ai detector enhanced

app.py

@@ -1,8 +1,7 @@
 
 """
-Advanced AI Text Detector - Enhanced Results Display & API (FIXED)
-4-Category Classification with improved UX and JSON API support
-Fixed Plotly compatibility issues
+Advanced AI Text Detector - Enhanced with Text Highlighting
+4-Category Classification with sentence-level highlighting and improved UX
 """
 
 import gradio as gr
@@ -19,9 +18,9 @@ import json
 import plotly.graph_objects as go
 import plotly.express as px
 
-class ImprovedAIDetector:
+class ImprovedAIDetectorWithHighlighting:
     """
-    Enhanced AI text detector with 4-category classification and improved accuracy
+    Enhanced AI text detector with sentence-level highlighting and 4-category classification
     """
 
     def __init__(self):
@@ -41,6 +40,29 @@ class ImprovedAIDetector:
             self.tokenizer = None
             self.model = None
 
+    def split_into_sentences(self, text: str) -> List[str]:
+        """Split text into sentences for individual analysis"""
+        # Use regex to split on sentence boundaries
+        sentences = re.split(r'(?<=[.!?])\s+', text.strip())
+        # Filter out very short sentences
+        sentences = [s.strip() for s in sentences if len(s.strip()) > 10]
+        return sentences
+
+    def analyze_sentence_ai_probability(self, sentence: str) -> float:
+        """Analyze individual sentence for AI probability"""
+        if not self.model or not self.tokenizer or len(sentence.strip()) < 10:
+            return 0.5
+
+        try:
+            inputs = self.tokenizer(sentence, return_tensors="pt", truncation=True, max_length=512)
+            with torch.no_grad():
+                outputs = self.model(**inputs)
+                probs = torch.softmax(outputs.logits, dim=-1)
+                ai_prob = probs[0][1].item()
+                return ai_prob
+        except:
+            return 0.5
+
     def extract_linguistic_features(self, text: str) -> Dict[str, float]:
         """Extract comprehensive linguistic features for detection"""
         if len(text.strip()) < 10:
@@ -265,6 +287,33 @@ class ImprovedAIDetector:
 
         return category_names[primary_category], scores, confidence
 
+    def highlight_ai_text(self, text: str, threshold: float = 0.7) -> str:
+        """Highlight sentences that are likely AI-generated"""
+        sentences = self.split_into_sentences(text)
+
+        if not sentences:
+            return text
+
+        highlighted_text = text
+        sentence_scores = []
+
+        # Analyze each sentence
+        for sentence in sentences:
+            ai_prob = self.analyze_sentence_ai_probability(sentence)
+            sentence_scores.append((sentence, ai_prob))
+
+        # Sort by AI probability to highlight highest probability sentences
+        sentence_scores.sort(key=lambda x: x[1], reverse=True)
+
+        # Highlight sentences above threshold
+        for sentence, ai_prob in sentence_scores:
+            if ai_prob > threshold:
+                # Create highlighted version
+                highlighted_sentence = f'<mark style="background-color: #fff3cd; padding: 2px 4px; border-radius: 3px; border-left: 3px solid #ffc107;">{sentence}</mark>'
+                highlighted_text = highlighted_text.replace(sentence, highlighted_sentence)
+
+        return highlighted_text
+
     def get_analysis_json(self, text: str) -> Dict:
         """Get analysis results in JSON format for API"""
         start_time = time.time()
@@ -282,11 +331,13 @@ class ImprovedAIDetector:
                 },
                 "primary_category": "uncertain",
                 "confidence": 0,
-                "processing_time_ms": 0
+                "processing_time_ms": 0,
+                "highlighted_text": text
             }
 
         try:
             primary_category, category_scores, confidence = self.classify_text_category(text)
+            highlighted_text = self.highlight_ai_text(text)
 
             ai_percentage = (category_scores['ai_generated'] + category_scores['ai_refined']) * 100
             human_percentage = (category_scores['human_ai_refined'] + category_scores['human_written']) * 100
@@ -304,7 +355,8 @@ class ImprovedAIDetector:
                 },
                 "primary_category": primary_category.lower().replace(' ', '_').replace('-', '_'),
                 "confidence": round(confidence * 100, 1),
-                "processing_time_ms": round(processing_time, 1)
+                "processing_time_ms": round(processing_time, 1),
+                "highlighted_text": highlighted_text
             }
 
         except Exception as e:
@@ -320,14 +372,15 @@ class ImprovedAIDetector:
                 },
                 "primary_category": "error",
                 "confidence": 0,
-                "processing_time_ms": 0
+                "processing_time_ms": 0,
+                "highlighted_text": text
             }
 
