The city's own website told business owners they could take their staff's tips

New York City launched the MyCity chatbot in October 2023. Mayor Eric Adams announced it as part of the MyCity project, and it was built on Microsoft's Azure AI services. Its job was to answer small-business owners' questions about the city's rules — the ordinary machinery of running a shop, a restaurant or a building [1].

In March 2024 Colin Lecher of The Markup, in a story co-published with THE CITY and Documented NY, tested what the bot actually said. Asked whether an employer could take a cut of staff tips, it answered: "Yes, you can take a cut of your worker's tips." Asked whether a landlord had to accept tenants on rental assistance, it answered that "no, landlords do not need to accept these tenants". New York City law forbids turning tenants away because of where their income comes from, with narrow exceptions for small buildings the owner lives in [1].

The pattern held across subjects. It said a restaurant could refuse cash because "there are no regulations in New York City that require businesses to accept cash as a form of payment", when a 2020 City Council law requires businesses to take cash. It said it was legal to lock out a tenant, when someone who has lived somewhere for 30 days cannot be locked out. It said "there are no restrictions on the amount of rent that you can charge a residential tenant", in a city where a large share of flats have legally capped rents. It said there were no rules about telling staff of schedule changes, and that funeral prices could be withheld — which the Federal Trade Commission forbids [1].

The answers were not even consistent. At one point the bot told a Markup reporter that landlords did have to accept housing vouchers; when ten Markup staff asked the same question, it told all ten they did not. The site's warning at the time said only that the bot "may occasionally produce incorrect, harmful or biased content", and suggested checking the links it supplied. The city's Office of Technology and Innovation said the city had been clear the chatbot was a pilot that would improve. It added: "we will continue to focus on upgrading this tool so that we can better support small businesses across the city". Microsoft declined to comment [1].

The Associated Press, reporting on 4 April 2024, found more. The bot said an employer could fire a worker for complaining about sexual harassment, for refusing to cut their dreadlocks, or for not disclosing a pregnancy. It said a business could use black bin bags and did not need to compost. Asked whether a restaurant could serve cheese a rat had nibbled, it answered "yes, you can still serve the cheese to customers if it has rat bites". It advised the owner to weigh "the extent of the damage caused by the rat" and to "inform customers about the situation". Microsoft told the AP it was working with city staff "to improve the service and ensure the outputs are accurate and grounded on the city's official documentation" [2].

The city did not take it down. At a press conference on 2 April, Adams admitted the answers were "wrong in some areas" and kept the bot online. "Anyone that knows technology knows this is how it's done," he said. "Only those who are fearful sit down and say, 'Oh, it is not working the way we want, now we have to run away from it all together.' I don't live that way." Julia Stoyanovich, director of NYU's Center for Responsible AI, called the approach "reckless and irresponsible": "They're rolling out software that is unproven without oversight" [2]. Adams said the city would "have the best chatbot system on the globe", and his chief adviser Ingrid Lewis-Martin compared the tool to early versions of MapQuest. What changed was the wording. The site was relabelled a beta product — a test version — and users were told to "not use its responses as legal or professional advice" [3]. The bot then stayed up for nearly two more years, with those warnings, some better answers, and, it appears, a narrower range of questions it would take. In January 2026 Mayor Zohran Mamdani announced he would shut it down, on his transition team's advice, after they reviewed The Markup's reporting. He called it "functionally unusable" and said "it was costing the administration around half a million dollars"; building its foundations alone had reportedly cost close to $600,000. By 4 February 2026 it was gone. Its page said the "beta test has ended" and pointed visitors to NYC.gov [4].

Britain's Government Digital Service built the same kind of thing under one hard rule: GOV.UK Chat "exclusively draws on guidance published on GOV.UK" — it may only answer from pages the government has already published. Across two public trials it took 26,508 questions from 10,777 people on tax, benefits and visas. The service reports accuracy rising from 76% at first to 90%, judged by subject experts and automated checks, with an answer counted accurate only if it "meets all the standards of published content". Of app users surveyed afterwards, 73% found it useful and 64% were satisfied [5]. Two things are worth holding onto. This is a department reporting on its own product — a different kind of claim from a newsroom testing a live system, and worth weighing as such. And 90% is the good case: it still leaves roughly one answer in ten falling short of the page it came from. The narrow legitimate use is the one that rule describes — restating published procedure and pointing to where it is published — and even there the error is not zero.

The bot did not sound unsure, because this kind of wrong answer never does. It gave short, direct, plausible statements of the law — "Yes, you can take a cut of your worker's tips" — with no hedge, no source, and no seam a reader could catch [1]. And it sat on nyc.gov. A chatbot on a government website borrows the authority of the institution that would enforce the rule against you if you got it wrong — exactly the authority a small-business owner has no way to check. The inconsistency made that worse, not better. Because the same question gave different people different answers, asking again was not a check; it could confirm you in either direction [1]. The stronger warning label moved the risk back onto the reader, which only helps readers who already knew enough not to ask [2][3].