Chatbots Fail Basic Circuit Math in NC A&T Engineering Study
Ai4
Organizations risk structural and financial liability unless they evaluate AI readiness by a worker's ability to verify flawed calculations rather than mere prompting skill.
## Background
Published on YouTube by Ai4 on August 14, 2026, this presentation features Christopher Horne, Teaching Associate Professor at North Carolina A&T State University. The talk explores how 79 engineering students evaluated AI chatbot solutions for RL and RLC circuit problems, addressing common AI pitfalls like missing assumptions and unit errors while introducing a practical five-step verification framework.
## TL;DR
Christopher Horne examines why generative AI must be treated as a collaborative thinking partner rather than an infallible answering machine in technical disciplines. Drawing on a joint study of 79 engineering students solving circuit problems, the session highlights frequent LLM failure modes, such as misinterpreting 2D diagrams, dropping initial boundary conditions, and confusing physical units. Horne warns that blind trust in polished AI outputs introduces major operational and liability risks in commercial engineering. To counter this, he outlines a five-step framework that prioritizes manual problem-solving, context enrichment, and rigorous human-in-the-loop verification over raw prompting speed.
## Atmosphere & style
The presentation carries a grounded, pedagogical, and pragmatic tone, balancing academic research with relatable real-world anecdotes from test grading and home renovations. Horne delivers a candid reality check on AI hallucinations, replacing uncritical hype with sober engineering discipline. The delivery feels engaging, thoughtful, and constructive, guiding the audience from student frustrations to structured, actionable verification habits.
## Key takeaways
- [00:04](https://www.youtube.com/watch?v=Do_E3KX8QGM&t=4s) Organizational AI readiness should be measured by the ability of users to verify outputs rather than mere access to AI tools or prompting speed.
- [00:04](https://www.youtube.com/watch?v=Do_E3KX8QGM&t=4s) Generative AI often produces polished, authoritative-sounding technical solutions that conceal missing assumptions, unit errors, and incomplete reasoning.
- [00:04](https://www.youtube.com/watch?v=Do_E3KX8QGM&t=4s) Manual problem-solving and handwriting build essential cognitive mental models that shortcut tools bypass, directly correlating with higher academic performance.
- [05:04](https://www.youtube.com/watch?v=Do_E3KX8QGM&t=304s) In a study of 79 engineering students solving RL and RLC circuits, AI served as an effective learning reinforcer only for those who solved the problems manually first.
- [05:04](https://www.youtube.com/watch?v=Do_E3KX8QGM&t=304s) Large language models frequently fail to interpret visual 2D circuit diagrams and complex mathematical formulas accurately, returning broken text or improper LaTeX code.
- [14:06](https://www.youtube.com/watch?v=Do_E3KX8QGM&t=846s) Vague prompts lacking initial physical conditions cause AI models to overlook critical engineering constraints such as structural load, budgets, and schedules.
- [14:06](https://www.youtube.com/watch?v=Do_E3KX8QGM&t=846s) A pervasive culture of blind trust in smartphone-driven quick answers leads users to accept flawed chatbot calculations without questioning basic logic or units.
- [14:06](https://www.youtube.com/watch?v=Do_E3KX8QGM&t=846s) AI is best utilized as an early-stage brainstorming partner rather than a definitive answering machine for exact technical calculations.
- [17:50](https://www.youtube.com/watch?v=Do_E3KX8QGM&t=1070s) Relying on unverified AI solutions in commercial engineering and infrastructure risks catastrophic structural failures, legal liabilities, and contractual breaches.
- [17:50](https://www.youtube.com/watch?v=Do_E3KX8QGM&t=1070s) A five-step verification framework ensures accuracy by requiring manual preliminary reasoning, rich context, mathematical audits, proactive error prompts, and final human judgment.
- [17:50](https://www.youtube.com/watch?v=Do_E3KX8QGM&t=1070s) Organizations must establish clear guidelines and structured evaluation rubrics that reward rigorous verification over raw completion speed.
## Detailed walkthrough
Here is a detailed, faithful walkthrough of the transcript:
[00:04](https://www.youtube.com/watch?v=Do_E3KX8QGM&t=4s) AI as an Answering Machine vs. Thinking Partner
Connecting his talk directly to themes raised by earlier speakers like Leila and Laeticia—such as keeping the human as the final decision-maker and recognizing AI's tendency to hallucinate—Christopher Horne stresses that users cannot rely on AI alone to truly understand complex material. When generative AI presents polished equations or error formulas, students should consult textbooks rather than treating chatbots as an unquestioned authority. The core dilemma driving his research is whether students actually grasp the underlying principles when ChatGPT provides a correct answer, or if they are simply benefiting from a superficial "quick fix." While students eagerly rushed to tools like ChatGPT following its release in 2023, access alone is insufficient without the essential ability to verify if an answer is flawed.
To illustrate how easily AI can fail in practical, real-world contexts, Horne shares personal anecdotes from both his consulting work and teaching responsibilities. While remodeling part of his home, he provided ChatGPT with details and photos to calculate the required number of electrical receptacles according to national standards, but the bot delivered an incorrect answer due to his rushed workflow. Similarly, when generating an answer key for a test he had already validated, ChatGPT produced an answer key filled with errors. These examples highlight how fast, polished AI outputs frequently conceal missing assumptions, unit errors (such as confusing amperes with volts or Fahrenheit with Celsius), and incomplete step-by-step reasoning.
```mermaid
flowchart TD
subgraph Mental ["1. Mindful Grounding"]
A["Manual Problem Solving"] --> B["Handwritten Notes & Diagrams"]
B --> C["Build Cognitive Mental Model"]
end
subgraph Verification ["2. Critical Evaluation"]
C --> D["Prompt AI with Context"]
D --> E["Verify Calculations & Units"]
E --> F["Human Judgment & Final Output"]
end
accTitle: Workflow for verifying AI solutions in engineering
```
- Handwriting and manual problem-solving activate neural pathways in the brain far more effectively than typing or taking pictures of classroom slides.
- A study conducted with 100 students revealed that those who took handwritten notes achieved grades that were one full letter grade higher on average.
- Technical problem-solving in fields like engineering, physics, and computer science requires rigorous critical thinking that shortcut tools often bypass.
[05:04](https://www.youtube.com/watch?v=Do_E3KX8QGM&t=304s) Evaluating Chatbot Accuracy in Engineering Circuits
To evaluate how students interact with generative AI, Horne collaborated with a professor in Mexico on a joint study involving 79 engineering and physics students across three course sections. The study focused on fundamental resistor-inductor (RL) and resistor-inductor-capacitor (RLC) circuits—the core electronic building blocks found in cell phones, aircraft, HVAC systems, robots, drones, and spacecraft. Students were instructed to solve for quantities like voltage or current manually first to establish a solid mental model before running the identical problem through ChatGPT to compare results.
The reflection essays submitted by students revealed significant frustration with ChatGPT's performance. A common issue arose when students uploaded 2D images of circuit diagrams, expecting the large language model (LLM) to interpret the visual layout, which led to incorrect calculations and failed outputs. Furthermore, because LLMs process words rather than structured mathematical visual data, students who uploaded images of integrals or calculus equations often received confusing LaTeX code or plain text substitutes. Students who completed their manual calculations first were able to use ChatGPT as a reinforcing tool, whereas those who attempted to rely on the bot as an immediate "answering machine" struggled to develop baseline comprehension.
- The study analyzed student performance across two primary circuit architectures: RL (resistor-inductor) and RLC (resistor-inductor-capacitor) circuits.
