Government should harness the power of AI to combat corruption while preserving managerial discretion.
In 2010, we found our roles and responsibilities for the City of New York converging as we sought to remedy the impact of one of the largest public corruption scandals in the city’s history. Both of us served previously as prosecutors; one of us became commissioner of the New York City Department of Investigation (DOI), and the other deputy mayor of operations in the Bloomberg Administration. During that time, DOI and the U.S. Attorney’s Office for the Southern District of New York exposed vendors that, while building a complex payroll modernization system called CityTime, also set up shell companies to bilk the City for fictitious work. The criminal investigation resulted in a record-breaking $500 million in restitution for taxpayers.
Exposing and tracking down this complex crime presented resource challenges even for DOI, the nation’s preeminent municipal investigation group with over 300 investigators, auditors, attorneys and support staff. The painstaking work necessary to identify potential public corruption across departments, contracts and programs exceeds resource capabilities in most, if not all, municipal jurisdictions.
As municipal governments experience new corruption scandals, the reaction can often be to pile on new rules and regulations atop what is already an intimidating bureaucracy. Instead, we suggest comprehensively applying increasingly capable artificial intelligence to deter and root out corruption. Unlike rigid rules, technology can more flexibly identify malfeasance while enhancing administrative efficiency and preserving managerial discretion. In addition to the established benefits of machine learning, which relies on structured data and predefined rules, generative AI can process both structured and unstructured data, analyze complex relationships and identify hidden fraud patterns through self-supervised learning — though this level of automation requires careful oversight. Additionally, AI’s natural language capabilities enable investigators to query data conversationally, generating insights without requiring extensive data expertise.
First, by constantly monitoring vast amounts of data, generative AI expands integrity oversight methodologies from the after-the-fact approaches that can occur too late in the lifespan of illegal conduct, typically depending on a tip and subsequent extraordinary amounts of forensic auditing. One can imagine all the red flags that AI might have generated in real-time during the CityTime project, for example, by analyzing unusual payment activity and transactions, alerting the City and DOI to patterns of cost overruns and indicators of shell companies.
Second, today there’s a fundamental tension in government systems between increased red tape and bureaucracy as a hopeful guard against corruption — but at the price of far slower and less responsive service. AI could alter that tension by making it difficult for bad actors to be bad without restricting the good actors’ flexibility — as one of us, Stephen Goldsmith, laid out along with Juncheng “Tony” Yang in “Expanding Discretion and Accountability in the Context of AI.”
Third, the ability of AI to review data will change the way we investigate public and private corruption in several fundamental ways. AI can objectively and broadly conduct preliminary “investigations” by reviewing structured and unstructured data. For example, it can review videos and pictures to see if work allegedly completed was done, analyze accounts payable, review 311 calls and dispatches and identify social media complaints about a particular contractor or regulated company.
Fourth, AI can greatly expand investigators’ reach and capacity by helping them analyze and visualize data through intuitive natural language searches. This capacity can simplify complex investigations by generating easier-to-understand explanations and associations from intricate datasets. By assisting with modeling, data visualizations and even coding, AI can speed up iterative analyses and discoveries.
We need an effort across cities to equip inspectors general, compliance monitors and internal auditors with the power and capacity of AI to make corruption easier to spot, and sooner, while acting as a force multiplier for those tasked with municipal oversight. The degree to which this new capacity can stem tax money from sluicing out the door to corruption and even detect and expose hazardous conditions makes the potential return on investment in AI exponential. Given estimates that corruption and fraud impose losses of 5% on municipal and state budgets, and given the incalculable costs of overly rigid bureaucracy, more precise integrity tools promise taxpayers tens of billions of dollars in savings. Below, we suggest ways AI trained on agency or project data could flag areas worthy of further review while increasing the oversight capacity of investigators and managers.
Modernizing contract oversight
The CityTime case helps us recognize the many ways AI could assist in reviewing contract terms in real-time, ensuring compliance and identifying risks. Internal auditors should utilize AI to continuously review open contracts for suspicious flags relating to the use of subcontractors and their bank accounts, prompting further review for previous investigations or disbarments.
Automation could examine invoices and compare them to the contract specifications. AI could identify ambiguous or excessively broad clauses allowing cost overruns or fraudulent activities. Regarding contract administration, AI could create lists of service-level agreements and then examine City data to verify compliance. Such reviews should flag material cost overruns, identify unusual activity and monitor whether Minority and Women-Owned Business Enterprise goals have been met with properly certified companies. These steps will enhance, not replace, the importance and reach of human investigators.
