Researchers surveyed faculty about the emotions they experience related to assigning failing grades. Responses show emotions ranging from disappointment to anger to sadness. Non-TT faculty were more likely to report fears about student course evaluation ratings and job instability.

The writer describes her experience using audio comments through a narrative of her own class. She provides her students' comments as evidence. Bauer used this at a point when students would be revising. She found students felt positively toward audio comments. She observed that commenting via audio allowed her to offer more praise and specific constructive critiques, that students responded to the tone more positively, and that it took no longer than writing (while covering more ground).

The researchers observed conflicting emotions among students engaged in revision. On one hand, students see revising as crucial to improving their work. On the other hand, they feel pressure to make the right kinds of revisions and worry they do not know what strategies to use. As Ballenger and Myers explain, "for experienced student writers who have invested time and emotional energy into their writing, revision can trigger a fear that they will be 'exposed,' creating emotional dissonance that often goes unrecognized and unspoken" (591). Experiencing these conflicting emotions–hope and worry–makes it more difficult for students to revise effectively. Ballenger and Myers recommend assigning students to reflect as they revise to make it easier for them to reckon with the conflicts and focus on revisions themselves.

The writer describes an assignment that requires students to reflect on and respond to instructor comments in order to "develop, analyze, articulate, reconsider, and explain their revision ideas."

The writer found that feedback and revision cycles are rife with negative and positive emotions regardless. Emotions often stem from students' perception of themselves as writers and tension between how they perceived they did and the feedback they received. The writer urges that building a trusting relationship between instructor and student is essential for creating an environment where students are open to feedback because trust helps students move beyond negative feelings associated with receiving feedback toward positive feelings that helps them integrate that feedback into their writing.

This recent study highlights the value of instructors giving less and more explicit/targeted feedback, and expecting students to integrate into their work during revision. The findings point to what many of us already know: students don't engage with summative feedback on already graded/finished essays as much as they do with the feedback we give them on drafts. They suggest we should spend less time on final essay feedback, and instead focus on feedback that supports the grading assessment and gives students work to explicitly engage with in a future assignment or end-of-semester revision.

The writer argues that instructors experience tensions between what they want to do in responding to student work, what they think they should do, and what they think their institutional contexts expect them to do. After surveying related literature, the writer analyzes the case of 1 instructor's experiences of these tensions.

The researchers found audio comments are less time intensive for instructors and often reduce the "cognitive load" students experience when trying to flip between written comments and their papers. More specifically, they found that students prefer audio comments for global, overall feedback but written feedback for micro-level advice.

The writers describe 11 ways students (and instructors) can respond to student writing. They explain contexts and purposes when each method might work best.

The writers emphasize that formative feedback (on in-progress assignments) must do two things in order to be effective: 1) identify gaps between the student's current work and assignment expectations and 2) provide specific guidance and explanation for how the student can close the gap. They urge instructors not to use feedback or comments to justify grades students earn.

The writer theorizes the work of responding in terms of the instructor-student exchange. He urges instructors to limit the quantity and focus of feedback to the essential skills/moves/ideas of a given assignment. The writer also describes a typology of 12 kinds of responders instructors might aim to be.

The writer encourages instructors to distinguish between grading and giving feedback. He proposes that the teaching of writing could aim to cultivate reflective conversations among students about their writing and the ways it fulfills or does not yet fulfill their particular aims or rhetorical purposes. He observes that grades (and even the expectation of being graded) can interfere in students' writing and learning processes.

The researchers argue that students need intentional, focused instructions about how to use reflective writing to engage instructor feedback and choose how to use it in their revisions. They found that the understandings of writing students describe in reflections often exceed what they demonstrate in their assignments themselves.

The writer observes that framing feedback in terms of ongoing dialogue and collaboration toward improvement enriches how students experience and react to it. Drawing on interviews with students and samples of instructor feedback and student revisions, she argues for instructors to give specific, engaged feedback that aims at initiating a dialogue between student and instructor.

The researcher found students prefer "specific," "clear," and "understandable" feedback that offers "help or direction but [does] not take control of the writing." He also found that students welcome comments that "pointed out problems and indicated ways to improve their writing." In summary, the research found that students expect feedback to "tell them how they had done on the writing and what they could do to improve it."

