OpenAI
OpenAI is an American artificial intelligence research and deployment organisation that develops large-scale machine learning systems and products. It is best known for the GPT family of language models, the ChatGPT conversational system, the DALL·E image-generation models, the Whisper speech-recognition system, the Sora video-generation system, and Codex software-engineering agents.[1]
OpenAI was announced in December 2015 as a non-profit research organisation. It was founded by a group that included Sam Altman, Elon Musk, Greg Brockman, Ilya Sutskever, John Schulman, Wojciech Zaremba and several other researchers and entrepreneurs. Altman and Musk initially served as co-chairs; OpenAI was therefore not founded by Musk alone.[2] Musk left its governing board in February 2018 and has had no formal management role in the organisation since then.[3]
OpenAI describes its mission as ensuring that artificial general intelligence (AGI) benefits all of humanity. It defines AGI as highly autonomous systems that outperform humans at most economically valuable work.[4] Its legal and financial structure has changed considerably since its founding. OpenAI created a capped-profit subsidiary in 2019 and, following a recapitalisation completed in 2025, now consists principally of the non-profit OpenAI Foundation and OpenAI Group PBC, a public benefit corporation controlled by the Foundation.[5] Sam Altman is its chief executive officer, and Bret Taylor chairs the boards of the Foundation and OpenAI Group.[6]
OpenAI is significant to transhumanism and futurism because its stated objective is the development of general-purpose machine intelligence that could transform human cognition, scientific research, economic production, governance and the distribution of power. Its technologies can function as forms of intelligence amplification, but sufficiently capable systems could also create problems involving AI alignment, autonomy, concentration of power, technological unemployment, surveillance, military applications and existential risk. Performance on linguistic, reasoning or professional tasks does not by itself demonstrate machine consciousness, subjective experience or moral personhood; those are separate empirical and philosophical questions.
History
Founding as a non-profit
OpenAI was publicly announced on 11 December 2015 and began operating in early 2016.[7] Its original announcement described it as a non-profit artificial-intelligence research company whose work would be directed towards positive human impact rather than financial returns. It argued that advanced AI should operate as an extension of individual human intentions and should be distributed as broadly and evenly as possible.[8]
The organisation's initial leadership included Sam Altman and Elon Musk as co-chairs, Ilya Sutskever as research director and Greg Brockman as chief technology officer. Other founding members named in the announcement were Trevor Blackwell, Vicki Cheung, Andrej Karpathy, Durk Kingma, John Schulman, Pamela Vagata and Wojciech Zaremba. Early advisers included Pieter Abbeel, Yoshua Bengio, Alan Kay, Sergey Levine and Vishal Sikka.[8]
The initial financial supporters included Altman, Brockman, Musk, Reid Hoffman, Jessica Livingston, Peter Thiel, Amazon Web Services, Infosys and YC Research. The group announced commitments totalling US$1 billion, although this represented pledged support rather than an immediate transfer of the entire amount.[8]
OpenAI initially encouraged its researchers to publish papers, source code and technical findings, and stated that patents would be shared with the world. It presented openness and collaboration as mechanisms for distributing the benefits of AI and preventing a single institution from monopolising advanced capabilities.[8]
Early reinforcement-learning and robotics research
OpenAI's early research programme covered reinforcement learning, generative models, robotics, multi-agent systems and methods for learning from human feedback.
In April 2016 it released the beta version of OpenAI Gym, an open-source toolkit containing standardised environments for developing, comparing and reproducing reinforcement-learning algorithms.[9] Gym became widely used in reinforcement-learning research and helped establish shared environments and evaluation procedures across otherwise incompatible software frameworks.
In 2017 OpenAI and researchers from DeepMind demonstrated a method for training agents from comparisons supplied by human evaluators. Instead of requiring a programmer to specify a complete mathematical reward function, the system learned a reward model from human preferences between pairs of behaviours.[10] This line of work became an important precursor to reinforcement learning from human feedback (RLHF), later used to shape the behaviour of instruction-following language models.