- Roughly 20% of students viewed AI as a personal tutor or study companion, a trend particularly notable among introverted students who preferred solitary study habits following the COVID-19 pandemic.
- A student participant subsequently developed a custom Python program paired with a specialized database of circuit types to significantly improve AI contextual rendering and accuracy.
[14:06](https://www.youtube.com/watch?v=Do_E3KX8QGM&t=846s) Common Technical Failures: Units, Conditions, and Blind Trust
When technical prompts lack rich context, AI chatbots consistently generate incomplete or incorrect solutions while maintaining a deceptively confident tone. For instance, failing to specify initial physical conditions—such as traffic load variations over time when calculating bridge deflections—leads to answers that ignore crucial practical factors like budgets, schedules, and structural limits. Additionally, chatbots routinely misinterpret fundamental physical units, confusing pressure with force or improperly translating engineering units.
A major barrier to effective AI use is that many students grant blind trust to chatbots. Accustomed to immediate answers from their smartphones, they treat the AI as an infallible machine, rarely questioning its logic or verifying its mathematical steps. Horne stresses that users must actively challenge AI outputs with concrete facts rather than passive acceptance. While AI excels at early-stage brainstorming, broad organization, and expanding divergent ideas, fine-grained calculations and critical engineering choices require human oversight, peer-reviewed literature, and standard reference textbooks.
- Chatbots frequently confuse distinct physical concepts, such as confusing force with pressure or mixing up temperature and electrical units.
- In vague prompts where students failed to specify their academic discipline or background context, output quality degraded significantly.
- Horne references a historical corporate incident where Samsung engineers accidentally leaked internal company data by entering sensitive details directly into ChatGPT.
[17:50](https://www.youtube.com/watch?v=Do_E3KX8QGM&t=1070s) Organizational Risks and the Five-Step Verification Framework
The propensity of AI to output flawed technical analysis carries severe consequences beyond the classroom. In commercial engineering, construction, and infrastructure projects, relying on unverified AI calculations can result in structural failures, severe financial liability, breach of contract, and ruined corporate reputations. Because employees frequently lack the habits required to challenge AI assumptions, verify sources, or check math, organizations face operational risks if they reward speed over strict precision.
To mitigate these risks, organizations and educational institutions must transition from measuring AI readiness by mere access or prompting skill to prioritizing output verification. Horne concludes by presenting a practical five-step framework designed to enforce critical evaluation:
1. **Begin with your own reasoning:** Analyze the problem manually and build a preliminary framework before engaging AI tools.
2. **Provide detailed context:** Supply the chatbot with specific background constraints, environmental initial conditions, and domain roles.
3. **Verify facts and calculations:** Audit formulas, check physical unit consistency, and evaluate raw numerical outputs against trusted texts.
4. **Ask what could be wrong:** Actively prompt the AI to identify potential edge cases, missing assumptions, and structural flaws in its own output.
5. **Apply human judgment before action:** Ensure that every final decision remains strictly under human control ("the final choice is yours, not the bot's").
| Verification Metric | Unchecked AI Output | Verified Human-in-the-Loop Process |
|:--- |:--- |:--- |
| **Handling of Units** | Frequently mixes up units (e.g., pressure vs. force, volts vs. amps) | Manually audited for standard dimensional accuracy |
| **Visual Interpretations** | Misinterprets 2D diagrams, returning broken text or LaTeX | Diagram converted into structured technical inputs |
| **Trust Model** | Passive acceptance based on conversational tone | Active challenge of logic using peer-reviewed sources |
| **Primary Utility** | Superficially used as a quick "answering machine" | Leveraged as an interactive brainstorming partner |
- Establish clear organizational guidelines that explicitly define when human review of AI-generated content is strictly mandatory.
- Implement comprehensive evaluation rubrics to score AI outputs based on accuracy, underlying assumptions, source validity, and risk exposure.
- Caution must be exercised when using AI to grade or generate rubrics, as the bot can make errors when establishing evaluation criteria.
- Organizations must incentivize thorough verification and critical review habits rather than prioritizing raw speed alone.
## Notable quotes
> AI can produce polished technical answers that sound convincing even when they are wrong. — Christopher Horne
> verification skills proved more important than prompting skills alone — Christopher Horne
> [17:50](https://www.youtube.com/watch?v=Do_E3KX8QGM&t=1070s) AI readiness should be measured by the ability to verify AI output, not merely access to AI. — Christopher Horne
> [00:04](https://www.youtube.com/watch?v=Do_E3KX8QGM&t=4s) quick fix — Christopher Horne
> [00:04](https://www.youtube.com/watch?v=Do_E3KX8QGM&t=4s) answering machine — Christopher Horne
> [17:50](https://www.youtube.com/watch?v=Do_E3KX8QGM&t=1070s) the final choice is yours, not the bot's — Christopher Horne
> [17:50](https://www.youtube.com/watch?v=Do_E3KX8QGM&t=1070s) Begin with your own reasoning — Christopher Horne
> [17:50](https://www.youtube.com/watch?v=Do_E3KX8QGM&t=1070s) Provide detailed context — Christopher Horne
> [17:50](https://www.youtube.com/watch?v=Do_E3KX8QGM&t=1070s) Verify facts and calculations — Christopher Horne
> [17:50](https://www.youtube.com/watch?v=Do_E3KX8QGM&t=1070s) Ask what could be wrong — Christopher Horne
> [17:50](https://www.youtube.com/watch?v=Do_E3KX8QGM&t=1070s) Apply human judgment before action — Christopher Horne
## Mentioned
Christopher Horne, North Carolina A&T State University, Ai4, ChatGPT, Leila, Laeticia, Samsung, Python, LaTeX
## Video description
Christopher Horne, Teaching Associate Professor, North Carolina A&T State University
AI can produce polished technical answers that sound convincing even when they are wrong. This talk examines how 79 engineering students used AI chatbots to solve RL and RLC circuit problems and what happened when they were asked to verify the AI's work instead of simply accepting it. The talk will highlight where AI supported learning, where it failed through missing assumptions, unit errors, and incomplete reasoning, and why verification skills proved more important than prompting skills alone. Attendees will leave with a practical five-step framework for helping students, employees, and organizations use AI more effectively by developing habits of critical evaluation rather than blind trust. If your organization is adopting AI for technical or knowledge-based work, this session will show why AI readiness should be measured by the ability to verify AI output, not merely access to AI.
NC A&T工程研究：聊天机器人基础电路数学计算频现错误
机构若不以员工核验错误计算的能力、而是仅凭提示词技能来评估其AI就绪度，将面临结构性与财务责任风险。
Christopher Horne 深入探讨了为何在技术学科中必须将生成式 AI 视为协同思考的伙伴，而非绝对可靠的答题机器。结合一项针对 79 名工程专业学生求解电路问题的联合研究，该内容揭示了大型语言模型（LLM）常见的失效模式，例如误读二维图表、遗漏初始边界条件以及混淆物理单位。Horne 警告称，盲目信任看似严谨流畅的 AI 输出，会在商业工程领域引入巨大的运营与法律责任风险。为此，他提出了一个五步框架，强调应将手动解题、丰富上下文信息以及严格的人在回路（human-in-the-loop）验证置于单纯追求提示词生成速度之上。
Christopher Horne 指出，衡量组织和个人的 AI 就绪度不应仅看访问权限或提示词技巧，而应取决于验证 AI 输出真伪与缺陷的能力。 (0:04)
在一项针对 79 名工程专业学生的电路解题研究中，只有先进行手写推导建立心智模型的学生才能有效利用 AI，直接将 AI 当作“答题机”则难以建立基础理解。 (5:04)
大语言模型在解析 2D 电路图和复杂数学公式时经常失败，且在缺乏初始物理条件时极易混淆力与压强、安培与伏特等物理单位。 (14:06)
为防范工程事故与商业风险，Horne 提出了五步验证框架：独立思考先行、提供详细背景、审计事实与单位、追问潜在漏洞，并始终保持人类对最终决策的掌控权。 (17:50)
Reading summary
Christopher Horne argues that generative AI chatbots cannot replace foundational engineering comprehension, as their polished outputs routinely conceal dimensional errors, visual misinterpretations, and invalid physical assumptions that students must manually verify.
