U.S. President Donald Trump at this year's Group of Seven summit (Asahi Shimbun file photo)

EDITORIAL: Trump must not be allowed to insert ideology into AI responses

（社説）米ＡＩ戦略　危ういトランプ流介入

U.S. President Donald Trump has announced an AI Action Plan with the goal of making the United States the world leader in development of artificial intelligence.

It is a major change of direction from the Biden administration, which emphasized control of the risks associated with AI.

The Trump administration wants to accelerate technology and infrastructure development through deregulation and increased investment.

The action plan will encourage the export of AI technology to friendly nations in a bid for supremacy in competition with an emerging China.

A surprising aspect of the plan is the level of intervention it envisages into the output from generative AI technology such as ChatGPT.

National guidelines for AI risk control were revised to eliminate from answers references to diversity, equity and inclusion, widely known as DEI, and climate change.

The federal government is supposed to sign contracts only with companies whose AI technology is objective and does not contain ideological bias.

While the Trump administration says such measures are needed to protect freedom of speech and American values, the action plan conversely carries the risk of gagging free speech.

To begin with, it is very difficult to establish neutral standards to help determine ideological bias. The bias could be defined according to the intent of the administration in power.

DEI initially sought to counter discrimination based on gender, race, nationality or disability and to ensure fair opportunity for those who had been treated unfavorably in the past.

Trump has long been negative about measures related to DEI and global warming. Since the start of his second term, he has implemented various steps reflecting that thinking.

We cannot ignore the attempt to interfere with generative AI, which is spreading rapidly in society and people's daily lives.

It is possible that the federal government's procurement conditions will become the standards regarding generative AI.

That would have a major impact on Japan as it depends on the United States for much of its generative AI technology.

Citing an example of bias in generative AI, Trump pointed to white male figures from American history who were transformed in AI answers into women or members of other races.

Generative AI undertakes deep learning using a massive amount of data and subsequently goes through repeated tuning to produce answers.

If there is bias in the data used for deep learning, the answers will reflect such bias and the tuning process could generate the wrong answer.

Put another way, if the learning data is selected arbitrarily and certain algorithms are intentionally chosen during the tuning process, it would be possible to produce answers that reflected a certain bias.

This is one characteristic of generative AI that we must always keep in mind. It could be used by iron-fisted regimes to gag free speech.

It was pointed out that the generative AI technology released this year by the Chinese startup DeepSeek avoided or rejected answers that were inconvenient to Beijing, such as regarding the Tiananmen Square incident.

Trump’s version of intervention that seeks to delete certain references is fundamentally no different.

--The Asahi Shimbun, Aug. 4

米トランプ政権が人工知能（ＡＩ）開発で世界をリードすることをめざす「ＡＩ行動計画」を発表した。
　ＡＩのリスク管理を重視したバイデン前政権の方針を転換し、規制緩和や投資により技術開発やインフラ整備を加速させる。友好国への技術輸出なども促進し、台頭する中国との覇権争いを制するという内容だ。
　驚くのは、チャットＧＰＴのような生成ＡＩの出力内容への介入だ。「ＤＥＩ（多様性、公平性、包摂性）」や「気候変動」への言及をなくすようＡＩのリスク管理に関する国の指針を改定し、連邦政府は「客観的でイデオロギー的偏向のない」事業者とだけ契約する方針を打ち出した。
　「言論の自由と米国の価値を守る」という大義を掲げるが、逆に言論統制につながりかねない危うさがある。そもそも「イデオロギー的偏向」を判断する中立的な基準をつくるのは困難で、時の政権で都合よく定義されかねない。
　ＤＥＩは、性別、人種、国籍、障害などによる差別の是正を目指し、不当な扱いを受けてきた人々に公正な機会を保障する考え方だ。トランプ氏はＤＥＩや地球温暖化対策に否定的で、就任以降さまざまな措置を講じている。
　その延長で、社会や生活のなかで急速に普及する生成ＡＩに介入することは看過しがたい。連邦政府の調達要件がそのまま「標準」となるおそれもあり、生成ＡＩの多くを米国に依存する日本にとっても無関係ではない。
　トランプ大統領は、生成ＡＩの「偏向」の例として、白人男性だった歴史上の人物の人種や性別を誤って出力した例を挙げた。
　一般的に生成ＡＩは「深層学習」という技術を用いて膨大なデータを使って学習を行った後、チューニングと呼ばれる調整作業を繰り返す。学習データに偏りがあれば回答も偏り、調整作業の結果、誤った出力をすることもある。
　裏返せば、学習データを恣意（しい）的に選別したり、調整の過程で意図的なアルゴリズム（計算手順）を加えたりすれば、出力結果を特定の方向に偏らせることもできる。これもまた知っておくべき生成ＡＩの特性であり、強権的な国家に都合のいい言論統制の手段にもなりうる。
　中国の企業ディープシークが今年公開した生成ＡＩは、天安門事件のような中国政府に都合のよくない情報に関する回答を拒否・回避するとの指摘がある。特定の記述を削除させるトランプ流の介入も本質的には変わらない。

(2025年8月4日)