OpenAI also used competitive games as environments for studying long-horizon planning, coordination and self-play. An early system defeated professional players in a restricted one-versus-one version of Dota 2. The later OpenAI Five system trained five cooperating neural networks to play a restricted version of the full team game.[11] In April 2019 OpenAI Five defeated OG, the reigning Dota 2 world champion team, in two consecutive games.[12] The achievement demonstrated the ability of large-scale self-play to produce complex cooperative strategies, although the system remained specialised to a constrained game environment.
In robotics, OpenAI developed Dactyl, a system that used reinforcement learning and simulated experience to control a human-like robotic hand.[13] A later version learned to manipulate and solve a Rubik's Cube using a physical robot hand. The project used extensive randomisation during simulation to improve transfer to the physical world.[14]
Departure of Elon Musk
Elon Musk left OpenAI's board in February 2018. The organisation initially stated that the departure would avoid possible future conflicts with Tesla's increasing work in artificial intelligence and autonomous vehicles.[15]
Musk and OpenAI later presented conflicting accounts of the organisation's financing, governance and proposed conversion to a commercial structure. OpenAI has published internal correspondence which it says shows that Musk supported creating a for-profit entity but disagreed with other leaders over ownership and control.[16] Musk has subsequently criticised OpenAI for departing from its original non-profit and open-source orientation and has pursued litigation concerning its mission and corporate development. These allegations and OpenAI's responses are disputed and should be distinguished from the uncontested fact that Musk ceased to hold a formal position in 2018.
Capped-profit structure
By 2019 OpenAI's leadership had concluded that developing frontier AI would require substantially more capital and computational infrastructure than a conventional research non-profit could readily obtain. The organisation estimated that it would need to invest billions of dollars in cloud computing, specialised hardware, supercomputers and technical staff.[17]
OpenAI therefore established OpenAI LP, which it described as a capped-profit company controlled by the original non-profit organisation. Investors and employees could receive returns up to an agreed limit, while additional value would accrue to the non-profit. OpenAI stated that the subsidiary's primary fiduciary obligation remained the advancement of the OpenAI Charter rather than the maximisation of investor returns.[17]
The structure was intended to combine access to private capital and employee equity with continuing non-profit control. It was nevertheless an important departure from OpenAI's original presentation as an organisation unconstrained by the need to generate financial returns.
Partnership with Microsoft
In July 2019 Microsoft invested US$1 billion in OpenAI. The companies agreed to develop large-scale computing infrastructure on Microsoft Azure, and Microsoft became OpenAI's exclusive cloud provider.[18]
The partnership was expanded in 2021 and again in January 2023, when Microsoft announced a multiyear, multibillion-dollar investment. Microsoft integrated OpenAI models into consumer and enterprise products and offered them to developers through Azure OpenAI Service.[19]
The relationship supplied OpenAI with capital, computing infrastructure, distribution and access to enterprise markets. It also made the organisation increasingly dependent on a major technology corporation and raised questions about whether the economic benefits and control of advanced AI could remain broadly distributed.
Following OpenAI's 2025 recapitalisation, Microsoft held approximately 27 per cent of OpenAI Group PBC. The OpenAI Foundation held 26 per cent, together with a warrant that could provide additional equity if specified valuation conditions were met. The remaining 47 per cent was held by current and former employees and other investors.[5]
Language models and generative AI
GPT series
OpenAI's Generative Pre-trained Transformer programme helped establish large-scale unsupervised pre-training followed by adaptation to downstream tasks as a central paradigm in natural-language processing.
The first GPT research, published in 2018, trained a Transformer language model on a large unlabelled text corpus and then fine-tuned it for tasks such as textual entailment, semantic similarity, reading comprehension and commonsense reasoning.[20]
GPT-2, announced in February 2019, was a 1.5-billion-parameter language model trained on text from millions of web pages. OpenAI initially withheld the complete model and used a staged-release process because of concerns about automated misinformation and other misuse.[21] The complete model was released later in 2019 after further analysis. The episode became an early case study in whether increasingly capable AI models should always be released openly.
