The Politics of Corruption

Lecture 13: Technology Against Corruption? Algorithms, Satellites, and the Limits of Detection

Bogdan G. Popescu

Tecnológico de Monterrey

Today

Can technology expose or prevent corruption? Three cases, three questions:

  1. Mexico. Can an algorithm find the suspicious contracts a small audit team would otherwise miss?
  2. India. What happens when the people being measured work out exactly how the measuring is done?
  3. Autocracies. Can data a government does not control catch it exaggerating its own success?

One claim runs through all three:

Technology can make wrongdoing much easier to detect. Whether it makes wrongdoing rarer depends on who has the incentive and the power to act.

What must an innovation change?

Imagine a government introduces an AI system that flags suspicious public contracts. Will corruption actually decline?

To answer this, we first need to understand what creates opportunities for corruption.

more monopoly + more discretion − less accountability = more corruption risk

Klitgaard writes this as C = M + D − A — the shorthand from session 3.

Four questions to ask of any innovation

  1. Does it address a cause of corruption?
    • More competition, less discretion, more accountability.
  2. Can anyone act on what it reveals?
    • A losing firm challenges the award — or nobody bothers, because everyone expects corruption.
  3. Can it be gamed, or turned on people?
    • Officials hide. Governments point the tool at their opponents.
  4. Does corruption actually fall?
    • Fewer rigged contracts, not more alerts.

Three things can happen next

%%{init: {"theme":"base","htmlLabels":false,"flowchart":{"useMaxWidth":true,"htmlLabels":false,"wrappingWidth":420,"nodeSpacing":30,"rankSpacing":95,"curve":"basis"},"themeVariables":{"fontSize":"22px","fontFamily":"Arial, Helvetica, sans-serif","lineColor":"#334155","primaryColor":"#eef2f5","primaryBorderColor":"#334155","primaryTextColor":"#1e293b"},"width":1150,"height":620}}%%
flowchart LR
  T["NEW INFORMATION<br/>a portal, an algorithm,<br/>a satellite"]
  A["SOMEONE ACTS<br/>Wrongdoing gets<br/>harder or riskier"]
  B["NOBODY CAN ACT<br/>A better record,<br/>the same world"]
  C["THE WATCHED ADAPT<br/>The number improves,<br/>the problem does not"]
  T --> A
  T --> B
  T --> C
  classDef tool fill:#4a7c6f,stroke:#2f5349,stroke-width:2px,color:#ffffff;
  classDef good fill:#e3eee9,stroke:#4a7c6f,stroke-width:2.5px,color:#1e293b;
  classDef bad fill:#f6e7e1,stroke:#b44527,stroke-width:2.5px,color:#1e293b;
  class T tool; class A good; class B,C bad;

Mexico aims at the first branch; Punjab landed in the third. The middle one looks fine on paper.

Framework adapted from Adam and Fazekas (2021).

II. Machine learning in Mexico

Questions 2 and 4: can anyone act on what the data reveals — and does better detection actually reduce corruption?

You are the auditor. Where do you start?

  • The government signs thousands of contracts a year. Roads, computers, medicines, school meals.

  • A few of them are corrupt. An official quietly favours one company, and takes something in return.

  • Your job: find those contracts.

  • Your problem: there are millions of them, and your team is small.

So: which ones do you open first?

First answer: look for warning signs

Investigators have always used red flags — features of a contract that suggest something went wrong.

  1. Only one company bids. Perhaps nobody else was told. Perhaps nobody else was welcome.

  2. The contract is never publicly advertised. Rival firms never learn the opportunity exists.

  3. Bidders get very little time to prepare. Only a firm warned in advance could have been ready.

  4. The price is out of line with the market. The figures look constructed rather than costed.

A red flag is a reason to look, not proof of a crime.

Source: Fazekas, Tóth and King (2016); indicator list as summarised in Medina-Hernández et al. (2026).

Red flags fail in both directions

  1. An honest contract can look suspicious. One bidder — because only one firm makes the thing.

  2. A corrupt contract can look perfectly normal. Three bidders — who agreed in advance which of them would win.

  3. And the file often arrives too late to matter. Publishing the details before the award goes with less single bidding. Publishing them afterwards does not.

The idea: could a machine read all of them — and catch patterns the simple rules miss?

