For years, Einstein believed the universe was static. It was a logical assumption. It fit the consensus. It felt right. It took Edwin Hubble’s 1929 discovery of the redshift to shatter that frame and prove the universe was expanding.
Einstein was not stupid. He was trapped in a plausible frame. If the sharpest mind in modern physics could get locked into a wrong model because it looked coherent, what does that say about the rest of us?
In business, we do not call it the Static Universe theory. We call it best practices or industry standard. We adopt ideas because they sound coherent, not because we have tested them against reality. This is epistemic debt — the accumulation of unverified assumptions in your decision-making. Like financial debt, it compounds quietly. You do not notice it until the market calls the loan and the operation collapses.
The Green Dashboard Lie (Ola, Mumbai)
I paid my epistemic debt in 2014.
I was overseeing Mumbai operations for Ola. The mandate was clear: aggressive growth. We were in a capital war with Uber. The dominant logic in the company, and across the industry, was simple: supply is the bottleneck, drivers care about money, therefore we must match or beat Uber’s incentives every 72 hours.
It was a logical deduction. We executed it without hesitation. My team spent our weeks adjusting incentive structures, shifting thresholds, and recalculating bonuses.
On paper, we were winning. My dashboard was green. Bookings were up. Supply hours were up. New cab enrolment was steady. Acceptance rates were stable. The logic held.
Then, late one Saturday, I booked an Ola without identifying myself. I asked the driver how it was going with Ola and Uber. He did not praise the high incentive payouts. He vented. The plans changed so often that he never understood what the current incentive was. Then he said the thing that destroyed my green dashboard: if he wanted to work for just four hours, he made nothing. The system only paid the people who drove all day. So on days he could not work full time, he just switched off the app.
I went back to the raw data. I stopped looking at averages and started looking at cohorts. He was right.
The Reality Check
Around 10% of our drivers were doing nearly 40% of the bookings. Those same 10% were absorbing approximately 70% of the total incentive payout.
We were not incentivising supply. We were overpaying a small group of full-time drivers who would have driven regardless. Meanwhile, the part-time fleet — the capacity we needed for peak hours — never showed up. They had quit because our logical incentive structure made it mathematically impossible for them to earn anything unless they drove all day.
The dashboard was green, but it never showed me the real story. We were buying growth inefficiently, and we were blind to it because the approach felt logical.
The Fix: Earnings Assurance
We did not tweak the numbers. We broke the logic.
We removed the complex trip-based targets that rewarded only the full-timers. We introduced a model called Earnings Assurance Guarantee. The logic was binary: during lean hours, if you drive, we guarantee a floor price of Rs 250 per ride. During peak hours, we guarantee Rs 350 per ride.
If a driver completed two peak rides and one lean ride, the guaranteed earning was Rs 950. If actual fares after commission came to Rs 800, we paid the Rs 150 difference. If they earned Rs 1,000 from fares, we paid nothing.
The impact was immediate. Peak supply jumped by roughly 20% because part-timers returned once the maths was safe. Peak hour bookings increased by approximately 15%. Total incentive payout dropped by nearly 10%. We grew while spending less.
Why? We stopped paying the full-time elite for work they were already doing, and we redirected a portion of that capital to the part-timers who actually needed the nudge to show up. We moved from a logical model built on trip targets to a truth model built on earnings assurance.
AI: The New Engine of Epistemic Debt
I tell this story because we are entering an era where sounding logical is about to become the default output for everything.
Generative AI is the ultimate green dashboard. It is trained on the internet. It is trained to be agreeable. It is trained to produce what researchers call plausible coherence. If you ask it for a marketing strategy, it will give you one that sounds right. It will use the correct terminology. It will structure the arguments logically. It will cite general principles that feel true. It will confirm your assumptions.
If I had used an AI assistant in 2014 and asked whether matching competitor incentives was a good approach for ride-sharing, it would have said yes, price parity and supply incentives are critical for market share. It would have validated the exact mistake that was breaking our supply chain.
AI is not optimised for truth. It is optimised for pattern matching. That is not a flaw — it is a capability. But if you use it to validate your thinking rather than to test it, you will build on sand while a confident machine voice reassures you that the foundation is solid. The real risk is not hallucination. It is validated assumptions.
The Epistemic Honesty Enforcer
I needed a system to kill the yes-man in my workflow. I do not need an AI that agrees with me. I need one that audits me.
I built a workflow called the Epistemic Honesty Enforcer. It is a prompt structure designed to strip away agreeable noise and force the system to classify its own claims.
Every sentence goes through a claim classification gate. Factual claims must be verifiable — if the source is not real, delete it. Causal claims — “X leads to Y” — must have a clear mechanism, or they get downgraded to speculation. Opinions are allowed, but must be labelled as judgment, not fact.
The hard stop rule: if the AI makes a claim it cannot verify, it is forbidden from writing it. It must either delete the sentence or rewrite it as a hypothesis. No shipping uncertain claims with a polite caveat. Either prove it, label it as a guess, or cut it.
Feel free to explore the workflow — plug it directly as context in a new chat or as a system prompt in a project.
Defence Against the Dark Arts
Intelligence does not protect you from wrong ideas. Einstein proved that. Data does not protect you from wrong ideas. My Ola dashboard proved that.
The only thing that protects you is a system that forces you to confront reality before you ship.
We are all going to use AI. The people who win will not be the ones who use it to write faster. They will be the ones who use it to think harder — who use it to hunt down their own epistemic debt before the market does it for them.
AI will help you scale the wrong idea with remarkable efficiency. Markets will not.