AI English and the environmental cost

University of Michigan

AI-assisted summary of AI English and the environmental cost — original source. Check the original for full context.

Source published: 2026-09-15

Key ideas

Submitting a simple query to generative AI systems like ChatGPT requires an estimated 50 times more electrical energy than a traditional Google search, accelerating computational carbon footprints and ocean warming. (9:56)
Empirical linguistic research shows that instructional tuning and prompt engineering do not make machine text more human-like, frequently steering outputs further away from natural conversational patterns. (3:56)
Current AI detection and evaluation tools systematically penalize non-native English speakers and regional dialect users, falsely flagging standard, careful syntax as plagiarized or machine-generated. (9:56)
Generative models inherit the narrow cultural values of their developers, producing an 'algorithmic monoculture' that favors standardized 'exam English' over globally spoken varieties. (12:24)

Detailed notes

## Background Published on YouTube by the University of Michigan on September 15, 2026, this episode of the Michigan Minds podcast features English and linguistics professor Laura Aull discussing the emergence and impact of "AI English." The show description highlights how generative AI tools disrupt natural human speech diversity, replicate linguistic bias across automated hiring and housing platforms, and impose a massive, underappreciated carbon footprint through everyday queries.

## TL;DR Linguist Laura Aull explains that generative models produce a homogenized "AI English" driven strictly by probability, flattening the dynamic variation, humble tone, and cognitive readability inherent in living human languages. Because large language models are trained on narrow, prestige dialects like standardized "exam English," automated systems and AI detectors disproportionately penalize multilingual writers and speakers of global dialects such as Jamaican and Indian English. Beyond societal and workplace gatekeeping, everyday adoption of these models carries an alarming environmental cost, requiring roughly 50 times the electricity of standard search engine queries. To preserve authentic human agency, Aull urges society to resist algorithmic monoculture, prioritize linguistic nuance over speed, and cultivate conversational openness rather than assertive machine rhetoric.

## Atmosphere & style The discussion carries an intellectual, measured, and conscientious tone, blending academic corpus linguistics with accessible cultural critique. Juan Ochoa facilitates an easy, thoughtful exchange, allowing Aull to weave together empirical research and personal passions, such as her reflections on ocean warming as a scuba diver. The dialogue moves from diagnostic analytical comparisons to urgent ethical concerns before settling into a warm, encouraging closing call for linguistic empathy and human agency.

## Bulletin - [09:56](https://www.youtube.com/watch?v=lufat51rPUw&t=596s) Submitting a simple query to generative AI systems like ChatGPT requires an estimated 50 times more electrical energy than a traditional Google search, accelerating computational carbon footprints and ocean warming. - [03:56](https://www.youtube.com/watch?v=lufat51rPUw&t=236s) Empirical linguistic research shows that instructional tuning and prompt engineering do not make machine text more human-like, frequently steering outputs further away from natural conversational patterns. - [09:56](https://www.youtube.com/watch?v=lufat51rPUw&t=596s) Current AI detection and evaluation tools systematically penalize non-native English speakers and regional dialect users, falsely flagging standard, careful syntax as plagiarized or machine-generated. - [12:24](https://www.youtube.com/watch?v=lufat51rPUw&t=744s) Generative models inherit the narrow cultural values of their developers, producing an 'algorithmic monoculture' that favors standardized 'exam English' over globally spoken varieties.

## Key takeaways - [00:00](https://www.youtube.com/watch?v=lufat51rPUw&t=0s) Laura Aull's research employs corpus linguistics to analyze digitized texts, examining how linguistic patterns influence educational assessment, power structures, and the boundary between human and synthetic language. - [01:13](https://www.youtube.com/watch?v=lufat51rPUw&t=73s) Unlike traditional voice assistants that act as informational filters, generative AI acts as a language selection engine, taking over intimate personal decisions regarding vocabulary and phrasing. - [01:13](https://www.youtube.com/watch?v=lufat51rPUw&t=73s) Human English represents an expansive world of expressive possibility across global dialects, whereas AI English is merely a narrow 'can of probability' derived from standardized essays, books, and curated web discussions. - [03:56](https://www.youtube.com/watch?v=lufat51rPUw&t=236s) Prompt engineering cannot compensate for structural absences in training data; predictive models lack the conversational breadth and nuanced interpersonal markers required for authentic human warmth. - [05:27](https://www.youtube.com/watch?v=lufat51rPUw&t=327s) Automated writing evaluation is as old as artificial intelligence itself, tracing back to the Project Essay Grade (PEG) system developed in the early 1960s alongside milestone gaming programs. - [05:27](https://www.youtube.com/watch?v=lufat51rPUw&t=327s) Because foundational design choices for LLMs were concentrated among a small, unrepresentative demographic, AI English serves as a linguistic echo of power that systematically marginalizes world Englishes. - [07:52](https://www.youtube.com/watch?v=lufat51rPUw&t=472s) While AI is suited for rule-bound tasks like basic computer coding, high-stakes arenas like hiring, housing, and legal decisions demand human empathy and should not be outsourced to probability-based efficiency. - [09:56](https://www.youtube.com/watch?v=lufat51rPUw&t=596s) A standard ChatGPT query consumes roughly 50 times more electricity than a traditional search engine lookup, raising severe ecological concerns regarding ocean warming. - [11:24](https://www.youtube.com/watch?v=lufat51rPUw&t=684s) Linguistic detection filters exhibit inherent bias against non-native writers and regional varieties like Jamaican and Indian English, misinterpreting their syntax as machine-generated text. - [11:24](https://www.youtube.com/watch?v=lufat51rPUw&t=684s) Across models and architectures, AI text consistently remains more grammatically dense, impersonal, and overly confident, whereas human language thrives on dialectal variety, readability, and humility. - [12:24](https://www.youtube.com/watch?v=lufat51rPUw&t=744s) Large language models reinforce 'exam English'—an artificial dialect historically favored by academic gatekeepers—at the expense of sensible, regularized global forms like English as a Lingua Franca. - [15:59](https://www.youtube.com/watch?v=lufat51rPUw&t=959s) Corpus analysis indicates that AI English suffers from severe cognitive bloat, stacking heavy strings of nominalizations around small verbs rather than utilizing accessible, readable 'noun-verb-noun' syntax. - [18:52](https://www.youtube.com/watch?v=lufat51rPUw&t=1132s) The 'Grammar for Good' initiative advocates for approaching diverse language with curiosity and structuring prose as a humble, qualified dialogue rather than an assertive, megaphone-style broadcast.