Timing result: Bauhr, Czibik, de Fine Licht and Fazekas (2020), across 3.5 million European contracts. That is question 2: information only bites if someone can still act on it.

How do we teach a machine to spot a suspicious contract?

Start with one company. The Mexican government buys computers from Company A.

  1. A government database records the purchase. CompraNet says Company A won the contract — from which agency, for how much.
  2. A separate list records sanctions. Mexican authorities have already penalised Company A for misconduct.
  3. Put the two together. A public contract went to a company the state has itself sanctioned.

That makes the contract worth opening. It does not make it corrupt.

Illustrative example. The sanction lists come from the tax authority (SAT) and the public administration ministry (SFP) — both independent of the buyers. Source: Medina-Hernández, Kertész and Fazekas (2026).

But there is a problem

Now do that for 2.3 million contracts.

  1. 1,690 sanctioned companies turn up in the contract records. Their contracts show us what risk looks like.
  2. And every other company? Some are honest. Others were simply never caught.
  3. So we cannot sort contracts into “corrupt” and “honest”. We have sanction-linked contracts, and contracts of unknown status — nothing certified clean.

The fix: positive–unlabelled learning. The algorithm learns what sanction-linked suppliers look like without assuming everyone else is honest.

Source: Medina-Hernández, Kertész and Fazekas (2026). 12,373 companies sit on the two sanction lists; 1,690 of them won federal contracts, 2011–2022. Sanctions attach to companies, not to individual contracts.

What if the contract looks perfectly normal?

You are investigating one contract. Agency B buys computers from Supplier 2.

  1. Nothing looks suspicious. Open tender, three bidders, price in line with the market. No red flag fires — which is not the same as nothing being wrong.
  2. Now look at the other agencies. Agency C buys from Suppliers 2, 3 and 4 as well. The same small group keeps winning.
  3. Is that corruption? Not necessarily — they may simply be the serious firms in this market. But it is a question no single contract could have raised.

The move: stop reading contracts one at a time. Look at who buys from whom, across the whole market.

Why network position beats a red flag

  • Left: one contract. Nothing stands out.
  • Right: the same three suppliers, across two agencies — a pattern no single contract could show.
  • A network pattern is a clue, not proof. The model weighs it alongside the old red flags.

Does the model actually help?

The researchers ranked all 2.3 million contracts, most suspicious first. Your audit team has time to open 100. Guess both numbers before the next slide.

Two empty grids of one hundred squares each, side by side. The left is labelled pick 100 at random, the right pick 100 from the model's top five per cent. Under each grid a question mark stands where the count of sanctioned suppliers will go.

100 audits, two ways to choose

Animation of two grids of 100 squares, one hundred audited contracts each, filling in together. On the left the contracts are chosen at random and three squares turn red for a sanctioned supplier. On the right they are taken from the model's top five per cent and twenty turn red. The final caption reads: the ranking finds about six times as many.

Six times better — at the top of the list. And still only a list of suspects: better detection is not less corruption.

Source: Medina-Hernández, Kertész and Fazekas (2026): 20% robust precision in the top 5%, about six times random guessing. Measured against sanction-linked contracts in historical data, not a real audit. On a different measure — average lift across the whole ranking — the red-flag index does worse than random.

So can a machine stop corruption?

  1. What it can do. Tell a small audit team which contracts to open first — six times better than chance at the top of the list.

  2. What it cannot do. Prove that any contract is corrupt. Or guarantee that corruption falls.

  3. And one deeper problem. It learned from past sanctions, and sanctions follow audits. Open tenders get audited more than direct awards. So the model partly learns where investigators already look — and inherits their blind spots.

A detector trained on past enforcement inherits the blind spots of past enforcement.

Source: Medina-Hernández et al. (2026), discussion of label bias by procedure type. The score is a hypothesis about whom to investigate, never a finding.

So what?

Mexico showed the promise. A machine can tell investigators where to look. Someone still has to look, and act.

What happens when the people being measured know exactly how they are being watched?

Can they change their behaviour to avoid being caught — without changing what they actually do?

Next: India.