Issuing early warnings
Oversight should not simply equal more bureaucracy, and AI can afford cities speed and flexibility without surrendering integrity. In 2024, the U.S. Attorney in the Southern District of New York charged the largest-ever number of individuals ever accused in a single agency, the New York City Housing Authority. The fraud, also investigated by DOI, involved purchase orders for repairs below $10,000, a threshold that permitted public workers to make purchases through an expedited process.
That is a perfect job for AI — to comb for patterns and transaction volume, including which employees were executing and which contractors were receiving those contracts. For low-income tenants waiting for hot water in their apartments, speed is essential, with this case presenting a good example of the need to balance discretion at the worker level with more advanced AI oversight that looks for suspicious patterns and verifies purchases and close-outs of completed recorded work. AI should continuously review work orders, resident feedback reports and other documents to look for problematic indicators.
Identifying potential risk areas
AI can provide predictive risk management by analyzing historical data and current trends to forecast fraud, inefficiency or regulatory noncompliance. Pattern recognition is a key to targeted investigation and provides yet another critical machine learning opportunity.
Think of public corruption as equivalent to burglary, where the police department uses AI to examine all burglary reports and crime scene photos, creating a catalog of characteristics that highlight for detectives those suspects most likely connected to the case they are investigating. AI can jump-start human review of events by identifying factors associated with illegal conduct, saving vast investigative resources and producing faster, more accurate results.
When the New York City Finance Department needed more audit capacity, it supplemented staff with an investment in analytics. Using additional and better data analytics, the Department did a more expeditious job of identifying cases to be assigned to field investigators. Now, applying AI — with ethics and privacy guardrails — the Department could review massive amounts of information, reduce the number of honest taxpayers selected for exams and assist officials in choosing the red-flagged and most suspicious individuals or companies for audits.
AI can identify those cheating on their municipal obligations and those officials corruptly thwarting laws and regulations. One previous New York City case involved expediters representing commercial property owners in negotiations with Department of Finance tax assessors. The tax assessors would take bribes from the expediters in exchange for improperly lowering the assessments for those buildings, draining the City’s coffers of an enormous amount of revenue. After decades of this illegal activity by a tightly knit group of co-conspirators, the U.S. Attorney’s Office eventually cracked open the wrongdoing leading to the prosecution. Today’s AI tools would have greatly facilitated this investigation, aiding the detection of the low assessment rates and patterns showing the connections between examiners and the examined. Moreover, that data could be analyzed without those insiders’ knowledge.
Analyzing networks and behavior
AI should be assigned to regularly review high-volume areas perceived as particularly risky, such as those where an employee’s exercise of discretion carries substantial financial implications. Such is the case with inspectors or expediters who help residents and commercial owners secure necessary public approvals. AI could look for patterns that expose behavioral differences, for example, identifying where a particular employee facilitated an unusual number of speedy license approvals for a single company. This type of bribery can often mean cutting corners when it comes to safety. AI can alert auditors to those relationships or outcomes that seem suspicious or outside the norm.
AI can integrate and synthesize information from various data sources — text, images and more — offering comprehensive insights that traditional analytics might miss. This ability is critical in identifying subtle indicators of corruption or inefficiency spread across disparate data systems. AI has the potential to set a new standard of network analysis by reviewing the relationships, interactions and flow of information or resources within a network involving departments, contractors, suppliers and other entities. By identifying relationships and behaviors indicative of a corrupt employee, vendor or permittee, AI can dramatically improve the ability of officials to focus their attention earlier and with more clarity.
For example, Shari Hyman, the Commissioner of the New York City Business Integrity Commission, with assistance from the Mayor’s Office of Analytics, cross-referenced layers of data from different departments to determine restaurants illegally disposing of grease by dumping it down city sewers. The Commission compared industry data on grease production with restaurant permit data and sewer backup data shared by the Departments of Health and Environmental Protection to better target enforcement and predict illegal activity. With AI, such comparisons across datasets should be routine.
AI can map relationships among contractors, subcontractors and project managers to detect patterns indicative of collusion or kickbacks. AI models can identify unusual decision-making patterns among project stakeholders, prompting investigations into potential conflicts of interest. By applying network analysis, municipal auditors and managers can better understand how different entities and individuals interact and identify potential risks, inefficiencies or opportunities for improving the City’s operations and integrity.
Clearly, using generative AI for public integrity presents risks and must be accompanied by human control and senior managerial oversight. In addition to continuously watching for misinformation, AI-supported investigations present serious ethical and privacy concerns. We suggest using the tools above to identify red flags from data in order to focus investigations and that its use to investigate a specific person occur only where there first is reasonable suspicion.
In sum, incorporating generative AI tools into the monitoring regimen will enhance the precision and capabilities of investors while saving precious City dollars and significantly improving government efficiency.