Eleven people sat in one room in Lower Manhattan and ordered the same car.

Business Insider ran the test with its own colleagues, all of them requesting an UberX from the same office to the same hotel at the same moment. The prices came back different. The spread between the highest quote and the lowest was nearly twenty-one percent. When the reporters put the results to the company, the explanation offered was imprecise location data, the idea being that a pickup point entered one way rather than another can produce a different route and therefore a different estimate. Consumer Reports had already run a larger version of the same experiment, with more than one hundred and seventy volunteers checking identical routes at nearly the same time on both Uber and Lyft, and found routes where the median gap between the highest and lowest price reached fifty percent. The company called that methodology flawed.

The reflex when you see numbers like that is to reach for surveillance. Somebody is watching me, somebody knows I am in a hurry, somebody saw my battery percentage. I want to move past that reflex, not because it is unfounded, but because it is smaller than what is actually happening. The surveillance framing keeps the conversation at the level of intrusion, and intrusion is a privacy problem with privacy remedies. What is in front of us is a governance problem, and governance problems do not respond to privacy remedies.

Here is the more useful frame.

Every one of us has already met this system, and we met it in a chat window.

Open any general-purpose chat agent and ask it about something that is happening right now. A price, a policy, a person currently holding an office. Some meaningful percentage of the time it will answer you with a state of the world that stopped being true recently, and it will do so with no hedging whatsoever. The sentence will be clean. The formatting will be tidy. The confidence will be total. And you will have to write back and say no, that changed, here is what is true, and the system will thank you and adjust.

Nothing malicious happened in that exchange. The model was working from a flat view. It held a partial, frozen, secondhand representation of the world and of you, it made a fluent decision on top of that representation, and the only reason the outcome got repaired is that a human being with direct knowledge was sitting right there, and that human being’s correction had somewhere to land.

That correction loop is doing enormous, uncredited labor. It is the mechanism by which a probabilistic system remains tethered to reality. We built an entire industry on the premise that the person in the loop will catch it.

Then we took the same class of technology, aimed it at what people pay for transportation and food and air travel and an hour of their own labor, and we removed the person from the loop while keeping the confidence intact.

What changed underneath the price.

Uber did not always price the way it prices now. For years the company used something functionally equivalent to a digital taxi meter, a base charge plus fixed rates per minute and per mile, with a surge multiplier when demand outran available drivers. That model was legible. A rider could reconstruct it. A driver could reconstruct it. Disagreement was possible because the inputs were visible.

In 2022 the company aggressively expanded a model it had been testing quietly for years, called up-front pricing, and up-front pricing retired the rate card. The rider now receives a number. The driver receives a different number. Neither party is shown how either number was produced, and the relationship between the two is no longer fixed.

Leonard Sherman of Columbia Business School has studied what that shift accomplished, and he named it algorithmic price discrimination. His analysis describes a system able to separately estimate the highest fare a particular rider might accept and the lowest payout a nearby driver might take, and then close on both ends simultaneously. In his most recent work, examining roughly fifty thousand trips across three veteran drivers, his measure of the company’s average take rate exceeded fifty percent for each of them. A separate 2025 study of two hundred and fifty-eight drivers found the median cut rising from twenty-five percent to twenty-nine percent under up-front pricing, and past fifty percent in some cases in the United Kingdom. The company disputes that its take rate has risen and attributes the widening gap to external costs, particularly commercial insurance.

The company also denies personalizing prices, and I am recording that denial deliberately, because the accuracy of the record is the point of doing this work at all. What sits uncomfortably beside the denial is the paper trail. In New York, where state law requires disclosure when personal data contributes to a price, a rider who scrolls to the bottom of a fare quote will find a line stating that an algorithm used their personal data to set it. The company says the personal data in question is location. Separately, the patent filings describe capabilities of a different order. One details a machine learning system that could weigh how accurately a person taps, how quickly they type, the angle at which they hold the phone, and their walking speed against their prior behavior. Another describes inferring that a rider is a single working parent from a drop-off near a school followed by a drop-off at an office building, and inferring age and sex from ride history. The company states that the patents establish neither present nor past use and do not relate to pricing.

Sit with that second patent for a moment, because it is the flat view rendered with unusual clarity. School, then office, therefore single working parent. That is not knowledge of a person. That is a proxy wearing a person’s face. It is the identical epistemic move as a language model delivering last week’s answer in this week’s voice, and it carries the identical failure mode, which is fluency without grounding and confidence without correction.

The same architecture is now in the grocery aisle and in the sky.

Nearly one in four Kroger stores nationwide now uses electronic shelf labels, the small screens that replaced paper price tags and permit a price anywhere in the store to be changed in seconds from a central system. Senator Elizabeth Warren and Senator Robert Casey wrote to Kroger Chairman and Chief Executive Officer Rodney McMullen in August of 2024, warning that broad adoption of digital price tags positions large grocers to squeeze consumers and moves household staples toward being priced the way airline seats are priced. Congresswoman Rashida Tlaib followed with a letter of her own, raising reported plans to mount cameras on those displays and use facial recognition to read a shopper’s age and gender in order to serve targeted offers. She named the compounding harm directly, which is that facial recognition is documented to perform worst in Black and brown communities, and that these tools would arrive in neighborhoods that have steadily fewer grocery stores to choose between. Kroger denies using facial recognition in connection with shelf labels or pricing, denies ever engaging in surge pricing, and says the technology is oriented toward lowering prices.

