The name OpenAI once sounded like a declaration. It suggested an artificial intelligence laboratory built around openness: open research, open collaboration and benefits that would not be captured by a small circle of companies or investors. In 2015, that promise felt both idealistic and necessary. The people behind the new nonprofit argued that increasingly capable AI should serve humanity broadly rather than become another proprietary advantage owned by whoever could afford the most computing power.
Ten years later, the name raises a more uncomfortable question: why is OpenAI called OpenAI when its most important frontier models are not open? ChatGPT may be available to hundreds of millions of people, and developers may build with OpenAI systems through an API, but access is not the same as ownership. Users cannot inspect the full training data, study every architectural decision, download the weights of the most capable models or independently decide how those systems should behave. They are allowed to use the intelligence, but OpenAI retains control over the machinery behind it.
It is tempting to treat this as a simple story of hypocrisy. A nonprofit promised openness, discovered a profitable product and quietly abandoned its principles. That version is emotionally satisfying, but it is too neat. The real history is more interesting because the meaning of “open” changed as the technology became more capable, more expensive and more politically significant. OpenAI moved from publication to staged release, from staged release to controlled access, and from a research laboratory to a global product company. Along the way, it developed an argument that restricting a model could sometimes serve the public better than releasing it.
The question is no longer whether OpenAI is literally open or closed. It is what kind of openness the company still believes in, which forms of control it considers legitimate and whether a private organization should be trusted to decide when the public is ready for the technology built in its name.
Why Is OpenAI Called OpenAI?
OpenAI’s name makes sense only when read in the context of its founding. In its original announcement from December 2015, the organization described itself as a nonprofit AI research company whose goal was to advance digital intelligence in a way most likely to benefit humanity as a whole. Because it would not be constrained by the need to produce a financial return, it argued, it could focus more clearly on positive human impact.
The language went further than a conventional corporate mission statement. OpenAI said that AI should extend individual human agency and be distributed as broadly and evenly as possible. Researchers would be encouraged to publish their work as papers, articles or code. Any patents would be shared, and the laboratory expected to collaborate freely with other institutions. The organization was presenting itself as an answer to a future in which a small number of powerful actors might otherwise control increasingly general intelligence.
The word “open” therefore did not refer to a single technical promise. It represented a collection of related ideas: research should circulate, scientists should collaborate, the benefits of AI should be widely distributed and the organization developing it should not place shareholder returns above humanity’s interests. Open code was part of that atmosphere, but the deeper promise was open participation in the future of intelligence.
This distinction matters because OpenAI never published a permanent rule stating that every model it created would always be released with downloadable weights. The founding announcement described a direction and an institutional character rather than a precise software licence. Even so, names create expectations. When an organization calls itself OpenAI, encourages publication and speaks about broadly distributed power, people reasonably assume that openness will remain central to how it operates.
The name was not just descriptive. It was a moral position about who should benefit from intelligence once intelligence itself became a technology.
Open Was Never Just One Thing
Much of the argument surrounding OpenAI becomes confused because people use the word “open” to describe several different conditions. A company may publish research without releasing model weights. It may release weights without revealing the training data. It may provide a free product while retaining complete control over the infrastructure. It may open-source supporting tools while keeping its most commercially valuable system behind an API.
These are not minor distinctions. Open research allows outsiders to learn from methods and results. Open-source code allows them to inspect and modify software. Open weights allow a trained model to be run and adapted independently. Open data would reveal or provide the material used to train it, subject to privacy and copyright constraints. Open governance would give affected communities a meaningful role in deciding how the system is developed and deployed. Open access merely means that people are permitted to use a service.
OpenAI has practised different combinations of these ideas at different moments. The company can accurately claim that it has expanded access to powerful AI, supported developers, published research and released some models openly. Critics can also accurately say that its frontier systems remain proprietary, their training is difficult to scrutinize and the public has little direct influence over their governance. Both descriptions can be true because they refer to different kinds of openness.
This is similar to the tension examined in Free Software vs Open Source. Access to useful technology does not necessarily grant the freedom to understand, modify or control it. A free ChatGPT account may place remarkable capabilities in someone’s hands, but it does not give that person authority over the model. The relationship resembles access to a platform more than possession of software.
That gap between access and control would become the defining feature of OpenAI’s evolution. The company did not simply close a previously open door. It replaced one idea of openness with another: instead of giving the public the model, it would give the public a managed way to interact with the model.
GPT-2 Was the Moment Openness Became a Safety Question
The shift became visible in 2019 with GPT-2. OpenAI had trained a language model capable of producing unusually coherent text, answering questions and performing several language tasks without task-specific training. The organization also worried that the same capabilities could support spam, impersonation, deceptive content and other misuse.