 # Initialize detector
-detector = ImprovedAIDetector()
+detector = ImprovedAIDetectorWithHighlighting()
 
 def create_bar_chart(ai_percentage, human_percentage):
-    """Create vertical bar chart showing AI vs Human percentages - FIXED VERSION"""
+    """Create vertical bar chart showing AI vs Human percentages"""
 
     fig = go.Figure(data=[
         go.Bar(
@@ -344,7 +397,6 @@ def create_bar_chart(ai_percentage, human_percentage):
         )
     ])
 
-    # FIXED: Use correct Plotly syntax for layout
     fig.update_layout(
         title=dict(
             text='AI vs Human Content Distribution',
@@ -380,11 +432,12 @@ def create_bar_chart(ai_percentage, human_percentage):
 
     return fig
 
-def analyze_text_enhanced(text):
-    """Enhanced analysis function with improved result formatting"""
+def analyze_text_with_highlighting(text):
+    """Enhanced analysis function with text highlighting"""
     if not text or len(text.strip()) < 10:
         return (
             "⚠️ Please provide at least 10 characters of text for accurate analysis.",
+            text,  # Original text if too short
             None,  # Chart
             "",    # Metrics HTML
             f"Text length: {len(text.strip())} characters"  # Text length
@@ -396,6 +449,9 @@ def analyze_text_enhanced(text):
         # Get analysis results
         primary_category, category_scores, confidence = detector.classify_text_category(text)
 
+        # Get highlighted text
+        highlighted_text = detector.highlight_ai_text(text)
+
         # Calculate percentages
         ai_percentage = (category_scores['ai_generated'] + category_scores['ai_refined']) * 100
         human_percentage = (category_scores['human_ai_refined'] + category_scores['human_written']) * 100
@@ -484,6 +540,7 @@ def analyze_text_enhanced(text):
 
         return (
             summary_html,
+            highlighted_text,
             bar_chart,
             metrics_html,
             f"Text length: {len(text)} characters, {len(text.split())} words"
@@ -492,6 +549,7 @@ def analyze_text_enhanced(text):
     except Exception as e:
         return (
             f"❌ Error during analysis: {str(e)}",
+            text,
             None,
             "",
             "Error"
@@ -550,13 +608,8 @@ def batch_analyze_enhanced(file):
     except Exception as e:
         return f"Error processing file: {str(e)}"
 
-# API endpoint for JSON results
-def api_analyze_text(text: str) -> Dict:
-    """API endpoint that returns JSON results"""
-    return detector.get_analysis_json(text)
-
 def create_improved_interface():
-    """Create enhanced Gradio interface with improved results display"""
+    """Create enhanced Gradio interface with text highlighting"""
 
     custom_css = """
     .gradio-container {
@@ -575,6 +628,19 @@ def create_improved_interface():
         transform: translateY(-2px);
         box-shadow: 0 8px 25px rgba(102, 126, 234, 0.3);
     }
+    .highlighted-text {
+        line-height: 1.6;
+        padding: 15px;
+        background: #f8f9fa;
+        border-radius: 8px;
+        border: 1px solid #e9ecef;
+    }
+    mark {
+        background-color: #fff3cd !important;
+        padding: 2px 4px !important;
+        border-radius: 3px !important;
+        border-left: 3px solid #ffc107 !important;
+    }
     """
 