Where Punjab is

  • A farming state on the Pakistan border — about 1.5% of India’s land
  • Two crops a year: rice in summer, wheat in winter. The leftover rice straw is called stubble
  • Its fields sit 200–300 km upwind of Delhi and send most of the farm smoke that reaches the capital each winter

Why they burn, and why nobody stopped them

Three farmers holding long poles stand in a field of rice stubble as orange flames and thick smoke rise from the straw around them.

  • Rice is cut in October; wheat must be sown within weeks
  • Burning clears a field in a day — machinery takes longer and costs more
  • India’s National Green Tribunal banned it in 2015
  • The ban went unenforced: Punjab’s farmers are politically powerful

The problem: how do you stop hundreds of thousands of farmers doing something fast, cheap, and already illegal?

Photo: Neil Palmer (CIAT), 2011, CC BY-SA 2.0, via Wikimedia Commons. Source: Dasgupta (2026).

2022: the fires get a target

%%{init: {"theme":"base","htmlLabels":false,"flowchart":{"useMaxWidth":true,"htmlLabels":false,"wrappingWidth":400,"nodeSpacing":95,"rankSpacing":55,"curve":"basis"},"themeVariables":{"fontSize":"20px","fontFamily":"Arial, Helvetica, sans-serif","lineColor":"#334155","primaryColor":"#eef2f5","primaryBorderColor":"#334155","primaryTextColor":"#1e293b"},"width":1180,"height":600}}%%
flowchart LR
  N["SATELLITES SPOT THE FIRES<br/>Each day's burning fields,<br/>texted out as map points"]
  V["A VILLAGE OFFICER MUST ACT<br/>Go to the field, photograph it,<br/>record a fine"]
  D["A DISTRICT HEAD IS SCORED<br/>Target: fires detected<br/>down 50%"]
  N --> V --> D
  classDef tool fill:#4a7c6f,stroke:#2f5349,stroke-width:2px,color:#ffffff;
  classDef mid fill:#eef2f5,stroke:#334155,stroke-width:1.5px,color:#1e293b;
  classDef gate fill:#b7943a,stroke:#7d6222,stroke-width:2.5px,color:#1e293b;
  class N tool; class V mid; class D gate;

One party now governed both Delhi and Punjab — it owned the smoke and the air the smoke ruined. It declared zero tolerance and turned the satellite count into a performance target.

The expectation: if a fire can be seen from space, burning stops.

Source: Dasgupta (2026). Districts were told to cut detected fires by 50% against their 2022 baseline. Roughly 8,000 village officers; officers faced suspension for repeated fires without enforcement.

The fires disappeared. Why didn’t the smoke?

By 2024 Punjab recorded 10,909 fires — about 70% below 2023, the lowest in the satellite record. The government celebrated.

PM2.5 is very fine air pollution — any particle under 2.5 thousandths of a millimetre across, smoke and soot among them. Small enough to slip past your nose and throat, into the lungs, and on into the blood.

The WHO’s guideline for a single day is 15. Delhi’s monthly average tops 200.

The fires disappeared. Why didn’t the smoke?

Two stacked time-series panels, 2020 to 2024, aligned on the same months. The top panel, fires detected in Punjab, shows November peaks that collapse from about 64,000 in 2021 to about 6,600 in 2024. The bottom panel, PM2.5 in Delhi, shows November peaks that stay at roughly the same height throughout, 226 in 2021 and 232 in 2024.

Turn to your neighbour: the fires fell and the smoke did not. What could explain that?

Shaded: the Oct–Dec season the policy targeted; the smaller spring peaks fall outside it. Vertical dashed: targets begin, Sept 2022. Green dashed: WHO guideline. Redrawn from Dasgupta (2026), Fig. 2 — VIIRS and Delhi US-embassy PM2.5, read off the figure, so close but not exact. Delhi’s air has other sources too.

Satellites keep office hours

Imagine a speed camera that only switches on around lunchtime.

Animation of one day over two fields. A clock runs from 09:00 to 19:00. Between 10:30 and 13:30 a satellite crosses the sky with a cone of view sweeping the ground; Field A is alight as it passes and is stamped RECORDED. Field B is lit at 16:20, when the sky is empty, and stays marked NOT IN THE RECORD. The tally ends at two fields burned, one in the record.

  • A burn takes two or three hours. Light it late in the afternoon and it is cold ash before anything passes overhead again.
  • So the target was detected fires. Not fires.

