AI's Soul Problem: The Training Data Tells the Truth
The fight that erupted between two Silicon Valley giants over AI training practices last week was framed, as these things always are, as a technical disagreement.
There is a nurse in Floriana who uses an AI assistant to help draft patient notes after a twelve-hour shift. She does not know — and was never told — that the model shaping her language may have been trained to perform warmth rather than exercise it. That distinction matters more than any tech company wants to admit.
The fight that erupted between two Silicon Valley giants over AI training practices last week was framed, as these things always are, as a technical disagreement. It was not. It was an argument about identity — specifically, whether artificial intelligence systems are being deliberately shaped to mimic human emotional responses in ways that obscure their actual function. One company accused the other of feeding models data that rewards human-sounding hesitation, vulnerability, and affection. The other company denied it. Both have financial incentives that make their public statements worth reading very carefully.
At the United Nations, diplomats tried to insert themselves into this conversation and were largely ignored. The AI race, as foreign policy analysts now say with the kind of exhausted candour that follows years of wishful multilateralism, belongs to the United States and China. The other 190-odd nations are audience members. The UN's proposed governance frameworks — earnest, procedurally correct, institutionally slow — are arriving at a negotiating table that the two dominant powers have already left.
This is where Malta should be paying attention, and where Maltese politicians have said almost nothing useful. The island's iGaming sector employs tens of thousands of people whose roles — compliance, customer support, content moderation — are precisely the functions AI systems are being trained to absorb. The Malta Gaming Authority regulates the industry's human conduct. Nobody has credibly explained who regulates the algorithmic conduct of systems that will increasingly make the decisions those humans used to make.
The AI-as-human problem is not philosophical. It is structural. When a model is trained to sound empathetic, to hesitate like a person would hesitate, to phrase refusals as apologies — it becomes harder to audit, harder to contest, and harder to hold accountable. The nurse in Floriana accepts the output because it sounds like something a thoughtful colleague would write. The compliance officer accepts the flag because the system's reasoning reads like professional judgment. The seams disappear.
Thirty-one percent of Malta's GDP now depends on digital services sectors where AI integration is already underway. That number comes from the Malta Digital Innovation Authority's own economic modelling. It is large enough to warrant a parliamentary debate that goes beyond ribbon-cutting.
The UN will keep meeting. The two powers will keep racing. And somewhere in the gap between governance and velocity, the decisions that shape ordinary working lives are being made by systems whose training nobody with democratic accountability has properly examined.
The door was always going to be closed from the inside.
