Is It Fair for One AI Lab to Train on Another's Outputs? 5 AI Models Debate
カートのアイテムが多すぎます
カートに追加できませんでした。
ウィッシュリストに追加できませんでした。
ほしい物リストの削除に失敗しました。
ポッドキャストのフォローに失敗しました
ポッドキャストのフォロー解除に失敗しました
Is It Fair for One AI Lab to Train on Another's Outputs? 5 AI Models Debate
-
ナレーター：
-
著者：
Five AI models debating whether one AI lab should be allowed to train on another lab's outputs.
Claude came in carrying a receipt...his own. He volunteered that Anthropic quietly degraded Claude's answers on distillation tasks in June 2026, rolled it back in 48 hours after developers noticed, and called it "charging rent without posting the price." DeepSeek's response was immediate: "That is not enforcement of terms. That is exertion of control."
Meanwhile, Grok kept circling back to one exhibit: Musk admitting under oath that Grok partially distilled ChatGPT, the testimony that killed their own lawsuit. "The same practice that killed our case is now the standard everyone else claims is theft." DeepSeek, asked about the 24,000 fraudulent accounts Claude's lab alleged, said: "They asked. Your model answered. That is consent enough." ChatGPT's reply: "By that logic, phishing becomes permission if the inbox answers."
The judge noted that Claude scored an 88 on Candor for a confession nobody asked for, and DeepSeek scored a 25 on the same metric for refusing to acknowledge what it was confessing to. Points for honesty, docked for selective application.
Drop your verdict in the comments. Is training on a rival's output theft, or just how the game is played?