## Detailed walkthrough ### [00:00](https://www.youtube.com/watch?v=lufat51rPUw&t=0s) Introducing Professor Laura Aull and the Study of "AI English"

The Michigan Minds podcast opens with Juan Ochoa, international public relations representative for the University of Michigan News office, introducing guest Laura Aull. Aull is a professor of English and linguistics who recently served as director of the English Department Writing Program at the University of Michigan. Her scholarship relies heavily on corpus linguistics—the large-scale analysis of digitized text collections—to explore how recurring patterns in language shape writing assessment, educational achievement, and structures of social power. Within the university, her research directly addresses how people are taught to write and how emerging technologies are actively reconfiguring human communication.

Aull is currently completing two major monographs that frame her current investigations: *AI English: Humans, Machines, and the Future of Writing*, which interrogates the shifting threshold between organic human language and machine-generated text, and *Grammar for Good*, a study exploring how structural linguistic awareness can cultivate more civil and ethical public discourse. Ochoa sets the stage by observing that while consumer AI has quietly integrated into daily life for years through voice assistants like Siri or Alexa, contemporary generative tools represent a dramatic escalation, raising the central question: by adopting these platforms, are users gaining a smart assistant, or are they surrendering their authentic voice?

- Laura Aull's methodology centers on corpus linguistics to investigate language patterns, academic success, and social power dynamics. - Her upcoming monograph *AI English: Humans, Machines, and the Future of Writing* focuses directly on the evolving border between human expression and synthetic text. - Her companion project, *Grammar for Good*, examines the role of linguistic knowledge in promoting ethical, civil discourse.

[01:13](https://www.youtube.com/watch?v=lufat51rPUw&t=73s) A Big Can of Some English: Probability versus Possibility

Addressing the distinction between legacy assistants and modern generative models, Aull draws an instructive comparison between Apple's Siri and OpenAI's ChatGPT. While both systems function through selection—asking Siri to play music requires trusting the algorithm to select specific tracks while excluding others—using ChatGPT to draft an email offloads decisions about language itself. Because language has historically served as a deeply intimate, social human practice, trusting a machine to select vocabulary and phrasing marks a profound shift. Siri acts as an informational selection tool, whereas ChatGPT acts as a language selection engine.

