團隊

大語言模型如今正日益滲透到知識生產、商業分析、法律諮詢與醫療決策等核心領域，一個過去常被視為技術性缺陷嘅問題，正迅速成為全球技術競賽嘅新焦點：大語言模型嘅「幻覺」（ Hallucination ）——即模型喺缺乏事實依據時，仍然以極高「信心」畀出流暢、卻虛假嘅答案。

人工智能（AI）日新月異，中國科企之成就令海外投資者對中資股刮目相看，迎來美國更強力的圍堵，務求拖慢中國的AI技術發展。香港大學經管學院創新及資訊管理學教授兼夏利萊伉儷基金教授蔣鎮輝表示，中國AI大模型的實力已領先全球，跟美國的差距也愈來愈近，華府的晶片管制不能完全解釋兩國AI技術差距，但接下來的挑戰也非光靠砸錢就能彌補。

Data donations, where individuals are encouraged to donate their personal information, have the potential to advance medical research and help limit the spread of pandemics, among other benefits. The decision to donate data is fundamentally a privacy decision. In this research, we build on the privacy calculus, a model describing privacy risks and benefits, and examine the impact of privacy concerns on data donation decisions, highlighting the role of societal benefits in privacy decisions. Based on two randomized experiments using the general context of data donation for medical research (experiment 1) and the specific context of data donation for COVID-19 research (experiment 2), we find that individuals who are highly concerned about privacy tend to donate less data (experiments 1 and 2). This effect holds under a variety of conditions and is consistent with prevailing research. However, this effect is contingent on the privacy calculus. When implicit or explicit societal benefits are perceived, particularly in the absence of privacy controls, the association between privacy concerns and data donation decisions is less salient, highlighting the significant role that societal benefits have in privacy decisions. We discuss the theoretical, practical, social, and ethical implications of these findings.

人工智慧圖像生成技術日新月異，推動市場營銷、廣告設計及藝術創作等領域的創新。港大經管學院創新及資訊管理學教授蔣鎮輝早前聯同其人工智慧大模型評測團隊對22款主流AI模型進行評估，深入剖析不同AI模型在圖像生成方面的表現及潛在風險。

Online video platforms face the challenge of balancing the needs of their users with those of their advertisers. Although users typically prefer to have less intrusive ads, advertisers aim to effectively catch user attention. This paper investigates how the provision of ad choice affects the effectiveness of video advertising. We argue that allowing users to choose an ad to view may trigger a “conjecture-formation-and-confirmation” process that motivates users to pay more attention to the selected ad. Two online experiments and four laboratory experiments are conducted to test the theorized underlying mechanism of the ad choice effect. Study 1 finds when users are unfamiliar (versus familiar) with the content of ad options (i.e., they need to make conjectures about ad content), ad choice is more likely to increase user attention to the chosen ad. Study 2 and Study 3 show that the impact of ad choice on user attention is more likely to be positive when users are enabled to make conjectures about ad content, such as when choice options provide more relevant information about ad content. Study 4a and Study 4b provide more direct support for the underlying mechanism by showing that the ad choice effect is attenuated when users cannot form conjectures about ad content at the choice stage. Study 5 further demonstrates that the positive effect of ad choice is robust across different ad settings. Taken together, these studies show ad choice is more likely to boost the effectiveness of video advertising when the “conjecture-formation-and-confirmation” process is triggered.

自 ChatGPT 問世以來，大語言模型LLM（Large Language Model）迅速成為全球科技競賽的焦點。LLM將被應用到越來越多領域，成为AI發展的新蓝海。目前，美國在基礎技術開發和創新應用上具有明顯優勢，其模型往往在技術前沿上表現出色；而中國的模型更強調針對本土語言環境的優化和實際應用的適應性。如何在兩國的競爭中找到平衡點並推動合作，將成為未來幾年的發展關鍵。

Many organizations have adopted internet monitoring to regulate employees’ cyberloafing behavior. Although one might intuitively assume that internet monitoring can be effective in reducing cyberloafing, there is a lack of research examining why the effect can occur and whether it can be sustained. Furthermore, little research has investigated whether internet monitoring can concurrently induce any side effects in employee behavior. In this paper, we conducted a longitudinal field quasi-experiment to examine the impacts of internet monitoring on employees’ cyberloafing and organizational citizenship behavior (OCB). Our results show that internet monitoring did reduce employees’ cyberloafing by augmenting employees’ perceived sanction concerns and information privacy concerns related to cyberloafing. The results also show that internet monitoring could produce the side effect of reducing employees’ OCB. Interestingly, when examining the longitudinal effects of internet monitoring four months after its implementation, we found that the effect of internet monitoring on cyberloafing was not sustained, but the effect on OCB toward organizations still persisted. Our study advances the literature on deterrence theory by empirically investigating both the intended and side effects of deterrence and how the effects change over time. It also has important broader implications for practitioners who design and implement information systems to regulate employee noncompliance behavior.

截至7月底，國內共推出超300個大模型。經過大模型數量之爭的上半場，下半場中國大模型該如何押寶？“持續穩定的政策支持、龐大的算力規模和廣闊的應用場景是中國獨特的競爭優勢和巨大的發展潛力。”蔣鎮輝教授認為，大模型將從數量到更重視應用端發展。

生成式人工智能（AI）工具發展日新月異，香港大學經濟及工商管理學院今日（12日）發表一項人工智能大語言模型評測綜合排行榜，通過港大深圳研究院建立評分系統，比較十多款大模型表現，顯示由中國科企百度開發的「文心一言」，在中文語境下綜合得分最高，但在「通用語言能力」卻跑輸「ChatGPT4-Turbo」，而大部分國產模型在英文語境下表現均處於「稍微劣勢」。