Timnit Gebru on deep unlearning, AI hype, and algorithmic racial bias

Artificial intelligence is often sold as magic: superhuman chatbots, smart cameras, and models that supposedly contain “all of human knowledge.” Timnit Gebru, one of the most influential voices in AI ethics, argues that this story is not just misleading—it’s dangerous.

In this article, we unpack her perspective on what AI really is, why today’s hype is so harmful, and how racial bias, environmental damage, and exploited labor sit at the core of many current AI systems.

What AI really is (and why the definition keeps moving)

For Gebru, AI is not a single thing. It’s a broad discipline made up of many different subfields and techniques—like large language models, computer vision, and expert systems—that often get lumped together under one buzzword: “AI.”

These subfields can be very different from each other. A chatbot like ChatGPT is not the same as a medical imaging system that helps detect cancer, even if both are called “AI.” They can use different data, different training methods, and have very different risks and benefits.

She also points out that what counts as “AI” changes over time. In the 1980s, “expert systems” were the hot AI technology. Today, almost no one calls them AI. The same thing happened with “deep learning,” which is now almost synonymous with AI, but used to be pushed to the margins under the name “artificial neural networks.”

This moving target makes it hard to have grounded conversations. When critics raise concerns about one kind of system—like chatbots trained on massive internet datasets—defenders often respond by pointing to completely different tools, such as narrow medical systems. That blurs the debate and avoids real accountability.

Why Timnit Gebru was fired from Google

Before founding her own institute, Gebru co-led Google’s Ethical AI research team. She was already well known in the field: she co-founded Black in AI, had published influential work on algorithmic bias, and was a visible critic of harmful uses of AI.

At Google, she and her co-lead Meg Mitchell (who was also later fired) worked in a context already filled with internal tension—like the Google walkout over the company’s handling of sexual misconduct allegations and the broader Black Lives Matter movement in 2020.

The breaking point came when Gebru co-authored a paper warning about the dangers of large language models—the same class of models that now power tools like ChatGPT and Gemini. The paper questioned the rush to build ever-larger models and highlighted several key risks. After internal conflict over this work, she was pushed out of Google. She describes that paper as “the icing on the cake” that led to her firing.

The hidden costs of large language models

Large language models (LLMs) like GPT-3, GPT-4, and Gemini are trained on vast amounts of text from the internet. Technically, they learn to predict the most likely next word in a sequence, which allows them to generate fluent, human-like text.

Gebru’s paper—and her ongoing work—highlights several core problems with the current LLM race.

Environmental and financial damage

Training giant models requires enormous computational power and energy. That means a large carbon footprint and massive infrastructure costs. Even when companies claim to use renewable energy, that electricity is still being diverted from other critical uses, like heating homes or powering local communities.

The financial cost also concentrates power. Only a handful of wealthy companies and institutions can afford to train and deploy these systems at scale. That shuts out researchers, communities, and countries without deep pockets, narrowing who gets to shape the future of AI.

Hegemonic views disguised as “all of human knowledge”

Companies often imply that because LLMs are trained on huge datasets, they somehow encode “all of human knowledge.” Gebru calls this out as false. The internet is not a neutral or complete representation of humanity—it reflects hegemonic views, dominated by Western, male, and English-speaking perspectives.

Even widely used resources like Wikipedia are known to be heavily biased in who writes, who is represented, and how topics are framed. When models are trained on this skewed data and then presented as objective or universal, they risk amplifying existing power imbalances and silencing marginalized voices.

Fluent text that deceives and automates bias

One of the most dangerous features of LLMs is their fluency. They produce text that looks coherent, confident, and grammatically correct—even when it’s wrong or harmful. That makes it easy for people to over-trust their outputs.

Gebru points to a real-world example from 2017: Facebook’s translation system turned a Palestinian’s Arabic phrase for “good morning” into “attack them.” The translation was grammatically correct in the target language, so there was no obvious sign of an error. Authorities arrested the man based on this mistranslation and only later released him.

This is a classic case of “automation bias”—our tendency to over-trust automated systems, especially when they appear confident and polished. Today’s chatbots are explicitly designed to sound like helpful, thoughtful assistants. That design encourages people to imagine a “mind” behind the text, making them even more vulnerable to deception, misinformation, and manipulation.

The cost of a single research path

Gebru also warns about the opportunity cost of pouring so much money, talent, and infrastructure into one narrow direction: ever-larger language models. When an entire field chases a single paradigm, other possibilities—different architectures, smaller and more accountable systems, or entirely different approaches to automation—get sidelined or starved of resources.

