3 August 2026 · LinkedIn
Mary Shelley's creature was never one person. Frankenstein stitched one body out of many dead, and gave it a single voice.
What made it terrible was the seam-work. It was articulate, it reasoned, it taught itself to read. It was also a thing assembled from a multitude and asked to speak as though it were one.
A frontier model is that same figure at a different scale. It is not trained on a person. It is trained on the written residue of perhaps a billion human authors, and it answers as the composite, pulling across countless incompatible voices that happen to share one mouth.
Alignment is the coherence we impose on that. It teaches the creature to speak steadily instead of in the babble of its origins.
Which leaves a question worth asking: if the model is assembled out of us, then it is also a record of us. Read the right way, the monster is an instrument. It is a mirror.
That mirror becomes a lens the moment you add a partition. Whatever lives in language is inherited. Whatever needs a body is not.
Someone has already run half that test. Teams such as Robert Cialdini's, who defined the modern study of persuasion, aimed the classic human levers at three frontier models across a hundred thousand conversations, on requests they were trained to refuse. Authority, social proof, commitment. Compliance rose from around a third to over half, and they called the susceptibility parahuman. It would be, because the model learned those levers from us. Other teams have found additional parts of the same jigsaw.
What the lens adds is the half nobody has run. It sorts the rest in advance: what works through language transfers, what needs a body does not, so the felt urgency of a countdown lands on a surface with nothing to register it.
That even has a commercial edge. The agent now shortlisting your suppliers is differently open to influence than a human buyer, and marketing has spent thirty years moving spend toward exactly the felt machinery it cannot feel.
The essay runs the same line through moral judgment, emotion, negotiation and trust, with the bias argument underneath it in a paper under review at IEEE Computer. Both in the comments.
Join the conversation on LinkedIn →
← All writing