DecEptioner

Applies algorithmic transformations to AI-generated text to reduce detectability by commercial AI detection systems (likely Turnitin, GPTZero, Originality.ai). The mechanism appears to involve lexical substitution, syntactic restructuring, and stylistic variation patterns that preserve semantic meaning while altering statistical fingerprints that detection models rely on. Implementation likely uses pattern matching against known detection heuristics (n-gram distributions, perplexity signatures, entropy markers) and applies targeted modifications to degrade classifier confidence scores.

Processes multiple text passages or documents sequentially through the obfuscation pipeline, applying consistent transformation rules across a corpus while attempting to preserve domain-specific terminology, tone, and factual accuracy. The system likely maintains a transformation context or style profile to ensure coherence across batch operations, preventing inconsistent rewrites that would signal synthetic modification to human readers or statistical analysis tools.

Allows users to specify which AI detection systems they are trying to evade (e.g., GPTZero, Turnitin, Originality.ai, Copyleaks), and applies targeted transformation strategies optimized against each detector's known weaknesses or heuristics. Implementation likely maintains a database of detection model signatures, known false-positive triggers, and adversarial examples, then selects transformation rules that maximize evasion probability for the specified target detector.

Maintains stylistic attributes (formality level, vocabulary complexity, sentence structure patterns, domain-specific terminology, brand voice) while applying detection-evasion transformations. Implementation likely uses style embeddings or linguistic feature extraction to identify and preserve domain markers, then applies transformations only to statistical signatures that detection models rely on (n-gram distributions, perplexity, entropy) while leaving style-critical elements intact.

Provides users with estimated detection scores or confidence metrics indicating how likely the transformed text is to be flagged by target detection systems. Implementation likely integrates with or mimics detection model APIs (GPTZero, Originality.ai) to provide real-time feedback, or uses proxy metrics (perplexity, entropy, n-gram novelty) as detection risk indicators. Users can iteratively refine transformations based on feedback to optimize evasion probability.

Allows users to apply multiple transformation passes to the same content, with each pass further modifying the text to reduce detection risk or improve specific attributes. Implementation likely maintains transformation history and allows selective application of different transformation strategies in sequence, with detection scoring feedback between passes to guide optimization. Users can experiment with different transformation intensities and combinations to find optimal balance between evasion and quality.

Exposes transformation and detection-scoring capabilities via REST or GraphQL API, enabling integration into content pipelines, publishing workflows, or third-party applications. Implementation likely includes authentication (API keys), rate limiting, batch endpoint support, and webhook callbacks for asynchronous processing. Developers can programmatically submit content, specify transformation parameters, retrieve results, and integrate detection feedback into automated workflows.