In her view, this isn’t just a technical choice; it’s a political and economic one. It locks us into a future shaped by a few corporations, built on massive data extraction and environmental harm, instead of exploring many possible futures for how AI could be used.

How AI systems encode racial bias

Long before GPT-3 and ChatGPT, Gebru was studying how AI systems treat people differently based on race and gender. One of her most influential projects, with researcher Joy Buolamwini, examined commercial facial analysis tools.

Their research showed that these systems had far higher error rates for darker-skinned women than for lighter-skinned men. The darker the skin, especially for women, the worse the performance. When such tools are used in contexts like CCTV surveillance or law enforcement, that means people with darker skin are more likely to be misidentified or falsely flagged as suspects.

This isn’t an abstract risk. It connects directly to existing policing practices, incarceration rates, and systemic racism. When biased tools are plugged into already unjust systems, they can supercharge harm under the guise of “objective” technology.

If you’re interested in how hype and power shape today’s AI race more broadly, you may also want to read this deep dive on AGI hype and big-tech bets on models like Grok 5.

Why representation in AI matters: the story of Black in AI

Gebru co-founded Black in AI around 2016–2017 after noticing two things at the same time: the near-total absence of Black people in AI research spaces, and the rise of AI tools being used in high-stakes areas like criminal justice.

At major AI conferences with thousands of attendees, she would see only a handful of Black researchers. Meanwhile, systems were being built and deployed to predict things like the likelihood of someone reoffending after prison. A 2016 ProPublica investigation showed that one such system was more likely to label Black people as high-risk than white people, and judges were already using these scores in decisions about bail and sentencing.

That contrast—the people most affected by these tools having almost no say in how they’re built—pushed her to create Black in AI. The organization focuses on increasing the presence, inclusion, visibility, and wellbeing of Black people in the AI field, so that those most impacted by these technologies can help shape them.

DAIR: building a different kind of AI research institute

After leaving Google, Gebru founded the Distributed AI Research Institute (DAIR). The goal is to create an independent space for AI research that centers the people most harmed by current systems, rather than the corporations that profit from them.

One of DAIR’s major projects is the Data Workers Inquiry, which focuses on the people whose invisible labor makes modern AI possible: data labelers.

The invisible workers behind AI

Large models don’t just learn from raw text scraped from the internet. They also rely heavily on human workers who label, clean, and structure data, and who provide feedback to “align” model behavior. These workers are often located in the global south and work under highly exploitative conditions.

Gebru describes data workers who are paid as little as one dollar an hour, with minimal breaks and little protection or recognition. Yet their painstaking labor is a big part of why these systems appear so “smart.”

Through the Data Workers Inquiry project, DAIR collaborates with these workers to research their own working conditions, organize collectively, and advocate for better treatment. One outcome has been the creation of the Data Labelers Association in Kenya, giving workers a platform to push back against exploitation.

“Deep unlearning”: rethinking our path with AI

Gebru’s forthcoming book, Deep Unlearning: The Rise of AI and the Radicalization of a Tech Idealist, plays on the term “deep learning” to make a point: we don’t just need better models; we need to unlearn the assumptions and power structures that got us here.

She traces her own journey—from growing up in Addis Ababa, Ethiopia, to coming to the United States as a refugee, to falling in love with math and physics as an escape from war and politics. Over time, she realized that technology is not separate from those messy realities. The same forces that drive conflict and inequality also shape who builds AI, who funds it, and who gets harmed by it.

For engineers and scientists, that means technical skill is not enough. Without understanding the social and political context of their work, they risk building products that harm the very communities they come from or claim to serve.

If you’re curious about how individual builders can rethink their relationship with AI, you might also find it helpful to read about hands-on experimentation, like in this story of building a game-playing AI from scratch, and compare that with Gebru’s call to question not just how we build, but why and for whom.

AI didn’t have to be this way—and still doesn’t

A core theme in Gebru’s work is that the current AI trajectory was never inevitable. The race toward bigger models, more data, and more centralized power is a choice, not destiny.

There were—and still are—many other paths we could take: smaller, more accountable systems; community-led projects; tools designed to support, not replace, workers; and research agendas driven by public interest instead of corporate profit.

“Deep unlearning” is about stepping back from the hype, listening to the people most harmed by current systems, and being willing to change course. That means centering environmental justice, labor rights, racial equity, and democratic control in how we design, deploy, and govern AI.

As AI tools like ChatGPT and Gemini cross a billion users, Gebru’s message is a reminder that scale without accountability is not progress. The real question isn’t how powerful our models can get—it’s what kind of world they are helping to build.