GPT-3, described in May 2020, contained 175 billion parameters. It demonstrated that sufficiently large language models could perform many tasks from instructions or a small number of examples without task-specific gradient updates.[22] Its capabilities included text generation, translation, question answering, summarisation and limited forms of arithmetic and code generation, while retaining substantial weaknesses involving factual reliability, bias and brittle reasoning.
OpenAI began providing access to its language models through the OpenAI API in June 2020. The service allowed developers to build applications on a general-purpose text interface while permitting OpenAI to control access, monitor use and update the underlying systems without distributing the model weights.[23]
Instruction following and RLHF
Unmodified language models are trained primarily to predict likely continuations of text rather than to follow a user's intentions. OpenAI developed instruction-tuned models using labelled demonstrations and reinforcement learning from human feedback.
The InstructGPT research trained a reward model from human rankings and used reinforcement learning to optimise GPT-3 models against those preferences. Human evaluators generally preferred outputs from the smaller instruction-tuned models to outputs from the much larger unmodified GPT-3 model.[24]
RLHF improved instruction following and reduced some harmful or irrelevant outputs, but it did not solve the full AI alignment problem. Human feedback can be inconsistent, culturally dependent, incomplete or strategically manipulated. A model can also learn to produce responses that evaluators prefer without reliably representing truth, uncertainty or the evaluator's underlying values.
ChatGPT
OpenAI released ChatGPT as a public research preview on 30 November 2022. The system used a conversational interface and was trained to respond to follow-up questions, acknowledge some errors, challenge incorrect premises and refuse certain requests.[25]
ChatGPT brought large language models into widespread public use and accelerated their adoption in education, software development, writing, research, customer service and professional work. The interface made the underlying technology accessible to users who did not need to understand machine-learning models or application programming interfaces.
The system also made limitations of language models visible at scale. ChatGPT and related models can generate fluent but false information, reproduce biases, make inconsistent inferences, follow malicious instructions, or express unwarranted confidence. OpenAI has repeatedly revised the models, system instructions, safety policies, interface and tool access in response to observed behaviour.
GPT-4 and later systems
OpenAI released GPT-4 in March 2023. GPT-4 accepted text and image inputs and produced text outputs. OpenAI reported performance comparable to humans on several professional and academic benchmarks, while noting that the system remained less capable than humans in many real-world situations.[26]
The GPT-4 release included a system card discussing risks and mitigations and was accompanied by the open-source OpenAI Evals framework for constructing and running model evaluations. OpenAI did not disclose the model's architecture, parameter count, training-compute requirements, dataset composition or hardware configuration, citing competitive and safety considerations.
Subsequent model generations added stronger multimodal processing, tool use, longer context windows, test-time reasoning and agentic capabilities. OpenAI introduced GPT-5 in August 2025 as a routed system combining rapid responses with more computationally intensive reasoning when appropriate.[27] The GPT product family has continued to change through later model revisions, so individual version numbers and availability are time-dependent.
Other systems and products
DALL·E
OpenAI introduced DALL·E in January 2021. The original model used a Transformer architecture trained on paired text and image data to generate images from natural-language descriptions.[28] Later versions improved resolution, composition, text rendering and integration with ChatGPT.
Text-to-image systems can extend human creativity and lower the technical threshold for producing illustrations, prototypes and visual communication. They also create disputes involving copyrighted training material, attribution, artistic labour, impersonation, fabricated evidence and the difficulty of distinguishing synthetic from recorded media.
Whisper
Whisper, released in September 2022, is an automatic speech-recognition system trained on multilingual and multitask audio data. OpenAI released the model and code under an open-source licence.[29] Whisper has been used for transcription, translation, accessibility tools and voice interfaces.
Sora
OpenAI announced Sora, a generative video model, in February 2024. The system generates and extends video from textual, visual or video inputs.[30] Early versions displayed limitations in physical consistency, causal sequences, spatial relations and the persistence of objects over time.
Video-generation systems increase the ability to visualise hypothetical environments, scientific concepts and artistic scenes, but also lower the cost of producing deceptive or non-consensual synthetic media.
