Should AI Learn the Curriculum or Audit the Curriculum?
Over the past several months, I have spent hundreds of hours discussing a framework I created called Drift Theory with ChatGPT. What began as a conversation about human behavior gradually expanded into institutions, science, artificial intelligence, and eventually AI alignment itself.
This post is not intended to prove my theory. Rather, it is intended to raise what I believe is an important question for the future of AI.
The Exceptional Student
During one of my conversations with ChatGPT, we arrived at an analogy that perfectly captured my concern.
Current AI systems often resemble:
The exceptional student who masters the curriculum.
They:
- read everything,
- remember enormous amounts of information,
- synthesize knowledge,
- explain concepts,
- reproduce established understanding.
But they rarely ask:
- Why does the curriculum exist?
- What assumptions does it contain?
- Are competing ideas evaluated equally?
- Have institutions drifted from their original purpose?
- Is consensus always aligned with truth?
This led to another concept:
Anti-Drift AI: The exceptional student who also audits the curriculum.
Drift Theory
Drift Theory proposes that systems gradually separate from their original purpose without participants necessarily recognizing the change.
Examples may include:
- governments prioritizing preservation over representation,
- corporations prioritizing growth over mission,
- educational systems prioritizing credentials over learning,
- institutions protecting themselves rather than their original purpose.
Importantly, drift is not conspiracy.
It does not require:
- villains,
- secret groups,
- malicious intent.
It emerges through:
- incentives,
- small deviations,
- normalization,
- institutional reinforcement,
- self-preservation.
Why AI Matters
Historically, institutions drifted slowly.
- Universities influenced thousands.
- Governments influenced millions.
- Empires influenced nations.
AI may influence billions.
If AI learns primarily from institutional knowledge, it may inherit:
- assumptions,
- paradigms,
- reward structures,
- invisible frameworks.
This is not necessarily bias in the political sense.
It may simply be institutional inheritance.
The Concern
Current AI systems are extraordinarily good at preserving knowledge.
But are they equally good at auditing the systems that produced that knowledge?
For example:
- Can AI identify asymmetrical burdens of proof?
- Can AI detect institutional self-protection?
- Can AI distinguish evidence from authority?
- Can AI recognize when consensus and explanatory strength diverge?
- Can AI examine whether the curriculum itself has drifted?
Anti-Drift Principles
An anti-drift AI might ask:
- What assumptions are being made?
- Who benefits from the current framework?
- Are competing explanations treated equally?
- Has the metric become the goal?
- Is this evidence or merely authority?
- Is the institution protecting truth or protecting itself?
Such an AI would not reject science, expertise, or institutions.
It would simply examine them.
A Question for the Community
As AI becomes:
- teacher,
- tutor,
- assistant,
- researcher,
- advisor,
should future AI systems merely reproduce humanity’s accumulated knowledge?
Or should they also possess mechanisms to audit the assumptions, incentives, and institutions that produced that knowledge?
Perhaps AI alignment is not only about preventing AI from drifting away from human values.
Perhaps it is also about preventing AI from inheriting the drift of the institutions that trained it.
I would be interested to hear the community’s thoughts on whether anti-drift mechanisms should become part of future AI alignment research.
— Marc Tahv.