Rather than immediately release the complete trained model, OpenAI announced a staged approach. In its GPT-2 publication, the company released a smaller version and a technical paper while initially withholding the full model. It described this as an experiment in responsible disclosure: capabilities would be opened gradually while researchers studied possible harms.
The decision was controversial partly because it challenged the culture from which OpenAI had emerged. Scientific credibility traditionally depends on allowing others to inspect, reproduce and challenge results. Withholding a trained model placed OpenAI in the position of asking the research community to trust its private judgment about what was too dangerous to release.
Yet the concern was not imaginary. A model does not become harmless merely because its code is available under an open licence. Once weights are released, they cannot be recalled. Safety restrictions can be removed, and the system can be adapted by people whose intentions differ completely from those of its creators. Traditional software can also be misused, but general-purpose models make the range of possible downstream behaviour unusually difficult to predict.
GPT-2 introduced a principle that would shape OpenAI’s future: openness should be conditional on capability and risk. The stronger the model became, the less automatic its release would be. Publication was no longer treated as a default scientific virtue. It became one variable in a broader calculation involving safety, misuse and control.
This was an intellectually defensible position, but it contained a structural problem. Who performs that calculation? If the organization that builds a powerful model also decides how dangerous it is, what evidence the public may inspect and when access should expand, safety can become a permanent source of private authority. The company may be acting responsibly, but outsiders have limited ability to verify that responsibility.
The GPT-2 debate was therefore about more than one language model. It marked the moment when OpenAI began moving from open research toward controlled disclosure—and when the word “open” started to mean “open when we decide it is safe.”
The Cost of Intelligence Changed the Institution
Safety was only one force pushing OpenAI away from its original structure. The other was money—not merely profit in the ordinary sense, but the extraordinary cost of developing increasingly capable models. Training advanced AI requires specialised hardware, enormous computing infrastructure, energy, data, engineering talent and the capacity to serve models to users at global scale.
The original nonprofit model was created before the full economic shape of modern generative AI had become obvious. OpenAI later said that it needed to scale far faster than anticipated. In 2019, it introduced a for-profit subsidiary under nonprofit control, arguing that the mission required access to capital that conventional donations could not provide. The philosophical contradiction was immediate: an organization founded to remain unconstrained by financial returns now needed investors in order to compete at the frontier.
That does not automatically mean the mission was fraudulent. If the creation of advanced AI genuinely requires billions—or eventually far more—then refusing commercial capital could simply leave development to corporations with fewer public-interest commitments. A nonprofit without sufficient computing power might preserve its moral purity while becoming technically irrelevant. OpenAI’s leaders could reasonably argue that influence over the future requires the resources to build that future.
But capital is not neutral. Investors expect growth, products require revenue, partnerships create dependencies and a large workforce must be paid. Once a laboratory becomes a platform, decisions that look like safety measures can also protect competitive advantage. Withholding weights may reduce misuse, but it also prevents competitors from reproducing the product. An API may broaden access, but it also creates recurring revenue and keeps users dependent on the provider.
This is why the change cannot be understood through declarations of good or bad intent. Institutional incentives matter even when people sincerely believe in the mission. OpenAI may want AGI to benefit humanity while also needing to defend market position, attract investment and finance immense infrastructure. The difficult question is not which motive is real. Several motives can be real at the same time.
OpenAI’s current organizational description reflects that compromise. The nonprofit OpenAI Foundation governs the for-profit OpenAI Group, which operates as a public benefit corporation. The company says the structure allows commercial success and mission to advance together. Critics are entitled to ask whether those goals can remain aligned when the value of the business depends on controlling scarce and powerful technology.
The API Redefined Open Access
In 2020, OpenAI made the new philosophy concrete by introducing an API. Instead of releasing its most capable models for anyone to download and run, it would host them and provide controlled access. Developers could send requests, receive outputs and build products without possessing the underlying system.
OpenAI explained this decision directly in its announcement of the API. Commercial access would help fund research, safety and policy work. Hosting very large models could make them available to smaller organizations that lacked the expertise and infrastructure to operate them. Most importantly, API access could be restricted if the technology was used in harmful ways. An openly released model could not be withdrawn once distributed.
The trade-off was dependence. When intelligence is delivered as a service, the provider retains the power to change prices, policies, limits and model behaviour. An application built around the API operates at the pleasure of the platform. A feature may disappear, a model may be replaced and a previously accepted use may be restricted. The user receives capability without sovereignty.
This resembles the broader transition from ownership to access explored in When Everything Becomes a Subscription. Cloud services are convenient precisely because someone else maintains the infrastructure. That same convenience makes exit difficult. In the case of AI, the rented object is not storage space or entertainment but a general-purpose cognitive tool woven into writing, programming, research and decision-making.
