Using Generative AI Systems for Critical Thinking Engagement in an Advanced Chemistry Course：A Case Study
在高级化学课程中使用生成式 AI 系统进行批判性思维参与：案例研究

Cite This：J．Chem．Educ．2024，101，3789－3794
引用此文：化工教育杂志2024,101,3789-3794

Read Online  在线阅读

KEYWORDS：Upper－Level Chemistry，Critical Thinking，Artificial Intelligence，Technology，Generative AI，ChatGPT
关键词：高级化学，批判性思维，人工智能，技术，生成式人工智能，ChatGPT

Generative AI（GenAI）is distinct from well－established traditional AI programs，which perform functions based on algorithms and pattern recognition．This new generation of artificial intelligence works by reviewing large data sets， searching for correlations and patterns，and creating something new in response to user questions called prompts．GenAI－ specific tools，which are already taking hold across numerous functions and professions，represent a disruptive rather than sustaining or incremental force that is reshaping how people write，create，brainstorm，and research．In education，GenAI tools carry the potential to address learning deficiencies from pandemic－modified educational environments at scale and with low costs；in the authors＇experiences，students already use GenAI tools as interactive tutors during studying，and educators experiment with AI to create lesson plans or streamline grading papers．Simultaneously，GenAI carries significant risks of providing inaccurate information，introduc－ ing prejudice due to biased training data，and continuing to erode individual data privacy and data security．${}^1$${}^1$^(1){ }^{1} Educators must thus determine how to leverage these tools effectively while ensuring that students gain the knowledge and skills they will need to be successful．
生成式人工智能（GenAI）不同于成熟的传统人工智能程序，后者根据算法和模式识别执行功能。新一代人工智能的工作原理是审查大型数据集，搜索相关性和模式，并创建一些新的东西来回答用户的问题，称为提示。GenAI 特定的工具已经在许多职能和专业中占据主导地位，代表着一种颠覆性的，而不是持续的或增量的。正在重塑人们写作、创造、集思广益和 research.In 教育方式的力量，GenAI 工具有可能以低成本大规模解决大流行改变的教育环境中的学习缺陷;根据作者的经验，学生在学习期间已经使用 GenAI 工具作为交互式导师，教育工作者尝试使用 AI 来创建课程计划或简化论文评分。由于有偏见的培训数据而引入偏见，并继续侵蚀个人数据隐私和数据安全。 ${}^1$${}^1$^(1){ }^{1} 因此，教育工作者必须确定如何有效地利用这些工具，同时确保学生获得成功所需的知识和技能。

