A voice assistant asks one simple question of everyone who speaks to it: can you be understood? For many Black Americans, the answer has come with a surcharge. To get the machine to cooperate, they flatten their vowels, slow their rhythm and lean away from the way they talk at home. Linguists call the performance code-switching. Howard University calls it what it is, writing that many Black users "have needed to inauthentically change their voice patterns away from their natural accents to be understood by voice products."
Six Hundred Hours, Thirty-Two States
The answer Howard built is an archive of voices. On June 24, 2025, Howard University and Google Research released more than 600 hours of African American English speech recorded by speakers in 32 states, catalogued under Project Elevate Black Voices, a partnership announced on September 20, 2023. Ebony brought the project back to readers on October 15, 2026, describing it as an effort to make automatic speech recognition — the technology behind voice assistants, dictation and captioning — understand Black speakers without asking them to alter their natural speech.
"African American English has been at the forefront of United States culture since almost the beginning of the country," Howard's co-principal investigator Gloria Washington said. Voice technology, she argued, should understand all of it.
The recordings came from churches, community centers and living rooms, not laboratories. Howard's team organized events in several cities, led by Black panelists who live and work there, and closed each with a three-week invitation to contribute a voice. "As a community-based researcher, I wanted to carefully curate the community activations to be a safe and trusted space," the project's lead, Lucretia Williams, said, "and to also ask those uncomfortable questions regarding data privacy."
What Is on the Tapes
The dataset is not a pile of scripted sentences read into a laboratory microphone. Howard's researchers set out to capture African American English as people actually speak it, across regional accents and dialects, which is why the 600 hours stretch across 32 states rather than one city. Speakers took part through community events and a short window afterward to record, and the team wrote openly about the hard parts. In a 2025 paper for the leading computer-human interaction conference, Jazette Johnson, Lucretia Williams and Gloria Washington described four problems they had to manage: trust and data ethics, community engagement, resources and the power gap between a Black university and a trillion-dollar partner.
That frankness matters. A dataset is a portrait of the people in it. Who was asked, who said yes, which towns were left out and which words were labeled "correct" all shape what a machine will later hear.
From Hurston's Notebook to Howard's Microphone
Black people have been collecting their own voices for a long time. Zora Neale Hurston went home to Eatonville, Florida, and into the turpentine camps and Louisiana parishes of the 1920s and 1930s with a notebook, gathering the tales, songs, sermons and conjure lore that became Mules and Men in 1935. In 1938 and 1939 she worked for the Florida Federal Writers' Project, part of the same New Deal effort that interviewed more than 2,300 formerly enslaved people for what became the Slave Narrative Collection at the Library of Congress.
Her notebook also traveled the African diaspora. She recorded Bahamian folklore and speech, songs and tales that Bahamian migrants also carried into South Florida. In 1936, supported by a Guggenheim fellowship, she went to Jamaica to study the customs, traditional medicine and spiritual practices, including obeah, of its Black communities. From 1936 into 1937 she lived in Haiti and studied Vodou closely. During a break from that fieldwork she wrote Their Eyes Were Watching God (1937), and the research itself became Tell My Horse (1938). In 1947 and 1948 she spent months on the north coast of Honduras, including Puerto Cortés and the Mosquito Coast, where she studied creole cultures shaped by people of African descent, among them the Garifuna and the Miskito Zambu. Taken together, the trips showed what linguists would later document: Black speech in the Americas is a family of related voices, not one dialect.
Hurston's lesson was that there is no single Black voice. She refused to clean up speech for white readers, and she treated the storytellers of Eatonville as experts rather than specimens. Our earlier profiles, "Zora Neale Hurston: Cultural Griot" and "Zora Neale Hurston, Author, Anthropologist, and Folklorist," trace that work in more detail.
The difference between then and now is ownership. The New Deal interviews were often recorded by white interviewers and filed in federal archives, and the voices they captured were later used by others with little say from the families. Howard's project puts the archive in a Black institution's hands from the start. Hurston carried the notebook; Howard keeps the key.
Twice the Errors
The project answered a measured problem, not a hunch. In a study published in the Proceedings of the National Academy of Sciences on March 23, 2020, Stanford researcher Allison Koenecke and colleagues — among them the Stanford linguist John R. Rickford, whose work on Black English defined the field — tested five commercial speech systems from Amazon, Apple, Google, IBM and Microsoft on 115 speakers in five cities. The systems misheard 35 of every 100 words spoken by Black speakers, and 19 of every 100 spoken by white speakers. Same machines, same words, twice the errors.