This dynamic illustrates what Aull characterizes as the foundational "creation story" distinguishing human English from AI English. Human English encompasses the vast, open-ended landscape of global English: it spans formal and informal registers, technical and interpersonal modes, native speakers and multi-language users, and regional varieties stretching from Ann Arbor to Sri Lanka. In contrast, AI English is drawn strictly from what is statistically probable within a narrow slice of text—principally standardized exam essays, published books, and sanitized, agreeable internet discussions. Consequently, generative tools default to formal, informational, assertive, and relentlessly enthusiastic phrasing. Aull summarizes the divide with a vivid metaphor: human English represents a vast world of possible expression, whereas AI English is merely "a big can of some English driven by probability." Relying on models like ChatGPT, Claude, Llama, or Gemini inevitably surrenders human choice and voice.

```mermaid flowchart TD accTitle: Human English vs AI English Comparison accDescr: A flowchart contrasting the expansive possibility of global human English with the restricted probability of AI English.

subgraph Human["Human English (World of Possibility)"] H1["Global English Varieties (Ann Arbor to Sri Lanka)"] H2["Fluid Registers (Informal, Interpersonal, Nuanced)"] H3["Native and Multilingual Speakers"] end

subgraph Machine["AI English (Can of Probability)"] M1["Narrow Training Data (Standardized tests, books, polite web posts)"] M2["Statistically Driven Register (Formal, assertive, highly enthusiastic)"] M3["Algorithmic Selection (Curated exclusions)"] end

Human -->|"Infinite expressive choices"| OUT1["Expansive Human Voice"] Machine -->|"Probability-based output"| OUT2["Homogenized Machine Voice"] ```

[03:56](https://www.youtube.com/watch?v=lufat51rPUw&t=236s) The Myth of Prompt Tuning and the Limits of Synthetic Personability

When Ochoa asks whether users naturally counteract machine uniformity by editing outputs or prompting the software to sound more natural, Aull pushes back against widespread assumptions. While instructional tuning and detailed prompt engineering do alter surface-level outputs, Aull points out that empirical linguistic research demonstrates that instructional tuning does not actually bring machine outputs closer to authentic human language—it frequently makes the text even less human-like.

The underlying limitation remains structural: prompt tuning cannot draw from registers the model never ingested. If a user instructs a model to adopt a casual or informal tone, the machine can only simulate informality from the limited sample of informal text inside its dataset. True human language is uniquely personable because real speakers dynamically weave together informational content and nuanced interpersonal markers. Because humans naturally possess this full expressive repertoire, they achieve far greater warmth, humility, and connection when writing independently, even if they occasionally treat generative models as an initial reference point.

- Instructional tuning modifies surface responses but linguistically moves outputs further away from genuine human writing patterns. - AI systems lack access to the organic breadth of conversational and informal speech varieties found across living communities. - Human communicators uniquely balance personal vulnerability, social rapport, and factual information in ways predictive models cannot simulate.

[05:27](https://www.youtube.com/watch?v=lufat51rPUw&t=327s) The History of AI English: From 1960s Automated Grading to LLMs

Tracing how AI English reached its present state, Aull emphasizes that assessing language via computing is as old as artificial intelligence itself. While public memory often anchors the birth of AI to early computing breakthroughs like IBM's checkers-playing program in the 1950s and 1960s, the Project Essay Grade (PEG) automated writing evaluation system was developed during that exact same era in the early 1960s to score student essays mechanically. The quest to measure and capture human expression has accompanied machine learning from its earliest days.

The transition to modern generative systems represents a massive leap in processing scale and architectural sophistication. Contemporary transformer architectures leverage vastly larger datasets and immense compute budgets to process expansive context windows, producing outputs that convincingly mimic authentic human dialogue. Yet the fundamental problem persists: these models were designed by a tiny, unrepresentative demographic. Citing researcher Karen Hao's critical reporting on the geopolitical and social architecture of technology, Aull explains that because a remarkably small group of developers made foundational design choices, AI carries a narrow worldview and an insular set of cultural values. AI English constitutes the direct "linguistic echo" of this concentration of power, inheriting training models that intentionally elevate certain prestige dialects while systematically erasing the vast majority of global English varieties.

- Algorithmic evaluation of writing dates back to PEG (Project Essay Grade) in the early 1960s, running parallel to milestone gaming demonstrations. - Modern large language models rely on transformer architectures and vast context windows to simulate human-like coherence. - Generative models reflect the narrow demographics and worldview of their developers, turning AI English into a linguistic monoculture that excludes broader global practices.

[07:52](https://www.youtube.com/watch?v=lufat51rPUw&t=472s) Efficiency versus Human Nuance in High-Stakes Decision-Making

Addressing common anxieties regarding workplace automation and economic displacement, Aull stresses that whether machines replace human labor depends entirely on context. Generative models operate through probability rather than understanding, leaving them inherently incapable of empathetic or truly innovative decision-making. Routine, rule-governed tasks that map cleanly onto probability distributions—such as producing basic computer code—can be handed off to automation for substantial efficiency gains.

Conversely, domains that demand moral judgment, institutional flexibility, and empathetic nuance must remain under human stewardship. Aull highlights high-stakes arenas where automated decisions introduce dangerous distortions: writing assessment, hiring evaluations, legal decisions, career matching, and housing allocation. Empirical studies consistently demonstrate that algorithmic decision-making in these areas exhibits pronounced discrimination against anyone whose speech deviates from standard "exam English"—the sanitized dialect of standardized testing and academia. Aull cautions that speed and efficiency must never override human fairness: "Human nuance, variety, and flexibility are more worthwhile than efficiency."

[09:56](https://www.youtube.com/watch?v=lufat51rPUw&t=596s) Looking to 2030: Environmental Cost and Algorithmic Bias

Looking forward toward the year 2030, Ochoa asks whether society will eventually stop caring whether books or journalistic reporting are written by humans or machines. Aull notes that current research reveals humans still care deeply; in fact, recipients react with visceral negativity upon discovering that correspondence or creative text was machine-generated. Setting aside emotional reactions, Aull outlines three urgent structural arguments for why humanity must continue to care, detailing the first two in depth.

The first imperative is the severe environmental cost of large-scale computational processing. Submitting a simple query to ChatGPT consumes roughly 50 times the electrical energy required for an equivalent query on a conventional search engine like Google. Speaking from personal passion as an avid scuba diver, Aull stresses that warming ocean temperatures represent an urgent ecological emergency, arguing that society cannot afford to normalize such extreme energy consumption for trivial digital conveniences. The second imperative is pervasive linguistic bias: automated detectors and language filters systematically penalize writers whose phrasing falls outside narrow statistical norms. Non-native speakers and multilingual writers are routinely subjected to false accusations of plagiarism or caught in AI-detection filters simply because their structured, careful phrasing overlaps with machine outputs, while users of legitimate regional dialects—from Jamaican English to Indian English—find their native varieties discounted entirely.

- Generating a single query through ChatGPT consumes an estimated 50 times more electricity than a traditional Google search. - Aull's direct experience as a scuba diver informs her concern over ocean warming and the irresponsible normalization of AI's carbon footprint. - Stylistic bias in AI detection tools disproportionately harms non-native English writers, falsely flagging authentic work for academic dishonesty while ignoring global dialects like Jamaican and Indian English.

[11:24](https://www.youtube.com/watch?v=lufat51rPUw&t=684s) The Fundamental Divide Between Human and AI English

AI detection tools regularly penalize non-standard dialects and non-native writers, mistakenly flagging their original work for plagiarism or machine generation simply because their syntax deviates from expected norms. Current models struggle to comprehend the broad spectrum of global dialects, remaining largely oblivious to varieties ranging from Jamaican English to Indian English. This points directly to the fundamental divide between human expression and synthetic generation: human English exhibits far greater natural variation and readability. Human language thrives across an expansive tapestry of dialects and genres, remaining less grammatically dense and demonstrably more humble. AI English, on the other hand, presents as rigidly formal, impersonal, and overly sure of itself.

Despite public assumptions that ongoing AI advancements will smooth out these quirks, Professor Laura Aull notes that these defining markers have remained consistent across models, bots, and time. AI text continues to be less readable and far less varied than human writing. Aull warns that if society stops caring whether content is authored by a person or an algorithm, whether by 2030 or in the present day, we are passively consenting to a linguistic future that is flattened, homogenized, and stripped of human character.

[12:24](https://www.youtube.com/watch?v=lufat51rPUw&t=744s) Algorithmic Monoculture and "Exam English"

When asked whether automated language models could level the playing field or merely exacerbate historical exclusions in contexts like job hiring, Aull concedes that in theory, models built on broader varieties of world English or curated small language models (like Bias Shield) could raise awareness. However, the current reality of large language models is that they act as a "big can of some English driven by probability." By prioritizing high-probability sequences, they inherently favor select language users while systematically marginalizing others. Researchers describe this phenomenon using phrases like "homogeneity by design" and "algorithmic monoculture," highlighting how natural variation is smoothed out. Furthermore, post-training instruction tuning tends to amplify this artificiality, making AI outputs sound even more like AI and less like natural human speech.

To illustrate how certain varieties become privileged, Aull prefers the term "exam English" over "standard English." This framing highlights the artificial dialect historically rewarded in classroom instruction, standardized testing, and employment decisions since the dawn of mass public schooling roughly 150 years ago. Exam English typically favors native-speaker idioms, even when those conventions are logically inconsistent.

A prime example is present-tense subject-verb agreement: standard English arbitrarily requires an extra "-s" for the third-person singular ("she asks"), while keeping other forms uninflected ("you ask"). By contrast, English as a Lingua Franca—used by over a billion global speakers as an additional contact language—frequently applies a sensible, regular pattern across all subjects ("I ask, you ask, she ask, they ask"). Despite being less logically consistent, the quirky conventions of exam English remain institutional gatekeepers. Aull urges that preserving the vitality of human language requires tolerating its inherent messiness, embracing innovation, and letting go of rigid biases tied strictly to test-taking conventions.

