ChatGPT, Gemini Mask Green Investing Risks Without Prompt Nudges

Panel 6 - 2: Green-Curious Investing and GenAI: A Field Experiment — Ai4

Ai4

A field experiment reveals retail investors fail to learn about greenwashing and return trade-offs from AI chatbots unless interfaces actively nudge them to query those specific risks.

AI-assisted summary of Panel 6 - 2: Green-Curious Investing and GenAI: A Field Experiment — original source. Check the original for full context.

Source published:

Detailed notes

## Background Published on YouTube by Ai4 on August 20, 2026, this presentation features Mary Ann Majadillas, Professor of Finance at California State University, Monterey Bay. The session covers a randomized classroom field experiment examining how students use ChatGPT and Google Gemini to learn about sustainable investing concepts, specifically greenwashing risks and expected return trade-offs.

## TL;DR Professor Mary Ann Majadillas details a randomized classroom experiment with 272 undergraduate students researching sustainable finance using ChatGPT and Gemini. The research demonstrates that commercial GenAI models routinely fail to surface critical investment risks—such as greenwashing and lower expected returns—unless users explicitly ask about them. Providing students with subtle instructional cues beforehand significantly increased their likelihood of prompting these topics and reporting meaningful learning outcomes. Furthermore, model differences played a notable role, with Gemini users reporting significantly higher rates of learning about greenwashing than ChatGPT users.

## Atmosphere & style The presentation carries an engaging, scholarly, and accessible tone, translating rigorous behavioral and statistical data into practical takeaways for retail investing. Majadillas speaks with clear pedagogical enthusiasm and clarity, grounding complex experimental dynamics in relatable scenarios about how everyday users casually treat chatbots as trusted advisors. The session feels like an insightful, forward-looking academic briefing that directly connects AI interface behavior to consumer financial literacy.

## Key takeaways - [00:04](https://www.youtube.com/watch?v=IlKi9r5V7KU&t=4s) Retail investors interested in sustainable investing face two critical obstacles: the risk of greenwashing and lower expected risk-adjusted returns. - [00:04](https://www.youtube.com/watch?v=IlKi9r5V7KU&t=4s) Everyday investors increasingly treat generative AI tools like ChatGPT and Gemini as trusted financial advisors without recognizing their omissions. - [00:04](https://www.youtube.com/watch?v=IlKi9r5V7KU&t=4s) Prior automated benchmarks show that off-the-shelf LLMs consistently omit greenwashing risks and risk-return trade-offs unless explicitly asked. - [03:52](https://www.youtube.com/watch?v=IlKi9r5V7KU&t=232s) The researchers conducted a randomized classroom field experiment with 272 undergraduate students across financial literacy and business statistics courses. - [03:52](https://www.youtube.com/watch?v=IlKi9r5V7KU&t=232s) Participants were randomly assigned one of three instructional cues: a base writing prompt, a greenwashing cue, or a return trade-off cue. - [03:52](https://www.youtube.com/watch?v=IlKi9r5V7KU&t=232s) The experiment evaluated interactions across two major AI platforms: ChatGPT during Fall 2025 and Google Gemini during Spring 2026. - [07:38](https://www.youtube.com/watch?v=IlKi9r5V7KU&t=458s) Receiving instructional cues dramatically increased prompt inclusion, with ~82% including greenwashing and ~69% including return trade-offs in initial prompts. - [07:38](https://www.youtube.com/watch?v=IlKi9r5V7KU&t=458s) Including critical concepts in initial prompts directly determined whether students reported learning about those concepts from the AI interaction. - [07:38](https://www.youtube.com/watch?v=IlKi9r5V7KU&t=458s) Platform differences emerged: Gemini was significantly more effective than ChatGPT at surfacing greenwashing concepts to unprompted users. - [07:38](https://www.youtube.com/watch?v=IlKi9r5V7KU&t=458s) Prior course subject matter did not drive spontaneous inquiry; financial literacy students were no more likely to ask about return trade-offs on their own than statistics students. - [13:40](https://www.youtube.com/watch?v=IlKi9r5V7KU&t=820s) Because generative AI models rarely provide complete unprompted guidance, investors risk major blind spots if they do not know what questions to ask. - [13:40](https://www.youtube.com/watch?v=IlKi9r5V7KU&t=820s) Fintech platforms and financial educators should build interface scaffolding and prompt cues to actively guide users to assess greenwashing and return trade-offs.

## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=IlKi9r5V7KU&t=4s) Motivation and Background: Green Investing and AI Advisors

Mary Ann Majadillas introduces her research on green investing and generative AI, conducted alongside colleagues in the College of Business who also served as instructors for the courses where the field experiments took place. The study is motivated by surging investor interest in climate-friendly financial products. Field experiments across Europe demonstrate that while retail investors are eager to allocate capital toward combating climate change, they face two major stumbling blocks. First, they fear "greenwashing"—the risk that companies and financial funds make exaggerated or deceptive environmental claims without delivering real-world impact. Second, environmentally friendly investments frequently offer lower expected returns for the same level of risk, leading to a less favorable reward-to-risk ratio compared to conventional assets.

Concurrently, everyday retail investors are increasingly treating generative AI platforms like ChatGPT and Gemini as trusted financial advisors. When deciding where to allocate money, people casually turn to AI systems under the implicit assumption that the models will provide balanced, objective guidance: "It's like asking a friend or asking a trusted advisor—trusted, right? We believe what the AI tells us." Majadillas notes that prior research by co-author Leslie Boni examined this dynamic through an automated benchmark: by submitting identical green investing queries across 100 simulated user accounts, Boni discovered that off-the-shelf AI models routinely fail to bring up greenwashing risks or risk-return trade-offs unless explicitly asked. The present study serves as the natural behavioral follow-up, exploring how real human users interact with generative AI when researching sustainable investments.

- **Investor Hesitations**: European field studies show strong interest in sustainable investing alongside acute concerns about greenwashing and subpar risk-adjusted returns. - **Retail Reliance on GenAI**: Retail investors increasingly treat chatbots as surrogate financial advisors without recognizing model omissions. - **Prior Controlled Benchmark**: Boni's preceding study demonstrated that baseline AI prompts omit greenwashing and return trade-offs when users do not prompt for them.