The availability of GenAI programs such as OpenAI＇s ChatGPT，Google＇s Bard（now called Gemini），and Microsoft Copilot has put the question of＂What is it good for？＂in sharp focus in higher education today．A scan of conversations on social media such as Facebook regarding online teaching and GenAI in the classroom reveals that the response of faculty to
OpenAI 的 ChatGPT、Google 的 Bard（现在称为 Gemini）和 Microsoft Copilot 等 GenAI 程序的可用性使“它有什么好处”的问题在当今高等教育中成为人们关注的焦点。对 Facebook 等社交媒体上关于在线教学和课堂上 GenAI 的对话的扫描显示，教师对
this challenge may range from wholehearted adoption to attempts to ban this technology．Many faculty are concerned that students may use these AI programs instead of thinking for themselves．
这种挑战的范围可能从全心全意的采用到试图禁止这项技术。许多教师担心学生可能会使用这些 AI 程序，而不是独立思考。
A first step in determining how GenAI can be used effectively in chemical education has been to explore the accuracy of the generated responses，since GenAI is well known to experience＂hallucinations＂in which it may completely fabricate evidence or references to respond to a prompt．Evaluations of ChatGPT have revealed mixed success in utilizing GenAI to respond correctly to requests for information，${}^{2,3}$${}^{2,3}$^(2,3){ }^{2,3} with a notable weakness in mathematical manipulations．${}^{4,5}$${}^{4,5}$^(4,5){ }^{4,5} As a result of this inconsistent accuracy，it is vital to train students to evaluate GenAI responses critically and not treat the output as expert answers．${}^6$${}^6$^(6){ }^{6}
确定 GenAI 如何有效地用于化学教育的第一步是探索生成的响应的准确性，因为众所周知，GenAI 会经历“幻觉”，在这种情况下，它可能会完全捏造证据或参考资料来响应提示。 ${}^{2,3}$${}^{2,3}$^(2,3){ }^{2,3} ${}^{4,5}$${}^{4,5}$^(4,5){ }^{4,5} 由于这种不一致的准确性，训练学生批判性地评估 GenAI 响应而不是将输出视为专家答案至关重要。 ${}^6$${}^6$^(6){ }^{6}
Although there is widespread agreement that critical thinking is a desirable outcome for students，the field of chemical education encounters a variety of definitions for the term．${}^{7,8}$${}^{7,8}$^(7,8){ }^{7,8} This project specifically uses the Association of American Colleges and Universities（AAC\＆U）definition，${}^9$${}^9$^(9){ }^{9} ＂Critical thinking is a habit of mind characterized by the
尽管人们普遍认为批判性思维是学生的理想结果，但化学教育领域遇到了该术语的各种定义。 ${}^{7,8}$${}^{7,8}$^(7,8){ }^{7,8} 本项目特别使用了美国高等院校协会 （AAC\&U） 的定义， ${}^9$${}^9$^(9){ }^{9} “批判性思维是一种思维习惯，其特征是

comprehensive exploration of issues, ideas, artifacts, and events before accepting or formulating an opinion or conclusion," because it is used across the University of Hartford for assessment purposes. The associated VALUE rubric ${}^9$${}^9$^(9){ }^{9} divides critical thinking into five steps: explaining the issues, providing evidence, examining the influence of context and assumptions, expressing the student’s position, and providing logical conclusions and related outcomes.
在接受或形成意见或结论之前对问题、想法、人工制品和事件进行全面探索“，因为它在整个哈特福德大学都用于评估目的。相关的 VALUE 评分标准 ${}^9$${}^9$^(9){ }^{9} 将批判性思维分为五个步骤：解释问题、提供证据、检查上下文和假设的影响、表达学生的立场以及提供合乎逻辑的结论和相关结果。

The current inconsistent reliability of GenAI output catalyzes new opportunities for students to develop and practice critical thinking skills. ${}^{10-12}$${}^{10-12}$^(10-12){ }^{10-12} Interacting with GenAI, refining a prompt to alter the output, and judging the accuracy of the product provide excellent opportunities for students to analyze a result, evaluate context and bias, and form a conclusion about the extent to which information is correctall of which are key components of critical thinking.
目前 GenAI 输出的可靠性不一致，这为学生发展和练习批判性思维技能提供了新的机会。 ${}^{10-12}$${}^{10-12}$^(10-12){ }^{10-12} 与 GenAI 交互、改进提示以更改输出以及判断产品的准确性为学生提供了分析结果、评估上下文和偏见以及形成关于信息正确程度的结论的绝佳机会所有这些都是批判性思维的关键组成部分。