The researcher analyzed instructor feedback and student revisions from approximately 80 essays composed in first-year writing classes. She found that students made more and better use of directive comments than general comments or open-ended questions. She found that "hedging" directive comments using "You might" or "What if" or "Perhaps" does not affect the degree to which students use them. She also found that students struggled most to act on comments that required complicated thinking and revisions. Treglia's findings suggest that providing scaffolded support for revision work might make it easier for students to act on instructor feedback.

The writers emphasize that summative feedback on final assignments does not accomplish as much as formative feedback on in-progress drafts. They encourage instructors to design the feedback and revision process to position students as initiators–writers asking for specific feedback–and curators–writers making meaning from the feedback they receive and choosing how to develop their work in response.

The researchers propose a model for helping students understand and make use of instructor feedback. Two of their recommendations are directly relevant to EWP pedagogy: prompting students to choose which feedback they use as they revise and explain assessment criteria in their own words and apply them to their drafts-in-progress.

This recommended reading doesn't yet have an annotation. But you might create one.

This recommended reading doesn't yet have an annotation. But you might create one.

This recommended reading doesn't yet have an annotation. But you might create one.

This recommended reading doesn't yet have an annotation. But you might create one.

This recommended reading doesn't yet have an annotation. But you might create one.

This recommended reading doesn't yet have an annotation. But you might create one.

This recommended reading doesn't yet have an annotation. But you might create one.

This recommended reading doesn't yet have an annotation. But you might create one.

This recommended reading doesn't yet have an annotation. But you might create one.

This recommended reading doesn't yet have an annotation. But you might create one.

This recommended reading doesn't yet have an annotation. But you might create one.

This recommended reading doesn't yet have an annotation. But you might create one.

This recommended reading doesn't yet have an annotation. But you might create one.

This recommended reading doesn't yet have an annotation. But you might create one.

This recommended reading doesn't yet have an annotation. But you might create one.

This recommended reading doesn't yet have an annotation. But you might create one.

This recommended reading doesn't yet have an annotation. But you might create one.

This recommended reading doesn't yet have an annotation. But you might create one.

This recommended reading doesn't yet have an annotation. But you might create one.

This recommended reading doesn't yet have an annotation. But you might create one.

This recommended reading doesn't yet have an annotation. But you might create one.

This recommended reading doesn't yet have an annotation. But you might create one.

This recommended reading doesn't yet have an annotation. But you might create one.

This paper considers the impacts of "imposter bias," an emerging cognitive bias arising from the rapid proliferation of AI-generated media. Imposter bias describes the tendency to misidentify human-generated media as AI-generated. Implicitly, the essay asks us to recognize distrust of factual evidence as a problem as serious as the unquestioned acceptance of false evidence. The problem isn't (only) that we may be fooled by doctored photos or texts, but that we'll increasingly mistrust the genuine.

A popular blogger explains "perplexity," one of the core concepts underlying most AI-detection tools. In short, perplexity describes a system's ability to predict the next word in a string of words. Underlying most AI-detection tools is the assumption that human-generated writing is somewhat less predictable than AI-generated writing. It's important to note that the concept of perplexity arose as a way to measure the quality of large-language models (because humans write with low-perplexity). (In my opinion, this necessarily disqualifies perplexity as a useful measure for detecting AI-writing).

This source contains the first references to "burstiness," which is the other concept underlying most AI-detection tools. Like perplexity (the other), burstiness was conceived as a measure of human-like writing. It's based on observations about the natural flow of human language. For example, if a text includes the word "cat," you'll probably see more instances of this word nearby. Our words are not distributed evenly over texts, but in clusters, or "bursts." For thirty years, language model engineers have strived to produce models that mimic the burstiness of human speech. In my own opinion, many things should tell us that "burstiness" is not a sustainable way to measure the differences between human and machine speech. For one, the concept of burstiness exists to close the gap between human and machine's speech patterns. For another, humans actually don't write similar to one another (or even similarly to themselves when writing on different subjects or in different genres). Burstiness is a function of genre, style, voice, manner, etc.

This article (written by students in my Spring 2024 AWE-Research class), examines the ability of Tandon EWP faculty to distinguish between human- and AI-generated writing. They observed that faculty easily identified purely AI-generated writing, but frequently misidentified human writing as AI-generated, and struggled to identify human-AI hybrid writing. It's not a perfect experiment—as the authors themselves acknowledge—but its conclusions are hardly meaningless, either. To me, they demonstrate Casu's point about "imposter bias." The risk of being fooled by fake media may be overshadowed by the risk of disbelieving genuine articles.

