Current AI research has shifted away from the idea of a single "intelligence dial" toward making models legible, accountable, and steerable—and the unifying discovery is that these properties don't compose for free and often actively trade against each other. The sharpest move is splitting transparency into "variable transparency" (can I read the model's state?) versus "algorithmic transparency" (can I reconstruct its reasoning?), revealing that a system can be fully inspectable in one sense yet opaque in another. Privacy, calibration, and fairness all show the same pattern: differential privacy and predictability-based privacy are formally incomparable, individually calibrated experts can assemble into a miscalibrated whole, and zero subgroup bias may demand more data than will ever exist.

Topics: model transparency, calibration, differential privacy, fairness tradeoffs, interpretability