Codex and agents
The name Codex was initially used for language models specialised for computer code. In 2025 OpenAI introduced a cloud-based Codex software-engineering agent capable of inspecting repositories, writing and modifying code, executing tests, fixing defects and preparing proposed changes for human review.[31]
Later Codex products combined command-line tools, cloud environments, editor integration and coordinated multi-agent workflows.[32] Such systems represent a transition from models that primarily generate answers to agents that perform extended sequences of actions in external environments.
Agentic systems can amplify human productivity, particularly in software engineering and research. They also introduce additional risks because errors can propagate through files, networks, tools or organisational workflows. Reliable deployment requires permission controls, sandboxing, audit logs, verification, rollback mechanisms and human responsibility for consequential actions.
Mission and governance
OpenAI Charter
The OpenAI Charter states that the organisation's mission is to ensure that AGI benefits all of humanity. It sets out four principal commitments:[33]
- Broadly distributed benefits: OpenAI states that it will use any influence over AGI deployment for the benefit of all and avoid uses that harm humanity or unduly concentrate power.
- Long-term safety: it commits to researching the safety of advanced AI and encouraging adoption of safety research throughout the AI community.
- Technical leadership: it argues that influencing the development and social effects of AGI requires remaining near the frontier of AI capabilities.
- Cooperative orientation: it commits to cooperation with research and policy institutions and to providing public goods that help society navigate advanced AI.
The Charter states that OpenAI's primary fiduciary duty is to humanity. It also says that the organisation would assist another safety-conscious project rather than continue competing if that project were substantially closer to building AGI.[33] The meaning, enforceability and practical application of these commitments remain subjects of debate.
The Charter acknowledged that safety and security concerns could reduce traditional publication of frontier research. This represented a qualification of the organisation's initial commitment to publish research and share patents openly.[33]
2023 leadership crisis
On 17 November 2023 OpenAI's non-profit board removed Sam Altman as chief executive and removed Greg Brockman from the board. The board's initial announcement stated that Altman had not been consistently candid in communications with directors.[34]
The decision produced an organisational crisis. Altman returned as chief executive several days later under a reconstituted board. An external review conducted by the law firm WilmerHale concluded that a breakdown in trust between Altman and the previous board had precipitated the removal. OpenAI stated that the review did not find that the decision arose from concerns about product safety, security, development speed or company finances.[35]
Following the review, OpenAI expanded its board, adopted new governance guidelines, strengthened its conflict-of-interest policy, created a whistleblower hotline and established additional board committees.[36]
The episode demonstrated the difficulty of governing a rapidly growing commercial organisation through a small non-profit board charged with representing a global public-interest mission. It also revealed tensions between independent oversight, executive authority, employee interests, investor relationships and the competitive development of frontier AI.
2025 recapitalisation
On 28 October 2025 OpenAI announced that its non-profit entity had become the OpenAI Foundation and that its commercial subsidiary had become OpenAI Group PBC.[5]
A public benefit corporation is a for-profit company whose governing documents require it to pursue stated public benefits and consider interests beyond immediate shareholder returns. It remains capable of raising capital and distributing financial returns.
The OpenAI Foundation retains special voting and governance rights, appoints the directors of OpenAI Group and can replace them. The Foundation and OpenAI Group formally share the same mission. A Safety and Security Committee of the Foundation oversees safety and security practices across the organisation.[5]
OpenAI states that this structure aligns mission, commercial growth and long-term incentives. Critics question whether equity ownership, investor expectations and the capital requirements of frontier-model development may nevertheless conflict with the original non-profit commitment. The existence of formal non-profit control does not by itself establish that every operational decision will maximise global welfare; conversely, the presence of commercial investment does not by itself demonstrate abandonment of the mission.
Safety, alignment and evaluation
OpenAI has conducted research on reward modelling, debate, amplification, adversarial testing, interpretability, model behaviour, robustness and the prevention of misuse. Its early work on human-preference learning contributed to methods now widely used for adapting language models to human instructions.