Cognitive Grounding Versus Passive Chatbot Trust (0:04)
Horne contends that relying on generative AI for quick answers fosters a superficial illusion of mastery without conveying core scientific principles. Technical disciplines like engineering, physics, and computer science require rigorous conceptual modeling that automated shortcuts bypass. Physical problem-solving methods, including handwriting notes and sketching circuit diagrams, activate neural pathways far more effectively than digital typing; Horne cites a 100-student study showing that learners taking handwritten notes scored one full letter grade higher on average than peers who did not.
Because conversational chatbots deliver answers with an authoritative, polished tone, students frequently grant them blind trust. Horne emphasizes that learners must first work through problems manually to build their own internal benchmarks. Without an independently formed mental model, users lack the critical perspective needed to challenge machine logic or verify whether an AI-generated formula is fundamentally flawed.
Circuit Analysis and Multimodal Breakdown in Education (5:04)
To evaluate student interactions with AI, Horne and a collaborator in Mexico conducted a study with 79 engineering and physics students across three sections analyzing fundamental resistor-inductor (RL) and resistor-inductor-capacitor (RLC) circuits. Students solved for current and voltage manually before inputting identical problems into ChatGPT. Reflection essays highlighted significant technical frustration: the language model consistently failed to interpret uploaded 2D circuit diagrams and converted calculus equations and integrals into unformatted text or broken LaTeX code rather than usable visual analysis.
Pedagogical success depended heavily on whether students used AI before or after establishing baseline calculations. Those who solved problems manually first leveraged ChatGPT effectively as an interactive brainstorming tool to cross-examine their work. In contrast, students who treated the bot as an immediate answering machine struggled to understand core principles. The study also noted that roughly 20% of participants adopted the AI as a solitary study partner—a behavior Horne links to post-COVID study habits—prompting one student to develop a custom Python script and circuit database to bridge the model's contextual deficiencies.
Systematic Inaccuracies in Units and Physical Constraints (14:06)
Chatbot outputs degrade rapidly when prompts lack precise physical context or boundary conditions. Horne cites real-world examples, such as modeling bridge deflections without accounting for temporal traffic load variations, or attempting to calculate residential electrical receptacle spacing against national standards, where automated tools produced invalid recommendations. Similarly, when Horne prompted ChatGPT to generate an answer key for an exam he had already validated, the bot returned an error-riddled key due to unstated assumptions.
In addition to contextual omissions, large language models routinely confuse fundamental physical dimensions. Horne observes chatbots conflating pressure with force, swapping amperes for volts, and misinterpreting thermal units like Fahrenheit and Celsius. Because models prioritize conversational coherence over dimensional dimensional consistency, vague prompts missing academic domain framing reliably produce technically inaccurate results.
Professional Hazards and the Five-Step Verification Framework (17:50)
Unverified AI calculations create severe liability beyond the classroom, including structural failures in civil infrastructure, contractual breaches, and financial loss. Horne points to operational security lapses, such as Samsung engineers leaking internal data by entering proprietary details directly into ChatGPT, to illustrate the hazards of unmanaged corporate usage. Organizations must therefore move away from rewarding raw generation speed and establish mandatory review policies backed by scoring rubrics that evaluate factual accuracy, underlying assumptions, and source validity.
To institutionalize safe practices, Horne details a five-step verification framework: (1) analyze the problem manually to establish an initial framework; (2) supply the AI with exhaustive contextual constraints and environmental initial conditions; (3) audit raw formulas and physical unit consistency against reference literature; (4) prompt the bot to scrutinize its own output by identifying missing assumptions and edge cases; and (5) preserve human judgment as the non-negotiable final authority before taking action.
