Generative AI: The Shaper Archetype and EU regulation

Take note that at the time of publishing this Insights article, the EU AI Act is published but not fully adopted yet. Please read the following webpage to get an outline of the key dates relevant to the implementation of the AI Act: AI Act Implementation: Timelines & Next steps | EU Artificial Intelligence Act. If you want to get notified about significant updates to the Act and its implementation, subscribe to the EU AI Act Newsletter.

We have created a series of articles focusing on one particularly popular type of AI, which is generative AI. The first article can be found here.

Generative AI

Generative AI, short for generative artificial intelligence, is a type of AI that can create new data, text, images, music, or even help write example computer code. In other words, generative AI can learn from existing content and use that knowledge to generate entirely new things. Examples of widely known generative AIs applications are ChatGPT and Midjourney. Organizations can optimize their use of generative AI by first defining a strategic approach tailored to their specific needs and requirements. This involves deciding whether to rent access to existing AI models, purchase pre-trained AI models with some customization options, or build their own AI models from scratch. There are three main archetypes when building generative AI: Takers (see first article), Shapers, and Creators (or Makers).

Considering this recent important event, we have created a series of articles focusing on one particularly popular type of AI, which is generative AI.

Shapers

Imagine you are looking for a new place to live, and you need to move in quickly. Chances are you will consider renting a furnished apartment. This is very convenient because you don't have to start furnishing from scratch. You can start living there immediately! The house has basic furniture, but it's not your home yet, so you decide to add personal items to your apartment. You decide to paint the walls and hang up family pictures, to make the apartment feel like your home.

This is an analogy for the Shaper archetype, where you shape pre-built AI technology to fit your needs. An example scenario is when companies customize existing AI models to integrate with their application or other use cases. This is unlike the Taker, where you don’t change the interior of your new apartment, and directly move in.

So, how do we translate this analogy to fit the technical view? Shaping an AI model is possible in many ways. Here, we highlight the three most common ways of technically shaping pre-built AI technology:

• Fine-tuning*: Adapting a pre-trained model to specific tasks. Achieved by continuing training on a smaller dataset, specific to the use case.
• Hyperparameter Optimization**: Adjusting fundamental model settings to customize the model for your specific application.
• Hypernetworks: Generating dynamic parameters for another neural network, enabling quick and affordable adaptation to new tasks.

To tailor an AI model to fit your needs, you can employ one or a mix of these methods.

* Also sometimes referred to as transfer learning.

** A hyperparameter is a configuration variable for training an AI model.

“Shaper” a generative AI tool: Cybersecurity considerations

When you use an existing generative AI model and shape it to fit your needs or the needs of your company, take note of the following security risks:

🔐 When you train the model yourself, actively seek out training data that is representative for your company. Because the generative AI model learns from the data that it’s fed, it will reflect those biases in its output, leading to discriminatory or unfair outcomes. Be extra careful with potential biases that are likely to affect the health or safety of people.

🔐 Beware of accidentally introducing new vulnerabilities. When shaping the generative AI model, vulnerabilities or weaknesses might be accidentally introduced. When you use training data from third parties or (unknown) sources, be careful not to accidentally ‘poison’ the model by using training data that was altered by hackers. These so called “model poisoning attacks” target AI models in the development or testing environments. Before you publish the AI model, have it thoroughly checked and (pen)tested for weaknesses and vulnerabilities.

🔐 Altering or shaping an existing AI model can implicate changes in liability and responsibility. According to the EU AI Act and the proposed Artificial Intelligence Liability Directive (AILD), if you make a substantial change to an AI model that alters its purpose or affects its compliance status, you could potentially be held liable. However, the question of liability in the context of AI is complex. It’s always recommended to consult with a legal expert before publishing a modified or shaped AI model.

Next up: Maker

This was the second article in a three-part series. What if you want to have full control and build your own AI model? In the last article, we will highlight the third generative AI type; the maker!