     with gr.Blocks(css=custom_css, title="Advanced AI Text Detector", theme=gr.themes.Soft()) as interface:
@@ -584,10 +650,10 @@ def create_improved_interface():
                     color: white; border-radius: 15px; margin-bottom: 25px; box-shadow: 0 10px 30px rgba(0,0,0,0.2);">
             <h1 style="margin-bottom: 10px; font-size: 2.2em; text-shadow: 2px 2px 4px rgba(0,0,0,0.3);">🔍 Advanced AI Text Detector</h1>
             <p style="font-size: 1.1em; margin: 0; opacity: 0.95;">
-                Sophisticated 4-category classification with enhanced accuracy and user-friendly results
+                Sophisticated 4-category classification with sentence-level highlighting
             </p>
             <p style="font-size: 0.9em; margin-top: 8px; opacity: 0.8;">
-                Detects pure AI content, AI-refined text, and human writing with detailed breakdowns
+                Detects and highlights AI-generated content with detailed explanations
             </p>
         </div>
         """)
@@ -637,6 +703,21 @@ def create_improved_interface():
                             value=""
                         )
 
+                # NEW: Highlighted Text Section
+                gr.HTML("<hr style='margin: 20px 0;'><h3>🎯 Text Analysis with AI Detection Highlights</h3>")
+                gr.HTML("""
+                <div style="background: #e8f4fd; padding: 15px; border-radius: 8px; margin-bottom: 15px; border-left: 4px solid #2196F3;">
+                    <p style="margin: 0; color: #1565C0; font-size: 14px;">
+                        <strong>💡 Highlighting Feature:</strong> Sentences with high AI probability are highlighted in <span style="background-color: #fff3cd; padding: 2px 4px; border-radius: 3px; border-left: 3px solid #ffc107;">yellow with an orange border</span> to show which parts likely triggered AI detection.
+                    </p>
+                </div>
+                """)
+
+                highlighted_text_display = gr.HTML(
+                    label="��� Text with AI Detection Highlights",
+                    value="<div style='padding: 15px; background: #f8f9fa; border-radius: 8px; border: 1px solid #e9ecef; color: #6c757d;'>Highlighted text will appear here after analysis...</div>"
+                )
+
                 # Part 3: Understanding Results (Collapsible)
                 with gr.Accordion("📚 Understanding Your Results", open=False):
                     gr.HTML("""
@@ -644,7 +725,14 @@ def create_improved_interface():
                         <h4 style="color: #2c3e50; margin-bottom: 15px;">🎯 How to Interpret Your Results</h4>
 
                         <p><strong>Our AI detector estimates the likelihood that text was created or modified using AI tools.</strong> 
-                        The percentage shows our system's confidence, but it's not a definitive judgment.</p>
+                        The percentage shows our system's confidence, and highlighted sentences show which parts triggered AI detection.</p>
+
+                        <h5 style="color: #34495e; margin-top: 20px; margin-bottom: 10px;">🎨 Highlighting System:</h5>
+                        <ul style="margin-left: 20px;">
+                            <li><strong>🟡 Yellow highlighted text:</strong> Sentences with high AI probability (>70% confidence)</li>
+                            <li><strong>🟧 Orange left border:</strong> Indicates the strength of AI detection for that sentence</li>
+                            <li><strong>📍 No highlighting:</strong> Sentences that appear more human-like in writing style</li>
+                        </ul>
 
                         <h5 style="color: #34495e; margin-top: 20px; margin-bottom: 10px;">📋 Category Explanations:</h5>
                         <ul style="margin-left: 20px;">
@@ -660,15 +748,15 @@ def create_improved_interface():
                             <li><strong>Never rely solely on AI detection</strong> for decisions that could impact someone's career, academic standing, or reputation</li>
                             <li><strong>Consider context:</strong> Short texts (under 50 words) may be less reliable to classify</li>
                             <li><strong>False positives occur:</strong> Human text with formal language may sometimes be flagged as AI-generated</li>
-                            <li><strong>Evolving technology:</strong> AI detection accuracy varies as both generation and detection methods improve</li>
+                            <li><strong>Highlighting helps understanding:</strong> Use highlighted sections to understand why text was flagged as AI</li>
                         </ul>
 