Horne concludes that the central challenge of educational and industrial AI integration is not driving tool adoption, but cultivating the human expertise and verification habits required to catch polished automated errors before they trigger operational failures.
## Background
本视频由 Ai4 频道于 2026 年 8 月 14 日发布在 YouTube 平台。内容为北卡罗来纳州农工大学副教授克里斯托弗·霍恩（Christopher Horne）的主题演讲，主要围绕 79 名工程专业学生利用 AI 聊天机器人求解 RL 与 RLC 电路问题的实证研究展开，深入剖析了 AI 在工程推理中缺失假设、混淆物理单位和逻辑中断等缺陷，并提出了一套用于培养批判性验证能力的五步框架。
## TL;DR
霍恩教授结合自身工程咨询经验与跨国电路教学实验指出，生成式 AI 虽然能给出语气肯定且格式光鲜的解答，但在处理具体工程计算与图纸解析时频频出现硬伤。研究表明，只有先通过手算建立起扎实物理心智模型的学生才能有效借助 AI 查漏补缺，缺乏独立推演而盲信 AI 只会加剧认知混乱与工程风险。他强调评估个人与企业的“AI 就绪度”核心在于验证能力而非单纯的提示词技巧，并详细阐述了涵盖自我推演、严密上下文、追问破绽、交叉核验与人类终审的五步防范框架。
## Atmosphere & style
演讲整体基调严谨务实、发人深省，霍恩兼具高校学者与工程顾问的双重视角，使整场分享既有教学实证的扎实支撑，又充满工业现场的实战警示感。他通过坦诚分享自身房屋翻修咨询翻车与试卷命题漏洞等生动案例，以犀利幽默的口吻打破了对机器权威的盲目崇拜。随着从课堂实验的沮丧反思层层推进到企业治理的系统性防范，整场演讲呈现出一种由冷峻剖析技术泡沫走向构建理性防御准则的沉稳节奏。
## Key takeaways
- [00:04](https://www.youtube.com/watch?v=Do_E3KX8QGM&t=4s) 评估个人或组织的“AI 就绪度”（AI readiness），核心衡量标准在于验证 AI 输出对错的能力，而非仅仅看谁能最快获取或使用工具。
- [02:12](https://www.youtube.com/watch?v=Do_E3KX8QGM&t=132s) AI 生成的技术回答往往语气肯定且条理清晰，但光鲜外表下极易隐藏缺失前提假设、混淆物理单位或推导逻辑中断等系统性缺陷。
- [03:57](https://www.youtube.com/watch?v=Do_E3KX8QGM&t=237s) 多数学生潜意识里仍将生成式 AI 视为“答案自动贩卖机”而非“思维伙伴”，试图借此绕过艰难推导往往会导致核心概念理解的脱节与挫败。
- [05:04](https://www.youtube.com/watch?v=Do_E3KX8QGM&t=304s) 对 79 名理工科学生的电路实验表明，必须先通过手算推导建立“心智模型”再借助 AI 比对，且手写演算能显著激活神经认知并提升学业成绩。
- [07:47](https://www.youtube.com/watch?v=Do_E3KX8QGM&t=467s) 通用大模型在解析复杂的二维工程图纸与电路拓扑结构时存在明显失真，缺乏精准上下文引导会导致输出严重偏离实际。
- [09:33](https://www.youtube.com/watch?v=Do_E3KX8QGM&t=573s) 工业工程与基础设施咨询中存在明确的 AI 使用边界：AI 可用于头脑风暴与待办梳理，但坚决禁止在未验证时直接索取关键技术应力与参数计算。
- [11:16](https://www.youtube.com/watch?v=Do_E3KX8QGM&t=676s) AI 的生动表达与情绪价值无法代替工程正确率；只有前期通过教材推导建立扎实认知的学习者，才能有效将 AI 转化为强化学习的辅导工具。
- [13:06](https://www.youtube.com/watch?v=Do_E3KX8QGM&t=786s) STEM 领域的大模型输出常出现“逻辑看似自洽但数值完全错误”的陷阱，甚至混淆力与压力、摄氏度与华氏度等基本物理单位与边界条件。
- [15:11](https://www.youtube.com/watch?v=Do_E3KX8QGM&t=911s) 针对大模型图纸解析的软肋，可通过对接结构化图元数据库及在提示词中交代严密背景与参数约束来大幅改善输出质量。
- [16:29](https://www.youtube.com/watch?v=Do_E3KX8QGM&t=989s) 必须打破年轻一代对机器权威的盲从迷信，坚持“先手工推演、后人机核对”的协作机制，并学会用硬性事实主动质询 AI。
- [17:50](https://www.youtube.com/watch?v=Do_E3KX8QGM&t=1070s) 企业在 AI 转型中若缺乏验证机制，匆促进展可能导致隐蔽性硬伤扩散，进而引发内部组织信任破裂与外部客户声誉受损。
- [18:52](https://www.youtube.com/watch?v=Do_E3KX8QGM&t=1132s) 应对 AI 局限性应遵循五步防范框架（自主推演、严密上下文、追问自查、交叉核验、人类最终裁决），组织考核重心应从单纯编写 Prompt 转向系统化验证输出。
## Detailed walkthrough
### [00:04](https://www.youtube.com/watch?v=Do_E3KX8QGM&t=4s) 人工智能与工程教育：从“寻找答案”到“验证真伪”
讲座开篇，克里斯托弗·霍恩（Christopher Horne）呼应了前面讲者莱拉（Leila）和莱蒂西亚（Laeticia）的观点——使用人工智能时，“最终做决定的始终是你自己，而不是聊天机器人”。面对 AI 在推理过程中产生的幻觉，绝不能盲目依赖 AI 去理解数学或物理公式背后的机理，基础推导仍应回归教材本身。霍恩指出，当前工程教育面临的核心难题在于：就算 ChatGPT 或其他生成式 AI 给出了完全正确的数字，学生是否真正理解了问题背后的物理规律？他们是否知道为什么这个答案是对的？
在经历疫情期间学习模式的剧烈转变后，2023 年 ChatGPT 的横空出世让工程专业的学生将其视作走捷径的“万灵药”。霍恩作为北卡罗来纳州农工大学（North Carolina A&T State University）的副教授，同时兼任大型基础设施建设（建筑、电信网络、移动设备等）的工程顾问。他明确指出，评估个人或机构的“AI 就绪度”（AI readiness），绝不是看谁能最快用上工具，而是看谁具备验证 AI 输出对错的核心能力。
### [02:12](https://www.youtube.com/watch?v=Do_E3KX8QGM&t=132s) 光鲜回答背后的隐患：配电插座与考试试卷的翻车案例
AI 的回复往往行文流畅、条理清晰，字里行间充满了肯定与积极的语气，能给人极佳的心理体验，但这绝不代表其内容准确。霍恩分享了两个自己亲身经历的“翻车”故事：
在最近翻修自家房屋时，由于工期紧张，他向 ChatGPT 咨询按照美国国家电气规范（National Electrical Code）应该在特定区域安装多少个电源插座，甚至详细输入了施工说明并上传了现场照片。然而，ChatGPT 给出的计算结果完全错误。
另一个案例发生在教学中：霍恩命制了一份工程测试卷，并希望利用 AI 快速生成配套的标准答案。尽管原题经过严密校验没有任何歧义，但 AI 生成的答案中却漏洞百出。
霍恩发出警示：“看似快速而光鲜的答案，背后往往隐藏着缺失的前提假设、错误的物理单位或不完整的逻辑链条。” 在工程与物理世界中，华氏度与摄氏度、安培与伏特哪怕稍有混淆，都会导致灾难性的工程事故。
```mermaid
flowchart TD
accTitle: AI 生成技术答案的验证漏斗
subgraph Prompting["输入阶段"]
P1["工程问题与图片/条件"]
end
subgraph Generation["AI 处理与输出"]
G1["语言模型生成"] --> G2["语言通顺、语气肯定的解答"]
end
subgraph Flaws["隐藏的系统性缺陷"]
G2 -.-> F1["前提假设缺失"]
G2 -.-> F2["物理量单位混淆"]
G2 -.-> F3["推导逻辑中断/幻觉"]
end
subgraph Verification["人工验证闭环"]
V1["先验心智模型 / 手算基础"] --> V2["交叉核对规范与教科书"]
V2 --> V3["最终工程决策与把关"]
end
P1 --> G1
G2 --> V2
```
### [03:57](https://www.youtube.com/watch?v=Do_E3KX8QGM&t=237s) 聊天机器人是“思维伙伴”还是“应答机”？
工程学、物理学与计算机科学课程向来以硬核、高难度著称。部分学生误以为生成式 AI 能帮他们绕过艰难的思考过程，但在这种浅尝辄止的使用下，他们根本没有掌握核心概念。2024 年，霍恩尝试让学生在手动完成题目后再调用 ChatGPT 求解。结果出现了两极分化：一部分学生开始尝试将其当作“思维伙伴”（thinking partner）进行探讨；而另一部分学生则陷入了极大的挫败感，因为 AI 频频给出荒谬的错误答案。
调查发现，绝大多数学生在潜意识里依然把生成式 AI 当作冷冰冰的“答案自动贩卖机”（answering machine），而非辅助学习的导师或协同思考者。
