Beyond the Chatbot: Towards an Invisible yet Explainable AI?
It's happening. Artificial intelligence is infusing our tools, promising a revolution in the way we work, create, and think. Yet, a nagging question remains: is the current integration of AI living up to its promises?
The observation is often that of a "visible" AI, embodied by chatbots added to our interfaces. An interaction that, as Andrej Karpathy recently pointed out, sometimes recalls the era of command terminals: we dialogue with the machine. Is this really the pinnacle of human-AI interaction? Worse still, this AI, particularly powerful LLMs, often functions as a "black box." Its internal mechanisms elude us, and the "explanations" it provides are sometimes post-hoc rationalizations rather than a faithful reflection of its functioning. How can we build a fruitful collaboration on such a mystery?
I believe another path is possible, more subtle, more powerful: that of an AI "engine," deeply integrated, acting in the background. An AI that does not manifest itself through a dialogue window, but through a relevant anticipation of our needs, a discreet automation of tedious tasks, an aid in structuring our ideas directly within our workflows. An AI that would become a true extension of our cognitive abilities, a silent partner.
The challenge is significant: how to make this AI both "invisible" and "explainable"? If complete algorithmic transparency of the most complex models remains a research horizon (as evidenced by the work of teams like Anthropic), interface design (UX/UI) has a fundamental role to play. It is not necessarily about understanding every neural weight, but about giving the user the keys to:
- Understand the AI's action: Through clear and discreet signals indicating when and how the AI intervenes.
- Keep control: By having the ability to easily accept, refuse, modify, or cancel the AI's suggestions or actions.
- Perceive the intention: Via indications about the source of information or the logic behind a suggestion (e.g., "Based on X," "Related to Y").
In short, it is about creating a transparency of use that builds trust and keeps the user in the driver's seat, even if the internal mechanics of the engine are unknown to them.
The future of artificial intelligence will not be determined solely by the raw performance of the models. It will crucially depend on our ability to design interactions that truly augment humans, in a fluid and respectful manner of their autonomy. The field of exploration is immense and exciting.
And you, how do you imagine the ideal integration of AI into your daily tools?