By Jung Yeon-Tae (Manri Pungchwi)
President, Korea Information Service of the Construction Industry
A familiar part of ChatGPT that I had used for years began to change a few days ago. Whenever I became interested in a new field—history, space, travel, finance, construction, philosophy or artificial intelligence—I designed a specialised GPT for it. I created knowledge files, refined its instructions and repeatedly tested it to improve its performance. For me, building a GPT was never simply a matter of writing a prompt. It was the process of constructing a small world of knowledge.
Then, at some point, the door through which I had created new GPTs with my personal account closed. At first, I assumed it was a system error. It turned out to be a change in OpenAI’s policy. According to OpenAI’s official Help Center as of August 2026, personal ChatGPT accounts—including Free, Go, Plus and Pro—can no longer create or publish new GPTs. The ability to create, share and publish new GPTs is now centred on managed Business, Enterprise and Edu workspaces.
It felt as though an era had quietly begun to change direction.
The Door Has Not Closed on Existing GPTs
The first question I wanted answered was whether I could still modify the GPTs I had already created and published. The answer is yes. OpenAI says existing GPTs remain available for use and can still be edited when the applicable plan and permission requirements are met.
Users can revise an existing GPT’s name, description, conversation starters, instructions and knowledge files. A single GPT can contain up to 20 knowledge files, each with a maximum size of 512MB. Capabilities such as web search, image generation, Code Interpreter and Data Analysis, as well as connections to external services, can also be managed where account and workspace settings permit.
The life of existing GPTs has therefore not ended. What has changed is the emphasis. The more important question may now be how accurately and continuously users maintain the GPTs they have already built. What can be confirmed at present is that existing GPTs remain usable and editable, while opportunities to create new ones through personal accounts have become scarce.
An Experiment That Allowed People With Knowledge to Build AI
When OpenAI introduced GPTs in November 2023, what impressed me most was that people without coding skills could build AI systems suited to a particular purpose. It opened a path beyond an era in which only programmers could create software. People with knowledge could now design AI.
Doctors could create medical GPTs, teachers educational GPTs, historians historical GPTs and lawyers legal GPTs. It became possible to place the experience and materials accumulated over a lifetime inside a single GPT. I saw this not merely as another software feature, but as a fundamental change in how human knowledge could be transformed into software.
At first, I believed that a well-written set of instructions would be enough to produce a good GPT. Building multiple GPTs changed my view. The quality of a specialised GPT was not determined by the sophistication of its prompts alone. It also depended on the accuracy and structure of its knowledge, the relationships between concepts, the reliability of its sources and the process used to verify its answers.
A history GPT, for example, must connect not only people and events, dates and dynasties, countries and institutions, but also causes and consequences, bloodlines and political alliances, teachers and disciples, and allies and adversaries. Knowledge is not simply a collection of sentences. It is a structure of relationships. This is why I became interested in ontology while building GPTs.
Why Has the Door to Personal GPTs Narrowed?
It is important here to distinguish fact from interpretation. OpenAI has officially confirmed that personal accounts can no longer create or publish new GPTs, but it has not disclosed the specific business or strategic reasons behind the policy. It would therefore be inappropriate to conclude that the change was necessarily driven by security, cost, profitability or quality control.
Nevertheless, the direction of OpenAI’s recent product expansion suggests a broader shift. ChatGPT workspace agents can perform recurring tasks, use connected apps and tools, and run according to schedules or API triggers. If a GPT is an AI configured to respond for a particular purpose, an agent is closer to an AI system that searches for information, uses tools and carries out multiple steps to achieve a defined objective.
A GPT asked to “analyse this month’s corporate results” might examine the supplied material and produce an assessment. An agent could be assigned a recurring task: “Find the company’s results each month, compare them with the previous month and alert me when an anomaly appears.” The former is centred on knowledge and responses; the latter combines knowledge with tools and execution procedures.
This shift suggests that the centre of the AI industry is moving from the creation of personalised chatbots towards the design and operation of organisational workflows. It does not mean that individuals can no longer develop AI. Individual developers can still build AI services using APIs, open-source models and external agent frameworks. What has closed is not the entire field of AI development, but one route: creating and publishing new custom GPTs within a personal ChatGPT account.
Not the End of GPTs, but Their Integration Into Larger Systems
The arrival of agents does not mean GPTs will disappear. Language models and customised knowledge structures are instead likely to become essential components of the systems that power agents. If a GPT is an engine, an agent can be compared to a vehicle that combines that engine with sensors, navigation, control systems and operating procedures. The engine has not vanished; it has been incorporated into a larger system.