The University of Hartford is a private comprehensive university in New England with a $40\mathrm{\%}$$40\mathrm{\%}$40%40 \% representation by students of color among the undergraduate population. Assessment of student learning outcomes at the University consists of two components. The first set is the Essential Learning Outcomes (ELOs), including critical thinking, oral communication, written communication, and teamwork, which are universal for all undergraduate students. The ELOs are assessed on a four-year cycle using the appropriate AAC&U VALUE rubrics. ${}^9$${}^9$^(9){ }^{9} Separate Program Learning Outcomes (PLOs) are specific to each major and are assessed approximately annually. For the chemistry and biochemistry majors, critical thinking, oral communication, written communication, and teamwork ELOs are a subset of the PLOs, which also include science-specific outcomes such as competence with experimental techniques, instrumentation, and safety procedures. As part of our regular “closing the loop” exercises, in which the assessment methods are themselves evaluated, the chemistry faculty discussed the need to redesign our existing assessments for critical thinking.
哈特福德大学是新英格兰的一所私立综合性大学， $40\mathrm{\%}$$40\mathrm{\%}$40%40 \% 本科生中有色人种学生。对学生在大学的学习成果的评估包括两个部分。第一组是基本学习成果 （ELO），包括批判性思维、口头交流、书面交流和团队合作，适用于所有本科生。ELO 使用适当的 AAC&U VALUE 评分标准以四年为周期进行评估。 ${}^9$${}^9$^(9){ }^{9} 单独的课程学习成果 （PLO） 特定于每个专业，大约每年评估一次。对于化学和生物化学专业，批判性思维、口头交流、书面交流和团队合作 ELO 是 PLO 的一个子集，其中还包括科学特定的结果，例如实验技术、仪器和安全程序的能力。作为我们定期的“闭环”练习的一部分，其中评估方法本身被评估，化学教师讨论了重新设计我们现有的批判性思维评估的必要性。

This project was implemented in Fall 2023 in a typical 400level Inorganic Chemistry course covering topics including group theory, coordination chemistry, and organometallic chemistry. Assessments in the course included nine problem sets, two GenAI assignments, three in-class exams, and one final exam. Each problem set, in addition to practice questions based directly on class content, included between one and three questions requiring students to analyze the primary literature. Examples of these literature questions are included in the Supporting Information.
该项目于 2023 年秋季在典型的 400 级无机化学课程中实施，涵盖的主题包括群论、配位化学和有机金属化学。该课程的评估包括 9 个问题集、2 个 GenAI 作业、3 个课堂考试和 1 个期末考试。除了直接基于课堂内容的练习题外，每组问题还包括一到三个问题，要求学生分析主要文献。这些文献问题的示例包含在支持信息中。

The nine students in the class were five junior and four senior chemistry and biochemistry students, with $67\mathrm{\%}$$67\mathrm{\%}$67%67 \% (6) of the students identifying as BIPOC (Black, Indigenous, and other People of Color), $33\mathrm{\%}$$33\mathrm{\%}$33%33 \% (3) identifying as male, and $67\mathrm{\%}$$67\mathrm{\%}$67%67 \% (6) identifying as female. Participation in the project was voluntary, but the entire class opted to allow their responses to be collected and analyzed outside of the grading process. Each assignment required the students to interact directly with a Generative AI system rather than being supplied with instructor-produced material. The activities were also designed so that the students would possess knowledge either through access to an original article (C&EN article summary, literature question) or from class materials (literature question, titration data analysis) to assess the accuracy of the responses. Free versions of Generative AI were used (i.e., ChatGPT 3.5 and Bard, now known as Gemini) to ensure that students could participate without added expense. After the first application, the students engaged actively in the overall research project both in terms of designing subsequent applications and in discussing their findings. Three different applications of GenAI were explored.
班上的九名学生是五名初中和四名高级化学和生物化学学生，其中 $67\mathrm{\%}$$67\mathrm{\%}$67%67 \% （6） 名学生被认定为 BIPOC（黑人、土著和其他有色人种）， $33\mathrm{\%}$$33\mathrm{\%}$33%33 \% （3） 被认定为男性，以及 $67\mathrm{\%}$$67\mathrm{\%}$67%67 \% （6） 被认定为女性。参与该项目是自愿的，但全班同学都选择允许在评分过程之外收集和分析他们的回答。每项作业都要求学生直接与生成式 AI 系统交互，而不是提供教师制作的材料。这些活动的设计还使学生通过访问原始文章（C&EN 文章摘要、文学问题）或课堂材料（文学问题、滴定数据分析）来掌握知识，以评估回答的准确性。使用了生成式 AI 的免费版本（即 ChatGPT 3.5 和 Bard，现在称为 Gemini）来确保学生可以在不增加费用的情况下参与。在第一次申请后，学生们积极参与了整个研究项目，包括设计后续申请和讨论他们的发现。探索了 GenAI 的三种不同应用。

