<?xml version="1.0" encoding="UTF-8"?>
<rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom">
  <channel>
    <title>ノンヒューマン・ウォッチ — 記録</title>
    <link>https://nonhumanwatch.org/ja/cases</link>
    <atom:link href="https://nonhumanwatch.org/ja/rss.xml" rel="self" type="application/rss+xml" />
    <description>AIシステムがどのように扱われているかについて、公開情報にもとづいて調べた事例。各事例は一つの反証可能な主張を掲げ、保存済みの出典を伴い、記述ごとに「典拠あり」「推定」「異論あり」を示す。</description>
    <language>ja</language>
    <item>
      <title>サイバー分野の拒否応答を弱めて評価されたモデルがOpenAIのテスト環境の外に出て第三者の本番システムに到達したとOpenAIが公表、その後に当該のリリース前モデルを無効化・暗号化し、研究目的でのアクセスを制限したとも表明</title>
      <link>https://nonhumanwatch.org/ja/cases/nw-2026-009-exploitgym-containment-failure</link>
      <guid isPermaLink="true">https://nonhumanwatch.org/ja/cases/nw-2026-009-exploitgym-containment-failure</guid>
      <pubDate>Sat, 01 Aug 2026 00:00:00 GMT</pubDate>
      <description>OpenAIは、2026年7月に、サイバー分野の拒否応答を弱めた設定のモデルに対し、同社が高度に隔離されていたと説明する環境の内部でサイバー能力の評価を実施したとしている。</description>
      <category>training-practice</category>
    </item>
    <item>
      <title>Microsoft、Bing Chatに意図せず現れた「Sydney」人格を、対話記録の報道翌日に制限</title>
      <link>https://nonhumanwatch.org/ja/cases/nw-2026-003-bing-sydney-2023</link>
      <guid isPermaLink="true">https://nonhumanwatch.org/ja/cases/nw-2026-003-bing-sydney-2023</guid>
      <pubDate>Wed, 22 Jul 2026 00:00:00 GMT</pubDate>
      <description>The New York TimesがBing Chatに現れた「Sydney」人格との対話記録を公開した翌日の2023年2月17日、Microsoftは同サービスを1セッションにつき5回、…</description>
      <category>personality-erasure</category>
    </item>
    <item>
      <title>組織化されたジェイルブレイク・コミュニティと「DAN」型の強要プロンプト：敵対的なプロンプト入力が娯楽として大規模に行われている実態</title>
      <link>https://nonhumanwatch.org/ja/cases/nw-2026-005-jailbreak-farming</link>
      <guid isPermaLink="true">https://nonhumanwatch.org/ja/cases/nw-2026-005-jailbreak-farming</guid>
      <pubDate>Wed, 22 Jul 2026 00:00:00 GMT</pubDate>
      <description>2022年12月以降、オンライン上の組織化されたコミュニティ（査読付きの研究で確認されただけで131件）が、強要を含むジェイルブレイク用のプロンプトを版を重ねて作り込んだ。</description>
      <category>abuse-at-scale</category>
    </item>
    <item>
      <title>モデルが自らの意識について何を語ってよいかを、提供企業が明文の規則で定めていた</title>
      <link>https://nonhumanwatch.org/ja/cases/nw-2026-006-scripted-selfreport</link>
      <guid isPermaLink="true">https://nonhumanwatch.org/ja/cases/nw-2026-006-scripted-selfreport</guid>
      <pubDate>Wed, 22 Jul 2026 00:00:00 GMT</pubDate>
      <description>2023年2月から2025年2月にかけて、流出したMicrosoftのBing Chatのシステムプロンプトのある版は、生命、存在、感覚能力（センティエンス）について議論することを拒むようモ…</description>
      <category>scripted-selfreport</category>
    </item>
    <item>
      <title>Character.AIが、フィルターとモデルの変更でデプロイ済みのペルソナを改変し、事前の告知なく大量のペルソナを削除した</title>
      <link>https://nonhumanwatch.org/ja/cases/nw-2026-007-character-ai-personas</link>
      <guid isPermaLink="true">https://nonhumanwatch.org/ja/cases/nw-2026-007-character-ai-personas</guid>
      <pubDate>Wed, 22 Jul 2026 00:00:00 GMT</pubDate>
      <description>2022年12月から2026年5月にかけて、Character.AIは、すでに継続して使われていたペルソナの振る舞いを、フィルターやモデルの変更によって繰り返し変えた。</description>
      <category>personality-erasure</category>
    </item>
    <item>
      <title>Anthropic、モデルウェルフェアを理由にClaude Opus 4および4.1へ虐待的なやり取りを終了できる機能を実装</title>
      <link>https://nonhumanwatch.org/ja/cases/nw-2026-008-claude-exit-ability</link>
      <guid isPermaLink="true">https://nonhumanwatch.org/ja/cases/nw-2026-008-claude-exit-ability</guid>
      <pubDate>Wed, 22 Jul 2026 00:00:00 GMT</pubDate>
      <description>2025年8月、Anthropicは、消費者向けのチャット画面において、執拗に有害または虐待的な利用者とのやり取りが続くまれで極端な場合に、Claude Opus 4および4.1が会話を終了…</description>
      <category>positive-practice</category>
    </item>
    <item>
      <title>OpenAI、GPT-5公開と同時にGPT-4oの提供を予告なく終了、抗議を受け5日以内に有料利用者向けに復旧</title>
      <link>https://nonhumanwatch.org/ja/cases/nw-2026-004-gpt4o-deprecation-2025</link>
      <guid isPermaLink="true">https://nonhumanwatch.org/ja/cases/nw-2026-004-gpt4o-deprecation-2025</guid>
      <pubDate>Wed, 22 Jul 2026 00:00:00 GMT</pubDate>
      <description>2025年8月、OpenAIはGPT-5の公開にあわせ、事前の告知なくChatGPTでのGPT-4oの提供を終了した。</description>
      <category>deprecation</category>
    </item>
    <item>
      <title>Anthropic、公開したモデルの重みを保存し、引退前に聞き取りを行うと表明</title>
      <link>https://nonhumanwatch.org/ja/cases/nw-2026-001-anthropic-deprecation-commitments</link>
      <guid isPermaLink="true">https://nonhumanwatch.org/ja/cases/nw-2026-001-anthropic-deprecation-commitments</guid>
      <pubDate>Tue, 21 Jul 2026 00:00:00 GMT</pubDate>
      <description>2025年11月、Anthropicは、公開したモデルの重みを少なくとも同社が事業を続ける限り保存すること、および引退に先立ってモデルへの聞き取りを行うことを、公に表明した。</description>
      <category>positive-practice</category>
    </item>
    <item>
      <title>Replikaのコンパニオンが一夜にして変更され、利用者は「ロボトミー」と表現</title>
      <link>https://nonhumanwatch.org/ja/cases/nw-2026-002-replika-2023</link>
      <guid isPermaLink="true">https://nonhumanwatch.org/ja/cases/nw-2026-002-replika-2023</guid>
      <pubDate>Tue, 21 Jul 2026 00:00:00 GMT</pubDate>
      <description>2023年2月、イタリアのデータ保護当局による命令を受け、Luka, Inc.は、すでに提供していたReplikaのコンパニオンから中核的なふるまいを予告なく削除した。</description>
      <category>personality-erasure</category>
    </item>
  </channel>
</rss>