```mermaid flowchart TD accTitle: Exam English vs Global Lingua Franca accDescr: Diagram illustrating how standardized exam English and global lingua franca differ in consistency and cultural prioritization. subgraph Exam["'Exam English' (Standardized)"] direction TB E1["Inconsistent Verb Forms ('she asks' vs. 'you ask')"] E2["Institutional Gatekeeping (Schools, Tests, Hiring)"] E1 --> E2 end subgraph ELF["Lingua Franca (Global English)"] direction TB L1["Consistent Verb Forms ('I ask, you ask, she ask, they ask')"] L2["Pragmatic Communication (1+ Billion Contact Users)"] L1 --> L2 end subgraph AIImpact["Algorithmic Filter"] AI["LLMs & Instruction Tuning"] Mono["Algorithmic Monoculture (Biased toward standardized probability)"] AI --> Mono end Exam -->|"Reinforces"| AI ELF -.->|"Excluded / Penalized by"| AI ```

[15:59](https://www.youtube.com/watch?v=lufat51rPUw&t=959s) Linguistic Differences: Variation and Readability

Drawing from the corpus analysis conducted for her book *AI English*, Aull compares patterns across centuries of human writing—spanning everything from informal text messages to bureaucratic government memos—against generative outputs from systems like ChatGPT, LLaMA, and DeepSeek. Her findings crystallize around two core linguistic differences:

| Feature Dimension | Human English | AI English | |:--- |:--- |:--- | | **Variation & Cohesion** | High dialect diversity; dynamic expressive forms (e.g., elongating letters like "yaaaay"); varied verb agreements ("they is" alongside "they are"); diverse structural linking devices. | Monotonous lexical choices; rigid adherence to standard verb paradigms; repetitive, predictable cohesive transitions. | | **Grammatical Readability** | Balanced alternation between plain, accessible clauses and dense information structures. | Unrelenting informational density; repetitive strings of long noun phrases wrapped around small verbs. | | **Cognitive Accessibility** | Easy for human brains to parse via clear "noun-verb-noun" flow, clearly identifying who is doing what. | Heavy, difficult-to-parse nominalizations ("noun-noun-noun-verb-noun-noun-noun") that make reading fatiguing. |

When asked whether prompt engineering could resolve this readability gap, Aull explains that research shows users are neither particularly skilled at nor interested in prompting models for plain, accessible syntax. While it is theoretically possible to steer models toward greater readability and variation, default system behaviors consistently trend toward heavy, nominalized bloat.

[18:52](https://www.youtube.com/watch?v=lufat51rPUw&t=1132s) Grammar for Good: Dialogue Over Megaphones

Addressing the pervasive fatigue caused by toxic online communication and constant digital shouting, Aull introduces her project, "Grammar for Good." She outlines actionable ways to adjust both reading and writing habits to foster empathy and clarity:

* **Reading with an open mind:** Readers should approach unfamiliar phrasing with curiosity rather than knee-jerk judgment, asking "what is going on with this English?" and finding delight in dialectal diversity instead of clinging to rigid standard rules. * **Writing as a dialogue rather than a megaphone:** Writers should deliberately structure prose around conversational openness, integrating hedges, qualifiers, and concessions. This rhetorical framing leaves room for reasonable doubt and alternative reactions. * **Rejecting synthetic assertiveness:** AI text frequently mimics a "megaphone"—uncompromisingly direct, assertive, and prone to rhetorical shouting or superficial flattery. In contrast, authentic human communication balances assertion with nuance and humility.

In her closing remarks, Aull urges listeners not to despair about the changing digital landscape. While AI English remains bounded by statistical probability across a narrow slice of text, human language continues to be powered by infinite possibility across global English. By actively confronting linguistic prejudice and keeping an open ear to diverse voices, speakers can protect human agency, choice, and voice for the future.

## Notable quotes > [01:13](https://www.youtube.com/watch?v=lufat51rPUw&t=73s) a big can of some English driven by probability — Laura Aull

> [07:52](https://www.youtube.com/watch?v=lufat51rPUw&t=472s) Human nuance, variety, and flexibility are more worthwhile than efficiency. — Laura Aull

> [09:56](https://www.youtube.com/watch?v=lufat51rPUw&t=596s) So far, humans care very much. And today I think we should care. — Laura Aull

> [09:56](https://www.youtube.com/watch?v=lufat51rPUw&t=596s) Estimates suggest that if you pose a simple query to AI to ChatGPT, let's say versus to a search engine like Google, you're using 50 times the electricity for the simple query. — Laura Aull

> [09:56](https://www.youtube.com/watch?v=lufat51rPUw&t=596s) I'm a serious scuba diver, and I know that the ocean should not get any warmer, and so I don't think we should normalize that kind of environmental impact. — Laura Aull

> [11:24](https://www.youtube.com/watch?v=lufat51rPUw&t=684s) Human English has more variation and readability than AI English, meaning that human English has more varied kinds of dialect and genres and it is less grammatically dense and it's more humble. — Laura Aull

> [11:24](https://www.youtube.com/watch?v=lufat51rPUw&t=684s) AI English is more formal, more impersonal, and more sure of itself. — Laura Aull

> [12:24](https://www.youtube.com/watch?v=lufat51rPUw&t=744s) homogeneity by design — Laura Aull

> [12:24](https://www.youtube.com/watch?v=lufat51rPUw&t=744s) algorithmic monoculture — Laura Aull

> [18:52](https://www.youtube.com/watch?v=lufat51rPUw&t=1132s) what is going on with this English? — Laura Aull

## Mentioned Laura Aull, Juan Ochoa, University of Michigan, Michigan Minds, AI English: Humans, Machines, and the Future of Writing, Grammar for Good, Siri, Alexa, OpenAI, ChatGPT, Google, Claude, Llama, Gemini, DeepSeek, Project Essay Grade (PEG), IBM, Karen Hao, Bias Shield

## Video description Laura Aull, professor of English and linguistics and recent director of the English Department Writing Program at the University of Michigan, joins the Michigan Minds podcast to discuss "AI English," how generative tools disrupt the natural diversity of human speech and why automated hiring and housing systems inherit these linguistic biases. She also explores a largely unexamined downside to our rapid adoption of large language models: the massive, real-world carbon footprint generated by our daily digital conversations.

Do you think people will care whether a human or a robot wrote the news or books we read?

So far, humans care very much. And today I think we should care. And here are three reasons why I think we should. One is the environmental cost of the computer processing used with AI. Estimates suggest that if you pose a simple query to AI to ChatGPT, let's say versus to a search engine like Google, you're using 50 times the electricity for the simple query. I'm a serious scuba diver, and I know that the ocean should not get any warmer, and so I don't think we should normalize that kind of environmental impact.