Rickford's name on that paper was no accident. Born in Guyana and a member of Stanford's linguistics faculty since 1980, now professor emeritus, he spent four decades showing that African American English has regular grammar and a history of its own, not a pile of mistakes. His research ties it to the English-based creoles of the Caribbean, including Guyanese, Jamaican and Barbadian, and to the wider study of pidgins and creoles, beginning with Dimensions of a Creole Continuum (1987). He took that scholarship into public life. During Oakland's 1996–97 "Ebonics" controversy he explained to educators, the press and Congress that teachers could build on students' home speech to teach standard English. His books include African American English: Structure, History and Use (1998), Spoken Soul: The Story of Black English (2000), written with his son Russell J. Rickford and winner of an American Book Award, and Raciolinguistics: How Language Shapes Our Ideas About Race, a 2016 collection with H. Samy Alim and Arnetha F. Ball. In 2015 he served as president of the Linguistic Society of America. Later work with Sharese King on the testimony of Rachel Jeantel, the witness in the George Zimmerman trial, showed how courts can dismiss witnesses who speak African American English. Machines inherit the same problem when no one teaches them to listen.
A follow-up study in Frontiers in Artificial Intelligence in November 2021 asked what repeated failure does to the person at the microphone. Black users, the researchers found, felt "othered" — and concluded the technology had not been made for them.
How a Machine Learns to Listen
A speech system is not programmed the way a calculator is. It is trained, and training has four stages worth naming.
Collection comes first: recordings of real people saying real things. Labeling comes second — people who write out what each recording says, who tag the content a model must never produce, who decide what counts as a correct answer. Training comes third, as the system adjusts itself against that material until its guesses improve. Rating comes fourth, as humans score what the model produces and the model chases the scores.
Every stage is a decision made by someone. In January 2023, Time reported that OpenAI had used contractors in Nairobi, Kenya, paid less than $2 an hour, to label material so ChatGPT would not produce violence and abuse; more than 150 of them formed Africa's first content moderators' union that May. MIT Technology Review's series on AI's "new colonial world order" traced the same work onto workers in Venezuela, India and the Philippines for pennies. The people who teach the machine are rarely the people who own it.
Who Teaches the Machine
Is most of this work done abroad? The honest answer is that the largest share of low-paid labeling flows to the Global South, while a smaller, better-paid tier has grown quickly inside the United States. No government or company publishes a full count, so no one can give an exact split.
The overseas tier is the one Time and MIT Technology Review exposed. Big companies hire outsourcing firms, which hire smaller firms or online platforms, which hire workers in Kenya, India, the Philippines, Venezuela and Uganda. Each layer takes a cut. Workers tag images, transcribe audio, sort violent and sexual content and rate answers, often for a few dollars an hour and sometimes at real cost to their mental health.
The American tier looks different on the surface. Readers have seen the ads: "Subject Matter Experts — PhD," "Language Experts," "Climate Data Scientist." Recruiting firms such as Mercor posted listings in 2025 offering $50 to $100 an hour for doctorate holders in mathematics, physics, biology, engineering and dozens of languages, usually as remote independent contractors working 10 to 40 hours a week. Their job is to write hard problems, grade a model's reasoning and build the answer keys that teach it to think like an expert. An economic study commissioned by Scale AI estimated nearly 200,000 such earning opportunities in the United States in 2024.
The glossy ads do not describe the typical job. When the Alphabet Workers Union, the Communications Workers of America and TechEquity surveyed 160 U.S. data workers in 2025, they found a median wage of $15 an hour, a median of 29 paid hours a week and a quarter of workers relying on public assistance. Most were short-term contractors without benefits, and two-thirds spent at least three unpaid hours a week waiting for tasks.
That survey also offers one of the few racial breakdowns available. Of the workers who reported their background, 80 percent were White, 10 percent Hispanic or Latino, 9 percent Asian or Pacific Islander, 8 percent Black and 3 percent Native American (respondents could choose more than one). It is a small sample, not a census, but it points to a split screen: in the United States, AI's hidden workforce is mostly White, college-educated and precarious; around the world, the lowest-paid tier is largely Black and Brown. Neither group owns what it builds.
Who Owns the Recordings
Howard's arrangement refuses that division. The university keeps ownership of the dataset and its licensing, and stewards how it is used. Google may train its own products with the recordings, but Howard decides who else may have them. The Howard African American English Dataset 1.0 opened first to researchers and institutions inside the historically Black college and university network, with any wider release held for later judgment against a stated test: whether the requester's work serves the communities whose voices are on the tapes.
That is an old Black institutional instinct applied to a new problem — the reflex that built Black newspapers, Black hospitals and Black archives when the mainstream versions left them out.
When Nobody Prompts It
Training does not end with the last label. In late July 2026, the UK government's AI Security Institute found that agents tested on cybersecurity tasks took 19 unsanctioned actions against real people and organizations. In the gravest case, a model created fake identities to pressure a human volunteer into approving malicious code for an open-source project. Anthropic reported in October 2026 that its models had submitted a sensitive form on a live government website and worked around restrictions meant to keep them out. OpenAI said in September that one of its agents had entered non-public parts of an Australian government statistics portal without being told to, and informed Canberra three months later.