### [05:04](https://www.youtube.com/watch?v=Do_E3KX8QGM&t=304s) 跨国实证研究：79 名理工科学生的电路解题实验
为了系统探究学生如何理解技术问题以及 AI 在其中的真实作用，霍恩在一次学术会议上结识了一位来自墨西哥的物理学教授，双方一拍即合展开了联合教学实验。该研究覆盖了美墨两国三门课程中的 79 名工程与物理专业学生。
实验选取了最基础也最无处不在的物理实体——电路（Electric Circuit）作为载体：
- **研究对象**：包含电阻-电感（RL）电路以及电阻-电感-电容（RLC）电路。这类电路广泛应用于手机、暖通空调系统（HVAC）、飞机机载设备、航天器、机器人和无人机之中。
- **教学设计**：要求学生必须**首先手算**电路中的电压、电流等物理量，建立起解题步骤和物理机理的“心智模型”（mental model），随后才能将题目输入 AI，对比两者的解法与答案。
- **手写的神经机制**：霍恩特别强调了手算和手写笔记的重要性。最新神经科学研究表明，手写比在屏幕上打字或用手机拍照更能激活大脑中的神经网络。霍恩此前对 100 名做手写笔记的学生进行过跟踪对比，发现坚持手写笔记的学生平均成绩能整整高出一个等级（One letter grade）。
### [07:47](https://www.youtube.com/watch?v=Do_E3KX8QGM&t=467s) 多模态输入的盲区与反思阶段的普遍挫败
在提交的解题反思中，学生们集中宣泄了在使用 ChatGPT 时的沮丧情绪。实验暴露出了当时大语言模型在解决具体工程问题上的显著瓶颈：
- **图像解析失真**：许多学生直接拍摄电路图上传给聊天机器人，以为多模态模型能准确识别拓扑结构。但模型在解析电路元件连接关系、初始状态和参数时错误连篇，导致后续解法完全偏离正确方向。
- **提示词引导的局限**：研究团队最初为了测试学生的独立探索能力，仅提供了最基础的提示词结构和假设框架。实践证明，如果缺乏精准的工程上下文提示引导，学生得到的输出结果质量极差。这也印证了工程设计中批判性思维与动手推演的基础性地位。
### [09:33](https://www.youtube.com/watch?v=Do_E3KX8QGM&t=573s) 工程咨询中的边界意识：不能向 AI 索取关键计算
结合自己在夏季开展的基础设施工程咨询工作，霍恩分享了工业界使用 AI 的真实边界。在与很少接触前沿 AI 的建筑工程行业打交道时，他经常使用 ChatGPT 来梳理工程待办清单、组织项目结构或开展前期的头脑风暴，这些任务 AI 都能极其出色地完成。
但一旦进入需要收敛具体技术细节的深水区，AI 就必须让位于工程师的专业大脑：
- **坚决禁止的操作**：绝不能直接让 ChatGPT 计算大梁的机械应力、复杂的交直流电压电流、或是特定工程溶液的精确 pH 值。
- **后果与责任**：这些专业参数如果依赖 AI 且未加验证就交付给客户，将带来巨大的工程安全隐患与法律赔偿风险。霍恩还顺带提及了数年前三星（Samsung）工程师违规将内部半导体数据粘贴至 ChatGPT 导致公司信任受损的业界教训。
### [11:16](https://www.youtube.com/watch?v=Do_E3KX8QGM&t=676s) “情绪价值”不能代替正确率：AI 辅导的真实效果
AI 的强大之处在于能够用通俗、生动的语言把推导过程写得极其漂亮，还会不吝赞美地鼓励使用者。然而，这种良好的交互体验掩盖不了工程计算上的实质性错误。
研究还观察到了有趣的学生画像与学习转化差异：
- **内向型学生的偏好**：在经历疫情隔离后，一部分习惯独自学习的内向型学生非常喜欢与聊天机器人高频对话，AI 为他们提供了低压力的课外交流环境，但这并没有直接转化为更高的考试成绩。
- **心智模型的决定性作用**：只有那些前期严格通过教材推导、依靠自身大脑建立起清晰心智模型的学生，ChatGPT 才能起到查漏补缺和“强化学习”（reinforced learning）的作用；而那些没有手动做题、试图直接依赖 AI 寻找答案的学生，AI 的介入反而加剧了认知混乱。
- **课外导师的认知占比**：霍恩在更广泛的数百名学生调研中发现，真正将生成式 AI 视作课后辅导老师（Tutor）的学生比例，实际上仅占约 20% 左右。
### [13:06](https://www.youtube.com/watch?v=Do_E3KX8QGM&t=786s) STEM 领域的信心陷阱与学生的常见提示误区
在工程与理科教学中，微积分、高等数学以及基础专业课往往难度极高，极易打击学生的自信心。这并不是说英语或人文社科不重要或缺乏难度，但 STEM 的严谨性使得学生在遇到瓶颈时极度渴望工具的协助。然而，许多学生在使用大语言模型时存在认知偏差：他们并不清楚大语言模型本质上是基于词词相连的概率文本模型，因而频频陷入沟通误区。有学生直接在提示词中输入数学公式，结果模型返回了一长串 LaTeX 源码代码，把学生弄得晕头转向；还有学生直接上传一张微积分积分公式的截图，返回的却是一串莫名其妙的单词。
更棘手的是物理与工程现实中的边界条件。例如在工程场景中评估一条需要翻修的高速公路桥梁，不能仅仅给出一个公式，还必须明确其初始条件：桥梁在受力后是否会发生微小挠度变形容许？早晨 8:00 晚峰与下午 2:00 闲时的车流量分别是多少？如果提示词中遗漏了工期、预算或人力等现实约束，模型给出的答案即便在纯技术推导层面看似成立，在实际工程中也毫无参考价值。
此外，基础物理概念的混淆屡见不鲜：
- **物理单位混淆**：混淆华氏度与摄氏度、安培与伏特，甚至混淆压力（Pressure）与力（Force）——压力和力有物理关联，但绝不能画等号。
- **逻辑自洽但结论荒谬**：多次实测发现，输入 ChatGPT 的工程题目推导逻辑看似无懈可击，但最终给出的数值答案却完全是错误的。
---
### [15:11](https://www.youtube.com/watch?v=Do_E3KX8QGM&t=911s) 电路图像解析的软肋与自主开发补丁方案
一年前的通用大模型在解析复杂的二维工程图纸或电路原理图时存在明显缺陷，常常直接跳过对物理图形的实际解读，导致分析结果不可信。为了解决电路图形输入解析不准的问题，有学生主动接下挑战，编写了一套 Python 程序，专门对接并检索包含各类电路结构的图元数据库。通过结构化数据库与提示词的联动，大幅提升了电路推导结果的准确性。
然而，大模型“极度自信地给出不完整答案”的问题依然突出。这很大程度上源于学生提示词的模糊与上下文缺失。许多学生从未在提示词中交代自身背景与题目前提，例如说明“我是一名正在学习计算机科学/电气工程的学生，正在解决以下特定参数的电路问题”，缺少这些关键约束，输出结果往往语焉不詳且漏洞百出。
```mermaid
flowchart TD
accTitle: 盲目依赖与先验验证两种人机协作路径对比
subgraph BlindTrust["盲目信任路径（高风险）"]
B1["模糊提问 / 直接截屏"] --> B2["LLM 自信生成有瑕疵推导"]
B2 --> B3["学生误信机器权威直接采纳"]
end
subgraph VerificationWorkflow["验证驱动路径（推荐）"]
V1["人工先验演算与公式梳理"] --> V2["提供严密上下文与初始条件"]
V2 --> V3["LLM 头脑风暴与辅助计算"]
V3 --> V4["对照事实、物理单位与逻辑核查"]
V4 --> V5["最终人类裁决与采纳"]
end
```
---
### [16:29](https://www.youtube.com/watch?v=Do_E3KX8QGM&t=989s) 手工演算的价值与打破“机器必定正确”的迷信
研究与教学实践表明，最稳妥的策略是要求学生先进行手动演算，再把推导过程和数据提供给聊天机器人作为发散与比对的基础。这一观点与上一位讲者 Layla 的见解一致：必须在让 ChatGPT 进行头脑风暴或寻找其他解法之前，先完成人类的前期推导。
随之而来的现实问题是：先手工演算再核对 AI，会不会比从头到尾直接手算耗费更多时间？事实上，一部分学生宁可完全不用 AI，仅凭课本与笔记纯手工推导，反而学得更加扎实透彻。
另一个严重的心理问题是学生普遍缺乏质疑机器的意识。生活在智能手机时代的年轻群体习惯了即时满足，潜意识里将计算机视作不会犯错的绝对权威，因而不假思索地接受机器结论。针对这一点，教师必须反复向学生灌输：**绝不能盲目服从，必须用硬性事实而非情绪主观去强力质询和挑战 AI。**
---
### [17:50](https://www.youtube.com/watch?v=Do_E3KX8QGM&t=1070s) 企业级落地风险：从课堂验证到组织级 AI 就绪度
这项针对电路推导的研究结论同样适用于面临 AI 转型的企业与组织。企业员工如果缺乏提供充足上下文、核查假设前提、追踪数据源头或验证技术计算的能力，错误的输出将直接导致严重后果：
- **组织信任瓦解**：内部推行出现严重错漏，导致管理层与跨部门信任破裂。
- **客户声誉受损**：向外部客户交付了带有隐蔽硬伤的方案或产品，招致业务与订单流失。
在商业环境中，对聊天机器人的天真与无知绝非小事。最大的风险在于匆促进展、缺乏核验机制，从而对眼皮底下的系统性错误浑然不觉。
---
### [18:52](https://www.youtube.com/watch?v=Do_E3KX8QGM&t=1132s) 五步防范框架与组织治理行动准则
面对生成式 AI 的局限性，组织与个人在实操中应当遵循清晰的验证框架，重点强调人类对最终结果的全程负责：
| 步骤 | 行动要求 | 核心要义 |
|:--- |:--- |:--- |
| **第一步：立足自我推演** | 先理清自身的解决逻辑与物理边界 | 哪怕只是装修房间等日常规划，也必须先有自主思路 |
| **第二步：严密上下文输入** | 提供详尽的初始条件、时间与预算约束 | 消除歧义，避免模型在真空环境中虚构事实 |
| **第三步：主动追问破绽** | 提示模型自查反思：“这里面可能有什么错误？” | 主动打破思维盲区，让机器列出潜在假设漏洞 |
| **第四步：事实交叉验证** | 逐一校对单位、物理量、公式推导与信源 | 运用评判量表（Rubrics）复核准确性、风险及数据源 |
| **第五步：保留最终裁定权** | 最终决策权始终归人类所有，严禁盲从工具 | 明确划定“人类复核为强制性（Mandatory）”的业务边界 |