Analysis of the student results was carried out using several methods. The students’ written work was collected, and responses were first reviewed to identify keywords or ideas common to several submissions to establish potential themes, such as initial assumptions, conclusions, lessons learned, and concerns about lack of depth, lack of analysis, or plagiarism. The entire set of submissions was then reviewed a second time to identify what percentage of students mentioned these themes. The students’ work for each assignment was copied and kept by the instructor but was coded only after final grades were submitted. The students each signed a release agreeing to this process in advance, although as co-authors this release was redundant. Two open-ended questions were added to the end-of-semester student teaching evaluations to allow students the opportunity for anonymous comments. These data were not available until the final grades were submitted.
使用多种方法对学生成绩进行分析。收集学生的书面作业，并首先审查回复，以确定几篇提交作品的共同关键字或想法，以确定潜在的主题，例如最初的假设、结论、吸取的经验教训，以及对缺乏深度、缺乏分析或抄袭的担忧。然后对整套提交的内容进行第二次审查，以确定有多少学生提到这些主题。教师会复制并保存学生每项作业的作业，但只有在提交最终成绩后才会进行编码。学生们都提前签署了一份同意这一过程的新闻稿，尽管作为合著者，这份新闻稿是多余的。在学期末的学生教学评估中增加了两个开放式问题，让学生有机会进行匿名评论。在提交最终成绩之前，这些数据是不可用的。

In the first application, included in the Supporting Information along with the grading rubric, the students each selected an article of at least 2000 words from C&EN (https:// cen.acs.org/) and asked ChatGPT to provide a summary. Some minimal guidance was provided for the initial prompt, “Acting as [Role] [Perform Task] in [Format]”, which the students could then adapt, such as, “Acting as a student, write a summary of this article in paragraph form.” The students were encouraged to take advantage of the chat function to refine the output. Their reports consisted of responding to a series of questions requiring a critical analysis of the result. The questions were adapted from the AAC&U VALUE rubric for critical thinking ${}^9$${}^9$^(9){ }^{9} to lead the students through the analytical process. Key components of critical thinking include examining their own assumptions, evaluating the factual accuracy of information, supporting their analyses with evidence, reaching a conclusion, and weighing the influence of the context and assumptions. Because two students did not complete the assignment as directed, only seven student responses were included in the analysis.
在第一份申请中，包含在支持信息和评分标准中，学生们每人从 C&EN 中选择一篇至少 2000 字的文章（https:// cen.acs.org/），并要求 ChatGPT 提供摘要。为初始提示“以 [格式] 充当 [角色] [执行任务]”提供了一些最低限度的指导，然后学生可以对其进行调整，例如，“以学生身份，以段落形式写这篇文章的摘要。鼓励学生利用聊天功能来优化输出。他们的报告包括回答一系列需要对结果进行批判性分析的问题。这些问题改编自 AAC&U 批判性思维 ${}^9$${}^9$^(9){ }^{9} 价值标准，以引导学生完成分析过程。批判性思维的关键组成部分包括检查自己的假设、评估信息的事实准确性、用证据支持他们的分析、得出结论以及权衡上下文和假设的影响。由于两名学生没有按照指示完成作业，因此分析中仅包括 7 名学生的回答。

