History With an API: When AI Industrializes the Past
Updated: Sep 8
If a system is trained on yesterday’s assumptions, rewarded by yesterday’s incentives, and constantly fed yesterday’s definitions of success, power, beauty, leadership, masculinity, femininity, wealth, authority, worth, and truth, then asking it to produce a better future is almost absurd.
It may produce newer language. It may produce faster decisions. It may even sound progressive as hell. But underneath, it can still be running the same old human code.
That is how history stops merely repeating and starts getting industrialized.
AI DOES NOT ESCAPE HUMAN HISTORY BECAUSE IT PROCESSES IT FASTER.
Scale changes the danger. A flawed belief that once shaped a family, company, institution, market, or generation can now be retrieved, repeated, recommended, ranked, scored, automated, and quietly presented as neutral. The interface can look futuristic while the definition underneath it is ancient.
The real question is not only whether a model can produce a biased output. The harder question is: what definition of reality did we encode before the model ever answered?
What assumptions produced the answer? Whose definition became the default? What history is being treated as evidence of what should happen next? What human value got translated into a metric and then mistaken for truth? Who disappears when the system optimizes for what already existed?
THIS IS NOT A HYPOTHETICAL PROBLEM.
NIST explicitly identifies systemic bias as bias arising from institutional procedures, practices, rules, and norms that advantage some groups and disadvantage others. NIST also warns that AI systems can increase the speed and scale of bias and perpetuate or amplify harm. That matters because the problem is not confined to a bad dataset. The institution, the metric, the workflow, the incentive, and the decision rule can carry history too.
NIST’s face-recognition testing found false-positive rates that, across demographic groups, often differed by factors of 10 to more than 100 depending on the algorithm. A system can therefore look technically impressive in aggregate while distributing its mistakes very differently across actual human beings.
A widely cited Science study exposed an even more dangerous mechanism: the proxy. A health-care algorithm used cost as a proxy for health need. Because unequal access and spending patterns meant less money was historically spent on Black patients, the algorithm could interpret lower cost as lower need. At the same risk score, Black patients were considerably sicker. Correcting the bias would have increased the share of Black patients receiving additional help from 17.7% to 46.5%. The math was doing what it was told. The definition was wrong.
Historical correlation is not automatically permission for future prescription.
WHEN HISTORY BECOMES TRAINING DATA, PROVENANCE BECOMES A GOVERNANCE ISSUE.
The Federal Reserve Bank of Chicago has documented lasting causal effects from twentieth-century redlining, including reduced homeownership and home values and increased segregation. That history matters to AI because modern datasets do not arrive from some sterile mathematical heaven. They are produced by institutions, policies, exclusions, incentives, access, and previous decisions.
If a model observes an outcome created partly by discrimination, exclusion, unequal access, or an earlier automated decision, then treats that outcome as clean evidence of preference, capability, risk, value, or deservedness, the past can become a self-validating instruction set.
That is where description and prescription have to be separated. ‘This happened frequently’ and ‘therefore this should guide what happens next’ are two completely different claims. Machines are extremely useful at detecting the first. Humans must remain accountable for authorizing the second.
FEEDBACK LOOPS MAKE THE PROBLEM WORSE.
Once automated systems influence hiring, lending, visibility, recommendations, health interventions, policing, pricing, education, or access, their outputs can help create the next dataset. The next model can then read those downstream outcomes as fresh evidence without recognizing that an earlier system helped manufacture them.
Separate research in Nature has shown another kind of recursive danger: models trained repeatedly on model-generated data can experience model collapse, progressively losing information about the original distribution. That research is not a study of social bias, but it reinforces a larger systems lesson: outputs fed back into inputs are not independent evidence. Recursion changes the thing being measured.
AGGREGATE ACCURACY CANNOT BE THE FINISH LINE.
A model can perform well overall while concentrating error, exclusion, or opportunity loss in particular groups. A business metric can improve while the underlying decision process becomes less defensible. A recommendation system can become more efficient while narrowing what the system considers normal, credible, beautiful, qualified, authoritative, safe, or worth showing.
If the only question is ‘Did the model predict the historical pattern accurately?’ we have already surrendered too much authority to the past.
WE NEED DELIBERATE NEW EVIDENCE.
Breaking a historical cycle does not mean mathematically forcing every population, person, company, or outcome to look identical. It means creating fair opportunity for genuinely new outcomes to exist, measuring those outcomes honestly, documenting the conditions that produced them, and allowing better evidence to challenge assumptions inherited from older systems.
History should inform. It should not become destiny disguised as data.
THE HUMAN PROBLEM UNDERNEATH THE MACHINE PROBLEM.
This applies to humans too. If trauma remains the foundation, healing gets filtered through survival. If domination remains the foundation, leadership becomes control. If scarcity remains the foundation, wealth becomes hoarding. If attention remains the foundation, identity becomes performance.
You cannot build a fundamentally different future from an unchanged foundation. Putting a faster processor on top of the foundation does not magically absolve it. Apparently silicon was not issued a moral reset button. Tragic.
SO WHAT IS AI ACTUALLY RESPONSIBLE FOR?
Not merely learning humanity’s patterns. Not merely predicting them more accurately. Not merely wrapping them in cleaner language.
The deeper responsibility is to make the assumptions visible enough that humans can examine which patterns deserve to survive us, which should remain historical evidence, which require correction, and which should never again be granted authority over another person’s future simply because they appeared often in the past.
If an AI system perfectly reproduces our existing definitions of success, authority, beauty, intelligence, leadership, value, and power, have we built intelligence, or have we built history with an API?
That is the standard I want us arguing about now. Not whether AI sounds modern. Whether the foundation it is scaling deserves the future we are giving it.
THE EVIDENCE BASE
NIST, Towards a Standard for Identifying and Managing Bias in Artificial Intelligence: https://nvlpubs.nist.gov/nistpubs/SpecialPublications/NIST.SP.1270.pdf
NIST, AI Risk Management Framework: https://www.nist.gov/itl/ai-risk-management-framework
NIST, Face Recognition Technology Evaluation, Demographic Effects: https://pages.nist.gov/frvt/html/frvt_demographics.html
Stanford HAI, 2026 AI Index, Responsible AI: https://hai.stanford.edu/ai-index/2026-ai-index-report/responsible-ai
Obermeyer et al., Science, Dissecting racial bias in an algorithm used to manage the health of populations: https://doi.org/10.1126/science.aax2342
Federal Reserve Bank of Chicago, long-run effects of redlining: https://www.chicagofed.org/research/content-areas/mobility/policy-brief-redlining
Shumailov et al., Nature, AI models collapse when trained on recursively generated data: https://www.nature.com/articles/s41586-024-07566-y
Heather Fricke | Frick-E Energy™
PUT THIS TO WORK
If humans, search systems, or AI can build the wrong version of you or your business, Frick-E Identity Engine™ builds the evidence-backed identity layer they should be able to retrieve, verify, and understand. WTFRICK-E™ functions inside that system as diagnostic intelligence, identifying identity drift, missing proof, conflicting signals, and interpretation gaps before they cost trust, consideration, or opportunity.
Comments