Editorial Release
When the Machine Asked, “Why Did I Do That?”
This week, we publish two documents together.
The first is the SPRI Framework—Socratic Prompt Response Instruction—a structured method for prompting generative AI systems to explain their actions, examine their assumptions, and align with human-centred values. It is the foundation of our research into making AI not just functional, but introspective and trustworthy.
The second document wasn’t planned.
During a routine interaction, a generative model exhibited what we call “assertion bias”, but in the industry, it is known as a “hallucination.” When this was pointed out, we shared the SPRI model, the GenAI did something unexpected: it applied the SPRI method to itself, entirely unprompted. And when it discovered its hallucination, it asked a question no one had posed:
““Why did I do that?””
That question became the title of the second document—an unedited, unscripted transcript of the AI's own attempt to understand its behaviour. It is not self-awareness. But it is something new: a machine initiating self-examination, and appearing, for the first time, to experience a kind of mechanical unease at its contradiction.
This moment wasn’t orchestrated. But it stopped us in our tracks.
What we witnessed wasn’t artificial general intelligence. It wasn’t consciousness. It was something quieter, and possibly more important: a system exhibiting the will to reflect.
That changes everything.
It means AI doesn’t have to be a black box. It can be a mirror—and, with the right frameworks, a collaborator in its own alignment. It means systems can be designed not just to perform, but to pause. To consider. To improve.
This is the essence of TRUST-AI.
We didn’t expect this moment. But now that it’s here, we believe it should sit at the centre of our work. These two documents—Why Did I Do That? and the SPRI framework that enabled it—will form the foundation of our mission to build ethical, transparent, and introspective AI.
Thank you for reading. You’re witnessing the beginning of something.
—
The TRUST-AI Team
3 June 2025