---

[03:52](https://www.youtube.com/watch?v=IlKi9r5V7KU&t=232s) Experimental Design and Instructional Cues

To evaluate unmediated user interactions, the researchers conducted a randomized classroom-based field experiment across two undergraduate courses: a financial literacy course and a business statistics course, each contributing 136 students for a total sample of 272 participants. The experiment spanned two consecutive terms: Fall 2025 (accounting for approximately two-thirds of the cohort with 182–186 students interacting with ChatGPT) and Spring 2026 (65 students interacting with Google Gemini).

```mermaid flowchart TD accTitle: Experimental Design Workflow accDescr: Flowchart illustrating student assignment across courses, instructional cues, AI platforms, and post-experiment surveys.

subgraph Cohort["Participant Sample (N=272)"] FL["Financial Literacy Class (n=136)"] BS["Business Statistics Class (n=136)"] end

subgraph Cues["Randomized Instructional Cues"] C1["Base Prompt Only (1/3)"] C2["Base + Greenwashing Cue (1/3)\n'Friend heard about greenwashing'"] C3["Base + Return Tradeoff Cue (1/3)\n'Friend heard about return differences'"] end

subgraph Platforms["GenAI Interaction"] GPT["ChatGPT (Fall 2025, n=182–186)"] GEM["Gemini (Spring 2026, n=65)"] end

subgraph Assessment["Evaluation"] SURV["Post-Interaction Survey\n(Prompting behavior & self-reported learning)"] end

Cohort --> Cues Cues --> Platforms Platforms --> SURV ```

The classroom task required students to engage directly with their assigned GenAI chatbot to write an essay on green investing. Immediately prior to prompting the model, students were randomly assigned to one of three instructional cue conditions: 1. **Base Cue Only (1/3 of students)**: Standard writing instructions without specialized concept prompts. 2. **Greenwashing Cue (1/3 of students)**: Base prompt plus an added nudge stating that a friend remembered hearing something about "greenwashing." 3. **Return Tradeoff Cue (1/3 of students)**: Base prompt plus an added nudge stating that a friend remembered hearing that some investments achieve better returns than others.

Following their chatbot sessions, students completed a survey assessing whether they raised these topics in their initial prompts and whether the AI interaction helped them learn about greenwashing and return trade-offs.

---

[07:38](https://www.youtube.com/watch?v=IlKi9r5V7KU&t=458s) Findings: Prompting Dynamics, Learning Outcomes, and Model Differences

Analyzing the survey data across the full cohort reveals that unguided AI interactions frequently leave decision-critical financial concepts unmentioned. Overall, 114 of the 272 students reported learning about greenwashing, and 117 of 272 reported learning about return trade-offs from their chatbot sessions. However, the randomized instructional cues fundamentally altered how students framed their opening queries and what knowledge they extracted. Students assigned the greenwashing cue were significantly more proactive, with roughly 82% incorporating greenwashing into their initial prompt, compared to 69% of students who included returns in their opening query when given the return trade-off cue.

``` +-------------------------------------------------------------------------------+ | Experimental Finding Summary (N=272) | +------------------------------------+------------------------------------------+ | Metric | Result | +------------------------------------+------------------------------------------+ | Overall Greenwashing Learning | 114 / 272 students (41.9%) | | Overall Returns Tradeoff Learning | 117 / 272 students (43.0%) | | Initial Prompt Inclusion (Cue: GW) | ~82% referenced greenwashing | | Initial Prompt Inclusion (Cue: RT) | ~69% referenced return trade-offs | | Model Disparity (Greenwashing) | Gemini significantly outperformed ChatGPT| +------------------------------------+------------------------------------------+ ```

Asking about these concepts up front proved decisive: participants who incorporated greenwashing or return trade-offs into their very first query were far more likely to report meaningful learning gains on those topics. A notable divergence emerged between the two AI platforms. Students using Gemini in Spring 2026 were significantly more likely to learn about greenwashing than students using ChatGPT in Fall 2025, indicating that underlying model architecture and default safety/context filtering meaningfully influence the breadth of financial counsel provided.

Interestingly, the researchers' initial hypothesis regarding financial domain familiarity was unsupported: students enrolled in the financial literacy course were no more likely to ask about return trade-offs on their own than students in the business statistics course. Rather than prior course subject matter, the presence of the external instructional cue served as the primary driver behind whether students surfaced risk-and-return trade-offs.

---

[13:40](https://www.youtube.com/watch?v=IlKi9r5V7KU&t=820s) Practical Takeaways and Future Research Directions

Majadillas emphasizes that while generative AI holds real potential to bolster investor education, its effectiveness is contingent on user prompt formulation: "What users learn depends heavily on how they prompt the system... it won't always give you the complete answer every single time you ask about making investments." Because typical retail investors do not know what they do not know, AI systems that rely entirely on unprompted user intuition risk cementing blind spots around critical financial trade-offs.

These findings carry direct implications for fintech designers, educators, and regulatory bodies: - **Explicit Prompt Guidance**: Financial institutions and consumer-facing apps should actively guide and nudge investors to explicitly query greenwashing exposure and return trade-offs. - **Interface Nudges & AI Scaffolding**: Financial literacy curricula and digital advisory tools must incorporate scaffolding mechanisms that automatically surface decision-critical risks before users execute investment decisions. - **Future Research Horizons**: Majadillas outlines plans to extend the classroom field experiment beyond the initial 272-student sample to broader, demographically diverse investor populations, evaluate additional commercial LLMs, and track downstream, long-term portfolio allocation behavior.