This paper puts to the test many of the top AI-detection tools. There are two key takeaways. The first is that the most advanced chatbots produced the least detectable texts. This means we won't know how much faith to put in the results of AI-detection tools, unless we know which chatbot produced the text being analyzed. The second takeaway is that the detection tools that produce the most true positives also produce the most false positives. Likewise, tools that produce the most true negatives also prudence the most false negatives. In this way, detection tools are like weighted coins.

This paper analyzes how students describe the experience of being accused of cheating. The researchers scoured Reddit for posts by students seeking advice about how to navigate cheating accusations and analyzed a total of 49 such conversation threads. Most of the original posters (38) claimed to be falsely accused. Researchers noted five patterns and trends. First, students adopted a legal posture and solicited legalistic advice. Commenters frequently advised denial, since AI-use is exceedingly hard to prove. Others advised escalating the situation to a dean or departmental director or chair. Commenters afraid of being falsely accused committed themselves to working in Google Docs, so as to have evidence to leverage against accusations. Many students who do not use AI describe using AI-detectors to check the likelihood of accusations. "Rewrite until it comes back human." "While some students accepted this as the new status quo, others resisted: "No, don't let it go. AI detectors are a scam, and it's not our responsibility to adjust our writing to make them happy".

This highly-cited paper documents AI-detection tools' tendency to misclassify texts by NNES as AI-generated. In this experiment, more than half of human written samples (some 61 percent) were misidentified. To put this in context, you'd get 20 percent better results by flipping a coin. But here's the big thing: The researchers then enlisted ChatGPT to make the human-generated texts sound more human. And it worked. With ChatGPT's help, only 11 percent of the samples (which were no longer human written) were categorized as AI-generated. This describes a bizarre situation where NNES might rely on AI-generated writing to prove their humanity. To put it another way, the ubiquity of detection tools can counterproductively make Chatbots more attractive, at least for some.

Like Church and Gale above, this paper also helpfully explains "burstiness," which is one two concepts underlying AI-detection tools. The authors here are offering a new mathematical model that better captures and represents the natural flow of human-written texts. Specifically, this new model captures the word distribution patterns that typify human-written texts. When a word (like, say, "cat") appears in a document, it's likely to appear again close by. Words are not typically distributed evenly or randomly, but in clusters or "bursts." That's the gist of "burstiness." My own feeling is that understanding the origins of this concept (existing to help engineers create more human-like machine-generated writing) reveals the inherent limitations of burstiness as a means of distinguishing between human- and machine- generated writing. For one, both humans and machines write "burstily." The concept exists in the first place to erase the gap between human and machine writing—not as any kind of watermark. Second, it should feel obvious that not all human-generated texts are similarly bursty, since burstiness is a function of genre, topic, voice, manner, style, etc. This suggests, to me, the inherent flaws of any AI-detection tool that relies on burstiness.

This paper introduces a new approach to AI-detection (which they call DetectGPT) that doesn't hinge on "perplexity" and "burstiness," which, as I've argued above, seem like inherently flawed measures. DetectGPT takes a novel approach. You show it a text, and it measures the changes required to make that text a likely product of a specific LLM. If it matches what the LLM could have produced, there's a high probability that the text in question was in fact produced by that LLM. While this new approach seems more promising, I see a potential flaw. Since all LLMs (ChatGPT, Claude, Gemini, etc) are trained on different models, their output is dissimilar from one another. In other words, the trait measured by DetectGPT is not generalizable across all LLMs. DetectGPT requires that the user already knows which model produced the candidate text. Of course, it's conceivable that a future version of DetectGPT runs multiple texts on each candidate text (on for each known LLM that might have produced it), but such an AI-detection tool doesn't yet exist. What's more, as Terry points out below, even a hypothetically perfect AI-detection tool could, at best, detect only AI-generated writing. It could not detect all instances of unauthorised AI use (for example, cases where a student relies on AI to generate ideas but expresses them in her own language.)