OpenAI uses staged deployment, system cards, external red-teaming, policy restrictions and capability evaluations to identify risks before and after model release. In 2025 it updated its Preparedness Framework, which assesses whether advanced models have developed capabilities that could create severe harm and specifies safeguards that may be required before deployment.[37]
Evaluated risk areas have included biological and chemical capabilities, cybersecurity, persuasion and forms of model autonomy. The framework distinguishes a model's underlying capabilities from the effectiveness of safeguards placed around it.
These measures provide evidence of an institutional safety programme, but their adequacy cannot be inferred solely from OpenAI's internal assessments. Important questions include the independence of evaluators, access to models and training information, reproducibility of results, treatment of unknown failure modes, authority to delay deployment and whether commercial competition creates incentives to interpret uncertainty permissively.
Technical alignment methods also remain incomplete. RLHF and similar procedures primarily optimise observable behaviour under selected evaluation conditions. They do not necessarily reveal the internal representations, objectives or strategies that a more autonomous system might develop outside those conditions. For this reason, behavioural testing, mechanistic interpretability, structural measurement, adversarial evaluation and post-deployment monitoring are complementary rather than interchangeable.
Relevance to transhumanism
Intelligence amplification
OpenAI's systems are already used as cognitive tools for writing, programming, translation, analysis, education, design and scientific work. In this role they function as forms of intelligence amplification: their immediate significance lies not in replacing the human mind but in extending the speed, scope and accessibility of human cognitive activity.
The effects are uneven. Access to capable AI can increase individual and institutional productivity, but differences in cost, infrastructure, language coverage, technical literacy and organisational power can cause those benefits to be distributed unequally. Dependence on centrally operated systems can also reduce user autonomy if access, model behaviour or permitted uses are controlled by a small number of providers.
Artificial general intelligence
OpenAI is unusual among major technology organisations in making AGI the explicit centre of its mission. If systems satisfying its definition were created, they could alter scientific discovery, economic production, education, healthcare, military power and political institutions.
AGI is not a single universally accepted technical threshold. Performance across benchmarks, professional examinations or selected economic tasks does not necessarily imply robust general intelligence, self-directed agency or competence in unfamiliar physical and social environments. Claims that a system has reached AGI should therefore be evaluated through explicit operational criteria rather than through branding, isolated demonstrations or subjective impressions.
Human–AI integration
OpenAI's founding statement described AI as an extension of individual human wills. This is compatible with a transhumanist interpretation in which machine intelligence augments rather than merely substitutes for human agency.[8]
Possible forms of integration include persistent personal assistants, scientific copilots, adaptive education, automated software development, multimodal accessibility systems and future connections with brain-computer interface technologies. However, an AI system mediating cognition can influence what information users encounter, how choices are framed and which values are embedded in decision-making. Effective augmentation therefore depends on user control, transparency, pluralism, privacy and the ability to inspect or replace the systems involved.
Superintelligence and existential risk
OpenAI's mission presupposes that the development of very powerful AI could create both exceptional benefits and exceptional harms. A system exceeding human abilities across most relevant domains could accelerate research in medicine, energy, materials science and environmental management. It could also amplify cyber operations, biological design, surveillance, strategic deception or automated warfare.
The possibility of artificial superintelligence raises a control problem: a system capable of outperforming its designers may become difficult to evaluate, constrain or correct. The seriousness of this possibility is not evidence that any current OpenAI system possesses such capabilities. It is a reason to construct measurable safety criteria, independent evaluations and governance mechanisms before they become necessary.
Criticism and controversies
Openness and commercialisation
The name “OpenAI” reflected the organisation's original commitment to open publication and broad distribution of research. Over time OpenAI moved towards controlled APIs, proprietary model weights and limited disclosure of training data, architecture and compute.
