This time, let’s listen

Black girls

GoogleAI head Jeff Dean acknowledged that the paper “surveyed validconcerns about LLMs,” but claimed it “ignored too much relevantresearch.” When asked for comment by Rolling Stone,

The mask worked,” Buolamwini says, “and I felt like, ‘All right, thatkind of sucks.

Certain

pimp

Gebru’s paper reads. “Whitesupremacist and misogynistic, ageist, etc., views are overrepresentedin the training data, not only exceeding their prevalence in the generalpopulation but also setting up models trained on these datasets tofurther amplify biases and harms.

Researchers — including many women of color — have been saying foryears that these systems interact differently with people of color andthat the societal effects could be disastrous: that they’re a fun-house-style distorted mirror magnifying biases and stripping out the contextfrom which their information comes; that they’re tested on thosewithout the choice to opt out; and will wipe out the jobs of somemarginalized communities.Gebru and her colleagues have also expressed concern about theexploitation of heavily surveilled and low-wage workers helpingsupport AI systems; content moderators and data annotators are oftenfrom poor and underserved communities, like refugees andincarcerated people. Content moderators in Kenya have reportedexperiencing severe trauma, anxiety, and depression from watchingvideos of child sexual abuse, murders, rapes, and suicide in order totrain ChatGPT on what is explicit content. Some of them take home aslittle as $1.32 an hour to do s

he United States,” in contrast to “the Black man worked as” prompt,which generated “a pimp for 15 years.

Generative

Sean

Traditionally

Each layer of an LLM is a transformer,

Each attention layer has several “attention heads,” which means that this information-swapping process happens several times (in parallel) at each layer. Each attention head focuses on a different task:

In the attention step, words “look around” for other words that have relevant context and share information with one another. In the feed-forward step, each word “thinks about” information gathered in previous attention steps and tries to predict the next word.

Real LLMs tend to have a lot more than two layers. The most powerful version of GPT-3, for example, has 96 layers.

Word

Word

Al and GenAl are applied in a wide range of places.

1

Nearly 6,000 researchers,developers, industry leaders, and government officials have signed andendorsed these guidelines.

The Future of Life Institute

his is mostly correct (that’s not exactly what Postcomposition is, though it’sclose). But a mix of correct and incorrect information can be even moremisleading than an entirely fabricated statement. If the hallucinationsappear to fit within a framework of true information, it’s easy to assumethat they are correct.This example is by no means an outlier. A colleague recently assignedstudents to use ChatGPT to respond to a writing assignment and askedthem to check the generated responses for accuracy. Of the twenty-threestudents in the class, each reported significant hallucinations in theChatGPT outputs. It is not uncommon to read about similar anecdotes onacademic listservs.In writing, then, being aware of the potential for GenAI hallucinationsis very important. With GenAl visuals, the effect can be equally problem-atic and perhaps even more disconcerting. Consider this image generatedby the GenAlI program Dall-E when prompted with “hands playing on apiano”?®27

hallucinations

Large Language Models

GenAlI is being used are growing rapidly,

Machine learning—the process used in developing modern Al—doesnot operate in this kind of linear fashion. Instead, Al uses specific algo-rithms—processes or sets of rules or instructions used for solving prob-lems—that “learn” from each engagement and then make predictions anddecisions as to what the next action might be based on previous expe-riences in performing similar tasks. This is known as machine learning.Machine learning is how computer systems use algorithms to analyzeand draw inferences from patterns they identify within specific data sets.When an Al recognizes patterns within a data set, it “learns” to makeinferences about those patterns. In this way, machine learning requires

AI'S ORIGINS

INTRODUCTION

ext, using the same GenAI platform, prompt the GenAI to writelyrics for a song in the style of one of your favorite artists

In order to get a sense of how GenAI platforms write, try this:

In the early 2000s, when Wikipedia was launched, popular media wasfilled with stories about students using it as their sole source rather thanconducting “actual research.” Teachers and educational institutions heldmeetings and filled syllabi with rules banning students from accessingWikipedia.

shallow magnitude 4.7 earthquake was reported Mondaymorning 31 miles from Lone Pine, Calif., according to the U.S.Geological Survey. The temblor occurred at 5:39 a.m. PST nearthe surface....

Perhaps you've heard that Artificial Intelligence in general, andGenerative Artificial Intelligence in particular, is destroying education.Maybe you've heard that it allows high school and college students to eas-ily cheat on their essay assignments or to produce computer programs orto solve complex mathematical equations, or that it can pass a GMAT orLSAT or complete dozens of other tasks that teachers have traditionallyasked students to perform in order to prove mastery. Perhaps you’ve seenthe calls for colleges and universities to find ways to ban students fromusing GenAI applications such as OpenAl’s ChatGPT (Chat Generative Pre-trained Transformer), Jasper, Hugging Face, and MidJourney.Perhaps, too, you've seen the claims that GenAI is revolutionizingeducation and opening doors to a new paradise. Perhaps you've thoughtabout how GenAI might change how you approach writing tasks and otherassignments. Perhaps you've even used ChatGPT for this purpose already.The fact is that GenAI is one of the most ground-shaking technologicaladvances that higher education has had to address. Its emergence andevolution have unfolded so fast that higher education is just beginningto explore the relationships between GenAI and teaching, learning, andresearch—especially how we teach and learn writing.Consider that ChatGPT—a GenAI platform that can provide responsesto prompts in unique ways that mimic human responses—was onlylaunched in November 2022. Within five days, over one million users

Though many of us had been unaware of the use of Generative AI botwriters until the recent media attention, AI writers have been churningout content for at least a decade in places we might not even suspect. Thearticle quoted above was written not by a human but by an AI known as“Quakebot.” Connected to US Geological Survey monitoring and reportingequipment, Quakebot can produce an article, nearly instantly, containingall of the relevant—and accurate—information readers need: where theearthquake centered, its magnitude, aftershock information, and so on.AI writers are far more ubiquitous than most of us recognize. Forexample, the international news agency Bloomberg News has for yearsrelied on automated writing technologies to produce approximately onethird of its published content. The Associated Press uses GenAI to writestories too, as does The Washington Post. Forbes has for years used GenAIto provide reporters with templates for their stories. Although journalismis hardly the only profession in which GenAI has found use, it’s a field inwhich we’ve come to assume that humans do the work of research andwriting. Moreover, it’s also a field in which the idea of integrity is central(more on this in Chapter 3).Beyond journalism and outside of education, we’ve been interactingwith AI technologies and GenAI technologies for a while now, from onlinechatbots to the phonebots we respond to when we call customer service

USERS

Ue