Unlike many papers here, interested in the effectiveness of AI-detection tools, Perkins and company are also interested in what faculty do with these analyses. The professors in the study, unsurprisingly I think, frequently did not report academic misconduct violations when detection tools advised. The authors bizarrely suggest that faculty need more training so that they might "rely on the raw results of the Turnitin tool." Their faith in AI-detection tools can't easily be explained, especially since they cite multiple sources (including Liang et al, cited above) that demonstrate these tools' tendency toward false positives. More bizarrely, the authors offer a list of tells (characteristics of machine-generated writing) including "overly complex language, papers not meeting length requirements, core texts and content discussed in class not being covered, and falsified references." Are they kidding? What veritable time-honored hallmarks of undergraduate writing! Anyway, it looks like the lead author's faith in detection tools was short-lived. A year after this study, he published another that demonstrated the ease with which leading detection tools can be fooled (see entry just below).

This preprint is all about strategies for fooling AI-detection tools, which is pretty easy to do. A student needs only to paraphrase bits of text or introduce a few spelling or grammar errors. What's most remarkable, I think, is that this trickery can reliably be outsourced to chatbots. Prompting ChatGPT to elevate the diction, for example, can drop true positive rates down to just 13 percent. Of course, a student can rinse and repeat until ChatGPT has laundered a text to the point of un-detectability. This suggests a strange new economy where the proliferation of AI-detection tools actually increases the market for the chatbots. As homework-laundering devices, AI-detection tools exist to make LLMs less risky, not more —their outputs more indistinguishable from human-generated writing. Universities worried about academic integrity would be mistaken to endorse them. This is partly why NYU's Vice Provost for Technology and AI, Clay Shirky, has instructed Turnitin to disable their AI-detection feature on Brightspace, and why the C-FSC's Technology and Administration Committee endorsed that move. Like Laing et al (cited above), this paper provides further evidence of detection tools' bias against NNES. Interestingly, perhaps strangely, the authors don't dismiss detection tools altogether. They suggest detection tools can helpfully "demonstrate to students how GenAI text may be integrated with their own writing," which I think means it can help students to hide their seams. No thanks.

This recommended reading does not yet have an annotation.

This recommended reading doesn't yet have an annotation. But you might create one.

This recommended reading doesn't yet have an annotation. But you might create one.

This recommended reading doesn't yet have an annotation. But you might create one.

The proliferation of artificial intelligence (AI) tools has transformed numerous aspects of daily life, yet its impact on critical thinking remains underexplored. This study investigates the relationship between AI tool usage and critical thinking skills, focusing on cognitive offloading as a mediating factor. Utilising a mixed-method approach, we conducted surveys and in-depth interviews with 666 participants across diverse age groups and educational backgrounds. Quantitative data were analysed using ANOVA and correlation analysis, while qualitative insights were obtained through thematic analysis of interview transcripts. The findings revealed a significant negative correlation between frequent AI tool usage and critical thinking abilities, mediated by increased cognitive offloading. Younger participants exhibited higher dependence on AI tools and lower critical thinking scores compared to older participants. Furthermore, higher educational attainment was associated with better critical thinking skills, regardless of AI usage. These results highlight the potential cognitive costs of AI tool reliance, emphasising the need for educational strategies that promote critical engagement with AI technologies. This study contributes to the growing discourse on AI's cognitive implications, offering practical recommendations for mitigating its adverse effects on critical thinking. The findings underscore the importance of fostering critical thinking in an AI-driven world, making this research essential reading for educators, policymakers, and technologists.

This article doesn't yet have an annotation, but you can add one!

This recommended reading doesn't yet have an annotation. But you might create one.

The rise of Generative AI (GenAI) in knowledge workflows raises questions about its impact on critical thinking skills and practices. We survey 319 knowledge workers to investigate 1) when and how they perceive the enaction of critical thinking when using GenAI, and 2) when and why GenAI affects their effort to do so. Participants shared 936 first-hand examples of using GenAI in work tasks. Quantitatively, when considering both task- and user-specific factors, a user's task-specific self-confidence and confidence in GenAI are predictive of whether critical thinking is enacted and the effort of doing so in GenAI-assisted tasks. Specifically, higher confidence in GenAI is associated with less critical thinking, while higher self-confidence is associated with more critical thinking. Qualitatively GenAI shifts the nature of critical thinking toward information verification, response integration, and task stewardship. Our insights reveal new design challenges and opportunities for developing GenAI tools for knowledge work.