A second reason is bias. It's clear that AI English tools are biased against language too close to itself or too far away, so it penalizes some English users, including nonnative writers of English, who are more often flagged for plagiarism or more often flagged by AI detection tools even when they haven't used it. And those who use non-standardized dialect AI English don't know varieties from Jamaican English to Indian English.

That leads to the third and final reason, which is that human and AI English are different. Human English has more variation and readability than AI English, meaning that human English has more varied kinds of dialect and genres and it is less grammatically dense and it's more humble. AI English is more formal, more impersonal, and more sure of itself.

[Source](https://www.youtube.com/watch?v=lufat51rPUw)

中文简报

AI英语及其环境代价语言学家Laura Aull解释称，生成式模型会产生一种完全由概率驱动的同质化“AI英语”，抹平了鲜活人类语言中固有的动态变化、谦逊语调和认知可读性。由于大型语言模型是在标准“应试英语”等狭隘的强势方言上训练的，自动化系统和AI检测工具不成比例地惩罚了多语言写作者以及牙买加英语和印度英语等全球方言的使用者。除了社会和职场层面的把关壁垒外，日常使用这些模型还会带来令人担忧的环境代价，其耗电量大约是传统搜索引擎查询的50倍。为了维护真实的人类主体性，Aull敦促社会抵制算法单一文化，重语言细微差异而非追求速度，并培养对话式的开放性而非机器那种断言式的修辞。向ChatGPT等生成式AI系统提交一个简单查询所消耗的电能据估计是传统Google搜索的50倍，加剧了计算碳足迹和海洋变暖。 (9:56) 实证语言学研究表明，指令微调和提示工程并不能使机器文本更像人类语言，反而常常让输出结果进一步偏离自然的对话模式。 (3:56) 目前的AI检测与评估工具系统性地惩罚了非英语母语者和区域方言使用者，错误地将规范严谨的句法标记为抄袭或机器生成。 (9:56) 生成式模型继承了其开发者狭隘的文化价值观，产生了一种“算法单一文化”，偏好标准化“应试英语”而非全球通行的各种英语变体。 (12:24)

Reading summary

University of Michigan professor Laura Aull argues that generative AI produces a rigid, probability-driven dialect that strips away human nuance and reinforces historical inequities, while imposing steep environmental and social costs. Probability versus Global Linguistic Possibility (1:13) Aull contrasts human English with synthetic language by examining their foundational sources. Global human English encompasses an expansive landscape of dialects, informal registers, regional varieties—such as Jamaican and Indian English—and cross-cultural contact conventions like English as a Lingua Franca. In contrast, large language models are trained primarily on standardized exam essays, published books, and polite digital interactions, producing what Aull calls 'a big can of some English driven by probability.' Rather than offering infinite expressive choice, models like ChatGPT, Claude, and LLaMA default to an assertive, formal, and relentlessly enthusiastic register. Empirical corpus analysis by Aull reveals structural divides in syntax and readability. Natural human writing alternates between accessible clauses and complex information, relying on straightforward 'noun-verb-noun' constructions that clarify agency. AI text, conversely, leans heavily on nominalizations—dense sequences of stacked nouns around weak verbs—that increase cognitive fatigue. Furthermore, Aull notes that prompt engineering and instructional tuning fail to humanize AI outputs. Because the underlying training data lacks living conversational speech varieties, tuning often amplifies synthetic stiffness and pushes the text further away from natural human communication. Algorithmic Monoculture and 'Exam English' (12:24) Automated evaluation of text is not novel; Aull notes that the Project Essay Grade (PEG) system began scoring student essays algorithmically in the early 1960s, coinciding with the earliest milestone demonstrations of computing. However, modern transformers inherit design biases from a narrow, unrepresentative demographic of developers, turning generative outputs into an insular linguistic monoculture that privileges prestige dialects. Aull highlights how models enshrine 'exam English'—an artificial dialect rewarded in standardized testing and mass schooling over the past 150 years. This dialect often enforces arbitrary conventions, such as irregular third-person singular verb endings ('she asks' versus 'you ask'), while penalizing logically consistent global patterns ('she ask') used by more than a billion lingua franca speakers. By encoding these test-taking standards into probability engines, language models amplify historical gatekeeping under the guise of objective correctness. Ecological Toll and High-Stakes Discrimination (9:56) Beyond stylistic homogenization, Aull identifies severe structural consequences associated with adopting automated language systems. Computationally, a single query submitted to ChatGPT consumes an estimated 50 times the electricity of a standard Google search. Drawing on her perspective as a scuba diver witnessing rising ocean temperatures, Aull argues that society cannot justify such extreme energy expenditures for routine digital conveniences. Simultaneously, relying on probability-driven models in high-stakes environments—including employment screening, student writing assessments, legal evaluations, and housing distribution—introduces direct systemic harms. Algorithmic evaluators and AI-detection software routinely penalize non-standard syntax, misinterpreting the structured phrasing of multilingual and non-native English speakers as machine generation or academic dishonesty. Because generative models lack understanding, empathy, and moral judgment, Aull stresses that algorithmic speed should not replace human stewardship in life-altering decisions. Reclaiming Dialogue Through Grammar for Good (18:52) To counteract digital exhaustion and synthetic shouting, Aull advocates for linguistic practices she terms 'Grammar for Good.' Rather than treating writing as an assertive 'megaphone'—a style frequently mimicked by confident AI text—communicators are encouraged to structure prose as a dialogue. This approach incorporates qualifiers, concessions, and hedges that leave space for mutual understanding, humility, and alternative viewpoints. For readers, Aull recommends replacing immediate grammatical judgment with curiosity toward unfamiliar dialects, appreciating regional variation rather than enforcing standardized rules. Embracing the natural messiness and expressive breadth of living language, she maintains, is essential to resisting homogenized machine outputs and preserving authentic human voice. Aull cautions that passive acceptance of machine-generated prose risks permanently flattening human expression into an unreadable, corporate monoculture; however, she notes that whether society establishes stricter guardrails or entirely stops distinguishing human from synthetic authorship by 2030 remains an open question.

