原則は三つです。第一に「問いの設定と最終判断は自分で行う」こと。何を問うかは価値観と文脈の理解に依存し、責任を伴う判断は委ねられません。第二に「学習が目的の作業は自分で行い、既に習熟した作業は委ねる」こと。認知科学では、努力を伴う想起や生成が記憶と理解を深めることが示されており、初学者が結論だけを受け取ると能力が育ちません。第三に「AIの出力を必ず自分の言葉で再構成する」こと。要約を読むだけでは理解した錯覚が生じます。情報収集、下書き、形式変換、反論の列挙はAIに委ね、問題の定義、優先順位付け、倫理的判断、最終文責は自分が担う、という線引きが成果と能力の両立に適しています。Three principles apply. First, keep problem framing and final judgment for yourself; what to ask depends on values and context, and accountable decisions cannot be delegated. Second, do the work yourself when learning is the goal, and delegate work you have already mastered; cognitive science shows that effortful retrieval and generation deepen memory and understanding, so novices who receive only conclusions never build capability. Third, always reconstruct AI output in your own words, because reading a summary creates an illusion of understanding. Delegate information gathering, first drafts, format conversion, and enumeration of counterarguments; retain problem definition, prioritization, ethical judgment, and final authorship. This division sustains both quality of output and personal competence.
認知科学の「望ましい困難（desirable difficulties）」の研究（Bjork ら）は、学習時に努力を要する処理の方が長期的な定着に優れることを示しています。また Sparrow らの2011年の研究は、情報が外部に保存されていると分かると人は内容より「どこにあるか」を記憶する傾向（グーグル効果）を報告しました。生成AIはこの外部化を思考のプロセスにまで広げるため、線引きなしに使うと、判断の根拠を説明できない状態に陥る危険があります。一方で人間の直観には系統的な弱点があり、AIによる反論の列挙は判断の質を高めます。線引きの本質は、AIを「答えを出す機械」ではなく「自分の思考を問い直す道具」として位置づけることにあります。Research on 'desirable difficulties' by Robert Bjork and colleagues shows that effortful processing during learning produces stronger long-term retention. Sparrow and colleagues' 2011 study on the 'Google effect' found that when people know information is stored externally, they remember where to find it rather than the content itself. Generative AI extends this offloading into the reasoning process, so using it without boundaries risks leaving people unable to explain the basis of their own decisions. At the same time, Kahneman's dual-process theory shows that human intuition has systematic weaknesses, and AI-generated counterarguments and blind-spot checks genuinely improve judgment. Mollick advocates human-in-the-loop collaboration with humans holding final responsibility. The essence of the line is treating AI as a tool for interrogating your thinking rather than a machine for producing answers.
「学習目的なら自分で行う」という原則は、納期に追われる実務では守りにくく、短期的な生産性が長期的な能力形成を常に押しのけます。また組織全体でAIへの依存が進むと、個人が線引きを守っても、周囲の速度に合わせる圧力で徐々に崩れます。何が「習熟済み」かの自己評価も過信に陥りやすい点が問題です。The principle of doing learning-oriented work yourself is hard to sustain under deadline pressure, where short-term productivity consistently crowds out long-term capability building. As organizations grow dependent on AI, individual boundaries erode under pressure to match everyone else's pace. Self-assessment of what counts as 'already mastered' is also prone to overconfidence.
週に一定の時間を「AIなしで考える」時間として確保し、重要な文書の骨子は自分で書いてからAIに推敲させる順序を習慣にします。チームでは「AIに委ねた部分」と「自分で判断した部分」を成果物に明記し、レビューで判断の根拠を口頭で説明する機会を設けることで、理解の錯覚を定期的に検出します。Reserve a fixed block of time each week for thinking without AI, and make it a habit to write the skeleton of important documents yourself before letting AI refine them. In teams, mark which parts of a deliverable were delegated to AI and which were human judgments, and use reviews where authors explain their reasoning aloud to detect illusions of understanding regularly.