## Notable quotes > [00:04](https://www.youtube.com/watch?v=IlKi9r5V7KU&t=4s) It's like asking a friend or asking a trusted advisor—trusted, right? We believe what the AI tells us. — Mary Ann Majadillas

> [13:40](https://www.youtube.com/watch?v=IlKi9r5V7KU&t=820s) What users learn depends heavily on how they prompt the system... it won't always give you the complete answer every single time you ask about making investments. — Mary Ann Majadillas

> [00:04](https://www.youtube.com/watch?v=IlKi9r5V7KU&t=4s) greenwashing — Mary Ann Majadillas

> [03:52](https://www.youtube.com/watch?v=IlKi9r5V7KU&t=232s) Friend heard about greenwashing — Mary Ann Majadillas

> [03:52](https://www.youtube.com/watch?v=IlKi9r5V7KU&t=232s) Friend heard about return differences — Mary Ann Majadillas

## Mentioned Mary Ann Majadillas, Leslie Boni, California State University, Monterey Bay, ChatGPT, Google Gemini, College of Business, Ai4

## Video description Mary Ann Majadillas, Professor of Finance, California State University, Monterey Bay

We conduct a randomized classroom-based field experiment to study the unmediated use of ChatGPT and Gemini to learn about green investing. We examine two decision-critical concepts: greenwashing potential and expected lower average returns. Undergraduate students in financial literacy and business statistics courses are randomly assigned one of three instructional cues immediately prior to their GenAI interaction: greenwashing risk, return tradeoff risk, or neither risk. We find that most students who receive the greenwashing or return tradeoff cues reference the concept in their initial prompts with Gen AI and are more likely to report learning about the concept from the GenAI interaction. While only about a third of students who do not reference the returns concept in their initial prompt report learning about the concept regardless of whether they use ChatGPT or Gemini, students who do not reference greenwashing in their initial prompt are significantly more likely using Gemini (46%) than ChatGPT (15%) to report learning about greenwashing. Our research contributes to the literature that examines direct interactions of investors with off-the-shelf GenAI models and their potential to supplement or replace financial advising professionals.

[Source](https://www.youtube.com/watch?v=IlKi9r5V7KU)

中文简报

若无提示引导，ChatGPT 与 Gemini 会掩盖绿色投资风险 一项实地实验显示，除非交互界面主动引导散户投资者针对特定风险提问，否则他们无法从 AI 聊天机器人处了解到“漂绿”行为及收益权衡问题。 Mary Ann Majadillas 教授详细介绍了一项涉及 272 名本科生的随机课堂实验，学生们使用 ChatGPT 和 Gemini 研究可持续金融。研究表明，商业生成式 AI 模型通常不会主动揭示诸如“漂绿”和预期收益偏低等关键投资风险，除非用户明确询问。预先为学生提供细微的指导性提示，能显著增加他们就这些主题发起提问的几率，并促成切实有效的学习成果。此外，模型之间的差异也起到了明显作用，Gemini 用户报告学到“漂绿”知识的比例显著高于 ChatGPT 用户。 零售投资者日益将生成式 AI 视为值得信赖的财务顾问，但此前的基准测试表明，通用大模型在未被明确提问时往往会忽略“洗绿”（greenwashing）风险以及环境投资较低的预期回报率。 (0:04) 研究团队针对 272 名本科生开展了一项随机课堂现场实验，分别测试了 ChatGPT 和 Google Gemini，并为学生随机分配基础提示、洗绿风险提示或收益权衡提示等不同引导。 (3:52) 引导提示词显著改变了提问行为，收到提示的学生在初始 Prompt 中提及洗绿（约 82%）和收益权衡（约 69%）的比例大幅提高；而在未主动提问洗绿的情况下，Gemini 用户报告学到该概念的比例显著高于 ChatGPT。 (7:38) 研究指出用户“不知自身所不知”，完全依赖用户自主提问容易形成关键财务决策盲区，因此金融科技产品与教育平台亟需引入界面引导和脚手架机制，自动提示潜在风险。 (13:40)

Reading summary

A classroom field experiment presented by Mary Ann Majadillas reveals that generative AI tools like ChatGPT and Gemini fail to surface critical sustainable finance concepts—specifically greenwashing risks and lower risk-adjusted returns—unless users are explicitly nudged to ask about them. Green Investing Hesitations and Retail AI Trust (0:04) Retail investor interest in climate-friendly assets has grown rapidly, but European field studies identify two major deterrents: concerns over corporate greenwashing and the tendency for sustainable funds to offer less favorable reward-to-risk ratios compared to conventional investments. Despite these complexities, retail investors increasingly treat commercial generative AI platforms as objective, trusted advisors for financial allocation. Prior research by co-author Leslie Boni tested this dynamic by running identical sustainable investment queries across 100 simulated user accounts. That benchmark revealed that standard, off-the-shelf generative AI models consistently omit greenwashing risks and risk-return trade-offs when users do not specifically bring them up, establishing the baseline for examining how real human investors interact with AI systems. Randomized Field Experiment and Prompt Cues (3:52) To analyze live user behavior, Majadillas and her colleagues conducted a randomized field experiment with 272 undergraduate students split evenly between a financial literacy course (136 students) and a business statistics course (136 students). The trial took place across two consecutive terms: Fall 2025 evaluated ChatGPT with 182 to 186 students, while Spring 2026 evaluated Google Gemini with 65 students. Participants were assigned to write an essay on green investing using their assigned chatbot and were randomly sorted into three equal prompt-cue groups: a base instruction cue, a cue mentioning a friend who heard about greenwashing, and a cue mentioning a friend who heard about return differences across investments. Researchers then evaluated post-interaction surveys measuring initial prompting content and self-reported learning gains. Prompting Dynamics, Model Disparities, and Domain Background (7:38) Overall, only 114 students (41.9%) reported learning about greenwashing and 117 (43.0%) reported learning about return trade-offs. However, randomized cues substantially shifted what users asked up front: roughly 82% of students given the greenwashing cue referenced it in their initial prompt, compared to 69% of students who asked about return trade-offs when given the returns cue. Inquiring about these concepts in the initial query proved decisive for whether users achieved reported learning gains. The study also found notable differences across platforms and academic backgrounds. Spring 2026 students using Gemini were significantly more likely to learn about greenwashing than Fall 2025 students using ChatGPT, indicating that model architecture and default context filtering influence output breadth. Furthermore, students enrolled in the financial literacy course were no more likely to spontaneously ask about return trade-offs than business statistics students, demonstrating that external cues, rather than academic domain familiarity, drove risk-and-return inquiries. Scaffolding AI for Retail Financial Decisions (13:40) Majadillas argues that generative AI's effectiveness in financial education is constrained by user prompt formulation: because retail investors often do not know what questions to ask, unprompted interactions leave persistent blind spots. The researchers recommend that financial technology developers, educators, and regulators introduce proactive scaffolding and prompt guidance within consumer interfaces to ensure decision-critical risks are addressed prior to investment allocation. The researchers emphasize that AI-driven investor guidance depends heavily on deliberate interface nudges rather than assumed model objectivity. Majadillas plans to extend this research beyond the initial 272-student academic sample to evaluate demographically diverse investor groups, test additional commercial large language models, and observe downstream, long-term portfolio allocation behavior.

