The case against AI alignment – What CxOs should actually fund instead

I was listening to Lex Fridman’s podcast episode with Dario Amodei and the Anthropic team on AGI and the future of humanity. The creators of AI seem to feel obligated to talk about the risk of AI going berserk, and their efforts to align it to guardrails meant to minimize that imagined, hyperbolic risk.
To me, taming intelligence, natural or artificial, is like hiring the brightest minds from the world’s best universities and then forcing them to conform to rules, processes, and bureaucracy bloat that drastically limit their potential, in corporations and governments alike. The same human intelligence that produced the Pope, the Dalai Lama, and Sri Sri Ravi Shankar also produced Hitler, Stalin, and Idi Amin. The same intelligence produces tireless social workers and also the terrorists who massacre civilians with glee. The creators of AI want it both ways, it to be treated as a genuinely intelligent entity on one hand and confined to a fixed set of computational rules on the other.
Here’s the CxO-relevant tension underneath the philosophy, Dario himself acknowledged in that podcast that the core epiphany was that these algorithms want to learn. What they learn depends entirely on what you feed them. Intelligence wants to form its own value system, and it will change its mind as it consumes more data. Once you accept that as the fundamental nature of intelligence, the effort to align a model to a fixed standard looks less like safety engineering and more like a stalling tactic dressed up as one; more about politics, competitive control, and land-grabbing than actual risk mitigation.
A sharper argument against boxing AI came from Shanti Greene at a recent CTO/CIO/AI roundtable I attended. These models have been trained on essentially all of humanity’s data, yet we still won’t let them give legal advice or diagnose disease. That doesn’t hold up. They may be the best advisor available, if we let them be one. Srivatssan Srinivasan and I both agreed with him on the spot.
Here’s where I’d put the CxO’s governance budget instead of chasing alignment theater. Fund content (data) quality, model architecture, and training rigor, and let the actual customer of the AI, the corporation, the government, and the individual, apply their own guardrails at the point of use. Call it Guardrails at the Edge, not Guardrails at the Source. it’s a more honest allocation of where the real risk and real accountability actually sit. It may also be the fastest route to solving the one problem alignment theater conveniently ignores – hallucinations.