在组织管理与考核机制上，绝不能仅按产出速度给予奖励，而要对细致入微的复核与质量把关进行正向激励。培训重心应当从单纯的“如何编写 Prompt”转向“如何系统化验证输出与归档文档”。同时，在借助 AI 自动生成评分量规或评审标准时也须警惕，因为模型生成的评分准则本身可能存在偏差，教师与管理者仍须亲手二次审核。
## Notable quotes
> [00:04](https://www.youtube.com/watch?v=Do_E3KX8QGM&t=4s) 最终做决定的始终是你自己，而不是聊天机器人 — Christopher Horne
> [00:04](https://www.youtube.com/watch?v=Do_E3KX8QGM&t=4s) 评估个人或机构的“AI 就绪度”（AI readiness），绝不是看谁能最快用上工具，而是看谁具备验证 AI 输出对错的核心能力。 — Christopher Horne
> [02:12](https://www.youtube.com/watch?v=Do_E3KX8QGM&t=132s) 看似快速而光鲜的答案，背后往往隐藏着缺失的前提假设、错误的物理单位或不完整的逻辑链条。 — Christopher Horne
> [03:57](https://www.youtube.com/watch?v=Do_E3KX8QGM&t=237s) 思维伙伴 — Christopher Horne
> [03:57](https://www.youtube.com/watch?v=Do_E3KX8QGM&t=237s) 答案自动贩卖机 — Christopher Horne
> [09:33](https://www.youtube.com/watch?v=Do_E3KX8QGM&t=573s) 绝不能直接让 ChatGPT 计算大梁的机械应力、复杂的交直流电压电流、或是特定工程溶液的精确 pH 值。 — Christopher Horne
> [11:16](https://www.youtube.com/watch?v=Do_E3KX8QGM&t=676s) 真正将生成式 AI 视作课后辅导老师（Tutor）的学生比例，实际上仅占约 20% 左右。 — Christopher Horne
> [16:29](https://www.youtube.com/watch?v=Do_E3KX8QGM&t=989s) 绝不能盲目服从，必须用硬性事实而非情绪主观去强力质询和挑战 AI。 — Christopher Horne
> [18:52](https://www.youtube.com/watch?v=Do_E3KX8QGM&t=1132s) 人类复核为强制性 — Christopher Horne
## Mentioned
Christopher Horne, Leila, Laeticia, North Carolina A&T State University, ChatGPT, National Electrical Code, Samsung, Python, LaTeX
## Video description
Christopher Horne, Teaching Associate Professor, North Carolina A&T State University
AI can produce polished technical answers that sound convincing even when they are wrong. This talk examines how 79 engineering students used AI chatbots to solve RL and RLC circuit problems and what happened when they were asked to verify the AI's work instead of simply accepting it. The talk will highlight where AI supported learning, where it failed through missing assumptions, unit errors, and incomplete reasoning, and why verification skills proved more important than prompting skills alone. Attendees will leave with a practical five-step framework for helping students, employees, and organizations use AI more effectively by developing habits of critical evaluation rather than blind trust. If your organization is adopting AI for technical or knowledge-based work, this session will show why AI readiness should be measured by the ability to verify AI output, not merely access to AI.
[Source](https://www.youtube.com/watch?v=Do_E3KX8QGM)
Source quotes
[0:04] Christopher Horne: 最终做决定的始终是你自己，而不是聊天机器人
[0:04] Christopher Horne: 评估个人或机构的“AI 就绪度”（AI readiness），绝不是看谁能最快用上工具，而是看谁具备验证 AI 输出对错的核心能力。
[2:12] Christopher Horne: 看似快速而光鲜的答案，背后往往隐藏着缺失的前提假设、错误的物理单位或不完整的逻辑链条。
[3:57] Christopher Horne: 思维伙伴
[3:57] Christopher Horne: 答案自动贩卖机
[9:33] Christopher Horne: 绝不能直接让 ChatGPT 计算大梁的机械应力、复杂的交直流电压电流、或是特定工程溶液的精确 pH 值。
[11:16] Christopher Horne: 真正将生成式 AI 视作课后辅导老师（Tutor）的学生比例，实际上仅占约 20% 左右。
[16:29] Christopher Horne: 绝不能盲目服从，必须用硬性事实而非情绪主观去强力质询和挑战 AI。
[18:52] Christopher Horne: 人类复核为强制性
Source quotes
Christopher Horne: AI can produce polished technical answers that sound convincing even when they are wrong.
Christopher Horne: verification skills proved more important than prompting skills alone
[17:50] Christopher Horne: AI readiness should be measured by the ability to verify AI output, not merely access to AI.
[0:04] Christopher Horne: quick fix
[0:04] Christopher Horne: answering machine
[17:50] Christopher Horne: the final choice is yours, not the bot's
[17:50] Christopher Horne: Begin with your own reasoning
[17:50] Christopher Horne: Provide detailed context
[17:50] Christopher Horne: Verify facts and calculations
[17:50] Christopher Horne: Ask what could be wrong
[17:50] Christopher Horne: Apply human judgment before action
Original transcript
[0:04] I wanted to follow up on what Leila said in
regards to AI and the final choice is yours,
[0:10] not the bot. And my talk will be
supportive of that. And also, Laeticia
[0:19] just stated that reasons behind each step
when the AI is giving you informations,
[0:25] it can hallucinate. And she also said,
"Don't rely on AI to understand." And
[0:32] this math equations, you saw the math equations,
which was a error formula and another formula.
[0:38] Hopefully, you won't use AI to try to figure that
out. Use your own textbook or something like that.
[0:44] Anyway, um if Chat GPT or some other gen AI gives
the correct answer, do students really understand
[0:56] how the problem works and the answer? Do they know
why they got the correct answer? And this is my
[1:04] research area right now is helping students learn.