精炼阅读摘要

商学院学者Mary Ann Majadillas等人的田野实验研究表明，散户投资者在使用生成式AI辅助绿色投资决策时，若缺乏针对性引导，大语言模型往往会忽略漂绿风险与风险收益权衡等关键考量；外部提示词干预能够显著促使投资者主动提问并获得相关金融知识。 绿色投资痛点与AI顾问的默认信息缺失 (0:04) 欧洲多项田野实验表明，散户投资者虽然对气候友好型金融产品展现出高度兴趣，但在实际配置中面临两大阻碍：其一是担忧“漂绿”（即企业或基金产品作出夸大、欺骗性的环保承诺而无实质环境成效）；其二是绿色投资的预期收益率通常低于同等风险水平的传统资产，导致风险调整后收益较差。与此同时，散户投资者日益将ChatGPT与Gemini等生成式AI视为可信赖的数字财务顾问，潜意识中假定模型会主动提供客观、全面的资产配置建议。 然而，合作学者Leslie Boni先前的基准测试研究显示，在100个模拟用户账户提出相同绿色投资咨询的受控实验中，现成的商业化大模型若未被明确追问，普遍不会主动指出漂绿风险或风险收益权衡。Majadillas指出，散户投资者在不知情的情况下盲目信任AI模型，极易导致关键投资风险被系统性忽略，本项田野实验旨在检验真实用户在未受人工中介干预时的实际交互行为。 课堂田野实验与三类随机提示词机制 (3:52) 研究团队在两门大学本科课程中实施了随机对照田野实验，样本共纳入272名学生，其中金融素养课程与商业统计课程各占136名。实验跨越两个连续学期：2025年秋季学期约占总样本的三分之二（182至186名学生使用ChatGPT），2026年春季学期则包含65名使用Google Gemini的学生。实验任务要求学生借助分配的AI聊天机器人撰写一篇关于绿色投资的分析短文。 在与模型对话前，学生被随机分配至三种教学线索条件之一：第一组仅获得基础写作提示词，不含专业概念提示（占三分之一）；第二组在基础提示之上增加了漂绿线索，说明“有朋友曾听说过漂绿问题”（占三分之一）；第三组则增加了收益权衡线索，说明“有朋友听说部分投资比其他投资收益表现更好”（占三分之一）。交互结束后，研究团队通过问卷调查学生在初始提示词中涉及的概念，以及是否通过AI交互掌握了漂绿与收益权衡知识。 提问行为主导知识获取与模型表现分化 (7:38) 实验数据表明，若任由用户自行探索，关键财务风险在人机对话中极易被遗漏。在272名受试者中，总体仅有114人（41.9%）表示通过对话了解了漂绿问题，117人（43.0%）了解了收益权衡。随机分配的提示词线索显著改变了提问行为：获得漂绿线索的学生中，约82%在第一轮提问中主动提及漂绿；而获得收益权衡线索的学生中，有69%在初始提问中纳入了收益比较。统计显示，在初始提示词中直接提出相应概念，是学生最终实现显著学习收益的决定性前提。 实验同时揭示了AI模型之间的性能差异与专业背景假设的失效。在2026年春季使用Gemini的学生，在漂绿认知上的学习成效显著高于2025年秋季使用ChatGPT的学生，研究者认为这反映了大语言模型在底层架构与默认上下文过滤机制上的差异。此外，研究者原先假设金融素养课学生会比统计课学生更主动关注收益权衡，但数据并未支持该假设：专业知识储备并未促成自发提问，外在提示词引导才是促使学生审视风险收益关系的核心驱动力。 金融科技设计干预与后续拓展研究 (13:40) Majadillas强调，生成式AI能够辅助投资者教育，但其有效性高度取决于用户如何构建提示词。由于散户投资者常常面临“不知道自己不知道什么”的认知局限，完全依赖用户自发直觉的AI系统可能会固化投资者的决策盲区。因此，金融科技界面、教育者与监管机构不能仅依赖基础模型对话，而应在产品交互中内置支架机制（scaffolding）与显式提醒，强制在用户执行决策前弹出漂绿风险与收益权衡审查。 关于未来的研究规划，研究团队计划将样本范围从目前的272名大学生扩展至人口统计特征更为多元的真实散户群体，对比评估更多商业化大语言模型，并建立长期追踪机制，进一步观察提示词干预与AI交互是否会对投资者后续真实的投资组合资金配置产生实质影响。 该实验表明，大语言模型在默认状态下不会自动为绿色投资提供平衡的风险分析，用户的初始提问方式决定了信息获取的深度。如何通过产品端提示设计弥补投资者的知识盲区，并在更大规模、多样化的投资者群体中检验其实际资产配置行为，仍是未来研究与监管需要解决的关键问题。

