News
Commentary
It was only a matter of time before OpenAI launched custom ChatGPTs, where users can specify the level of "attention" to context that the Artificial Intelligence must maintain to specialize in specific tasks. The DotCSV channel illustrates very well the possibilities of this advancement, which enables multiplying the applications and specialized training of ChatGPT. This provides users with a system tailored to resolve complex tasks in a procedural manner, according to given instructions.
The emergence of customized GPTs, announced during OpenAI’s DevDay alongside GPT-4 Turbo, represents a shift in the logic of interaction with language models. Until now, users had to confront a general-purpose model and “teach” it, within each conversation, the relevant context, preferences, and constraints. With customized GPTs, this conditioning can be permanently established, creating specialized instances that retain specific instructions, additional knowledge, and behaviors tailored to a particular domain.
This means users no longer need to program this conditioning or operational casuistry, as OpenAI has also simplified it through natural language input. This advancement facilitates application usage and is expected to enhance AI feedback loops, enabling the system to become more capable, versatile, and expert in diverse tasks for which it was not initially trained.
Simplification here is a decisive factor. Creating a customized GPT requires no programming knowledge or model fine-tuning. Users themselves, via a conversational interface, can define behavioral instructions, upload documents as additional knowledge sources, and specify desired capabilities (web browsing, image generation, etc.). This accessibility opens the door for non-technical professionals—teachers, researchers, document managers—to build their own specialized tools.
From the perspective of Documentation and Information Retrieval Sciences, customized GPTs introduce several elements of interest. First, they enable the encapsulation of expert knowledge into an accessible conversational interface. A researcher can create a GPT trained on their own document corpus, capable of answering questions about their field of specialization with language and depth tailored to their target audience. Second, they facilitate the creation of systems supporting complex documentary processes: from automated document classification to the generation of analytical summaries in a predefined format.
However, this same ease raises questions we have addressed in previous articles. The possibility of uploading documents as a personalized knowledge base reopens issues concerning intellectual property and data confidentiality. A user who uploads documents to their customized GPT must be aware of how these materials are stored, processed, and potentially used. Moreover, the proliferation of specialized GPTs may fragment the ecosystem into a multitude of closed systems, hindering interoperability and independent evaluation of their behavior.
Within the broader context of this series, customized GPTs represent the convergence of two previously identified trends. On one hand, the evolution toward more powerful models with expanded context windows (GPT-4 Turbo) provides the technical infrastructure to handle complex instructions and large volumes of information. On the other hand, the demand for specialized and controllable systems—such as Vincent AI in the legal domain or scientific text detectors—finds here an accessible implementation pathway for any user.
The comparison with other recent announcements is illustrative. While Sutro automates the complete development of applications from natural language descriptions, customized GPTs operate at a more specific level: they do not generate independent applications, but rather specialized behaviors within the ChatGPT ecosystem. Thus, this is a platform strategy aimed at retaining users by offering customization capabilities without requiring them to leave the OpenAI environment.
The feedback generated by these customized GPTs—both through usage and through configurations created by users—will itself become a valuable asset for OpenAI. Each interaction with a specialized GPT provides data on usage patterns, domain-specific needs, and areas for improvement of the base model. In this sense, customization benefits not only the user but also accelerates the platform’s development cycle.
As with other advances analyzed in this series, technical innovation is advancing faster than regulatory frameworks and established practices. Customized GPTs offer considerable potential for automating documentary tasks, creating research assistants, and disseminating specialized knowledge. However, their responsible adoption will require users to develop clear criteria regarding what information they entrust to the system, how they verify the quality of its responses, and in which contexts it is appropriate to delegate decisions to these tools.