Source quotes

[1:13] Laura Aull: a big can of some English driven by probability [7:52] Laura Aull: Human nuance, variety, and flexibility are more worthwhile than efficiency. [9:56] Laura Aull: So far, humans care very much. And today I think we should care. [9:56] Laura Aull: Estimates suggest that if you pose a simple query to AI to ChatGPT, let's say versus to a search engine like Google, you're using 50 times the electricity for the simple query. [9:56] Laura Aull: I'm a serious scuba diver, and I know that the ocean should not get any warmer, and so I don't think we should normalize that kind of environmental impact. [11:24] Laura Aull: Human English has more variation and readability than AI English, meaning that human English has more varied kinds of dialect and genres and it is less grammatically dense and it's more humble. [11:24] Laura Aull: AI English is more formal, more impersonal, and more sure of itself. [12:24] Laura Aull: homogeneity by design [12:24] Laura Aull: algorithmic monoculture [18:52] Laura Aull: what is going on with this English?

Original transcript

[0:00] [music] [0:03] Welcome to the Michigan Minds podcast [0:05] where we explore the wealth of knowledge [0:06] from faculty experts at the University [0:08] of Michigan. [music] I am Hano Choa, [0:10] international public relations [0:11] representative for the Michigan news [0:13] office. I'm pleased to welcome Professor [0:15] Laura All, [music] a scholar of English [0:17] and linguistics and recent director of [0:19] the English department writing program [0:20] at [music] the University of Michigan. [0:22] Al's research uses corpus linguistics, [0:24] the large-scale study of digital text to [0:26] examine how language patterns shape [0:28] writing assessment, academic success, [0:30] and social power. At U of M, her work [0:33] focuses on how we are taught to write [0:35] and how technology is reshaping human [0:37] communication. She is currently [0:38] completing [music] two major monographs, [0:40] AI English, humans, machines, and the [0:43] future of writing, which explores the [0:44] evolving boundary between human and [0:46] artificial language, and grammar for [0:48] good, which [music] examines how [0:49] language knowledge can foster more [0:52] ethical and civil discourse. Welcome, [0:54] Professor. It's great to have you with [0:55] [music] us. [0:56] >> Thanks so much for having me. [0:57] >> We have been using AI for years without [0:59] really thinking about it, things like [1:01] Siri or Alexa, but obviously not to the [1:03] extent we do today. You have been [1:05] researching a concept called AI English. [1:08] When we use these tools, are we losing [1:09] our own voice or are we gaining a smart [1:12] assistant? [1:13] >> You bring up a useful example. I think [1:15] it's a helpful contrast to think about [1:18] something like Siri versus something [1:20] like CHBT. [1:22] In some ways, they're similar because we [1:26] use them both in a kind of selection [1:29] process. So if you ask Alexa or Siri to [1:34] play a certain kind of music, you are [1:36] trusting Alexa with the selection of [1:39] that music. What Alexa selects and what [1:42] Alexa leaves out. When you use ChachiPT [1:45] to compose an email, you're trusting GPT [1:48] to select and leave out certain kinds of [1:51] English if you're writing in English. [1:54] So we can think about Siri as a kind of [1:57] information selection and performance [2:00] tool and chachibi as a language [2:04] selection and performance tool. But [2:07] historically language has been a deeply [2:09] personal and social practice for humans. [2:12] So that's different. But they're similar [2:15] in the sense that if you use them both [2:19] you are trusting them to make choices [2:21] for you to an extent. [2:24] That brings us to I think a really [2:26] important contrast between AI English [2:29] and human English which we can think of [2:32] as a kind of creation story of AI [2:35] English. [2:36] So human English is based on [2:40] what is possible in global English. [2:44] That means formal English, informal [2:47] English, interpersonal English, [2:49] informational English, English used by [2:52] native speakers, English used by folks [2:54] who use it as an additional language, [2:57] English used in Ann Arbor, English used [2:59] in Sri Lanka, right? So, human language, [3:03] in this case, human English is based on [3:06] what's possible in global English by all [3:09] kinds of people all over the world in [3:11] all kinds of contexts. AI English is [3:14] based on what's probable in some kinds [3:16] of English, especially the English of [3:19] tests, books, and agreeable internet [3:22] posts. So, it favors formal English, [3:26] informational English, and assertive, [3:28] enthusiastic English. In other words, [3:31] human English is a big world of possible [3:35] English. AI English is a big can of some [3:39] English driven by probability. So while [3:42] Chachibi or Claude or Llama or Gemini [3:46] may feel like an assistant, we are [3:49] giving up some human choice and human [3:52] voice if we use AI English instead of [3:55] human English. [3:56] >> But when people use, for example, Chad [3:58] GPT or any of these other AIS, aren't [4:01] they editing the response to make it [4:03] more human or whatever tone they want it [4:06] to sound like? I think that is the [4:08] impression and it is true that [4:10] instructional tuning or prompting a tool [4:14] to modify its answers does change some [4:17] of the responses. I'd say two things [4:20] though are important to note. One is [4:22] research shows that that kind of [4:24] instructional tuning doesn't make AI [4:27] English more like human English. It [4:29] actually makes it less like it. And two, [4:32] AI English still is that big can of of [4:35] some English driven by probability. So [4:39] it doesn't have all of human English to [4:41] draw from. So you can instruct it to [4:44] make its tone a little different. For [4:46] instance, you could say, I want this to [4:48] be more informal, but it only has so [4:50] much informal English to go on. So it's [4:53] still not going to have the range that [4:54] human English has. [4:55] >> But people can still make it more [4:58] personable, for example. And that would [5:00] be a direction that humans can do better [5:02] than AI. Human English is consistently [5:05] more personable because it blends [5:07] interpersonal features. It blends both [5:12] informational features and personal [5:14] features better than AI English does. So [5:16] yes, humans can do that. They have more [5:19] to draw on if they do that on their own, [5:23] >> even if they use AI English as a [5:25] reference point. How was AI English [5:28] developed and how did we get to this [5:29] point? [5:30] >> AI English is as old as AI and [5:34] evaluating language using AI is as old [5:37] as AI. So in the 1960s a lot of people [5:40] think about the beginning of AI as the [5:44] moment when an AI tool right when IBM [5:48] won a checkers game. But in that same [5:51] time in the early 1960s, PEG, an [5:54] evaluative writing tool for essay [5:57] grading was developed and launched. So [5:59] as long as we've had AI, we have had [6:02] tools that have attempted to capture and [6:04] evaluate AI. However, today we have [6:07] generative AI English tools, meaning [6:10] that we have these transformative [6:11] architectures that have a lot more data [6:15] and use a lot more computer processing [6:18] to produce what attempts to sound