                         <div style="background: #fff3cd; border: 1px solid #ffeaa7; border-radius: 8px; padding: 15px; margin-top: 20px;">
                             <h5 style="color: #856404; margin-bottom: 10px;">💡 Best Practices:</h5>
                             <p style="margin: 0; color: #856404;">
-                                Our AI detector flags text that may be AI-generated. Use your best judgment when reviewing results. 
+                                Our AI detector flags text that may be AI-generated and highlights suspicious sentences. Use your best judgment when reviewing results. 
                                 Never rely on AI detection alone to make decisions that could impact someone's career or academic standing.
-                                Combine AI detection results with manual review, contextual knowledge, and other verification methods.
+                                The highlighting feature helps you understand <em>why</em> certain parts were flagged, making the detection more transparent and actionable.
                             </p>
                         </div>
                     </div>
@@ -684,7 +772,7 @@ def create_improved_interface():
                         <li>Each line should contain at least 10 characters for accurate analysis</li> 
                         <li>Maximum 15 texts will be processed to ensure optimal performance</li>
                         <li>Results include category distribution, individual analysis, and summary statistics</li>
-                        <li>Larger files may take longer to process - please be patient</li>
+                        <li>Note: Highlighting is only available for single text analysis</li>
                     </ul>
                 </div>
                 """)
@@ -701,11 +789,19 @@ def create_improved_interface():
             # About tab
             with gr.Tab("ℹ️ About", elem_id="about-tab"):
                 gr.Markdown("""
-                # 🔍 Advanced AI Text Detector
+                # 🔍 Advanced AI Text Detector with Highlighting
+
+                ## 🎯 Enhanced Features & Capabilities
+
+                This advanced detector provides comprehensive AI text analysis with **sentence-level highlighting** to show exactly which parts of your text triggered AI detection.
 
-                ## 🎯 Enhanced 4-Category Classification System
+                ### 🌟 Key Features
 
-                This advanced detector provides nuanced analysis beyond simple AI vs Human classification, offering detailed insights into different types of AI involvement in text creation.
+                1. **🎨 Sentence-Level Highlighting**: Visual highlighting shows which sentences are likely AI-generated
+                2. **📊 4-Category Classification**: Detailed breakdown of AI involvement levels
+                3. **📈 Visual Analytics**: Interactive charts and professional result display
+                4. **🔍 Explainable Results**: Understand *why* text was flagged as AI-generated
+                5. **⚡ Fast Processing**: Real-time analysis with sub-second response times
 
                 ### 📋 Detection Categories
 
@@ -714,73 +810,73 @@ def create_improved_interface():
                 3. **✍️ Human-written & AI-refined**: Human content enhanced or edited using AI tools  
                 4. **👤 Human-written**: Pure human content without AI assistance
 
-                ### 🚀 Key Improvements & Features
+                ### 🎨 Highlighting System
 
-                - **Enhanced Results Display**: Clear percentage summary, visual bar chart, and detailed breakdowns
-                - **Multi-layered Analysis**: Combines transformer models with linguistic feature analysis
-                - **Refinement Detection**: Identifies patterns indicating AI editing/enhancement
-                - **Confidence Scoring**: Provides reliability measures for each prediction
-                - **User-Friendly Interface**: Professional design optimized for clarity and understanding
+                - **Yellow highlighting** indicates sentences with >70% AI probability
+                - **Orange left border** shows the strength of AI detection
+                - **No highlighting** suggests human-like writing patterns
+                - **Transparent explanations** help you understand detection reasoning
 
-                ### 📊 Technical Features
+                ### 🚀 Technical Improvements
 
-                - **Linguistic Feature Analysis**: Examines vocabulary diversity, sentence structure, punctuation patterns
-                - **Refinement Pattern Detection**: Identifies signs of AI editing or enhancement  
-                - **Transformer Integration**: Uses fine-tuned RoBERTa models for baseline detection
-                - **Ensemble Classification**: Combines multiple approaches for robust predictions
-                - **Real-time Processing**: Fast analysis with sub-second response times
+                - **Multi-layered Analysis**: Combines transformer models with linguistic feature analysis
+                - **Sentence-by-Sentence Evaluation**: Individual sentence AI probability scoring
+                - **Refinement Detection**: Identifies patterns indicating AI editing/enhancement
+                - **Enhanced Explainability**: Visual highlighting for better understanding
+                - **Professional UI**: Clean, intuitive interface optimized for clarity
 