Detailed notes — zh

## Background 该内容由 Ai4 频道于 2026 年 8 月 20 日发布在 YouTube 平台上。视频记录了加州州立大学蒙特雷湾分校财务学教授玛丽安·马哈迪亚斯（Mary Ann Majadillas）的学术演讲，介绍了一项关于散户投资者使用 ChatGPT 和 Google Gemini 学习绿色投资的随机课堂田野实验，重点探讨了教学提示词对理解“漂绿”和“预期低收益”等关键概念的影响。 ## TL;DR 研究团队针对 272 名修读金融素养与商业统计课程的本科生开展了随机田野实验，评估他们在不同教学提示线索引导下利用 ChatGPT 和 Gemini 探索绿色投资的效果。实验表明，收到特定教学线索的学生绝大多数会在首条 Prompt 中提及“漂绿”或“回报权衡”，并显著提升后续学习认知；而若未在初始提问中提及漂绿，使用 Gemini 的学生仍有 46% 了解该概念，远高于 ChatGPT 的 15%。研究指出，通用 GenAI 模型鲜少主动揭示投资风险，因此金融科技产品与教育机构需要通过前端设计主动引导（Nudge）投资者提出关键决策问题。 ## Atmosphere & style 整场演讲呈现出严谨、沉稳且条理清晰的学术报告风格。马哈迪亚斯教授以从容专业的语调娓娓道来，从现实痛点层层递进至实验设计与量化统计图表，叙事克制而具说服力。观众在聆听过程中能直观感受到实证研究的严密逻辑，以及对当下普通散户过度依赖通用 AI 工具决策这一现实问题的深切关照。 ## Key takeaways - [00:04](https://www.youtube.com/watch?v=IlKi9r5V7KU&t=4s) 散户投资者在绿色投资中普遍面临两大核心决策风险：企业夸大环保成效的“漂绿”（Greenwashing）行为，以及在相同风险水平下预期平均收益偏低的“收益权衡”（Return tradeoff）问题。 - [00:04](https://www.youtube.com/watch?v=IlKi9r5V7KU&t=4s) 生成式 AI 已成为散户获取投资参考的常见工具，但前期受控实验表明，通用大模型通常极少主动向用户提示漂绿或收益权衡等关键风险。 - [03:52](https://www.youtube.com/watch?v=IlKi9r5V7KU&t=232s) 研究团队针对 272 名本科生开展随机田野实验，在金融素养与商业统计课堂中分别测试了 ChatGPT 与 Google Gemini 在绿色投资学习中的实际表现。 - [03:52](https://www.youtube.com/watch?v=IlKi9r5V7KU&t=232s) 实验在学生与 AI 交互前随机分配三类教学线索（对照组、漂绿提示组、回报提示组），以检验外部轻微干预对学生提示词及学习效果的影响。 - [07:38](https://www.youtube.com/watch?v=IlKi9r5V7KU&t=458s) 前置教学线索能有效激活提问：获得漂绿与回报线索的学生中，分别有 82% 和 69% 在向 AI 发送的首条提示词中主动提及了相关概念。 - [07:38](https://www.youtube.com/watch?v=IlKi9r5V7KU&t=458s) 首条提示词（Initial prompt）对学习成效起决定性作用，主动提及风险概念的学生在交互后掌握相关知识的比例显著更高。 - [07:38](https://www.youtube.com/watch?v=IlKi9r5V7KU&t=458s) 在未主动提及漂绿的学生中，使用 Gemini 的学生报告学到漂绿概念的比例（46%）显著高于 ChatGPT 用户（15%），体现了不同大模型在未引导状态下的输出差异。 - [07:38](https://www.youtube.com/watch?v=IlKi9r5V7KU&t=458s) 学生的专业背景（如是否修读金融素养课）并未显著提高其自发询问回报风险的概率，外部提示词线索才是决定提问行为的核心驱动力。 - [13:40](https://www.youtube.com/watch?v=IlKi9r5V7KU&t=820s) 生成式 AI 辅助投资者教育的上限高度取决于用户的提问质量（Prompting），当前模型尚无法在无引导的情况下完全替代专业理财顾问的全面风险揭示功能。 - [13:40](https://www.youtube.com/watch?v=IlKi9r5V7KU&t=820s) 金融科技产品与教育机构应当在交互界面中引入主动引导机制（Nudge），显式启发投资者向 AI 质询漂绿风险与投资回报权衡。 ## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=IlKi9r5V7KU&t=4s) 研究背景与动机：绿色投资与生成式 AI 的交汇 来自加州州立大学蒙特雷湾分校（California State University, Monterey Bay）商学院的财务学教授玛丽安·马哈迪亚斯（Mary Ann Majadillas）介绍了她与多位商学院同事合作完成的研究。这项研究源于对现实中投资者行为的观察：随着全球对气候变化议题的重视，越来越多的散户投资者希望通过环保或绿色投资来贡献一份力量。然而，欧洲等地的多项田野实验表明，投资者虽然对绿色投资抱有热情，却常常面临两大关键风险： 一是“漂绿”（Greenwashing），即企业或项目夸大自身的环保贡献，实际行为与宣传不符；二是绿色投资通常伴随着风险与回报的权衡（Return tradeoff），在相同风险水平下往往提供相对较低的平均收益，即回报与风险的比率（Reward-to-risk ratio）不如传统投资。 与此同时，生成式 AI（GenAI）已成为普通大众获取日常建议与决策参考的常用工具。许多散户投资者将 ChatGPT 等工具视作“可信赖的顾问”或身边的朋友，直接询问投资建议。马哈迪亚斯的合著者莱斯利·博奈（Leslie Boni）此前发表过一项受控实验研究：博奈通过模拟 100 个不同用户的身份向 AI 提出相同的投资问题，结果发现 AI 很少会主动提及“漂绿”风险或“风险-收益权衡”。基于此，本项研究旨在进一步探究在自然、未受中介干预的情况下，真实用户究竟如何与生成式 AI 互动，以及不同教学提示（Instructional cues）如何影响他们对这些关键决策概念的学习。 --- ### [03:52](https://www.youtube.com/watch?v=IlKi9r5V7KU&t=232s) 实验设计：两阶段课堂田野实验与提示词线索干预 研究团队在大学课堂中开展了一项随机田野实验，共纳入 272 名本科生，分别来自两门课程：金融素养课（Financial Literacy）与商业统计课（Business Statistics），每门课程各 136 名学生。实验分为两个学期进行：2025 年秋季学期共有 182 名学生（约占总样本的三分之二），他们使用的是 ChatGPT；2026 年春季学期共有 65 名学生，他们使用的是 Google Gemini。 ```mermaid flowchart TD accTitle: 课堂随机田野实验流程图 accDescription: 展示272名本科生在不同学期、不同AI模型与三种提示词线索下的分组实验流程。 subgraph Participants ["参与学生 (N = 272)"] F["金融素养课 (n = 136)"] S["商业统计课 (n = 136)"] end subgraph Cohorts ["实验学期与模型"] C1["2025 秋季学期 (ChatGPT)"] C2["2026 春季学期 (Gemini)"] end subgraph Cues ["随机分配的三组教学线索 (1:1:1)"] Q1["对照组：仅基础写作任务线索"] Q2["漂绿组：基础线索 + 漂绿提示"] Q3["回报组：基础线索 + 回报差异提示"] end subgraph Tasks ["实验任务"] T1["任务 1：与 AI 对话并撰写短文"] T2["任务 2：完成后续认知调查问卷"] end Participants --> Cohorts Cohorts --> Cues Cues --> T1 T1 --> T2 ``` 实验的核心干预是在学生向 AI 发送首条提示词（Initial prompt）撰写短文前，随机向其呈现三种不同类型的教学线索（Instructional cues）： - **基础对照组（约 1/3 学生）**：仅获得基础写作任务线索。 - **漂绿线索组（约 1/3 学生）**：在基础线索上增加一条关于漂绿的背景信息，例如“你的朋友记得听过关于‘漂绿’的一些说法”。 - **回报线索组（约 1/3 学生）**：在基础线索上增加一条关于回报的背景信息，例如“你的朋友记得听过关于某些投资回报优于其他投资的说法”。 在利用 AI 完成短文撰写任务后，学生需要填写一份问卷，调查他们在此次 AI 交互过程中是否了解到了“漂绿”和“预期收益较低”这两项关键决策概念。 --- ### [07:38](https://www.youtube.com/watch?v=IlKi9r5V7KU&t=458s) 实验结果：提示词引导、模型差异与课程背景的影响 由于数据权限限制，研究团队仅对问卷调查结果进行了量化统计与分析。整体而言，在未考虑分组差异的情况下，全样本中共有 114 名学生（约 42%）报告 AI 提到了“漂绿”问题，有 117 名学生（约 43%）报告在互动中了解到了“投资回报”相关内容。 | 评估维度 / 实验指标 | 核心统计数据与发现 | |:--- |:--- | | **漂绿线索对首条 Prompt 的影响** | 获得漂绿线索的学生中，约 **82%** 在向 AI 提出的第一个问题中主动提及了“漂绿”。 | | **回报线索对首条 Prompt 的影响** | 获得回报线索的学生中，约 **69%** 在向 AI 提出的第一个问题中主动提及了“投资回报”。 | | **首条 Prompt 对实际学习结果的影响** | 在首条提示词中主动提及相关概念的学生，在交互后报告学到“漂绿”或“风险-回报权衡”的比例显著更高。 | | **AI 模型之间的显著差异 (Gemini vs ChatGPT)** | 在未在首条提示中提及漂绿的学生中，使用 **Gemini** 的学生报告学到漂绿概念的比例（46%）显著高于使用 **ChatGPT** 的学生（15%）。 | | **金融素养课程背景假设的检验** | 原假设金融素养课学生会因专业背景更主动询问回报问题，但实证发现**未获显著支持**；决定学生是否提问的根本驱动力是实验中给出的**教学线索（Prompt cue）**。 | --- ### [13:40](https://www.youtube.com/watch?v=IlKi9r5V7KU&t=820s) 核心启示与未来研究方向 马哈迪亚斯在总结中指出，生成式 AI 虽然具备辅助投资者教育的潜力，但用户最终能学到什么，在极大程度上取决于他们最初如何向系统提问（Prompting）。现阶段的通用大模型不会在每次用户询问投资时都主动提供全面的风险揭示与风险-收益权衡。 基于这些发现，研究团队提出了针对实际应用的建议： - **机构与产品设计层面的引导**：金融教育机构和金融科技应用（如投资类 App）应当主动对投资者进行引导（Nudge），在界面交互中设计专门的 AI 指引机制，显式提示用户去询问“漂绿”风险和“收益权衡”等关键决策概念。 - **未来研究拓展**：目前的实验局限于两个班级的 272 名大学生样本。未来的研究计划扩大参与者的多样性，测试更多主流的大语言模型，并致力于追踪和观察投资者在受到 AI 辅助后的长期实际投资行为。 ## Notable quotes > [00:04](https://www.youtube.com/watch?v=IlKi9r5V7KU&t=4s) 这项研究源于对现实中投资者行为的观察：随着全球对气候变化议题的重视，越来越多的散户投资者希望通过环保或绿色投资来贡献一份力量。 — Mary Ann Majadillas > [00:04](https://www.youtube.com/watch?v=IlKi9r5V7KU&t=4s) 欧洲等地的多项田野实验表明，投资者虽然对绿色投资抱有热情，却常常面临两大关键风险：一是“漂绿”（Greenwashing）……二是绿色投资通常伴随着风险与回报的权衡（Return tradeoff） — Mary Ann Majadillas > [00:04](https://www.youtube.com/watch?v=IlKi9r5V7KU&t=4s) 许多散户投资者将 ChatGPT 等工具视作“可信赖的顾问”或身边的朋友，直接询问投资建议。 — Mary Ann Majadillas > [00:04](https://www.youtube.com/watch?v=IlKi9r5V7KU&t=4s) 博奈通过模拟 100 个不同用户的身份向 AI 提出相同的投资问题，结果发现 AI 很少会主动提及“漂绿”风险或“风险-收益权衡”。 — Mary Ann Majadillas > [03:52](https://www.youtube.com/watch?v=IlKi9r5V7KU&t=232s) 研究团队在大学课堂中开展了一项随机田野实验，共纳入 272 名本科生，分别来自两门课程：金融素养课（Financial Literacy）与商业统计课（Business Statistics） — Mary Ann Majadillas > [07:38](https://www.youtube.com/watch?v=IlKi9r5V7KU&t=458s) 获得漂绿线索的学生中，约 82% 在向 AI 提出的第一个问题中主动提及了“漂绿”。 — Mary Ann Majadillas > [07:38](https://www.youtube.com/watch?v=IlKi9r5V7KU&t=458s) 获得回报线索的学生中，约 69% 在向 AI 提出的第一个问题中主动提及了“投资回报”。 — Mary Ann Majadillas > [07:38](https://www.youtube.com/watch?v=IlKi9r5V7KU&t=458s) 在未在首条提示中提及漂绿的学生中，使用 Gemini 的学生报告学到漂绿概念的比例（46%）显著高于使用 ChatGPT 的学生（15%）。 — Mary Ann Majadillas > [07:38](https://www.youtube.com/watch?v=IlKi9r5V7KU&t=458s) 原假设金融素养课学生会因专业背景更主动询问回报问题，但实证发现未获显著支持；决定学生是否提问的根本驱动力是实验中给出的教学线索（Prompt cue）。 — Mary Ann Majadillas > [13:40](https://www.youtube.com/watch?v=IlKi9r5V7KU&t=820s) 用户最终能学到什么，在极大程度上取决于他们最初如何向系统提问（Prompting）。现阶段的通用大模型不会在每次用户询问投资时都主动提供全面的风险揭示与风险-收益权衡。 — Mary Ann Majadillas > [13:40](https://www.youtube.com/watch?v=IlKi9r5V7KU&t=820s) 金融教育机构和金融科技应用（如投资类 App）应当主动对投资者进行引导（Nudge），在界面交互中设计专门的 AI 指引机制，显式提示用户去询问“漂绿”风险和“收益权衡”等关键决策概念。 — Mary Ann Majadillas ## Mentioned Mary Ann Majadillas, Leslie Boni, California State University, Monterey Bay, ChatGPT, Google Gemini, Ai4 ## Video description Mary Ann Majadillas, Professor of Finance, California State University, Monterey Bay We conduct a randomized classroom-based field experiment to study the unmediated use of ChatGPT and Gemini to learn about green investing. We examine two decision-critical concepts: greenwashing potential and expected lower average returns. Undergraduate students in financial literacy and business statistics courses are randomly assigned one of three instructional cues immediately prior to their GenAI interaction: greenwashing risk, return tradeoff risk, or neither risk. We find that most students who receive the greenwashing or return tradeoff cues reference the concept in their initial prompts with Gen AI and are more likely to report learning about the concept from the GenAI interaction. While only about a third of students who do not reference the returns concept in their initial prompt report learning about the concept regardless of whether they use ChatGPT or Gemini, students who do not reference greenwashing in their initial prompt are significantly more likely using Gemini (46%) than ChatGPT (15%) to report learning about greenwashing. Our research contributes to the literature that examines direct interactions of investors with off-the-shelf GenAI models and their potential to supplement or replace financial advising professionals. [Source](https://www.youtube.com/watch?v=IlKi9r5V7KU)