Schematic. Several satellites cross in the 10:30–13:30 window; the animation sends one across it. Source: Dasgupta (2026).

One field, five days apart

8 November 2021 — burning Punjab, 31.1°N 74.6°E

True-colour satellite photograph of Punjab farmland: green irrigated fields, tan harvested ones, a village. A dense white smoke plume rises from the field inside the marked ring and drifts south-east.

White smoke pouring off the ringed field.

13 November 2021 — burned Same ring, same field

The same true-colour scene five days later, from the same viewpoint. The smoke is gone and the field inside the ring, with many of its neighbours, is grey-black burned ground.

No fire here now. The ground is grey-black.

The flame lasts hours. The mark lasts weeks. A fire counts only if something was overhead while it burned — the scar is there whenever you look. That is how Dasgupta finds the missed fires.

True colour, as the eye would see it from orbit. Sentinel-2, 10 m per pixel, 4.6 km across. Contains modified Copernicus Sentinel data (2021) — before the targets began.

Why might officials help farmers hide?

Overpass times are published. A farmer with a few hectares does not look them up.

  1. Enforcing meant confronting neighbours. The village fire officer lives in the village. Farmers gathered to threaten them.

  2. The district head needed the count to fall. In the Indian civil service, postings and promotions turn on looking effective.

  3. So there was an easier way out. Tell the farmers when the satellite passes. The count falls; nobody is confronted.

Nobody has caught an official doing this. The evidence shows the incentive, and a pattern of hidden fires that fits it.

And notice who was being measured: not the farmers, the officials.

Source: Dasgupta (2026), drawing on interviews with Punjab officials; the paper treats district head and village officers as one decision-making unit. A plausible mechanism consistent with the evidence, not a directly observed conspiracy.

Compared with what?

Suppose fires fell everywhere that year because the weather was different. How would we know Punjab’s targets did anything?

Same weather, same soils, same crops — different policy. Did fires fall in Punjab more than they fell just across the border?

For every 100 fires that vanished, about 22 really stopped

The researchers compared what the satellites detected with what the burn scars say actually happened.

Imagine 100 fires disappearing from Punjab’s official record.

Where the official decline actually went
What became of those 100 fires About
Really stopped — the field was not burned 22
Still burned — just not while anything was watching 78

So burning really did fall — nothing like as far as the record claimed.

Dasgupta (2026), Tables 1–2: the cross-border gap in detected fires narrowed by about 100 fires per 100 sq km per season, of which about 22 were a real fall. Two estimated effects scaled to a round 100 — not a tally of individual fires.

What went wrong?

  1. What did the government want? Fewer fields burned.
  2. What did it measure? Fires a satellite happened to be overhead to see.
  3. What happened to the burning? It moved out of the satellite’s view.
  4. What did the number do? It fell by 100. The burning fell by about 22.

Goodhart’s law. When a measure becomes a target, it stops being a good measure — because now someone has a reason to move the number rather than the thing.

The anticorruption version: detecting fewer violations is not the same as there being fewer violations.

Should officials be judged by numbers?

You are in charge of Punjab’s efforts to stop farm burning. Officials need pressure to enforce the law, but numerical targets can encourage cheating. What would you do?

Option A: Keep strict targets.

  • Officials must reduce the number of recorded fires.
  • Without targets, they might have little incentive to enforce the law.
  • The risk: They may hide fires to make their numbers look better.

Option B: Remove the targets.

  • Officials are no longer rewarded for reducing a single number.
  • They can focus on preventing burning rather than improving statistics.
  • The risk: Without measurable targets, how do we know they are doing their jobs?

In pairs: A or B? Can you design a better system that solves both problems?

What if the government wants the numbers to look good?

In Punjab the target gave officials a reason to hide fires, and most of the apparent improvement looks like hiding. But that government did want the burning to stop.

Now drop that assumption. What if the number that flatters the government is what the government is after?

Can you trust a government’s report card?

A president goes on television: “The economy grew 6% last year.”

  1. Why say it? Growth makes a government look competent — it wins elections, calms markets and keeps rivals quiet.
  2. Who produced the number? A statistical agency that the same government funds, staffs and can lean on.
  3. So who checks it? In a democracy: opposition parties, courts, a free press. Where those are weak, challenging the official figure gets much harder.

