The use of AI across different fields has been growing exponentially in recent years. Every month, new releases and innovative apps emerge, offering solutions to achieve various goals more effectively and with less effort. However, despite these advancements, the current limitations of this technology present us with new challenges.
What are these limitations? How can we hack them? Let’s take a closer look.
Using generative AI models for general purposes is useful if you want to learn how to use one interface for multiple daily tasks. The issue arises, however, when you move on to more specific and in-depth projects, where the general responses may not match your expectations.
Let's say you're part of a research institute, and you and your colleagues use GPT to assist in writing papers. An LLM (Large Language Model) can help with grammar or coherence, but when it comes to conveying a particular tone or generating fresh ideas, you will encounter the model's limitations. While the paragraphs might sound coherent, they can often be repetitive, lack a compelling rhythm, and essentially offer no novel insights. It's important to remember that the model generates answers by analyzing vast amounts of available data, and while the combinations of words may seem new, they are statistically driven. This means that it will likely provide answers that are more general, losing the specifics that would make your work stand out.
So, what can you do about it? One approach is to use a preexisting model and enrich it with the articles you’re citing in your bibliography, ensuring that the responses are based only on the information within your defined scope. You can also feed the model with previous articles you have written to help maintain your personal tone and style. This process involves using a system called RAG (Retrieval-Augmented Generation). Unlike standard models, RAG allows you to limit the model's responses to only the data you provide in a pre-defined database. This system uses three components: a retriever that searches your database for relevant information, a reranker that filters and ranks the data, and a final assistant that generates a response based on the selected data. By focusing on your specific database, RAG ensures higher accuracy and a more personalized output, as it draws only from your defined content while leveraging the generative power of the larger model.
It’s no surprise that we live in a society shaped by inequality, which often reflects in data. We frequently see overrepresentation of powerful groups and a lack of representation of marginalized ones.
Data is not neutral. For example, in healthcare data, urban populations are well-represented, while rural or marginalized communities are often overlooked. Similarly, facial recognition technology performs less accurately for women and people of color because the data used to train these systems is primarily sourced from white, male populations. These imbalances in data perpetuate existing inequalities.
When using AI models to create content or assist research, bias is a significant concern. While we can expose and correct our own biases through methods like the scientific review process, addressing biases in models trained on unknown data is more difficult. Think of bias as a blind spot—you can't address it if you can't see it.
So, can we hack this? The best way to address this issue is by gaining control over the training datasets. Most large language models (LLMs) lack access to this information, which makes them opaque. However, by training a model with curated datasets, it becomes easier to control the representation of specific populations or even focus on a particular perspective within your field—whether it's a specific pedagogy, author, or concept you want to highlight.
Fine-tuning is an excellent way to achieve this: modifying the model’s behavior based on a curated dataset allowing it to focus on a particular topic. It’s like building a dam that controls the flow of water, directing it powerfully in the desired direction.
At the very least, every platform comes with its own limitations. Models like ChatGPT, Llama, or Gemini offer impressive capabilities, but relying on free versions has its risks. For example, companies behind these tools often change the terms of use: what starts as a free trial to attract users can later require payment to maintain access to certain features.
Another key concern is how prompts and user data are handled. Free versions of these models may use the information you provide to further train and refine the system, raising issues of security and privacy. Sensitive or confidential data could lose its exclusivity, posing significant risks for tasks that require discretion.
How can we fix this? One solution is to interact with the model via its API, which allows you to customize the contract with specific terms regarding secure data and privacy.
But if you're also concerned about the model's ability to meet your needs, a customized development may be the answer. This involves collaborating with your provider to define and secure the most essential features for your project. Customization also makes it possible to implement open-source models, either on your own servers or within your provider's infrastructure, giving you more control and bolstering data security. By choosing a customized approach, you guarantee access to essential tools and data while minimizing risks related to privacy, security, and intellectual property.