Published well before ChatGPT 2.0's release in Nov. 2022, this meta-analysis was ahead of the curve, but already feels dated. The researchers compared how 30 articles defined and measured "AI literacy" in the context of K-12 education. They cleverly borrow and adapt Bloom's taxonomy to propose a six-tiered hierarchy of AI literacy, which begins with memorizing basic facts about AI and culminates in the ability to create novel AI platforms. As useful as this is, I worry the researchers are conflating AI literacy with AI proficiency. Although the former requires a bit of the later, it feels important to keep these ideas distinct. The difference is as important as that between sex education and sex lessons. What I think we want most for our students is the ability to make good choices about AI use, not the ability to do it better. The reviewed studies generally agree AI literacy means "having the essential abilities that people need to live, learn and work in our digital world through AI-driven technologies" (Ng 2021) though there's little agreement on what these specific abilities are. For me, a key ability is recognizing the potential for learning loss, given that AI's uptake and impact on education is more clear and direct than any other industry.

I've been scanning literature to see how researchers define "AI Literacy," and this Oct 2024 paper by a joint MLA-CCCC Task Force is far and away the most useful resource to date. While most conflate AI Literacy with AI Proficiency (University of Florida's AI Across the Curriculum Initiative is among the worst offenders), these researchers helpfully define AI Literacy as a specific set of skills and habits that promote learning and growth in the age of AI. For example, their "AI Literacy" includes: The ability to explain one's reasons for using or not using AI for critical reading, writing, research, and learning processes. The ability to distinguish between one's own contribution to a final product and that of AI. The ability to monitor one's GenAI use in relation to their learning goals, including recognizing when an over-reliance on GenAI tools prevents learning. The ability to monitor how GenAI habits affect their development as writers and identify when GenAI use limits growth and potential. The development of writing skills necessary to engineer prompts successfully and to identify the limitations of AI-generated prose. These goals feel very crucial, concrete, and achievable—and many of them are already deeply embedded in our classes, because they boil down to metacognition.

The article introduces UF's 2023 AI Initiative called AI Across the Curriculum. Their expressed goal is to give students in all majors experience using AI tools. The article doesn't clearly define their goals, or describe how they will measure their success. "[A]ll students are provided with a suite of AI opportunities and are encouraged to engage," they write. There's not a lot in the way of rationale or thoughtful implementation other than to say "[q]ualified human capital is essential to face the challenges of the 21st-century, and UF is positioning itself to lead in meeting this global societal need." Here's how the initiative works: Certain classes will be designated as part of the initiative using five tags, which refer to different competencies: Understanding AI, Using and applying AI, evaluating AI, and Evaluating and Creating AI. (Notice these roughly follow the Bloom taxonomy, recommended by Ng above.) Like Ng, therefore, UF's definition of AI literacy looks a lot like AI proficiency, where the apex is the creation and evaluation of AI tools—rather than the creation and evaluation of new knowledge, as per Bloom. That feels like a problem. Even as it tries to embrace and apply good pedagogical theory, the initiative is fundamentally AI centered rather than learning-centered.

This recommended reading doesn't yet have an annotation. But you might create one.

Abstract: As artificial intelligence (AI) became a part of daily life, it has become important to determine user competence in using AI technology. Here, we propose the concept of AI literacy and develop a quantitative scale for obtaining accurate data regarding the AI literacy of ordinary users. We first identified the primary core constructs of AI literacy, including awareness, use, evaluation, and ethics. Next, we generated 65 items to capture these four constructs; only 31 items were retained after a three-step content validation process. Then, we conducted a survey, and collected two samples of data. By reducing the number of items using the first sample and performing reliability and validity tests on the second sample, we obtained a 12-item instrument for the quantitative measurement of AI literacy. The results confirmed that the proposed four-construct model is an adequate representation of AI literacy. Further, AI literacy is significantly related to digital literacy, attitude towards robots, and users' daily usage of AI. This study will not only aid researchers in understanding how user competence in using AI technology affects human–AI interactions but will also help designers develop AI applications that are aligned with the AI literacy levels of the target users.

A Director of Academic Integrity at UC San Diego, Bertram Gallant designed this guide as a choose your own adventure, with multiple entry points depending on user's level of experience with AI and pedagogical training. Click here to see NYU's adaptation (or "remix") of the same guide.

