NYC MyCity business chatbot

The MyCity Business chatbot was a generative AI chatbot that New York City launched in 2023 on Microsoft Azure AI. The city's Office of Technology and Innovation ran it. It answered business owners' questions about city rules, and it was confidently and repeatably wrong about what the law required.

What the reporting found

On 29 March 2024, The Markup published an investigation with THE CITY and Documented NY. It found the chatbot routinely advised businesses in ways that would break the law.

It said employers could take a cut of workers' tips. It said landlords could refuse tenants who pay with Section 8 housing vouchers or other sources of income. New York City law forbids that. It said tenant lockouts were allowed, and that there were "no restrictions" on rent charges. It said stores could go cashless, against a 2020 city law. It said funeral homes could conceal their prices, against the federal funeral rule.

The reporting tested specific questions. It did not measure an error rate across all answers.

When ten staffers at The Markup asked the housing-voucher question, all ten got the same wrong answer. An earlier test had returned the correct answer, so the answer had changed over time.

How the city responded

Mayor Eric Adams declined to take the chatbot offline and called it a pilot. "It's wrong in some areas, and we've got to fix it," he said. He also said: "we're going to have the best chatbot system on the globe."

The city relabeled the chatbot a beta product and added a disclaimer. The Office of Technology and Innovation said the chatbot had "already provided thousands of people with timely, accurate answers." That claim was not independently confirmed. Microsoft said it was working with the city on fixes.

About two weeks after the tips error came to light, the city applied a patch. It sent questions outside the chatbot's topics to NYC.gov instead of answering them.

In March 2025, then chief technology officer Matt Fraser said the city still planned to expand the chatbot. It would cover all the content of the city's 311 service, not small business alone. He said a patch had cut hallucinations, meaning made-up answers, "exponentially." He gave no figure.

What the audit found

On 30 December 2025, the Office of the New York City Comptroller issued a performance audit of the MyCity system. Brad Lander was then the city's Comptroller. The audit reviewed the system's cost, its planning, and whether it did what was promised, including the chatbot's answers. It found the chatbot "appears to be unable to provide accurate or consistent information."

An internal weekly production report, reproduced in the audit, showed the chatbot did not answer 23 of 48 tested government questions. People asked more than 2,200 questions in July and August 2025. Of the 70 users who left thumbs-up or thumbs-down feedback, 71.4 percent were negative (50 of 70). The city disputes that share, putting it at roughly 2.25 percent of all responses. The audit rebuts the city's way of counting.

The auditors' own testing found the same question answered two different ways. Answers also changed with trivial edits, such as "NYC" versus "nyc."

The audit also found the wider MyCity system had cost over $100 million, across more than 120 agreements with about 50 vendors. It lacked a system development plan. It had not delivered the promised single form for reaching city benefits.

The audit made seven recommendations, including one to conduct AI red-teaming, where testers deliberately try to make a system fail. The Office of Technology and Innovation disagreed with all seven.

How it ended

Mayor Zohran Mamdani announced the chatbot's shutdown in late January 2026, as a budget cut. He called it "functionally unusable." It was not shut down to fix its accuracy.

Reporting in 2024 corroborates a development cost of roughly $600,000. The figure of roughly $500,000 a year for maintenance, and the description of the chatbot as waste, come from the new administration.

As of early 2026, the chat.nyc.gov page reads: "The Chatbot beta test has ended."

Who bears the harm

The public asked the chatbot directly and could act on its answers. No caseworker or other professional stood between them.

The chatbot flagged no one and denied no one. It handed out instructions. Whoever followed a wrong one took on its legal risk. The harm could fall on people outside the conversation, such as workers and tenants, who had no channel to contest anything.

Why this case matters

Every model is wrong sometimes. The case file argues that the defining failure here was institutional, not technical. Reporters published the errors, an incident database listed them, and a formal audit documented them again. The city kept the chatbot online and planned to expand it, until a budget cut ended it. Exposure is not correction.

The case file contrasts this chatbot with two careful benefits chatbots in this Lab: Nava's assistive chatbot and the Imagine LA Benefit Navigator. In both, a caseworker checks the chatbot's answer before passing it on. Here the public asked directly. So there is no checking habit to protect, and no professional to train.

The case file points to four moves instead. Check each answer against an authoritative legal source before the public sees it. Label a generated answer as generated, and as outside what it may speak to. Agree in advance on a trigger that pauses a chatbot shown to be wrong. Make the outside audit binding rather than advisory.

What this network is drawn from

This network follows the pattern the case file describes. It does not reconstruct the actual chatbot. It shows the chatbot, the city's business pages, the public who asked, the disclaimer, and the outside reviewers. The workers and tenants who bear the risk are outside it.

A separate part of MyCity, a childcare benefits application flow, is not this chatbot. Its figures, including a roughly 46 percent application-ineligibility rate, belong to that flow. Being deemed ineligible is an application outcome, not proof of an algorithm's error.

What the available tools can and cannot address

A failure pathway is a link between two parts of the network where a mistake made by one part can be passed on to the other. Here a mistake is, for example, an answer telling a landlord he may refuse a housing voucher. A pathway counts as closed once it passes on only a few mistakes. The work along it goes on.

This case has a budget of 10 units. Each tool costs the same at every target level. Before any tool is used, four failure pathways are open. They are Answers to the public, Chatbot draws on the business pages, Public asks a question, and One chatbot answers everyone. Mistakes also build on one another across the network, so the targets are not met at any level that sets them.

Explore (No Targets) sets no targets. Under Service Targets Only, Escalate checks meets the targets on its own, for 2 units. When monitoring flags trouble, it raises how closely the chatbot's answers are checked, instead of waiting for the next review. On this network it has Business owners and the public check the chatbot's answers more closely. The case file argues a checking habit has no purchase here, because no professional stands between the chatbot and the person who asks. The sources describe no such checking by the public. Many other combinations within the budget also meet these targets. One is Mark AI-written records with Keep prompts neutral, for 4 units.

Under Service and Safety Targets and All Governance Targets, the targets can be met within the budget. Both levels ask you to close every failure pathway, among other targets. Each open pathway needs its own tool. Escalate checks closes Answers to the public. Mark AI-written records closes Chatbot draws on the business pages. Keep prompts neutral closes Public asks a question. Check with a second model closes One chatbot answers everyone.

Together those four cost 9 units. Every combination within the budget that meets the targets at these two levels includes all four. Counting each tool's settings, four such combinations meet them under Service and Safety Targets, and three under All Governance Targets.

Pause AI on alarms also closes Answers to the public, by halting the chatbot's answers until a person clears a review. But the chatbot then adds too little to the work. No combination within the budget that includes it meets the targets at any level that sets them.

More is not better here. Every tool at its strongest setting at once closes every failure pathway, but costs 34 units, more than three times the budget. It also leaves the chatbot adding too little to the work, so it misses the targets at every level that sets them.