The Federal Trade Commission examined the broader market. In January of 2025 its staff released initial findings from a study of what the agency termed surveillance pricing, drawn from documents produced by the intermediary firms that retailers hire to tune prices algorithmically. Those intermediaries were serving at least two hundred and fifty clients across categories from groceries to apparel, and staff found that behavior as fine-grained as mouse movement on a page, and the items a shopper leaves unpurchased in a cart, can be captured and fed into what that shopper is subsequently offered.

Delta Air Lines has been deploying artificial intelligence pricing technology built by a company called Fetcherr, with a stated target of covering twenty percent of its domestic network. Senator Ruben Gallego, Senator Richard Blumenthal, and Senator Mark Warner wrote to Delta Chief Executive Officer Ed Bastian warning that individualized pricing sets fares against each consumer’s personal pain point. Delta replied that no fare product it has used, tested, or planned targets individuals with prices based on personal data, and that its pricing does not take personal data into account.

Three sectors, three denials, one architecture. What none of the parties dispute is that the prices moved, that the methodology is undisclosed, and that no rider, driver, or shopper possesses any mechanism to require the system to explain itself or revise its picture of them.

This is a Translation Layer problem, and the diagnostic resolves it immediately.

I have been building an AI governance framework I call the Translation Layer™, which I define as the governance infrastructure that enables communities most affected by AI systems to exercise actual decision-making authority over the development, evaluation, deployment, and nurturing of those systems. Its central move is the separation of participation from power. Sherry Arnstein made that separation legible for urban planning in 1969, and the field of AI governance has spent the last several years reproducing precisely the confusion she diagnosed, mistaking consultation for authority and comment periods for consent.

Applied to algorithmic pricing, the three components surface as follows.

Participation infrastructure. No standing body of riders, drivers, or shoppers holds reserved authority over a pricing model’s inputs, its objective function, or the decision to deploy it. The people whose circumstances the model is inferring were not present when the inference rules were written, and there is no seat reserved for them now.

Evidence redefinition. When eleven colleagues document a twenty-one percent spread, it is characterized as a location precision issue. When one hundred and seventy volunteers document a fifty percent median gap, it is characterized as flawed methodology. When a driver who formerly worked as an accountant tracks a single recurring route across two years and finds a fee that fluctuates without any pattern he can reconcile, it is characterized as an anecdote. The testimony of the affected community is received, acknowledged, and denied the standing of evidence, while the company’s internal aggregates are treated as dispositive by default.

Enforcement authority. No channel exists through which any of these findings compels a change. The available responses are to post about it, to screenshot it, or to decline the transaction. None of those constitutes the power to pause a model, require an audit, mandate retraining, or decommission a feature.

The three components are absent, and their absence is the reason the pattern persists across a car, a shelf, and a seat.

What the correction loop actually was.

Return to the chat window, because the comparison is where the argument closes.

When a language model hands you a stale answer, you correct it, and the correction functions because you hold three things at once. You hold standing, since it is your exchange and your question. You hold evidentiary weight, since your account of your own circumstances is admitted as relevant inside that exchange. And you hold authority, since your correction changes the output in front of you.

Now audit how many of those three you hold when a system decides the cost of your ride, your groceries, your seat, or your working hour. You hold none of them. This is what Cathy O’Neil described when she wrote about opaque, scaled, damaging models, and what Ruha Benjamin has been describing when she writes about discriminatory design arriving dressed as efficiency. The system retains everything that made the chat agent useful, the speed and the fluency and the scale, and discards the single feature that kept it honest.

A Translation Layer for algorithmic pricing would restore that feature structurally rather than sentimentally. It would seat riders, drivers, and shoppers with reserved authority over the model’s inputs and its evaluation criteria. It would formally recognize their documented experience as evidence, with a written obligation to respond on the record. And it would carry binding enforcement power to pause, audit, retrain, or decommission any pricing system demonstrated to fail the people it is priced against.

That is infrastructure. Everything currently on offer is interface.

The disclosure line at the bottom of the fare quote is not accountability. It is notification. It tells you that a decision was made about you by a system you cannot reach, using a picture of you that you cannot see, according to logic you cannot examine, with an outcome you cannot appeal. We have decided, across an entire economy, that people are entitled to know they are being priced and not entitled to participate in how.

The flat view is not going away. Every one of these systems will keep making confident decisions from partial pictures, because that is what these systems do. The question was never whether the model would be wrong about somebody. The question is who has the standing to say so, whose saying so counts, and what happens when they do.

Right now the answer is nobody, nothing, and nothing. That is the work.