Cryer argues that the burden of enforcing Academic Integrity policies has shifted unfairly from universities to students. He points out that political scientists call this "responsibilization," which means "the moral burdening of the entity at the end of the pipeline [of power and authority]." However, others might actually view this as a return to a norm that characterized education before Turnitn. The assumption that universities can easily detect and punish cheating is an outlier in the history of education. Nevertheless, Cryer's larger point is that it's neither realistic nor fair to expect students to bear this responsibility very well—unless we give them the resources and training they need to make good choices. "no previous generation has been faced with the ever-present option to offload their work, at no cost, with a low likelihood of immediate negative consequences." Therefore, it's never been more necessary to explain clearly explain to students the value and utility of well-designed assignments.

In this article, a team of Australian researchers study the 39 academic integrity policies of 39 universities, and come away with a list of five best practices. In short, they conclude that the most effective academic integrity policies are accessible, detailed, supported, and grounded in process and values. Also, they must outline the responsibilities of all parties. Although this does not pertain explicitly to Artificial Intelligence policies, these characteristics nonetheless provide a helpful framework for evaluating AI policies.

Complied by University of Missouri Faculty Developer Lance Eaton, this Google spreadsheet is a repository of classroom AI policies shared by hundreds of college faculty members all over the world. {URL] Generally speaking, they range from unhelpful to very unhelpful. In my opinion, this document is useful only as an artifact documenting universities' early awkward attempts to come to terms with generative AI. However, there are a few exceptions.

Kolb & Co. study students' reasons for not cheating. Although they found that the majority of students choose not to cheat (good news!), they were discouraged by the students' motivations for uprightness. Around one third of the students say they chose not to cheat because they felt cheating wasn't feasible or impractical. In other words, they suspected cheating would involve more work than not cheating, or they felt certain they'd be caught. The authors conclude with a final worry that "constructing the "cheat proof" classroom may—inadvertently—impede students' ethical development." In other words, the best classroom is one that teaches students how and why to overcome motivations to cheat even when provided ample opportunity. I should note that this research diverges from the majority of research in this area, which suggests that cheating is not actually strongly tied to opportunity.

Developed by the students in Alexander Landfair's Fall 2024 "International Writing Workshop Introduction", this document names the ethical and pedagogical principles that guide students' decision-making about using generative AI tools.

This article explains that the promise of learning can provide stronger disincentives to cheating than punishments or consequences. However, it might also follow that perceived lack of education value of an assignment creates the strongest incentives toward cheating. Therefore, it falls on teachers to state explicitly the value and utility of assignments, which are never as obvious or clear as instructors think. Stating learning goals on each assignment is a key way to reinforce academic integrity policies.

This article compares the efficacy of Academic Integrity Education programs at 44 Australian and New Zealand universities. They find that many programs suffer from lack of clarity about the values and practical skills that give rise to rules and regulations. They recommend that academic integrity programs be well-timed, so as to meet students when most needed, and collaborative, providing students with opportunities to check and improve understanding. Although they don't mention AI policies in particular, this framework is clearly useful for instructors crafting and evaluating classroom policies and supporting lessons.

With help from educators across the nation, the US Dept. of Education developed these recommendations to help prepare post-secondary schools prepare for changes brought on by generative AI: Establish clear policies; create infrastructure for support and advising; rigorously test and evaluate AI tools and services; form partnerships for testing AI models across educational applications; and reconsider programs in light of growing impact of AI on disciplines and job market.

Watters, a popular ed-tech theorist, offers some strong opinions on the larger trend toward surveillance and policing in the classroom. Having read much of her work, I think it's fair to say that Watters takes serious problems like academic misconduct and low student engagement that lead to surveillance and policing. It's only that these are poor answers to those problems. Not only do things like detection tools fail to address the root causes of academic misconduct and low-engagement, they actually exacerbate these problems by eroding the assumption of good faith and trust on which student-centred learning rests.