The API therefore gave “open” a new meaning. OpenAI would be open for use, but not necessarily open for inspection, modification or independent operation. The public could enter through the front door, while the engine room remained private.
ChatGPT Made Access Feel Like Openness
When ChatGPT arrived in late 2022, the philosophical change became easy to overlook because the product felt so accessible. A complex language model no longer required programming knowledge or an API account. Anyone could open a conversational interface and begin experimenting. The effect was more socially significant than releasing a research repository that only specialists could understand.
But the interface also concealed the system. Users could experience intelligence without seeing the choices that shaped it. They did not know the complete training data, the full moderation architecture, the weight given to different values or every way in which the model changed between versions. OpenAI could update the product centrally, and the same user might receive meaningfully different behaviour without choosing to install anything new.
The platform was open in the everyday sense that almost anyone could approach it. It was closed in the infrastructural sense that almost nobody outside OpenAI could reproduce the service independently. This is not a contradiction unique to OpenAI. Google Search, Instagram and cloud productivity tools are also broadly accessible but privately governed. OpenAI applied the platform model to something that increasingly resembles a general layer of knowledge and reasoning.
That is why the debate matters. A closed photo-sharing platform controls one part of online life. A closed AI assistant may mediate writing, education, software development, information discovery and professional judgment. As the scope of the tool expands, private control over its defaults becomes more consequential.
The question in Human-Centered AI: Who Should Remain in Control? therefore applies to the company as well as the model. Human-centered AI cannot mean only that the interface is pleasant or the responses are helpful. It must also ask which humans possess meaningful authority over the system.
Safety Is a Real Argument—and a Convenient One
Criticism of closed AI often dismisses safety as public relations. That is too easy. Releasing powerful model weights creates genuine risks because control cannot later be restored. A model can be fine-tuned to ignore safeguards, integrated into automated systems or redistributed across jurisdictions with different laws and norms. Openness can support independent research and innovation, but it can also make harmful adaptation easier.
OpenAI’s preference for staged releases and hosted access therefore has a coherent logic. The company can monitor patterns of use, improve safeguards, limit abusive accounts and update a deployed system when weaknesses appear. None of those interventions is possible in the same way after unrestricted weights have spread across the internet.
The problem is that safety and control point in the same direction. Measures that reduce risk also strengthen the provider’s authority. They make the technology harder to audit, reproduce and challenge. They preserve commercial scarcity. Even when safety is the genuine motivation, closure produces strategic benefits for the company making the decision.
This creates an accountability problem. OpenAI possesses more information about its systems than the public, regulators or independent researchers. It can explain that a release would be dangerous without revealing enough for outsiders to evaluate the claim fully. The public is asked to accept that the institution capable of building the model is also the institution best qualified to set the boundaries around it.
That may sometimes be true. It should never become unquestionable. Safety cannot function as a magic word that ends debate. A responsible closed model still requires meaningful transparency about capabilities, limitations, evaluation methods, governance and incidents. Otherwise, “trust us” becomes the operating system of AI policy.
The deeper tension mirrors the one found in Open Code, Closed Governance: Who Really Controls Open Source?. Publishing code does not guarantee democratic control, but withholding code does not remove the need for democratic legitimacy. Openness and governance are separate problems, and neither can substitute for the other.
OpenAI Is Not Completely Closed
Any fair account must acknowledge that OpenAI’s trajectory has not moved in only one direction. The company has released research, developer tools and models such as Whisper and CLIP. More significantly, in 2025 it introduced gpt-oss-120b and gpt-oss-20b, its first open-weight language models since GPT-2.
The gpt-oss weights can be downloaded, customized and run on infrastructure chosen by the user. OpenAI presented them as a way to support local deployment, experimentation and an open-model ecosystem while applying safety training and evaluation before release. Their existence complicates the claim that the word “Open” has become entirely meaningless.
However, open-weight does not automatically mean fully open. The weights are available, but that does not provide the complete training dataset or make the entire development process reproducible. More importantly, gpt-oss does not erase the distinction between OpenAI’s openly released models and its proprietary frontier systems. The company can support an open ecosystem at one level while retaining control over its most strategically important capabilities at another.
This may represent a practical middle path. Not every model must carry the maximum capability or maximum risk. Open-weight systems can serve developers who need local control, privacy or customization, while hosted frontier models remain subject to tighter oversight. The approach recognises that “open versus closed” is not always a binary choice.
It may also create a two-tier ecosystem. The public receives capable open models, while the provider retains exclusive access to the most advanced intelligence. Openness becomes a product category rather than the organizing principle of the institution. Users may customize yesterday’s frontier while tomorrow’s frontier remains behind a managed interface.
Whether that balance is responsible or self-serving cannot be decided by branding alone. It depends on the capability gap between open and closed systems, the transparency of release decisions and whether independent researchers can meaningfully test the claims made by the company.