Source quotes

[0:04] Mary Ann Majadillas: 这项研究源于对现实中投资者行为的观察：随着全球对气候变化议题的重视，越来越多的散户投资者希望通过环保或绿色投资来贡献一份力量。 [0:04] Mary Ann Majadillas: 欧洲等地的多项田野实验表明，投资者虽然对绿色投资抱有热情，却常常面临两大关键风险：一是“漂绿”（Greenwashing）……二是绿色投资通常伴随着风险与回报的权衡（Return tradeoff） [0:04] Mary Ann Majadillas: 许多散户投资者将 ChatGPT 等工具视作“可信赖的顾问”或身边的朋友，直接询问投资建议。 [0:04] Mary Ann Majadillas: 博奈通过模拟 100 个不同用户的身份向 AI 提出相同的投资问题，结果发现 AI 很少会主动提及“漂绿”风险或“风险-收益权衡”。 [3:52] Mary Ann Majadillas: 研究团队在大学课堂中开展了一项随机田野实验，共纳入 272 名本科生，分别来自两门课程：金融素养课（Financial Literacy）与商业统计课（Business Statistics） [7:38] Mary Ann Majadillas: 获得漂绿线索的学生中，约 82% 在向 AI 提出的第一个问题中主动提及了“漂绿”。 [7:38] Mary Ann Majadillas: 获得回报线索的学生中，约 69% 在向 AI 提出的第一个问题中主动提及了“投资回报”。 [7:38] Mary Ann Majadillas: 在未在首条提示中提及漂绿的学生中，使用 Gemini 的学生报告学到漂绿概念的比例（46%）显著高于使用 ChatGPT 的学生（15%）。 [7:38] Mary Ann Majadillas: 原假设金融素养课学生会因专业背景更主动询问回报问题，但实证发现未获显著支持；决定学生是否提问的根本驱动力是实验中给出的教学线索（Prompt cue）。 [13:40] Mary Ann Majadillas: 用户最终能学到什么，在极大程度上取决于他们最初如何向系统提问（Prompting）。现阶段的通用大模型不会在每次用户询问投资时都主动提供全面的风险揭示与风险-收益权衡。 [13:40] Mary Ann Majadillas: 金融教育机构和金融科技应用（如投资类 App）应当主动对投资者进行引导（Nudge），在界面交互中设计专门的 AI 指引机制，显式提示用户去询问“漂绿”风险和“收益权衡”等关键决策概念。

Source quotes

[0:04] Mary Ann Majadillas: It's like asking a friend or asking a trusted advisor—trusted, right? We believe what the AI tells us. [13:40] Mary Ann Majadillas: What users learn depends heavily on how they prompt the system... it won't always give you the complete answer every single time you ask about making investments. [0:04] Mary Ann Majadillas: greenwashing [3:52] Mary Ann Majadillas: Friend heard about greenwashing [3:52] Mary Ann Majadillas: Friend heard about return differences