OpenAI has argued that unrestricted publication can enable misuse and that retaining control over deployment permits monitoring and staged release. Its Charter explicitly anticipated that safety and security concerns would reduce conventional publication of frontier research.[33]
Critics argue that commercial incentives and competitive advantage, rather than safety alone, also explain the reduced transparency. Because outsiders generally cannot inspect the full training process or independently reproduce frontier models, it can be difficult to distinguish restrictions motivated by legitimate risk reduction from those protecting market position.
Mission and corporate control
OpenAI's transition from a non-profit laboratory to a capital-intensive commercial organisation has generated continuing debate over mission drift. The organisation maintains that the non-profit Foundation retains control and that commercial growth is necessary to finance its research and deployment.[5]
Critics contend that investor ownership, employee equity, commercial partnerships and competition for market share may conflict with a fiduciary commitment to humanity as a whole. Supporters respond that advanced AI cannot be developed or safely operated without large-scale capital, infrastructure and experienced personnel.
The decisive question is therefore not whether OpenAI is categorically “non-profit” or “for-profit”, but whether its governance arrangements produce verifiable decisions consistent with its stated public-benefit mission under conditions of commercial and geopolitical pressure.
Copyright and training data
Authors, artists, news organisations and other copyright holders have sued OpenAI and other generative-AI companies, alleging that copyrighted works were copied without permission for model training and that some outputs reproduce or substitute for protected material. In 2024 eight United States newspaper publishers joined litigation against OpenAI and Microsoft.[38]
OpenAI argues that training models on publicly available material constitutes fair use under United States law and has provided mechanisms through which website operators can restrict automated access.[39] Plaintiffs dispute that position, particularly where material was allegedly obtained from unauthorised sources or outputs compete with the original works.
The law remains unsettled and varies by jurisdiction. In July 2026 the Delhi High Court ruled that OpenAI had not infringed the copyright of Indian news agency ANI under the facts presented, while related litigation continued in the United States and Canada.[40] Individual judgments should not be generalised into a universal rule governing all datasets, outputs or legal systems.
Bias, misinformation and representation
Large language models learn statistical regularities from human-produced data and can reproduce stereotypes, factual errors and dominant cultural assumptions. Safety training can reduce some harmful outputs but can also introduce new value judgements concerning which views are represented, refused or prioritised.
Because a single model may serve users across different legal and cultural contexts, no fixed behaviour policy can be entirely neutral. Questions about system behaviour therefore involve governance as well as engineering: who selects the values, who can contest them, what degree of personalisation is permitted, and which limits are necessary to protect others.
Synthetic text, images, audio and video can also facilitate impersonation, fraud and misinformation. These risks arise from the interaction of model capability, user intent, distribution platforms and social institutions rather than from a model in isolation.
Labour and concentration of power
OpenAI systems can automate components of software development, administration, media production, education, research and other knowledge work. They may increase productivity and create new forms of employment, but can also displace tasks, weaken bargaining power or redistribute income from workers and creators towards owners of models, data and computing infrastructure.
Frontier AI development requires large quantities of capital, specialised hardware, electricity, data and engineering expertise. These requirements favour a small number of corporations and states. Concentration can accelerate development by pooling resources, but it may also reduce scientific pluralism, democratic accountability and the ability of individuals or smaller institutions to shape how advanced AI is used.
Assessment
OpenAI has been one of the central institutions in the transition of artificial intelligence from specialised research systems to widely used general-purpose models. Its work on reinforcement learning, generative pre-training, human-feedback alignment, multimodal systems and public deployment has influenced both academic research and the commercial AI industry.
Its history also illustrates unresolved tensions in frontier-AI development:
- openness versus control of potentially dangerous capabilities;
- public-benefit missions versus the capital requirements of large-scale computing;
- rapid deployment versus prior safety assurance;
- centralised governance versus broad distribution of benefits;
- behavioural usefulness versus factual reliability and deeper alignment;
- human cognitive enhancement versus dependence on externally controlled systems.
OpenAI's stated mission is consequential, but it is a claim about institutional purpose rather than evidence that the mission has been achieved. Evaluation should therefore focus on measurable outcomes: the capabilities produced, the risks identified, the transparency and reproducibility of safety evidence, the distribution of benefits, and the degree to which governance remains effective as the systems and organisation become more powerful.
