What If the Pattern Was Created by the Problem?
We keep calling artificial intelligence the future while feeding it the past.
Not only our knowledge. Our outcomes. Our records of who got funded, hired, promoted, treated, believed, protected, arrested, approved, denied, celebrated, ignored and considered successful.
Then we ask a machine to find the pattern.
What if the pattern was created by the problem?
That question is where this work begins for me, because historical data can be perfectly accurate and still be dangerous when we forget what produced it.
The data can be real. The interpretation can still be wrong.
Redlining is a brutal example. Human beings classified neighborhoods as financial risk. Capital moved accordingly. Investment changed. Homeownership changed. Property values changed. Decades later, those changed economic conditions existed as real data.
Federal Reserve Bank of Chicago research found lasting effects on homeownership, house values, rents and segregation, and estimated that redlining could explain roughly 40 to 50 percent of the house-value gap across affected neighborhoods from 1950 to 1980. Read the Federal Reserve Bank of Chicago research.
Now remove the history. Give only the outcome to a machine and ask it what represents a strong investment. The machine may identify the pattern correctly while understanding absolutely nothing about the system that helped create the pattern.
That is not a small distinction.
We have already seen the modern version.
Researchers studying a widely used health-care algorithm found that it used health-care spending as a proxy for health need. Spending, however, was shaped by unequal access to care. At the same risk score, Black patients were considerably sicker than White patients. Correcting the disparity would have increased the percentage of Black patients identified for additional help from 17.7 percent to 46.5 percent. Read the Science study indexed by PubMed.
The algorithm did not need a rule that said one group deserved less care. The inequality was already hiding inside the measurement.
A system does not need to hate anybody to reproduce the consequences of something humans built through discrimination. It only needs the outcome.
NIST already recognizes that AI bias is not limited to bad data or prejudiced developers. Its work identifies systemic, computational/statistical and human-cognitive forms of bias and explicitly places AI inside the institutions and social conditions surrounding it. Read NIST's standard on identifying and managing AI bias.
So what are we authorizing the past to do?
This is the question I think the AI conversation has not pushed hard enough.
If systemic conditions can create historical outcomes, historical outcomes cannot automatically be treated as neutral ground truth for future decisions.
History is evidence. It is not automatically instruction.
What happened matters. Why it happened matters too. A repeated outcome might describe human behavior. It might also describe access, exclusion, policy, geography, previous automation, institutional design or an old decision that kept reproducing itself.
Those are different realities. A machine that cannot distinguish them can be technically accurate while helping the old reality survive.
And we are scaling this fast.
Stanford's 2026 AI Index reports organizational AI adoption at 88 percent. It also reports 362 documented AI incidents in 2025, up from 233 in 2024, while responsible-AI benchmark reporting continues to lag capability benchmarking. Read the Stanford 2026 AI Index.
Capability is accelerating. Adoption is accelerating. The authority we give machine interpretation is accelerating. That makes the foundation matter more, not less.
Breaking the cycle requires new evidence, not prettier language.
I do not want AI to forget history. I want us to understand history well enough to stop confusing a historical pattern with a permanent rule.
That means tracing consequential data back to the conditions that produced it. It means separating description from prescription. It means interrogating proxies that measure something easier instead of the human reality we actually care about. It means measuring feedback loops when a model's decisions help create the next dataset. It means preserving meaningful human correction instead of treating it as noise.
And then it means doing the part human beings are notoriously good at discussing and terrible at sustaining: changing reality.
Create fair opportunity for outcomes the old system rarely allowed. Measure what happens. Keep the evidence. Let new evidence challenge inherited assumptions. If the system changes the conditions and a different outcome becomes possible, that new outcome belongs in what the future learns too.
Nature researchers have separately shown that generative models trained recursively on model-generated data can lose information about the original distribution, with less common parts disappearing first. That research is about model training rather than social inequality, and I will not pretend it proves more than it does. But it should make us extremely serious about what information survives repeated machine interpretation. Read the Nature research on model collapse.
The challenge
Before a consequential AI system uses a historical pattern to influence a future human outcome, we should be able to answer one question:
What created the pattern?
Not because every historical pattern is wrong. Not because every unequal outcome is proof of discrimination. And not because machines should be forced to produce whatever result humans prefer.
Because prediction is power. And if we do not understand what produced the evidence behind the prediction, we can accidentally turn yesterday's conditions into tomorrow's infrastructure.
AI does not escape human history merely because it processes it faster. Scale can turn inherited assumptions into infrastructure.
The future should learn from history.
It should not be sentenced to repeat it.
Comments