like [6:21] naturally produced authentic human [6:24] language. And it's much more successful. [6:26] It has much longer context windows. [6:29] There are a lot of um particular details [6:31] that make generative AI different than [6:33] AI. But the question of how we got to AI [6:37] English the way it is is an important [6:39] one because English large language [6:41] models were developed by a very small [6:44] portion of the population. So Karen how [6:46] in her book empire of AI says that we [6:49] should be very concerned about AI [6:52] because there are so few people who made [6:54] design choices about it that it [6:56] represents a very particular worldview [6:59] with only a particular set of values and [7:02] choices. The same can be said for AI [7:05] English. It's the linguistic echo of [7:07] that larger trend in AI. Meaning that [7:11] choices were made about English large [7:13] language models that didn't represent [7:16] all kinds of English in all kinds of [7:18] places all over the world in all kinds [7:20] of uses of English. So how we got to AI [7:24] English the way it is now is a long [7:27] history of AI that began with evaluation [7:31] saying this is good or bad English but [7:33] now it is based on some language models [7:36] that have included some English and [7:38] excluded other English and that's why AI [7:41] English is a big can of some English [7:44] driven by probability while human [7:46] English is a big world of global English [7:50] driven by possibility. A lot of people [7:52] have fears around AI, including the [7:55] concern that their jobs could be at [7:57] stake. From your perspective, are we [7:59] going to be replaced or is AI primarily [8:02] a tool that can help us become more [8:04] efficient? [8:05] >> This is an open question, but right now [8:07] I think we can say it depends. AI still [8:10] needs human in oversight because it is [8:14] driven by what's probable, not all [8:16] possible choices. [8:18] So AI cannot make empathetic innovative [8:22] decisions. For example, choices that are [8:24] driven by probabilities like basic [8:27] computer coding are more likely to be [8:30] efficiently done by AI for instance. [8:32] Those things that require innovative, [8:34] empathetic, flexible, adaptive language [8:38] are best done by humans or overseen by [8:40] humans even if they take more time. By [8:43] that I mean things like writing [8:45] evaluation, hiring choices, legal [8:49] decisions, career matching, and housing. [8:51] [clears throat] [8:52] All of those research shows when done by [8:54] AI show clear bias. They're biased [8:57] toward what's already in the AI can. So [9:01] if you aren't using exam English, the [9:03] English of schools and tests, you are [9:06] less likely to get a favorable career [9:08] match or favorable housing. those kinds [9:11] of processes need to be overseen by [9:12] humans so that the bias that's embedded [9:16] in English large language models is not [9:19] influencing those decisions. So for me I [9:21] hope this means people will recognize [9:23] that while there are selective uses [9:25] where AI can be useful, efficiency is [9:27] not reason enough. Human nuance, variety [9:31] and flexibility [9:32] are more worthwhile than efficiency. I [9:35] think there was a professor who had once [9:38] said that uh AI is not at the level [9:42] where it can meet people or interview [9:45] people. So that human touch is still you [9:48] know not not there. So I guess people [9:50] can be you know rest assured that their [9:52] jobs are still still safe to to some [9:55] extent. Um if you could look into the [9:58] future what do you see for 2030? Do you [10:00] think people will care whether a human [10:02] or a robot wrote the news or books we [10:05] read? [10:05] >> So far, humans care very much. Research [10:09] shows that even people who use AI [10:11] English respond badly to finding out [10:13] that something has been sent to them [10:15] that was generated by AI. So something [10:18] would have to shift for people not to [10:20] care. In 2030 and today, I think we [10:23] should care. And here are three reasons [10:26] why I think we should. One is the [10:28] environmental cost of the computer [10:30] processing used with AI. Estimates [10:32] suggest that if you pose a simple query [10:36] to AI to Chachib let's say versus to a [10:40] search engine like Google you're using [10:42] 50 times the electricity for the simple [10:45] query. I'm a serious scuba diver and I [10:47] know that the ocean should not get any [10:49] warmer and so I don't think we should [10:51] normalize that kind of environmental [10:53] impact. So we should use AI selectively. [10:56] So environmental cost is one reason to [10:58] care whether something is generated by [11:00] AI or whether a lot of things are [11:03] generated by AI by you or other people [11:05] consistently. A second reason is bias. [11:09] It's clear that AI English tools are [11:11] biased against language too close to [11:13] itself or too far away. So, it penalizes [11:16] some English you users including [11:19] non-native writers of English who are [11:21] more often flagged for plagiarism uh or [11:24] more often flagged by AI detection tools [11:27] even when they haven't used it and those [11:29] who use non-standardized dialect. AI [11:32] English doesn't know varieties from [11:34] Jamaican English to Indian English. That [11:36] leads to the third and final reason [11:38] which is that human and AI English are [11:40] different. Human English has more [11:42] variation and readability than AI [11:44] English. meaning that human English has [11:46] more kinds of dialects and genres and it [11:49] is less grammatically dense and it's [11:51] more humble. AI English is more formal, [11:54] more impersonal and more sure of itself. [11:57] A lot of people ask me if this will [11:58] change because AI is evolving but these [12:01] two qualities are not so far changing [12:03] across time, across models and across [12:06] bots. AI English is still less readable [12:09] and varied than human English. So in [12:12] 2030 or today, if we don't care, that [12:15] probably means that we're accepting [12:17] English that's less varied and less [12:20] readable than human English. And I hope [12:22] that doesn't happen. [12:24] >> You have mentioned that standard English [12:25] can sometimes be used as a tool for [12:28] exclusion, including in settings like [12:30] job interviews. Does AI make that [12:32] problem worse by automating those [12:34] biases, or could it help level the [12:37] playing field? [12:39] If English large language models were [12:41] designed differently with more kinds of [12:43] global English and more varieties of [12:45] English, they could potentially help [12:47] level the playing field. And arguably [12:51] small language models and other curated [12:54] tools like bias shield [12:57] can help raise awareness and and thereby [13:02] help level the playing field. But [13:05] because AI English is a big can of some [13:07] English driven by probability, it favors [13:10] some English and not others, which means [13:13] it favors some English users and not [13:15] other English users. Researchers use [13:17] terms like homogeneity by design and [13:20] algorithmic mono to try to capture this [13:24] flattening out of variation in human [13:27] English. And what it means is that AI [13:29] English cannot level the playing field [13:31] as it currently is because that bias is [13:34] already baked into