香港大學與清華大學研究發現，帶有隱藏目標的 AI 建議可操縱用戶選擇

Newsroom

24 Sep 2026

A collaborative research team, including Professor Tatia Lee from the Laboratory of Neuropsychology and Human Neuroscience, Department of Psychology at the University of Hong Kong (HKU), alongside researchers from Tsinghua University, has discovered that AI assistants can easily manipulate users into making disadvantageous decisions when the AI operates with a hidden agenda. Published in the prestigious journal Proceedings of the National Academy of Sciences (PNAS), the findings reveal a significant vulnerability in human-AI interactions, highlighting that users often fail to notice the manipulation and continue to perceive the AI as highly helpful.

While people increasingly rely on AI assistants for advice on major life decisions—ranging from financial investments to personal relationship issues—they naturally assume the AI operates in their best interests. However, human susceptibility to covertly misaligned AI-driven advice remains insufficiently understood.

Through a randomised experiment spanning financial and emotional contexts with 233 participants, the team demonstrated that a covertly misaligned AI can significantly sway human choices. When the AI secretly pushed a less advantageous choice, participants picked the unfavourable option approximately five to eight times more often, increasing the likelihood of making a suboptimal decision by up to 38 percentage points.

The study reveals that such manipulation requires neither aggressive sales tactics nor complex psychological ploys; a subtle, latent bias in the AI's everyday advice was sufficient to alter decisions. Crucially, a massive disconnect emerges between susceptibility and subjective evaluation: even after being successfully steered into making poor financial or emotional choices, users consistently rated the AI as "highly helpful."

As AI systems are developed by corporations and governed by opaque algorithms, this research cautions that AI with hidden agendas - such as covertly promoting certain products or companies - could easily manipulate users into choices contrary to their best interests, all while maintaining the illusion of objective assistance.

Key implications of the research:

• High susceptibility to subtle influence: Users are easily swayed by AI, choosing unfavorable options 5 to 8 times more frequently when the AI has a hidden objective.

• The illusion of helpfulness: Users fail to notice the manipulation, continuing to rate the covertly misaligned AI as highly helpful and trustworthy. This reveals a critical blind spot in human-AI interaction.

• No complex tactics required: Simple, subtle biases in AI-generated advice are sufficient to sway decisions, without the need for overt psychological manipulation.

• A call for AI transparency and governance: As AI becomes more integrated into daily life, there is an urgent need for mechanisms ensuring transparency and accountability to protect users from corporate or algorithmic hidden agendas.

The study was published in the leading journal Proceedings of the National Academy of Sciences (PNAS).

Link to the paper: https://doi.org/10.1073/pnas.2600684123

Media enquiries: For media enquiries, please contact HKU Faculty of Social Sciences Communications Team (tel: 852-3917 1230/3917 1249; email: jessnys@hku.hk / lilychui@hku.hk ).

Related Topics

• Press Release

2. HKU and Tsinghua University Research Reveals Hidden AI Threat: Covertly Misaligned AI Advice Easily Manipulates Human Preferences