The author, a leading voice in AI in education, describes two major advances in generative AI between 2022 and early 2025. In late 2024, researchers have bypassed the key obstacle in the way of building more powerful LLMs. AI developers are no longer constrained by the size of a model's training data. Instead, they effectively program models to check their thinking as they go. "This is a big deal," Mollick explains "because previously the only way to make AIs perform better was to train bigger and bigger models, which is very expensive and requires a lot of data." If I understand correctly, they are effectively automating effective prompt engineering, so users can get in one prompt the kind of quality results ordinarily only achievable through a series of high-quality prompts. As a result, OpenAI's latest "reasoner" (as these are called) is capable of near-PhD-level analysis of topics. "These aren't the usual AI hallucinations or misquoted papers," Mollick claims, "they're legitimate, high-quality academic sources… When I click the links, they don't just lead to the papers, they often take me directly to the relevant highlighted quotes."

The rise of Generative AI (GenAI) in knowledge workflows raises questions about its impact on critical thinking skills and practices. We survey 319 knowledge workers to investigate 1) when and how they perceive the enaction of critical thinking when using GenAI, and 2) when and why GenAI affects their effort to do so. Participants shared 936 first-hand examples of using GenAI in work tasks. Quantitatively, when considering both task- and user-specific factors, a user's task-specific self-confidence and confidence in GenAI are predictive of whether critical thinking is enacted and the effort of doing so in GenAI-assisted tasks. Specifically, higher confidence in GenAI is associated with less critical thinking, while higher self-confidence is associated with more critical thinking. Qualitatively GenAI shifts the nature of critical thinking toward information verification, response integration, and task stewardship. Our insights reveal new design challenges and opportunities for developing GenAI tools for knowledge work.

This recommended reading doesn't yet have an annotation. But you can add one.

Jaswinder Bolina, a poetry professor at University of Miami, says we can save our jobs—and hearts and minds—from the clutches of A.I. by learning to be more poetic.

Ted Chiang, a New Yorker staff writer, describes the inherent un-impressiveness of generative AI tools.

Ted Chiang, a New Yorker staff writer, explains why generative-AI tools are inherently incapable of producing art.

Noam Chomsky—a cognitive scientist known as "the father of modern linguistics"—says machine learning is based on wrong ideas about knowledge and probability.

This recommended reading does not yet have an annotation. But you might create one.

Angela Duckworth, University of Pennsylvania psychology professor, believes ChatGPT won't erode students' ability to think, but might actually enhance it.

Irina Dumitrescu, New York Times and New York Review of Books contributor, says students who don't write don't learn.

Yuval Harari, a history professor at Hebrew University of Jerusalem, argues that AI may cost humans our power to shape our cultures and societies.

Henry Kissinger, a former U.S. Secretary of State, says the AI revolution is changing humanity's relationship with knowledge, warning that artificial intelligence may soon advance beyond human's ability to question it.

Rohan Mehta—a high-school senior from Pennsylvania— suggests highschools should not forbid ChatGPT but instead show students how to use it ethically and productively.

Ethan Mollick, a business professor at The Wharton School of University of Pennsylvania, says college classes must instruct students how to use ChatGPT as a learning tool.

Adrienne LaFrance, The Atlantic's executive publisher, says surviving the AI revolution requires us to hold on to the things that make us human, including ephemerality, uncertainty, and physical touch.

Victoria Livingstone, a veteran professor of literature, explains why she quit her job because of ChatGPT-fueled cheating.

NYT columnist and poet Margaret Renkl wonders whether robots who find our right words for us aren't taking from us a pleasurable agony that makes us human.

Chad Raymond, a political science professor, argues that Project-based classrooms are the solution to the problems posed by generative AI tools.

This recommended reading doesn't yet have an annotation. But you might create one.

Inara Scott, an associate dean of teaching and learning at Oregon State University, says ChatGPT will doom higher education unless professors reinvent the college curriculum.

Lawrence Shapiro, a law professor at University of Wisconsin-Madison, refuses to abandon writing assignments, even though some students may use GPT to cheat.

Owen Terry, a computer science major at Columbia University, says college professors must design better "AI proof" writing assignments.

This recommended reading doesn't yet have an annotation. But you might create one.

This recommended reading doesn't yet have an annotation. But you might create one.

This recommended reading doesn't yet have an annotation. But you might create one.

Jonathan Zimmerman, a history professor at UPenn, believes that generative AI can not only prevent students from learning, but also from discovering themselves.

This recommended reading doesn't yet have an annotation. But you might create one.

This recommended reading doesn't yet have an annotation. But you might create one.

This recommended reading doesn't yet have an annotation. But you might create one.