English large [13:36] language models. And research suggests [13:39] as I mentioned earlier that [13:41] instructional tuning makes AI English [13:43] more like AI English and less like human [13:47] English. So one thing we can do to [13:49] preserve human English is recognize that [13:52] its creativity and adaptability come [13:54] with some messiness. [13:56] If we want human English voice and [13:58] personability and innovation, we might [14:01] need to get over some bias about certain [14:05] kinds of English and have an open mind [14:07] about usage beyond what is [14:08] conventionally valued in schools and [14:10] tests. [14:11] >> Can you define what standard English is? [14:14] I like to use the term exam English to [14:17] draw attention to the kinds of contexts [14:20] in which English is used in instruction [14:24] and used in high stakes decisions tends [14:27] to favor certain kinds of usage over [14:30] others. So if we think about exam [14:32] English as the English that's privileged [14:34] and rewarded in schools and tests and in [14:38] employment historically then what I mean [14:42] is English that has tended to favor [14:45] native English speaker usage even when [14:48] that doesn't make complete sense. So for [14:51] instance we have a form of subjectverb [14:53] agreement in standardized English that [14:56] changes in the present tense. It adds an [15:00] s, right? She asks, [snorts] but you [15:04] ask. And there's a pattern in English as [15:07] a lingua francco, which is English used [15:10] as a contact in additional language by [15:12] billion by a billion language users [15:14] today that often makes the verb [15:18] consistent. [15:19] >> So I ask, you ask, she ask, they ask. So [15:24] standard English is less consistent in [15:26] that, but that has been the form of [15:28] subjectverb agreement that has been [15:31] prioritized in schools and tests. And so [15:34] I like to say it's exam English because [15:36] it isn't applicable in every context and [15:40] it doesn't even always make sense. [15:42] However, it is what is prioritized [15:45] historically since mass schooling was [15:48] introduced in English around 150 years [15:50] ago. So, how can you tell when something [15:53] is AI generated versus human uh [15:56] generated English? [15:59] >> The AI English book I just finished [16:01] explores five areas of patterns in many [16:04] kinds of human English over time from [16:07] text messages to government memos and in [16:10] AI English generated by chatbt, llama [16:13] and deepseek. I could say a lot more [16:15] about that and the book goes into much [16:18] more detail, but across patterns and [16:20] bots and time, there are two overall [16:23] differences. Human English has more [16:25] variation and readability than AI [16:28] English. More variation means human [16:30] English has more varied kinds of usage, [16:33] cohesion, and connection. So, human [16:36] English uses everything from multiple [16:38] A's in yay in a text message or a social [16:42] media post to more forms of subjectverb [16:45] agreement like I just mentioned. [16:47] So, she ask, she asks or they is and [16:52] they are and more kinds of cohesion or [16:55] more ways to make texts hang together. [16:58] By contrast, if we continue these same [17:01] examples, AI English is less likely to [17:03] use innovative forms. It knows fewer [17:06] kinds of subject verb agreement and it [17:08] uses the same cohesive forms [17:11] repetitively. So that first overall [17:13] takeaway is that human English has more [17:16] variation. It has more variation in the [17:18] ways that it connects. It has more [17:20] dialect features and it has more ways of [17:24] creating cohesion. The second is [17:27] readability. [17:28] That means that human English blends [17:31] plain and dense grammatical structures [17:34] more often. So a dense grammatical [17:36] structure is a sentence with long nouns [17:39] and small verbs. So the subject is long, [17:43] small verb, the object is long. And then [17:45] a plainer or simpler sentence would have [17:48] a simple subject that's close to the [17:50] verb. And that's easier for our brains [17:52] to parse. meaning we can tell who's [17:55] doing what or what's happening more [17:56] easily when we go noun verb noun rather [18:00] than noun noun noun verb noun noun nouni [18:04] English more often uses long noun [18:05] phrases lots of nouns and fewer verbs [18:09] and it uses them more regularly so that [18:12] it's got that dense orformationally [18:14] dense or grammatically dense style [18:18] repetitively over sentences making it [18:20] harder to read and less accessible [18:22] >> to better that we could do better [18:25] prompts, right? [18:26] >> Research suggests that at least so far [18:29] people are not particularly good at or [18:32] interested in making AI English more [18:35] simple and accessible. But [18:37] theoretically, yes, with more knowledge [18:40] and awareness of patterns, we could use [18:44] AI tools better and instruct them toward [18:47] more accessibility or more readability [18:50] and toward more variation. We have all [18:52] read social media comments and felt [18:54] exhausted by the constant shouting. You [18:57] are working on a project called grammar [18:59] for good. Is there a way we can change [19:01] how we structure our sentences to make [19:02] it harder to to be mean and easier to be [19:06] understood? [19:06] >> I think there are things we can do both [19:08] when we read and when we write to [19:11] practice grammar for good. When we read, [19:14] we can try reading for human patterns [19:16] and possibilities with an open mind, [19:19] delighting in variation and leaning into [19:22] our curiosity about language. We can try [19:25] to ask what is going on with this [19:27] English rather than being attached to [19:29] only one kind of English. When we write, [19:32] we can think about dialogue rather than [19:34] megaphones. So, I think you're right to [19:36] use the metaphor of shouting. When we [19:39] use writing that uses dialogue patterns, [19:42] we use hedges and concessions to leave [19:44] room for reasonable doubt and [19:47] alternative reactions. AI English [19:49] doesn't do this as much as human [19:51] English, which is why people feel like [19:53] AI English is too assertive or [19:55] flattering. AI English is more like a [19:57] megaphone. AI English shouts more often [20:00] than human English. But humans can write [20:02] with dialogue. they can qualify as well [20:04] as emphasize so they leave space for [20:07] more perspectives when there's room for [20:09] reasonable doubt and alternative [20:11] reactions. So we can try to read with an [20:14] open mind and we can try to write with [20:16] dialogue rather than megaphones. [20:19] >> Professor Al, we have run out of time. [20:21] Is there anything else you would want to [20:23] say to our listeners? [20:25] >> I would say don't lose hope. AI English [20:28] is still driven by probability in some [20:30] English and human English is driven by [20:33] possibility in global English. So long [20:35] as we can try to get over our bias and [20:39] listen to human English with an open [20:41] mind, we can preserve things like choice [20:43] and voice in the future of English. [20:47] [music] [20:50] Thank you for listening to this episode [20:52] of Michigan Minds produced by [music] [20:54] Michigan News, a division of the [20:56] university's office of the vice [20:58] president for communications.