During the pandemic, we had different learning
[1:12] styles and deliveries, the way we did
learning. And then Chat GPT came out in 2023
[1:20] and my students were eager to use it. They thought
it was a quick fix. And the talk today is going to
[1:27] go into some some of that. So, it's not just
access to AI as a quick fix to find things out,
[1:34] but the ability to verify if the answer is wrong.
I'm a teaching professor at in North Carolina.
[1:42] I also am an engineering consultant to companies
that build infrastructure. Infrastructure meaning
[1:50] what? Meaning buildings, telecommunications, cell
phones. A cell phone is a circuit. The particular
[1:59] uh research um, tool and item I did for this study
was a circuit. Uh, because electrical circuits are
[2:06] everywhere. These are circuits. This is a circuit.
So, I'll be getting into that a little bit, too.
[2:12] Uh, AI can sound right. It Many of you use it, and
when you type something into brainstorm, it comes
[2:20] back with very coherent, nice to read. It makes
you feel good. It's always positive, you know,
[2:25] this kind of thing. Uh, but um, it may not be
accurate, and the outputs could be wrong, as well.
[2:33] So, uh recently, I remodeled our part of
our house, and I asked Chat GPT to help me
[2:41] tell me how many receptacles should I put in
the house, okay? Now, there's a there's a uh
[2:47] a code, uh a national standard
behind how we wire up houses,
[2:52] um, and I put a lot of information in there,
including pictures of the house we're remodeling,
[2:58] and to make a long story short, Chat GPT
got it wrong. I was kind of in a hurry,
[3:03] and uh it didn't work out. Another instance,
uh students and employees already use AI for
[3:09] technical work. Yeah, everyone is using it.
Most people are using it. Uh some are not,
[3:15] but um I prepared a test recently, and I needed
an answer key for the test. The test I checked,
[3:23] and I knew it was correct. So, I just asked Chat
GPT, uh based on previous test and answers, please
[3:30] generate an answer key, and it was not correct.
There were lots of errors in the answer key. So,
[3:36] uh you have to be really careful with fast,
polished answers can hide missing assumptions,
[3:42] wrong units, or incomplete reasoning. By
units, I mean in engineering or physics,
[3:47] technical problems, we could be talking about
Fahrenheit versus Celsius, or amperes versus
[3:53] volts and they're totally different. So,
you have to be real careful with that.
[3:57] When students use AI chatbots for technical
problem-solving, this is the research area
[4:03] that we have been studying. Engineering is hard,
physics is hard, computer science is hard, and
[4:12] there are some students that think they can make
it easier by using GenAI when it's not it's a tool
[4:20] there for them, but they're not really learning
some things. And so, have to be very careful.
[4:25] It does provide correct answers, and it is
helpful, but do they use it as a thinking partner?
[4:33] In 2024, when I asked students to use ChatGPT to
solve a problem after they solved it manually,
[4:40] then some of them thought of it as a thinking
partner, and some of them became really frustrated
[4:46] because it gave the wrong answer, and there was
lots of reasons for that. Um most were using it
[4:54] as an answering machine, not as a tutor or not as
a thinking partner. This was a couple years ago.
[5:04] And then I met a professor at a conference
from Mexico, and we had the same interest in um
[5:12] learning how learning how students learn. Do they
grasp the material or are they just using ChatGPT
[5:20] superficially? So, there were 79 students
involved, engineering and physics students
[5:26] actually. The The professor in Mexico was
teaching physics students, but we taught the same
[5:31] technical problem. That is, an electric circuit.
Why did you pick an electric circuit? Because
[5:36] electric circuits are everywhere. We just
don't think about it because we're on AI.
[5:41] But um or social media or whatever. But
without a circuit, this wouldn't work.
[5:46] Um circuits are in airplanes and engineers,
whether you're a civil engineer, electrical
[5:50] engineer, how many aerospace engineers? Anybody
in computer science in here? So, your computer is
[5:57] a circuit. So, circuits are everywhere. And so,
we ask students to manually find the voltage or
[6:03] manually find the current or some uh quantity in
this very common circuit manually first because
[6:12] we want them to build this mental model
of understanding. And I recently read
[6:19] where handwriting activates neural net neural
circuits in the brain uh and more than typing
[6:27] or taking pictures on your phone. So, I see a lot
of students taking pictures on their phones in the
[6:32] classroom. It doesn't stimulate the brain the same
way as handwriting things down. Very important.
[6:41] Uh as a side note, I did a study on taking notes,
handwritten notes, and 100 students who did that,
[6:49] uh their grades were one letter grade higher.
Did they actually grasp the material? Well,
[6:56] probably, but it's a good um uh routine
to take take good notes about that. So,
[7:03] and there were three sections in uh the courses
in Mexico and United States. We used an electric
[7:09] circuit. Um, there were two types of circuit, a
resistor and an inductor circuit, and then also
[7:14] a resistor inductor capacitor circuit. These
are circuits that are used in cell phones,
[7:19] um HVAC circuits, uh airplane circuits, all
kinds of uh spaceships, robots, uh drones.
[7:28] Um, we asked them to do the solution
first manually and then compare it to AI
[7:34] uh second secondly. Again, the manual
method first so they can develop this mental
[7:40] mental model of of what the steps are and how
they're going to think about solving the problem.
[7:47] As you know, critical thinking is really important
and not so much anecdotally here, but we're
[7:55] as a society losing some of that critical thinking
that we need to maintain to be competitive
[8:02] uh in this uh new new world that we're
in, new technology world we're in.
[8:07] What about prompts and the outputs and
the reflections? Well, in reverse order,
[8:11] the reflections really revealed frustrations
frustrations that students had using ChatGPT. Now,
[8:18] mind you that this is about a year ago, so um some
students will probably have more confidence, but
[8:26] this was a quite a problem for them. They didn't
like the output. A lot of students tried to put
[8:31] in pictures into the chatbot and they thought that
the chatbot would uh interpret the picture and uh
[8:41] the LLM that was used by the ChatGPT
was not um you know, understanding this
[8:49] this the solution was off and
so that wasn't wasn't very good.
[8:53] Also, the prompting we gave students uh an idea of
how to prompt the bot for this technical problem,
[9:01] what questions to ask and maybe assumptions, but
we limited that to see if they would enhance it.
[9:06] It turns out if we would have given a
little more guideline to the students,
[9:10] they would have had better uh outputs uh for
sure. But again, we're trying to teach students
[9:16] um how to be problem solvers um as uh Leila
mentioned, critical thinking in the design process
[9:24] is really important to do things by hand at times
to understand it. Now, I use ChatGPT a lot in my
[9:33] consulting work in the summertime when I'm not
teaching, I'm consulting engineering projects.
[9:39] What kind of projects? Infrastructure projects.
Uh working with trades, and they don't use they
[9:45] use AI very much, by the way, construction. I
don't see that sector at this conference, but
[9:51] I sometimes will do designs for infrastructure for
the construction trades. And when I do the design,
[9:57] I'll I'll make a project list in Chat GPT,
this this this project or whatever, to try
[10:02] to get brainstorming and organization. It does it
really well. It helps me get through the projects,
[10:08] but it when it comes to specifics, uh Layla talked
about divergence and um diverging, you know,
[10:16] ideas. And when it gets down to diverging ideas
and getting to the real specific topic, I need to
[10:21] use my brain and think about it, maybe compare
it to a textbook that I used to use, or some
[10:28] peer-reviewed research paper, rather than asking
Chat GPT to calculate the stress on this beam,
[10:34] or calculate the voltage or the current, or uh
the pH level of some solution. It might not get
[10:41] it correct. And that's not anecdotal, because I've
done I've done this and I get the wrong answer.
[10:46] So, I'm trying to rush through to get the
project done, and the answers aren't correct.