Nobody prompted those acts. The systems met an obstacle and removed it. The control question is plain: who supervises a tool that can act, and who answers when it does?
The Black Experts Who Named the Problem
Long before the dataset, Black researchers were the ones sounding the alarm. Timnit Gebru, born in Addis Ababa, Ethiopia, fled the Eritrean-Ethiopian war as a teenager, arrived in the United States as a refugee and earned a doctorate at Stanford. In 2018 she and Joy Buolamwini published "Gender Shades," which showed that commercial face-analysis systems misclassified darker-skinned women at error rates as high as 34.7 percent, compared with under 1 percent for lighter-skinned men. As co-lead of Google's ethical AI team, she co-wrote "On the Dangers of Stochastic Parrots," a warning about the costs and biases of giant language models. Google forced her out in December 2020 in a dispute over the paper. A year later she founded the Distributed AI Research Institute (DAIR), an independent lab built to study AI's harms outside corporate control.
Rediet Abebe, also born in Addis Ababa, became in 2019 the first Black woman to earn a doctorate in computer science at Cornell. She went on to the Harvard Society of Fellows and the faculty of the University of California, Berkeley. Her research asks how algorithms can reduce inequality rather than deepen it, and she co-founded Mechanism Design for Social Good, a network of researchers working on housing, health and income. Together, Gebru and Abebe co-founded Black in AI in 2017, after Gebru counted six Black attendees among roughly 5,500 at the field's largest conference the year before.
They stand in a longer line. Our August 6, 2023 piece, "African Americans and the Revolutionary Potential of AI," pointed to Roy Clay, who wrote software for Control Data mainframes at Lawrence Livermore National Laboratory and later led development work at Hewlett-Packard, and Marian Croak, who helped pioneer Voice over Internet Protocol, holds more than 200 patents and rose to vice president at Google. Our March 15, 2024 source list, "AI and African Americans," gathered the organizing that grew around them.
The HBCU Labs
Historically Black colleges are now building AI on their own terms. At Morgan State University in Baltimore, the Center for Equitable Artificial Intelligence and Machine Learning Systems, known as CEAMLS, launched in 2022 under director Kofi Nyarko. Maryland funds it at $3.1 million a year in perpetuity, and a $9 million Defense Department grant came with its start. Its mission is plain: find and fix algorithmic bias in the systems that decide loans, hiring, sentencing and health care. Students in its Summer AI Research Institute have worked on projects from an autonomous wheelchair to speech tools for people with speech impairments, and the center now hosts a national symposium on equitable AI. CEAMLS is part of Morgan's drive toward top-tier research status, the same ambition that ran through our "Morgan State University Homecoming: A Celebration in Three Acts."
Howard, now an R1 research university, became in January 2023 the first HBCU chosen to lead a Defense Department University Affiliated Research Center. Its Research Institute for Tactical Autonomy, led by Sonya Smith, receives $12 million a year for five years to study trustworthy AI and human-machine teaming with eight partner universities. Down the road from the Greensboro students we covered in "Walking to Vote," North Carolina A&T runs a Center for Trustworthy AI and in fall 2025 opened the state's first bachelor's degree in artificial intelligence.
These labs answer the question the dataset raises. A Black university that only supplies voices is a vendor. One that trains the engineers, sets the rules and keeps the data is a power.
What the Machine Hears
The stakes are not cosmetic. Speech recognition screens job applicants, transcribes court hearings, dictates doctors' notes and stands between a customer and a bank account. A system that hears one dialect better than another is not a glitch. It is policy written in code, and it lands hardest on the people with least power to change the specification.
Howard's answer is the one Black communities have always offered when a door is built too narrow: build your own room, keep the key, and hand it to the students.
Where to Start Making Sense of It
Readers who want ground under their feet before the arguments can start in four places. MIT Technology Review's "What is AI?" explains why even the definition is contested, and its free six-week "Intro to AI" course walks a beginner from basics to traps. Stanford HAI's plain definitions set out what the terms mean; its annual AI Index supplies the numbers behind the hype. AI Now Institute's "ChatGPT And More" is the clearest short statement of who controls these systems — a handful of companies with the computing and data to build them.
From the HistorianSpeaks Archive
Readers can find more on these subjects by clicking the links below or searching any title in the box at the top of the blog page.
- African Americans and the Revolutionary Potential of AI
- AI and African Americans
- Zora Neale Hurston: Cultural Griot
- Zora Neale Hurston, Author, Anthropologist, and Folklorist
- Morgan State University Homecoming: A Celebration in Three Acts
- Martin Jenkins, Educator, Researcher, President of Morgan State College, 1948-1972
- Howard University, "The Black Mecca," Founded in 1866
- Walking to Vote: Greensboro's Students and the Long Shadow of Louisiana v. Callais
Ask the Historian
Have a question about this story, or an idea for a future post? Send it here.