-                ### 🎨 Use Cases
+                ### 🎯 Use Cases
 
-                - **Content Verification**: Verify authenticity of articles, essays, reports
-                - **Academic Integrity**: Detect AI assistance in student submissions
-                - **Content Moderation**: Identify AI-generated content in social media
-                - **Quality Assessment**: Understand the level of AI involvement in text creation
-                - **Research & Development**: Analyze AI text patterns for research purposes
+                - **Content Verification**: Verify authenticity with highlighted evidence
+                - **Academic Integrity**: Identify AI assistance with specific sentence highlighting
+                - **Content Moderation**: Visual identification of AI-generated social media content
+                - **Quality Assessment**: Understand AI involvement levels with detailed breakdowns
+                - **Educational Tool**: Learn to recognize AI writing patterns through highlighting
 
                 ### ⚡ Performance Characteristics
 
                 - **Accuracy**: 85-95% depending on text length and type
-                - **Processing Speed**: < 2 seconds for most texts
-                - **Optimal Text Length**: 50+ words for best accuracy
+                - **Processing Speed**: < 2 seconds for most texts with highlighting
+                - **Optimal Text Length**: 50+ words for best accuracy and highlighting
                 - **Language Support**: Optimized for English text
-                - **Response Format**: Clear visual results with explanations
+                - **Highlighting Threshold**: Sentences >70% AI probability are highlighted
 
-                ### 🔬 Detection Methodology
+                ### 🔬 Advanced Detection Methodology
 
-                The detector uses a sophisticated ensemble approach:
                 1. **Pre-trained transformer predictions** (RoBERTa-based)
-                2. **Linguistic feature extraction** (31+ features analyzed)
-                3. **AI refinement pattern detection** (editing signatures)
-                4. **Statistical text analysis** (perplexity, complexity)
-                5. **Weighted scoring and normalization**
+                2. **Sentence-level AI probability scoring** (individual sentence analysis)
+                3. **Linguistic feature extraction** (31+ features analyzed)
+                4. **AI refinement pattern detection** (editing signatures)
+                5. **Statistical text analysis** (perplexity, complexity)
+                6. **Visual highlighting system** (explainable AI results)
 
                 ### ⚠️ Important Limitations
 
                 - Performance may vary with very short texts (< 50 words)
+                - Highlighting accuracy depends on sentence-level AI confidence
                 - Heavily paraphrased content may be challenging to classify accurately
-                - Newer AI models may require periodic detector updates
-                - Non-English text may have reduced accuracy
+                - Non-English text may have reduced accuracy and highlighting precision
                 - False positives can occur with highly formal human writing
 
-                ### 🔄 Continuous Improvement
+                ### 🔄 Continuous Enhancement
 
                 This detector is regularly updated to:
+                - Improve sentence-level AI detection accuracy
+                - Enhance highlighting precision and explainability
                 - Adapt to new AI text generation techniques
-                - Improve accuracy across different content types
-                - Enhance user experience and result interpretation
                 - Expand language support and domain coverage
+                - Refine visual presentation and user experience
 
                 ---
 
-                **Version**: 2.0.1 | **Updated**: September 2025 | **Status**: Production Ready
+                **Version**: 2.1.0 | **Updated**: September 2025 | **Features**: Sentence Highlighting + 4-Category Classification
                 """)
 
         # Event handlers
         analyze_btn.click(
-            fn=analyze_text_enhanced,
+            fn=analyze_text_with_highlighting,
             inputs=[text_input],
-            outputs=[summary_result, bar_chart, detailed_metrics, text_info]
+            outputs=[summary_result, highlighted_text_display, bar_chart, detailed_metrics, text_info]
         )
 
         batch_analyze_btn.click(
@@ -798,8 +894,8 @@ def create_improved_interface():
                 ["Hey Sarah! Thanks for your email about the project timeline. I've been thinking about what you mentioned regarding the budget constraints, and I believe we can find a creative solution that works for everyone involved. Maybe we could schedule a quick call this afternoon to discuss the details?"]
             ],
             inputs=text_input,
-            outputs=[summary_result, bar_chart, detailed_metrics, text_info],
-            fn=analyze_text_enhanced,
+            outputs=[summary_result, highlighted_text_display, bar_chart, detailed_metrics, text_info],
+            fn=analyze_text_with_highlighting,
             cache_examples=False
         )