GPTZero：识别AI生成文本的利器

import mathimport refrom typing import List, Dict class GPTZeroDetector:    def __init__(self, perplexity_threshold: float = 50.0,                 burstiness_threshold: float = 0.5):        self.ppl_threshold = perplexity_threshold        self.burst_threshold = burstiness_threshold     def tokenize(self, text: str) -> List[str]:        return text.split()     def split_sentences(self, text: str) -> List[str]:        sentences = re.split(r'[.!?\u3002\uff01\uff1f]+', text)        return [s.strip() for s in sentences if s.strip()]     def compute_perplexity(self, text: str,                            token_probs: Dict[str, float]) -> float:        tokens = self.tokenize(text)        if not tokens:            return float('inf')        log_sum = 0.0        for token in tokens:            prob = token_probs.get(token.lower(), 1e-6)            log_sum += math.log(prob)        avg_neg_log_prob = -log_sum / len(tokens)        perplexity = math.exp(avg_neg_log_prob)        return perplexity     def compute_burstiness(self, text: str) -> float:        sentences = self.split_sentences(text)        if len(sentences) < 2:            return 0.0        lengths = [len(self.tokenize(s)) for s in sentences]        mean_len = sum(lengths) / len(lengths)        if mean_len == 0:            return 0.0        variance = sum((l - mean_len) ** 2 for l in lengths) / len(lengths)        std_dev = math.sqrt(variance)        burstiness = std_dev / mean_len        return burstiness     def detect(self, text: str,               token_probs: Dict[str, float] = None) -> Dict:        ppl = self.compute_perplexity(text, token_probs or {})        burst = self.compute_burstiness(text)        ppl_score = min(1.0, max(0.0,                      (self.ppl_threshold - ppl) / self.ppl_threshold))        burst_score = min(1.0, max(0.0,                        (self.burst_threshold - burst) / self.burst_threshold))        ai_prob = 0.6 * ppl_score + 0.4 * burst_score        if ai_prob > 0.8:            verdict = "HIGH_AI"        elif ai_prob > 0.5:            verdict = "LIKELY_AI"        elif ai_prob > 0.3:            verdict = "UNCERTAIN"        else:            verdict = "LIKELY_HUMAN"        return {"perplexity": round(ppl, 2), "burstiness": round(burst, 4),                "ai_probability": round(ai_prob, 4), "verdict": verdict}

import mathimport refrom typing import List, Dict, Tuple  class GPTZeroDetector:    """GPTZero式AI文本检测器：基于困惑度和突发性"""     def __init__(self, perplexity_threshold: float = 50.0,                 burstiness_threshold: float = 0.5):        self.ppl_threshold = perplexity_threshold        self.burst_threshold = burstiness_threshold     def tokenize(self, text: str) -> List[str]:        return text.split()     def split_sentences(self, text: str) -> List[str]:        sentences = re.split(r'[.!?\u3002\uff01\uff1f]+', text)        return [s.strip() for s in sentences if s.strip()]     def compute_perplexity(self, text: str,                            token_probs: Dict[str, float]) -> float:        tokens = self.tokenize(text)        if not tokens:            return float('inf')        log_sum = 0.0        for token in tokens:            prob = token_probs.get(token.lower(), 1e-6)            log_sum += math.log(prob)        avg_neg_log_prob = -log_sum / len(tokens)        perplexity = math.exp(avg_neg_log_prob)        return perplexity     def compute_burstiness(self, text: str) -> float:        sentences = self.split_sentences(text)        if len(sentences) < 2:            return 0.0        lengths = [len(self.tokenize(s)) for s in sentences]        mean_len = sum(lengths) / len(lengths)        if mean_len == 0:            return 0.0        variance = sum((l - mean_len) ** 2 for l in lengths) / len(lengths)        std_dev = math.sqrt(variance)        burstiness = std_dev / mean_len        return burstiness     def detect(self, text: str,               token_probs: Dict[str, float] = None) -> Dict:        ppl = self.compute_perplexity(text, token_probs or {})        burst = self.compute_burstiness(text)        ppl_score = min(1.0, max(0.0,                      (self.ppl_threshold - ppl) / self.ppl_threshold))        burst_score = min(1.0, max(0.0,                        (self.burst_threshold - burst)                        / self.burst_threshold))        ai_prob = 0.6 * ppl_score + 0.4 * burst_score        if ai_prob > 0.8:            verdict = "HIGH_AI"        elif ai_prob > 0.5:            verdict = "LIKELY_AI"        elif ai_prob > 0.3:            verdict = "UNCERTAIN"        else:            verdict = "LIKELY_HUMAN"        return {            "perplexity": round(ppl, 2),            "burstiness": round(burst, 4),            "ai_probability": round(ai_prob, 4),            "verdict": verdict        }     def simulate_token_probs(self, text: str,                              is_ai: bool = True) -> Dict[str, float]:        tokens = self.tokenize(text)        unique = set(t.lower() for t in tokens)        probs = {}        if is_ai:            for i, t in enumerate(unique):                probs[t] = 0.3 + 0.5 * (1.0 / (1.0 + i * 0.1))        else:            import random            random.seed(42)            for t in unique:                probs[t] = random.uniform(0.01, 0.8)        return probs  if __name__ == "__main__":    detector = GPTZeroDetector()    ai_text = ("The analysis reveals that the implementation of "               "machine learning algorithms has significantly "               "improved the accuracy of predictive models. "               "Furthermore the results demonstrate that the "               "proposed approach outperforms existing methods.")    human_text = ("So here is the thing. I tried running the code "                  "last night — wow, it crashed! Three times. "                  "But then, suddenly, it worked. Weird, right?")    ai_probs = detector.simulate_token_probs(ai_text, is_ai=True)    human_probs = detector.simulate_token_probs(human_text, is_ai=False)    r1 = detector.detect(ai_text, ai_probs)    r2 = detector.detect(human_text, human_probs)    print(f"AI文本检测: {r1}")    print(f"人类文本检测: {r2}")