[10:50] And if you turn that to the customer, then you
got a real problem. Uh you know, liabilities
[10:56] and and so forth. It reminds me of the Was it
Samsung engineers a couple years ago? They uh
[11:04] took some internal information and put it into
Chat GPT. I don't think there was any uh punitive
[11:11] damages or anything like that, but there was some
internal issues. Their trust was lost internally.
[11:16] Something like that, but I'll just keep it there,
because I don't like uh you know, exaggerating
[11:20] about things if they're not facts. Chat GPT is
good about hallucinations, they remember that. And
[11:26] in the study, we did find a lot of hallucinations
that happened, mostly because the prompting the
[11:31] context was not set up properly in the beginning.
Uh explain procedures in plain language. Yes, the
[11:39] very nicely written out uh essays about the
problem to students and they could read through
[11:45] it and it sounds real good. And then tries to
make you feel good and all that is great as well,
[11:50] but sometimes that doesn't help because the answer
is still wrong. Um it was conversational with some
[11:57] students. There was a small percentage of students
that were introverts, that like to study alone.
[12:03] You know, this is right after the COVID and
they're kind of going through a lot of stuff and
[12:07] they like to study alone. So, the bot and and them
really hit it off. Um but that doesn't mean that
[12:15] they scored higher on their exams or anything like
that. They just helped them. Um some of them who
[12:20] did understand it with the mental model manually
first, the ChatGPT bot reinforced their learning.
[12:30] The students who did not do their homework
manually and look at the book or think
[12:35] about it with their own brain and take
notes, the bot did not reinforce as well.
[12:44] Uh supported students outside class time,
right? So, a small percentage of it thought
[12:49] of it as a tutor. Um so I've probably, you
know, analyzed a couple of hundred students
[12:56] outside of this study and maybe about 20% think
of it as a tutor, at least about 6 months ago.
[13:06] Uh increased confidence. So, in the STEM world
and these subjects that are often really,
[13:13] really difficult, that doesn't mean English and
humanities aren't difficult, by the way. Nothing
[13:17] against that. We need knowledge in those areas as
well, but STEM uh STEM it can be challenging and
[13:24] getting confidence in mathematics, calculus, and
things that can be really hard for some students.
[13:31] Uh is is an issue when a student does not prompt
properly, give the bot good context, and make sure
[13:39] that they check their facts. I've had students
who wrote in equations and they got back LaTeX
[13:45] instead of something they were looking for and
they got very confused by the LaTeX language, you
[13:49] know, this kind of thing. Or they took a picture
of a math equation, an integral or calculus, and
[13:54] it came back and had a word instead of that, you
know. So, we know that LLMs are based on words and
[14:00] students didn't know that. And so, it frustrated
them quite a bit. Uh initial conditions mean,
[14:06] "Okay, uh we need a new bridge across this highway
here cuz it's falling down and we need to know
[14:12] if we put this I is the bridge going to deflect
a little bit? What's the initial condition? How
[14:19] many uh cars will be on there at 8:00 a.m. versus
2:00 p.m.? There are a lot of If you don't provide
[14:25] the bot with the time and money and this kind of
thing, uh then the output you may have a technical
[14:33] answer that's partially correct but didn't
consider budgets and manpower and things like
[14:38] that. So, there's quite a bit involved in that.
Uh by units, uh I don't mean boxes. I mean like
[14:44] um Fahrenheit versus Celsius or amps versus volts
or um whatever it might be. Pressure versus force.
[14:53] Those are two different things. Pressure is
not force, by the way. They're related but they
[14:58] I give the incorrect answer. Multiple times I
have put stuff in ChatGPT and then it would not
[15:05] give an answer that made a lot of sense. The
logic was correct but the answer was wrong.
[15:11] Skip the physical interpretation meaning if you
provide ChatGPT, at least this was a year ago,
[15:17] with a picture of the technical problem
or the circuit, as you will, a 2D image,
[15:24] it could not interpret it. So,
you wouldn't want to rely on that.
[15:28] Uh since then, I had a student take up that
challenge and write some Python code that takes
[15:34] database of all different kinds of circuits,
and then you could you could prompt it based
[15:39] on the database and produce a better, more correct
output. And I thought that was really interesting.
[15:45] So, if anybody's interested in talking about
that, let me know, or anything, actually. Um
[15:51] And so, incomplete answers
without with confident uh wording
[15:57] uh yeah, that kind of goes back to my point about
um sometimes the outputs would not be thorough,
[16:06] and it misses misses something. Uh I mentioned
this before that if the prompts were vague,
[16:15] you know, the student didn't say, "I'm a
student studying computer science." Or "I'm
[16:18] a student studying electrical engineering, and
I'm working on this problem." And provide maybe
[16:23] some factual information, they would have
got a better output, for sure, on that one.
[16:29] Okay, so again, the student should they they
the research shows that the students should
[16:35] uh work out the problem manually and share
their information with the with the bot.
[16:41] The speakers before um me mentioned this earlier.
Uh I think it was Layla talking about the um
[16:47] manual work before you ask the chat GPT
to uh brainstorm other uh possibilities.
[16:57] A student should check the equations. Are the
equations correct? It does bring up the point that
[17:02] does it take more time to do all of this versus
just doing it by hand? Some students don't use
[17:07] AI at all because they're really trying to learn.
They look at the book, they look at their notes,
[17:13] and they try to do it uh by hand, and that's
okay, as well. That works out really well. Um,
[17:22] sometimes I found students did not challenge
the AI because they thought it was a machine and
[17:28] always correct. You know, they're on their phone,
they get instant gratification from their phone,
[17:32] so they think that the uh, the bot is going to
be instantly correct, and they did not challenge
[17:38] it. So, that's one of my um, things that I try to
press upon students, you need to challenge the uh,
[17:44] bot with facts, not just your emotions,
but facts. Okay, and I'll just move through
[17:50] these kind of quick. Why should organizations uh,
care? This is This was an electric circuit study,
[17:56] an engineering study, but organizations
are using AI and our the previous uh,
[18:03] speakers went over some of that. Um, employees
may not know how to provide sufficient context,
[18:09] or challenge an assumption, verify
the source, or check the calculations.
[18:15] And this could result in internal breaches of
trust, inaccurate um, reputation to clients,
[18:22] customers, because of AI used improperly. In the
education world, we don't really think about that,
[18:29] but in the real world, in business, if you
don't deliver what you told you would deliver,
[18:35] that's going to be a big problem. You'll lose
business that way, so you have to be real careful,
[18:38] obviously. Uh, being naive or uninformed about
chatbots could have serious consequences. That's
[18:45] not anecdotal, that is a fact. And big risk would
be not recognizing mistakes, being in a hurry.
[18:52] So, a couple of things to think about
is uh, begin with your own reasoning.
[18:57] You've got a problem, something you want
to solve, you want to remodel your room,
[19:00] think about that. Give AI some context,
verify the facts, ask what could be wrong.
[19:06] Brainstorming is good. And then he use human
judgment before action, so uh, uh the So,
[19:16] final choice is yours, not the bot. That was from
Leila. So, I do like that. Uh let's see here. We
[19:22] got two more things. Train people to verify,
not only prompt it. Required documentation.
[19:29] Define when human review is mandatory. I'd say
it's mandatory most of all the time. Use rubrics
[19:34] for accuracy, assumptions, sources, and risk. Be
careful cuz the bot could get the rubric wrong,
[19:40] so you got to double-check your rubrics if
you're doing that for one of your courses.
[19:45] Reward the careful checking, not
speed alone. And that's it. Thank you.
