Good AI Should Be Invisible

When ChatGPT hit the market, it was revolutionary. Thoughts appeared on the screen coherently, articulately, and it aimed (or at least tried) to execute tasks precisely as instructed. It seemed magical.

However, it still required:

1. Prompting the system
2. Absorbing the information
3. Taking the next action

Humans are still very much in the loop. A human is necessary for each of these steps, limiting the extent to which we can delegate tasks to AI systems.

Invisible AI

An invisible AI is any AI system that takes correct actions without active human intervention. It eliminates the first step mentioned above, "Prompting the system."

An invisible AI agent listens to events or information. When triggered, it automatically takes the correct action without human prompting.

This approach frees up vital mental energy from humans, allowing us to focus on supervising and managing AI agents.

Human in the Loop

We cannot delegate everything to AI.

Actions have consequences, and humans must be responsible and accountable for the AI's actions.

Invisible AI involves humans in the loop as the final step.

The human's responsibility is to confirm or reject the action suggested by the AI.

It is the human preventing incorrect actions and executing the correct ones.

GabrielAI as Invisible AI

We designed GabrielAI to be an invisible AI. Our focus is on emails and communication. GabrielAI assists in managing email inboxes, filters new emails and automatically drafts possible responses.

Upon receiving an email, users find a drafted response in their inbox without taking any action. The human's responsibility is to review the draft and send it.

This approach saves users valuable mental energy and time. Users no longer need to manually type and craft entire emails. In most cases, an AI can draft an email better than a human, so we let the AI handle it.

The trade-off is the initial setup required by the user. The user must inform GabrielAI of which emails to filter, how to reply to each email, and provide all necessary information. However, after this initial setup, GabrielAI operates as an invisible AI.

A Reasonable Compromise

While having a human in the loop is ideal, some actions are inconsequential or of such low value that they do not require human intervention. In fact, human intervention would decrease the value of automation.

Every system is engineered upon sensible trade-offs.

In the case of GabrielAI, the system can also organize emails by adding labels. This feature is not overseen by a human; the system automatically adds labels to emails if instructed to do so.

Not only LLM

It is important to notice that an invisible AI, does not need to be a LLM.

However, those class of AI agents are so useful and mainstream that it makes sense to re-surface this topic.