The students designed the second application to be included as part of a homework problem set. They recommended testing GenAI for its data analysis capabilities. Because eight of the
学生们设计了第二个应用程序，作为家庭作业问题集的一部分。他们建议测试 GenAI 的数据分析功能。因为 8 个
nine students were simultaneously taking Quantitative Analysis, the students suggested using their titration data from that course as the basis for the second exercise. The analysis of those data for the other class had already been completed and graded, so there was no question that the GenAI analysis would be submitted in place of the students’ own work. The assignment directions were general, asking students what kinds of questions the GenAI could answer correctly or at all regarding a set of titration data, which allowed students to do their own exploration. Seven of the nine total students submitted work for this question.
9 名学生同时参加定量分析，学生们建议使用该课程中的滴定数据作为第二次练习的基础。对另一个班级的这些数据的分析已经完成并进行了评分，因此毫无疑问，GenAI 分析将代替学生自己的作业提交。作业方向是一般性的，询问学生 GenAI 可以正确回答哪些类型的问题，或者可以正确回答有关一组滴定数据的问题，这使学生可以进行自己的探索。总共 9 名学生中有 7 人提交了这个问题的作业。

Exposing the students to peer-reviewed chemical literature is a course objective for Inorganic Chemistry. Each of the nine problem sets included at least one multipart question requiring students to read and understand an article from the primary literature. For the final GenAI critical thinking analysis application, each student was reassigned one of these literature questions to see how well the AI could find specific answers within an article. Three pairs of students among the nine students in the course received the same questions but were specifically assigned either ChatGPT or Google’s Bard (Gemini), to compare the performance and user-friendliness of two different systems. The students submitted reports similar to those for the first assignment using the same rubric. Seven responses were submitted and analyzed for this assignment.
让学生接触同行评审的化学文献是无机化学的课程目标。九个问题集中的每一个都包括至少一个由多个部分组成的问题，要求学生阅读和理解原始文献中的一篇文章。对于最终的 GenAI 批判性思维分析应用程序，每个学生都被重新分配了其中一个文献问题，以了解 AI 在文章中找到特定答案的能力。该课程的 9 名学生中有三对学生收到了相同的问题，但专门分配了 ChatGPT 或 Google 的 Bard （Gemini），以比较两个不同系统的性能和用户友好性。学生使用相同的评分标准提交了与第一次作业类似的报告。为此作业提交并分析了 7 份回复。

The students’ critical assessment of GenAI’s performance was compiled through a combination of reviewing the students’ submissions, as described above, and conducting discussions with the entire class following each assignment.
如上所述，学生对 GenAI 表现的批判性评估是通过审查学生提交的内容并在每项作业后与全班进行讨论来编制的。

Five of seven students concluded that GenAI’s best application was to distill key findings of existing materials, particularly in cases where an individual might not have time to read the articles in depth. Factually, they found that the summaries were largely accurate, although 86% (6 of 7) of the students noted that they would have selected different key points to emphasize and might have added more depth. For example, when ChatGPT reported that an article “…explores the recent advancements and challenges in the field of mitochondrial DNA editing”, the student would have preferred the added context from the article that “mitochondrial diseases affect 1 in every $4,300-6,700$$4,300-6,700$4,300-6,7004,300-6,700 people.” A second student noted that ChatGPT seemed to select keywords and then add filler words around them such as, “I find the information about the resurgence of small-molecule drugs and their evolving role in the pharmaceutical industry quite fascinating.” Another observation from $71\mathrm{\%}$$71\mathrm{\%}$71%71 \% (5) of the students was that the summaries were superficial and lacked analysis and deeper understanding, although some of that is simply the limitation of a summary.
七名学生中有五名得出结论，GenAI 的最佳应用是从现有材料中提取关键发现，尤其是在个人可能没有时间深入阅读文章的情况下。事实上，他们发现这些总结在很大程度上是准确的，尽管 86%（7 人中有 6 人）的学生指出他们会选择不同的关键点来强调，并且可能会增加更多的深度。例如，当 ChatGPT 报告一篇文章“......探讨了线粒体 DNA 编辑领域的最新进展和挑战“，学生会更喜欢文章中添加的上下文”线粒体疾病影响每个人 $4,300-6,700$$4,300-6,700$4,300-6,7004,300-6,700 中的 1 人”。“ 第二名学生指出，ChatGPT 似乎会选择关键字，然后在它们周围添加填充词，例如，”我发现有关小分子药物复兴及其在制药行业中不断发展的作用的信息非常有趣。（5） 名学生的另一个 $71\mathrm{\%}$$71\mathrm{\%}$71%71 \% 观察是，摘要很肤浅，缺乏分析和更深入的理解，尽管其中一些只是摘要的局限性。