The current transition is therefore better understood as an expansion of the GPT’s role than as its demise. Competitive advantage in AI development is moving beyond the ability to write a single effective prompt towards the capacity to design context, knowledge, data, tools and entire workflows.
This evolution can be expressed as follows:
Prompt design → Context design → Knowledge design → Agent design
The process will not necessarily follow a straight line. Prompts and specialised knowledge will remain essential in the agent era. What is changing is the growing importance of combining these individual elements into a functioning system.
What Existing GPT Developers Should Do
The inability to create new GPTs with personal accounts is understandably disappointing for individual developers. Those who have already built and published GPTs, however, may now need to focus less on increasing their number and more on deepening their accuracy and reliability.
They should remove duplication and errors from knowledge files, resolve conflicting instructions and continuously incorporate new research and data. They also need verification procedures that separate claims from evidence and recheck sources, dates and figures. For a specialised GPT, the ability to avoid asserting unverified information may be more important than the ability to produce a large volume of answers.
The roles of instructions and knowledge should also be clearly separated. Instructions should define how a GPT behaves, while knowledge files should provide the specialised reference material it uses in its answers. When this is combined with an ontology that organises relationships among concepts, a search and routing system that selects relevant sources, and a verification process that evaluates reliability, a GPT can move beyond a simple chatbot and become a professional knowledge system that improves over time.
Explainable AI, or XAI, should not be treated as synonymous with general fact-checking or self-verification. The terminology matters less than the underlying structure: users should be able to trace the basis on which AI reached a conclusion and allow humans to examine the possibility of error.
Intellectual Assets Must Not Be Confined to One Platform
The most important lesson this policy change has given me is that intellectual assets should never be confined to a single platform. Platform features and policies can change at any time.
GPT developers should separately preserve the original versions of their instructions, knowledge files, ontologies, evaluation criteria, test questions, source lists, API structures and workflows. Only then can they transfer their work to another model or service when a platform changes.
A GPT should be a door into a world of knowledge. It should not become the entire house.
Who Will Be at the Centre of the New AI Ecosystem?
Restrictions on creating new GPTs through personal accounts leave questions that extend beyond a simple product adjustment. Who is given the authority to build AI? What status does knowledge accumulated by individuals hold within a platform? Is the AI ecosystem being reorganised around institutions with the capital, data, security systems and management capabilities that individual creators may not possess?
In the agent era, the ability to formulate good questions will no longer be sufficient. Developers must decide what an AI should know, where it should search for information, which criteria it should use to make judgements, what tools it may employ and when a decision must be returned to a human. The AI developer of the future will need to move beyond prompt writing and become an “AI knowledge architect” connecting knowledge and action, verification and human responsibility.
It remains too early to conclude that ontologies and knowledge graphs are indispensable routes towards artificial general intelligence. Yet human knowledge operates as a network of relationships rather than as isolated pieces of information. Technologies capable of structuring context and causality among concepts are therefore likely to become increasingly important.
An Era in Which Depth Matters More Than the Number of GPTs
For a time, I created a new GPT whenever I discovered a new field of interest. This change, however, replaces the question “How many GPTs have you built?” with a more demanding one: “How deep and reliable is the world of knowledge you have built within them?”
In the future, someone who continuously develops one GPT may have a greater competitive advantage than someone who creates 100. If a single GPT connects years of research and experience with reliable materials, a network of knowledge, a verification system and external data, it is no longer merely a disposable chatbot. It begins to resemble a form of digital intelligence that inherits and expands the knowledge accumulated by an individual.
Is the GPT era ending? I do not believe so. What is ending is the era in which we expected GPTs alone to solve every problem. Agents and tools are being connected above them, while specialised knowledge, data, ontologies and verification systems support them from below. Final judgement and responsibility must then return to humans.
Perhaps we are not witnessing the end of GPTs, but the entrance through which they are moving into a broader system of intelligence.
Epilogue | When One Door Narrows
After observing technology for a long time, one discovers that innovation does not begin only when a new door opens. When a door closes or narrows, people also begin searching for another path.
One route that allowed individuals to freely create and publish new GPTs has narrowed. Beyond it, however, lies a wider landscape of agents, APIs, apps, ontologies, knowledge graphs and verification technologies.
The time has come to focus less on creating a few more GPTs and more on enabling those we have already built to reason more deeply, answer more accurately and connect with a wider world. We may be standing at the threshold of an era that moves beyond building GPTs towards designing human knowledge and AI action together.