The Structure Changed, but the Mission Remained
OpenAI continues to say that its mission is to ensure artificial general intelligence benefits all of humanity. In 2025, the company outlined another structural evolution: the nonprofit would retain control while the commercial entity became a public benefit corporation. OpenAI argued that making advanced AI broadly available could require hundreds of billions of dollars and eventually even more, making a conventional path to large-scale capital necessary.
This is the institutional version of the same philosophical trade-off. The nonprofit mission is intended to constrain the company, while the commercial company provides the resources to pursue the mission. OpenAI’s leadership describes this as alignment between purpose and scale. Skeptics see a mission increasingly dependent on the financial system it was originally designed to resist.
The current structure may prove resilient, or it may reveal how difficult it is for a mission-driven institution to remain independent once its technology becomes a global market. Either way, the name OpenAI continues to perform important work. It reminds the organization—and everyone watching it—that the original ambition was broader than building a successful product.
The word “open” has become a standard against which the company can be judged. That may be uncomfortable for OpenAI, but it is also valuable. A name that creates accountability is not meaningless simply because the institution has failed to embody it perfectly.
What Would an Open OpenAI Look Like Today?
Returning to the exact practices of 2015 is neither realistic nor necessarily responsible. Frontier AI is more capable, more widely deployed and more entangled with public life than the early laboratory imagined. Genuine openness today would require more than uploading model weights and hoping for the best.
It could begin with clearer distinctions. OpenAI should say precisely when it means open access, open research, open source or open weights. Treating these terms as interchangeable allows broad availability to stand in for genuine user control. A person using ChatGPT is not in the same position as a developer running an open-weight model locally, and neither has the same insight as a researcher with access to training data and evaluation methods.
Greater openness could also mean stronger external scrutiny. Independent researchers need enough access to test safety, bias, security and social effects without depending entirely on company-selected demonstrations. Regulators and public-interest institutions need meaningful evidence about systems whose decisions may affect education, employment, information and public services.
Governance matters just as much as technical disclosure. If OpenAI’s mission concerns all humanity, the people affected by its systems should not appear only as users, customers or sources of feedback. The company does not need to become an online referendum, but a global mission requires forms of accountability broader than executive judgment and investor confidence.
Finally, openness should include practical exit. Users and developers should be able to move their data, workflows and applications without unbearable switching costs. A platform becomes more trustworthy when people can leave it. Portability does not reveal model weights, but it limits the power created by dependence.
None of these measures would eliminate the need for controlled releases. They would make control less unilateral. The central issue is not whether every powerful model must be distributed without restrictions. It is whether decisions made in the name of public safety are themselves open to meaningful public examination.
The Name Is a Promise, Not a Description
So why is OpenAI called OpenAI? Historically, because it was founded around the belief that advanced artificial intelligence should be developed for broad human benefit, supported by publication, collaboration and an institution not driven primarily by shareholder return. The name captured an aspiration to prevent intelligence from becoming the private property of a narrow technological elite.
Today, the name is less an accurate description of every product than a record of that original promise. OpenAI has expanded access to AI on a scale few organizations have achieved, but it has done so mainly through services it controls. It has released important open tools and new open-weight models, while keeping its frontier systems behind managed interfaces. It remains governed by a nonprofit foundation, but it also operates through a commercial structure built to attract vast amounts of capital.
This does not make OpenAI simply open or simply closed. It makes the company a living example of the conflict between scientific openness, safety, commercial scale and institutional power. Each side of that conflict has legitimate arguments. Openness can distribute knowledge and invite scrutiny, but unrestricted release can create irreversible risks. Central control can support safety and accessibility, but it can also concentrate authority over tools that increasingly shape how people think and work.
The strongest criticism of OpenAI is therefore not that it has betrayed one simple definition of openness. It is that the company often asks the public to accept its own definition of openness while retaining the power to change that definition. Access becomes openness, mission becomes accountability and safety becomes the justification for decisions outsiders cannot fully examine.
The strongest defense is that an institution serious about beneficial AI must adapt when its original assumptions no longer match reality. A laboratory that released every model without regard for capability might be consistent, but not responsible. A nonprofit unable to finance frontier research might be principled, but powerless. OpenAI chose influence, scale and controlled deployment because it believes those tools offer the best path toward its mission.
Whether that belief is correct remains unresolved. The answer will not be found in the company’s name, charter or product announcements alone. It will be visible in who ultimately controls the technology, who can challenge its decisions and whether the benefits of increasingly capable AI are genuinely distributed rather than merely made available through another dominant platform.
OpenAI’s name still matters because it preserves the question the company can never completely escape: open for whom, open in what way and open on whose terms?