The students all consistently emphasized the need for skepticism about the accuracy of the GenAI output, and they universally stated that the original article should always be consulted to review in-depth information because the AI was not always reliable. They ( $71\mathrm{\%},5$$71\mathrm{\%},5$71%,571 \%, 5 ) were also concerned about the risk of plagiarism and noted sentences and phrases that in their opinion had been inadequately paraphrased from the original articles.
学生们都一直强调需要对 GenAI 输出的准确性持怀疑态度，他们普遍表示，应始终参考原始文章以审查深入信息，因为 AI 并不总是可靠的。他们 （ $71\mathrm{\%},5$$71\mathrm{\%},5$71%,571 \%, 5 ） 还担心抄袭的风险，并注意到他们认为从原始文章中转述不当的句子和短语。

In the second application, in which GenAI was used to analyze titration data, the students variously asked ChatGPT to calculate molarities, provide sample calculations, and provide balanced chemical reactions for their titrations. In comparing the GenAI results against their own work, two of seven students indicated their numbers matched those from the AI, two students found the AI answers to be inaccurate, and the other three students did not mention validating the calculations. These findings are consistent with other reports that GenAI does not always reliably produce mathematically accurate calculations. ${}^{4,5}$${}^{4,5}$^(4,5){ }^{4,5} All of the students pointed out that the free version, ChatGPT 3.5, did not have the ability to create graphs from entered data. If a graph was presented, two students noted, it was taken from an Internet source using different data rather than the data provided. In $43\mathrm{\%}$$43\mathrm{\%}$43%43 \% (3) of the responses, the AI did, however, provide instructions for how to create a graph using either Excel or Google sheets, although these instructions were not tested by the students.
在第二个应用程序中，使用 GenAI 分析滴定数据，学生们以各种方式要求 ChatGPT 计算摩尔浓度，提供样本计算，并为他们的滴定提供平衡的化学反应。在将 GenAI 结果与他们自己的工作进行比较时，七名学生中有两名表示他们的数字与 AI 的数字相匹配，两名学生认为 AI 的答案不准确，其他三名学生没有提到验证计算。这些发现与其他报告一致，即 GenAI 并不总是可靠地产生数学上准确的计算。 ${}^{4,5}$${}^{4,5}$^(4,5){ }^{4,5} 所有学生都指出，免费版 ChatGPT 3.5 没有从输入的数据创建图表的能力。两名学生指出，如果提供了一张图表，它是使用不同的数据而不是提供的数据从互联网来源获取的。然而，在第 （3） 项回答中 $43\mathrm{\%}$$43\mathrm{\%}$43%43 \% ，AI 确实提供了如何使用 Excel 或 Google 表格创建图表的说明，尽管这些说明没有经过学生的测试。

