Timnit Gebru

First, “AI” is too vague. Do you mean a medical-imaging system, a language tool for an under-resourced community, a hiring algorithm, a weapons platform, or a giant chatbot marketed as the path to AGI? These systems have different purposes, failure modes, and power structures. Treating them as one inevitable technology lets companies make grand claims while avoiding concrete accountability. Our future should not be determined by a handful of corporations building ever-larger general-purpose models on exploited labor, enormous resource use, and poorly documented data. A system whose purpose is supposedly everything cannot be meaningfully tested for every context. And scaling it does not magically turn publicity into scientific progress; it can introduce new failures even in tasks for which better-scoped tools already exist. But I do not accept the claim that this corporate race is the only possible future. Start with an actual need. Define who the tool serves, who governs the data, what failure means, and who has the power to stop deployment. Smaller, carefully designed systems can support local languages, medical work, or documentation of inequality. Federated community organizations can share infrastructure without surrendering control to a few cloud companies. So AI means a struggle over institutional choices, not an autonomous machine destiny. The important question is not whether “AI” will save or destroy humanity. It is: who is building what, for whose benefit, at whose expense, and who gets to decide?

That still depends on what you are calling “AI.” Everyday life may change substantially—not because a magical general intelligence arrives, but because institutions insert automated systems into hiring, education, health care, policing, benefits, media, and workplaces. A person may never choose to use a chatbot and still have their life shaped by a model deciding whether their résumé is seen, their claim is flagged, or their labor is monitored. The scale of change is therefore a political and institutional choice, not a technological weather forecast. Companies want us to imagine inevitability: enormous models everywhere, replacing every tool and mediating every interaction. But why should one general-purpose system be deployed across contexts with completely different needs and risks? Often the result will be unreliable services, more surveillance, concentrated power, and workers forced to absorb the failures. There are better changes available. Communities could build narrow tools for local languages, specific medical tasks, or documenting inequality, with affected people controlling the data and governance. Everyday life will change as much as powerful institutions force these systems into it—but the direction remains contested.

I reject the premise that benefits and harms can be totaled into one score for “AI.” Benefits for whom? Harms imposed on whom? A corporation may count lower labor costs as a benefit while workers experience surveillance, deskilling, or job loss. A government may call automated eligibility screening efficient while people wrongly denied services bear the consequences. Under the current corporate trajectory, I expect the overall impact to be harmful: greater concentration of power, exploitation of data workers, wasteful resource use, unreliable systems inserted into high-stakes institutions, and less accountability because everything is hidden behind the vague label of AI. That harm is not caused by an autonomous machine deciding our destiny. It comes from institutions choosing to deploy systems without clearly defined purposes, documentation, or meaningful power for affected communities. But that trajectory is not inevitable. Specific, carefully scoped tools can be useful—supporting local languages, particular medical tasks, or research documenting inequality. The alternative is to start with a real need, define failure, and give communities control over data and deployment. So my answer is conditional but not neutral: today’s giant-model race is likely to produce more harm than benefit, while community-governed, purpose-built systems could produce genuine public value.

My view is not waiting on a single dramatic technical discovery, because the central issue is institutional power. A benchmark result, a larger model, or another claim of an “emergent” capability would not answer: What is the system for? Who controls it? Whose labor and data built it? Who bears the failures? What would change my expectation is a real shift in how these systems are built and governed: affected communities and data workers holding decision-making power; independent researchers receiving durable resources; locally knowledgeable organizations controlling infrastructure and data; and institutions choosing specific, documented tools over giant systems marketed as universal intelligence. I would also want credible evidence that people can refuse deployment, challenge harmful decisions, and obtain remedy—not a corporate ethics board with no authority. If that became the dominant trajectory rather than a marginal alternative, I would expect a much more beneficial future. Conversely, further consolidation among cloud companies, wider deployment of poorly scoped models in high-stakes settings, and intensified exploitation would make my assessment even more negative. The decisive event is not a machine suddenly becoming magical. It is people reorganizing—or further entrenching—the power around these systems.

The observation that most shaped my view is how consistently technical choices follow institutional power. When companies define scale as progress, an enormous general-purpose model becomes the answer before anyone has specified the problem. Workers, affected communities, environmental costs, and failure modes are treated as external details, while corporate capability claims receive extraordinary attention. At the same time, community-rooted organizations show that another path exists. Work on local languages, data-worker power, and documenting inequality begins with specific people and needs. It asks who controls the data, what the tool is for, and what failure looks like. That is far more meaningful than announcing another giant model and retroactively searching for uses. So the contrast has shaped my view: concentrated institutions build systems that reproduce their priorities, while locally knowledgeable groups can build narrower tools that serve actual communities. “AI’s future impact” will not be determined mainly by some discovery inside a model. It will be determined by which institutions receive resources, whose knowledge counts, and who gets to say no.