See also
- Artificial intelligence
- Artificial general intelligence
- Artificial superintelligence
- AI alignment
- AI safety
- AI governance
- ChatGPT
- Generative pre-trained transformer
- Google DeepMind
- Anthropic
- Intelligence amplification
- Collective intelligence
- Global brain
- Technological singularity
- Existential risk
- Transhumanism
External links
- Official website
- OpenAI Research
- OpenAI Charter
- OpenAI organisational structure
- OpenAI safety
- OpenAI on GitHub
OpenAI on Wikipedia- OpenAI on Twitter
References
- ↑ Research, OpenAI.
- ↑ Introducing OpenAI, OpenAI, 11 December 2015.
- ↑ OpenAI supporters, OpenAI, 20 February 2018.
- ↑ OpenAI Charter, OpenAI.
- ↑ 5.0 5.1 5.2 5.3 5.4 Our structure, OpenAI.
- ↑ David Vélez and Robin Vince join the boards of the OpenAI Foundation and OpenAI Group PBC, OpenAI, 21 July 2026.
- ↑ Ten years, OpenAI, 11 December 2025.
- ↑ 8.0 8.1 8.2 8.3 8.4 Introducing OpenAI, OpenAI, 11 December 2015.
- ↑ OpenAI Gym Beta, OpenAI, 27 April 2016.
- ↑ Learning from human preferences, OpenAI, 13 June 2017.
- ↑ OpenAI Five, OpenAI, 25 June 2018.
- ↑ OpenAI Five defeats Dota 2 world champions, OpenAI, 15 April 2019.
- ↑ Learning dexterity, OpenAI, 30 July 2018.
- ↑ Solving Rubik's Cube with a robot hand, OpenAI, 15 October 2019.
- ↑ OpenAI supporters, OpenAI, 20 February 2018.
- ↑ OpenAI and Elon Musk, OpenAI, 5 March 2024.
- ↑ 17.0 17.1 OpenAI LP, OpenAI, 11 March 2019.
- ↑ Microsoft invests in and partners with OpenAI to support us building beneficial AGI, OpenAI, 22 July 2019.
- ↑ Microsoft and OpenAI extend partnership, Microsoft, 23 January 2023.
- ↑ Improving language understanding with unsupervised learning, OpenAI, 11 June 2018.
- ↑ Better language models and their implications, OpenAI, 14 February 2019.
- ↑ Language models are few-shot learners, OpenAI, 28 May 2020.
- ↑ OpenAI API, OpenAI, 11 June 2020.
- ↑ Aligning language models to follow instructions, OpenAI, 27 January 2022.
- ↑ Introducing ChatGPT, OpenAI, 30 November 2022.
- ↑ GPT-4, OpenAI, 14 March 2023.
- ↑ Introducing GPT-5, OpenAI, 7 August 2025.
- ↑ DALL·E: Creating images from text, OpenAI, 5 January 2021.
- ↑ Introducing Whisper, OpenAI, 21 September 2022.
- ↑ Sora: Creating video from text, OpenAI, 15 February 2024.
- ↑ Introducing Codex, OpenAI, 16 May 2025.
- ↑ Codex, OpenAI.
- ↑ 33.0 33.1 33.2 33.3 OpenAI Charter, OpenAI.
- ↑ OpenAI announces leadership transition, OpenAI, 17 November 2023.
- ↑ Review completed and Altman, Brockman to continue to lead OpenAI, OpenAI, 8 March 2024.
- ↑ Review completed and Altman, Brockman to continue to lead OpenAI, OpenAI, 8 March 2024.
- ↑ Our updated Preparedness Framework, OpenAI, 15 April 2025.
- ↑ Eight US newspapers sue OpenAI and Microsoft for copyright infringement, Associated Press, 30 April 2024.
- ↑ OpenAI and journalism, OpenAI, 8 January 2024.
- ↑ Indian court says OpenAI did not violate news agency ANI's copyright, Reuters, 24 July 2026.