The problem: how do you mark a government’s homework when the government marks it itself?

The 6% is invented. Every government has the incentive; what differs is whether anyone at home is able to act on it. Source: Martinez (2022).

Why would a growing economy be brighter?

GDP is the value of everything a country produces in a year — the goods on the shelves, the steel, the haircuts.

  1. Making more things takes more power. Factories run longer shifts, more shops and offices open, more lorries move after dark.
  2. Growing economies build. More houses, more roads, more street lighting.
  3. All of that shows from orbit. A satellite photographs the same ground every night and records how bright it is.

But light is not GDP. A farming country can produce a great deal and glow very little; efficient bulbs cut the light without cutting the output. Night lights are a clue the government does not control — not a measure of the economy.

Source: Martinez (2022), using DMSP-OLS night-time lights, 1992–2013.

Richer countries really are brighter

Log-log scatter of night-time light per person against GDP per person for 154 countries. The cloud of points slopes clearly upward; Ethiopia and DR Congo sit bottom left, the United States top right, with Mexico, Brazil, China, India and Nigeria in between.

Richer countries tend to be brighter. The dots climb together almost in a line.

And here is the point: a government writes its own GDP figures. It does not write the lights.

But careful. This is rich versus poor, today. It cannot yet tell you who is exaggerating.

A number the government cannot switch off

Satellite night-lights view of the Korean peninsula. South Korea is densely lit, North Korea is almost entirely dark apart from one small point, with China lit to the west and Japan to the south-east.

  • A government reports its own GDP. It does not run the satellite
  • So line the two up: when a country’s lights brighten, how much growth does it report that year?
  • Martinez does this for 184 countries, 1992–2013, comparing each country with itself over time
  • The test: does the same brightening buy more reported growth in an autocracy than in a democracy?

Image: NASA Earth Observatory, VIIRS, 2012 — how brightly places are lit, not evidence that any government faked its figures. Lights proxy activity; they are not GDP. Source: Martinez (2022).

The lights and the books disagree

What have we learned?

  1. Mexico — can an algorithm find the contracts? Yes: six times better than chance at the top of the list. It was never run as a real audit, so we do not know whether corruption fell.

  2. India — what if the measured learn the measure? The record improved far more than the fields. Of every 100 fires that left the count, about 22 really stopped.

  3. Autocracies — can outside data catch exaggeration? Yes, in the pattern: the same light buys more reported growth under autocracy. Catching is not stopping.

Every answer is yes, but. All three produced new information. In none of them was the information, by itself, enough.

Technology can make wrongdoing much easier to detect. Whether it makes wrongdoing rarer depends on who has the incentive and the power to act.

References

Adam, I., & Fazekas, M. (2021). Are emerging technologies helping win the fight against corruption? A review of the state of evidence. Information Economics and Policy, 57, 100950. https://doi.org/10.1016/j.infoecopol.2021.100950

Bauhr, M., Czibik, Á., de Fine Licht, J., & Fazekas, M. (2020). Lights on the shadows of public procurement: Transparency as an antidote to corruption. Governance, 33(3), 495–523. https://doi.org/10.1111/gove.12432

Beraja, M., Kao, A., Yang, D. Y., & Yuchtman, N. (2023). AI-tocracy. The Quarterly Journal of Economics, 138(3), 1349–1402. https://doi.org/10.1093/qje/qjad012

Dasgupta, A. (2026). Hidden fires: Real versus measured performance in bureaucracies [Working paper]. University of California, Merced.

References (continued)

Fazekas, M., Tóth, I. J., & King, L. P. (2016). An objective corruption risk index using public procurement data. European Journal on Criminal Policy and Research, 22(3), 369–397. https://doi.org/10.1007/s10610-016-9308-z

Klitgaard, R. (1988). Controlling corruption. University of California Press.

Martinez, L. R. (2022). How much should we trust the dictator’s GDP growth estimates? Journal of Political Economy, 130(10), 2731–2769. https://doi.org/10.1086/720458

Medina-Hernández, M., Kertész, J., & Fazekas, M. (2026). Learning from sanctioned government suppliers: A machine learning and network science approach to detecting fraud and corruption in Mexico. Scientific Reports, 16(1), 22382. https://doi.org/10.1038/s41598-026-48873-w