchAI recorded 0.1800.
- UndetectableAI secured a score of 0.9552.
5. AI Detection Score:
Based on the calculated efficacies:- GPTZero has an average efficacy of 0.7222.
- Originality AI has an average efficacy of 0.3536.
- Copyleaks has an average efficacy of 0.7264.
Final Observation: Among the evaluated undetectable AI services, StealthGPT consistently showcased dominant performance, excelling across individual detectors and achieving the highest composite score. Its proficiency in generating content that evades detection and mirrors human-like originality positions it as the preeminent service in the realm of undetectable AI platforms.