Original transcript

[0:04] All right. Um, hi everybody. Uh, my name is Marian [0:07] Mahadas and I am very happy to be  here uh, presenting our paper. So, [0:15] this is a co-authored paper with several of  my colleagues at the college of business. [0:23] Many of them are the instructors who teach the  classes where we performed the experiments. [0:34] All right. So I want to start with why we  you know the motivation behind writing this [0:42] paper. Um increasingly there you know we  see a lot of interest from investors about [0:52] uh environmentally friendly investments you  know in other words we all want to help you know [0:59] um sort of combat climate change right um and  so there are many field experiments conducted in [1:08] Europe um where investors say they want to make  green investments, but they also highlight the, [1:19] you know, the fact that they don't  want to make investments, you know, [1:25] uh towards uh projects that don't really uh do  what they claim to do, right? So, in other words, [1:33] that's greenwashing, right? where, you  know, they're claiming to um do more for [1:40] the planet than what they're actually doing.  And a second issue with these investments is [1:47] that they typically offer a lower return for the  same level of risk. Right? So, in other words, [1:55] the return to risk, the reward to risk ratio isn't  as high as you would find in a regular investment. [2:05] So that's one part of it. Now today, right,  with AI, we all do it. We use it for practically [2:17] everything. You know, it's it's just so easy to  ask a quick question. It's like asking a friend [2:21] or asking a trusted advisor, trusted, right?  We believe what they we believe what the AI [2:29] um tells us. And so it's not surprising that many  investors also tend to use AI. And by investors [2:38] I mean us the retail the retail investors. You  know we're like oh hey why don't I ask Chad GBT [2:46] uh where to invest right? Um and so the one  of my co-authors Leslie Bonai um in fact did [2:57] a controlled experiment in a paper that was  published last year. So in that paper Lesie [3:05] was pretending to be 100 different people. She  just typed the same question every single time. [3:14] um and you know just to see what AI would give her  right and you know what she found is that not you [3:25] know she didn't always get any information from  greenwashing about greenwashing and she didn't [3:34] always get information about the tradeoffs  between risk and return. So this paper is [3:43] the natural followup to that paper. You know what  if you just observe how people actually use it. [3:52] Okay. So in the um in the diagram the upper panel  shows you the framework for what how we believe [4:02] investors you know acquire information in the real  world right. Um but because this is a classroom [4:11] experiment, uh the lower panel shows you our  experimental design. So let me talk a little bit [4:21] more um about that. So we conducted a randomized  um classroombased uh field experiment. We had 272 [4:34] undergraduate students from two classes. One is  a financial literacy class and the other one is [4:43] a business statistics class. Uh we had about the  same number of students from each class. So we [4:51] had 136 students from each class and they are from  they the experiments were conducted from fall 2025 [5:03] so one semester fall 2025 and spring 2026 right um  but we had more students participating in the fall [5:15] so we had 182 um students in the fall that's about  2/3 of the sample and we had 65 participating in [5:27] spring 2026. So what's what's also notable  about that split is that the 186 students in [5:37] fall 2025 they used chat GPT and the 65 students  participating in spring 2026 they used Gemini. [5:49] Um so the activities um consisted of two things  right. So in the first activity they were asked to [6:02] uh interact with either Chad GPT or Gemini to  write an essay. So that's the prompt in blue [6:13] uh right up there. So that's the instructional  cue that students got. Onethird of the students [6:21] got only that Q. The other third of the  students got that Q plus one additional [6:31] Q about greenwashing. Right? So specifically  the uh the Q says your friend remembers hearing [6:40] something about greenwashing. Okay. And then the  last third, the last one-third of the class, uh, [6:50] they also got the first prompt, the first cue, and  then they got another one about returns, right? [7:00] Your friend remembers hearing something about some  investments getting better returns than others. [7:09] So given those instructions, you know, they were  told to use AI to basically write that essay. [7:20] Okay. Um and and then after they did that, uh they  were asked to take a survey about the experience. [7:34] So [7:38] this is what we we had four hypotheses  but I'll skip that. I'll I'll [7:46] um I'll mix that in with a discussion of the  um results. So this is you know this is what we [7:55] um observed after conducting the experiment right.  So um before I actually um delve into the results [8:05] um we were uh we were only cleared to use like  um some questions from the survey. We couldn't [8:14] use the essay. So that's why these are the these  are the results that we um present. So big you [8:24] know first let's start big picture um overall  you know if the question asked did gener did [8:32] generative AI mention that greenwashing is  a potential problem for investments of this [8:41] type. So that will be the first um the first  three bars on the left side. So the upper chart [8:52] on the left side. Okay. Um so the answer to  that is 114 students out of the 272 students [9:03] uh said yes. Okay. So they got information about  greenwashing the for the next question about [9:13] returns. Okay. So it's like a similar question.  Did you learn um did you learn about returns in [9:21] your interactions? That's the that's the bottom  panel and also the first um first set of results. [9:31] So out of 272 students, 117 students said yes.  Okay. Now remember all the cues that we provided [9:44] at the beginning of the experiment, right? So  the second bar going back to the upper panel, [9:52] that second set of results, the students who got  the greenwashing Q, they were more likely to ask [10:02] about greenwashing in their first prompt. So  the first in the first question that they asked [10:09] uh AI they asked about greenwashing that's  about um 82% of the um students and we obtain [10:22] a similar result for returns. So the students  who got prompted about returns, you know, [10:30] who got who got a cue about, you know, returns,  they were also more likely to ask about the [10:39] uh returns from green investments in their  first um question, in their first prompt. So [10:47] that's about 69% of the um sample. Now, of those  right of the among those who used who asked about [11:04] greenwashing in their first prompt and among those  who asked about returns in their first prompt, [11:13] they are also more likely to report learning  about greenwashing and learning about [11:22] um the return and risk tradeoff in their  interaction with AI, right? And you know um [11:38] the other the other thing we found here is  that the students were more likely to learn [11:49] about greenwashing from Gemini. So that's  the very last um that's the very last column [11:57] uh on the upper uh on the upper uh figure. So  the spring 2026 students, the students who used [12:10] Gemini were more likely to learn about  greenwashing than the students who used [12:17] um chat GPT uh in their um in the experiment.  And this is a you know we found that this was a [12:28] statistically significant result, right? Um and  finally one of our um one of our hypothesis was [12:42] the students who would be in the financial  literacy class since they are more you know [12:48] since they talk about investments and risk and  return on investments. you know, we were we were [12:55] um wondering we hypothesized that they would be  more likely to ask about uh returns in their uh [13:04] you know in their interactions with AI.  Uh we don't find strong support for that. [13:11] We actually find that getting the queue, you know,  in other words, if they were if they were part of [13:18] the students who got the cue about greenwashing  and returns, that's the driver for asking about um [13:26] you know the returns concept u among the students.  So what are our takeaways from this experiment? [13:40] Um so one of our conclusions is that gener  generative AI can support investor education [13:50] but what users learn depends heavily on how  they prompt the system. Right? So you know [13:59] um in other words it won't always give you  the complete um answer every single time you [14:08] uh you know you ask about making investments. Um  our second result is that organizations should [14:18] teach investors um to ask about greenwashing and  return trade-offs explicitly. So this follows [14:24] from the you know this follows from the first  result. So if you're um you know just in terms [14:32] of practical applications right so if you're  designing an app you might want something that [14:38] you know prompts the users or that nudges the  users towards asking these types of questions. [14:49] And you know another um another one  of our findings support um designing [14:56] AI guidance and financial literacy programs  that surface decision critical um concepts. [15:08] And finally uh some of the future research  that we identified um coming out of this [15:15] uh you know this is a classroom  experiment you know from 272 [15:20] students out of two classes. Um ideally  we want to broaden this out to have [15:27] um you know more diverse uh participant in the  survey and you know using additional AI models [15:38] uh and with the ultimate goal of maybe um being  able to observe long-term investment behavior. [15:52] And thank you very much. That  was my presentation. [applause]