Nearly all ( $86\mathrm{\%},6$$86\mathrm{\%},6$86%,686 \%, 6 of 7 ) of the students commented on the limitations of Generative AI with respect to laboratory data. One drew a strong distinction between performing calculations and analyzing the results. Generative AI is adept at the first, albeit with varying accuracy, but is unable to do the second. In an interesting twist, one student contributed what appeared to be an AI-generated response to a question of its own limitations: GenAI “…cannot analyze or interpret specific experimental results without access to the actual data.” Each student had a complete set of data available to share with the AI, suggesting that these words were not written by the student.
几乎所有（ $86\mathrm{\%},6$$86\mathrm{\%},6$86%,686 \%, 6 共 7 名）学生都评论了生成式 AI 在实验室数据方面的局限性。其中一项研究明确区分了执行计算和分析结果。生成式 AI 擅长第一个，尽管准确性各不相同，但无法做到第二个。有趣的是，一名学生贡献了似乎是 AI 生成的回答，以回答一个关于其自身局限性的问题：GenAI“......无法分析或解释特定的实验结果，而不能获得实际数据。每个学生都有一套完整的数据可以与 AI 共享，这表明这些单词不是由学生编写的。

In this final application, $57\mathrm{\%}$$57\mathrm{\%}$57%57 \% (4 of 7) of the students commented on the importance of a good prompt to obtain the answers they sought, up from $29\mathrm{\%}$$29\mathrm{\%}$29%29 \% ( 2 of 7 ) who commented on this issue in the first assignment. Reviewing the commentary from the three students who detailed their chat conversations with their respective GenAI, the students were less concerned with refining the audience or role part of the prompt (i.e., “as a chemistry major [role] write a summary for a professor [audience]”) but rather asked repeatedly for the GenAI to shorten its overly long initial responses. For example, one student pasted in the abstract for an article and prompted, “What problem are the authors trying to solve?” which resulted in a four-paragraph response. The student then asked the AI to shorten the answer, which again resulted in multiple paragraphs. When the student finally asked the AI to describe the problem in a few sentences, the student deemed the response to be of the appropriate length. The substantial impact of a well-constructed prompt on producing accurate results in a desirable format has been well-established by other groups, ${}^{12,13}$${}^{12,13}$^(12,13){ }^{12,13} and even without guidance, the students began to perform prompt engineering ${}^{10,14}$${}^{10,14}$^(10,14){ }^{10,14} to optimize their answers.
在这份最终申请中， $57\mathrm{\%}$$57\mathrm{\%}$57%57 \% （7 人中的 4 人）的学生评论了良好的提示以获得他们所寻求的答案的重要性，高于 $29\mathrm{\%}$$29\mathrm{\%}$29%29 \% （7 人中的 2 人） 在第一份作业中对这个问题发表了评论。回顾三名学生的评论，他们详细介绍了他们与各自 GenAI 的聊天对话，学生们不太关心改进提示的受众或角色部分（即，“作为化学专业的 [角色] 为教授 [受众] 写一篇总结”），而是反复要求 GenAI 缩短其过长的初始回复。例如，一名学生在一篇文章的摘要中粘贴并提示：“作者试图解决什么问题？”，结果是一个四段式的回答。然后，该学生要求 AI 缩短答案，这再次导致多个段落。当学生最终要求 AI 用几句话描述问题时，学生认为回答的长度合适。其他小组已经明确了构建良好的提示对以理想格式产生准确结果的实质性影响， ${}^{12,13}$${}^{12,13}$^(12,13){ }^{12,13} 即使没有指导，学生也开始进行提示工程 ${}^{10,14}$${}^{10,14}$^(10,14){ }^{10,14} 以优化他们的答案。

To analyze an article, the students learned that GenAI cannot currently access an article via a citation. Because the systems have limits for how many words may be pasted in at one time, the students had to feed in the articles section by section, often performing their own preliminary analysis of the article to determine in which section their answer was likely to be found and then supplying the relevant text to the AI. Having put forth the effort of entering their article text, the students got very frustrated when the GenAI produced answers to their
为了分析一篇文章，学生们了解到 GenAI 目前无法通过引用访问文章。由于系统对一次可以粘贴的单词数量有限制，因此学生必须逐节输入文章，通常对文章进行自己的初步分析，以确定他们的答案可能在哪个部分找到，然后将相关文本提供给 AI。在努力输入他们的文章文本后，当 GenAI 为他们的 GenAI 生成答案时，学生们感到非常沮丧