【英語論文の書き方】第93回　「AI（人工知能）の危険性」について 論文翻訳・英文校正　研究者専門の翻訳会社　ワールド翻訳サービス

第92回では「誤解を招く指標を避けるために、何を、なぜ測っているのかを理解する」を取り上げました。
 
第93回（今回）のテーマは「AI（人工知能）の危険性」についてです。
 
人工知能（AI）は、フォーム入力や文献引用チェック、データ分析など、
煩雑で反復的な作業を効率化する可能性があります。

しかしGeoffさんは、こうした便利さの一方で、
AIにはいくつかの課題があると指摘しています。
たとえば、相関と因果関係を区別できなかったり、
回答に一貫性がなかったり、
剽窃の問題（表現の幅が限られる研究手法の記述などで、
検出ツールが誤って剽窃と判定することがある）や
事実の誤生成（ハルシネーション）、
結論の導出過程が不透明なブラックボックス問題、
データ品質のばらつきなどです。

さらに、AIができることは既存情報の組み合わせにとどまり、
革新的な発見につながる新たな解釈を見出すことは難しいとしています。
 
Geoffさんは、AIは人間の洞察力の代わりにはならないこと、
AIソフトウェアが提示する結果を厳密に検証する責任を持ち、
その使用に伴う倫理的影響を十分に認識しておくことが
重要だと述べています。
 
この記事が、AIの活用やリスクを理解するうえで
少しでも参考になれば幸いです。

    Artificial intelligence (AI) holds the promise of helping us deal with a variety of tedious and repetitive tasks, such as filling out forms, checking literature citations in a manuscript, or analyzing data. Unfortunately, many scientists, engineers, and other professionals are using it to replace humans in work that is intellectually, emotionally, and economically satisfying. This is clearly unethical: computers and their software should be our assistants, not our replacements. But careless use of AI also leads to many problems, some of which may be quite subtle and difficult to detect. Here, I’ll summarize the major problems with the “large language models” that are increasingly used for writing and editing.
    Mistaking correlation for causality
    AI software does not think and therefore cannot understand our writing; it only searches for correlations based on the frequency with which words appear close together. Even when these words are correlated (i.e., they appear together), the correlation may not be based on a causal relationship. Never forget: “correlation does not imply causality”.
    Inconsistent “translation”
    If you ask AI software the same question repeatedly, or ask using different words with the same meaning, you often get different answers. Thus, it’s necessary to “triangulate” by comparing two or more answers to see whether they agree. More importantly, always do a reality check by employing your skepticism and your analytical skills to ensure that you agree with the software’s conclusions.
    Plagiarism
    There are many tools for detecting plagiarism (use of another person’s writing without their permission); you’ve probably already encountered this software if you submit manuscripts to a journal publisher. To minimize the risk of plagiarism, use one of these tools or search for parts of any AI-generated text using Google or another search engine.
    One problem you’ll encounter is that there are relatively few ways to describe specific research methods, which means that everyone ends up using the same words to describe those methods. Plagiarism software doesn’t understand this, and often flags these words as plagiarized. You may need to work around this problem by speaking with a journal’s editor.
    Hallucinations
    When AI software cannot find what you’re looking for, it often invents “facts”. This is particularly common for literature citations, which are easy but time-consuming to check. It’s harder and takes much more work to find instances of hallucination in the Results, Discussion, and Conclusions of a journal article. Always rigorously review any results, interpretations, and conclusions provided by AI software.
    The black box problem
    It’s difficult and sometimes impossible to learn the logic AI used to reach a conclusion. This remains true even with software that attempts to show the chain of logic that it used to reach a conclusion. Without knowing the logic, it’s difficult to know when an AI solution can and cannot be trusted.
    Quality control
    Even today, the input data used to create large language models is usually not screened for quality, thus unreliable information gets equal weight to the truth. This undermines the software’s potential to find the correct information and correctly interpret and use that information.
    Nothing new
    AI can only repeat information that has already appeared in its database. It cannot detect possible new interpretations that would lead to breakthroughs. Future AI software may learn how to examine the current state of knowledge, identify gaps, and provide suggestions on how to fill those gaps. But current software cannot do any of these things.
    Be cautious!
    Now, and for some long time ahead of us, AI software will not be trustworthy. If you use it, remember that it’s not a substitute for human insight and that it’s your responsibility to rigorously scrutinize its findings. But also remember the ethical implications of its use: the more often you use the software, the more free (unpaid) training you are giving that software and the sooner that software will acquire the “skills” required to replace your colleagues—or you.
    To learn more
    Bergstrom, C.T.; West, J.D. 2025. Modern-day articles or bullshit machines? https://thebullshitmachines.com/

メールアドレス
お名前

メルマガ登録