docketing@steinip.com
By Lauryn Bishoff
Artificial intelligence (AI) programs work by combining datasets with complex algorithms to produce a “human-like” output. Elements such as the dataset and algorithm are fundamental to how an AI system operates, and organizations make choices about how openly these components are shared. While some companies maintain full transparency, others protect their data and model architectures closely. Openness in AI exists on a spectrum, as many organizations adopt hybrid approaches, sharing certain elements while keeping others private. [2] For the purposes of understanding the AI landscape, comparing open and closed AI systems based on their general characteristics is important.
Open AI programs are typically public in nature, meaning that the model architecture, training data, and source code are available to the community. [1] This transparency enables researchers, developers, and organizations to examine how the system is built, assess its outputs for bias, and use the existing models as foundations for new applications. One major advantage of open AI is the collaborative potential offered. By making AI accessible and downloadable, open platforms promote innovation and empower smaller organizations to build AI tools that reflect their own needs. [1] However, open AI systems often suffer from slower update cycles, limited computing power, and weaker data security. Since their datasets and code are exposed, they can be more vulnerable to misuse or intellectual property concerns. [2] Meta’s Large Language Model Meta AI (LLaMA) is an example of a successful open AI service. [1]
In contrast, closed AI programs are created with data, models, and training methods kept confidential. [1] As a result, closed models often demonstrate superior performance, faster outputs, and better security. Their private nature allows developers to build safety mechanisms, moderate content, and respond to misuse in a controlled environment. At the same time, this secrecy limits collaboration and external verification. Without insight into what data was used or how outputs are generated, users have fewer tools to identify bias or harmful training methods in these models. [2] Closed systems also tend to centralize power in a few major companies (because these companies keep their capital, data, and control concentrated), raising concerns about accessibility and monopolization. [2] Still, closed AI programs avoid some of the legal risks associated with open systems and are often better suited for commercial use. Arguably the most well-known AI platform, OpenAI’s ChatGPT, is a closed AI service. [1]
The choice between open and closed AI approaches depends on the goals, resources, and values of the organizations developing or adopting the technology. Open AI prioritizes transparency to the wider community, while closed AI emphasizes performance, control, and security. However, experts anticipate a shift away from open AI to prevent sharing powerful models with competitors and to reduce the risk of exploitation by hackers. [3] As the field evolves, the most effective AI development may involve combining elements of both to leverage openness where possible, while also maintaining necessary safeguards to ensure ethical use and fair competition.
References
[1] George Lawton, Attributes of Open vs. Closed AI Explained, TechTarget (July 8, 2024), https://www.techtarget.com/searchEnterpriseAI/feature/Attributes-of-open-vs-closed-AI-explained.
[2] Angela Luna, The Open or Closed AI Dilemma, Bipartisan Pol’y Ctr. (May 2, 2024), https://bipartisanpolicy.org/blog/the-open-or-closed-ai‑dilemma/.
[3] John Werner, Open AI Systems Lag Behind Proprietary and Closed Models, Forbes (Nov. 6, 2024), available at https://www.forbes.com/sites/johnwerner/2024/11/06/open-ai-systems-lag-behind-proprietary-and-closed-models/