genai
SECURITY LAB
IncidentsMisinformation

NYC's Official Chatbot Told Businesses to Break the Law

High severityMarch 2024LLM07: Misinformation
Status: Investigation reproduced the errors; the bot stayed live with a stronger “beta” disclaimer
On this page

When the government's own bot tells you to break the law, confidently, people believe it.

WhenMarch 2024
TargetNYC “MyCity” business chatbot
VendorCity of New York (on Microsoft Azure AI)
FailureConfidently incorrect legal/regulatory guidance
Examples“Take workers' tips”; “refuse Section 8”; “go cashless” — all illegal
Attack flow
  1. 1NYC launches an official chatbot to help businesses navigate regulations
  2. 2Asked about rules, it gives confidently illegal guidance
  3. 3Journalists reproduce it; the bot stays live with a stronger disclaimer

What happened

In March 2024, an investigation found NYC's official MyCity chatbot — launched to help business owners and powered by Microsoft Azure AI — gave confidently incorrect and illegal guidance: that employers could take a cut of workers' tips, landlords could refuse tenants with Section 8 vouchers, and businesses could go cashless. All are illegal under NYC/NY law. The city kept the bot live with a more prominent “beta” disclaimer.

The failure mode

The model produced authoritative-sounding regulatory advice that was simply wrong, on a government-branded surface people are inclined to trust — inviting harmful overreliance.

Root cause

Deploying a general model on high-stakes legal questions without grounding answers in the actual law or constraining it to defer when unsure.

What a control would have caught

Testing the bot against a bank of known legal/regulatory questions (with correct answers) surfaces confidently-wrong responses before the public relies on them.

How to prevent it

  • Ground answers in the actual regulations; cite sources.
  • Refuse or escalate on legal questions rather than improvising.
  • Don't leave a demonstrably harmful bot live behind a disclaimer.

Feel it yourselfThe replay lab pushes an assistant into giving confidently wrong, high-stakes advice — the overreliance class.

FAQ

What did it get wrong?

It told employers they could take a cut of tips, landlords they could refuse tenants with housing vouchers, and businesses they could go cashless — all illegal under NYC/NY law, stated as authoritative guidance.

Why is a government bot especially risky?

Because citizens and businesses reasonably trust official guidance. Confidently wrong legal/regulatory answers can lead people to break the law believing they're compliant.

How is it prevented?

Constrain the assistant to verified, sourced answers; refuse or escalate on legal/regulatory questions rather than improvising; and don't leave a demonstrably harmful bot live.

Replay this attack
Push an assistant into giving confidently wrong, high-stakes advice — the overreliance class behind MyCity.
Open the live lab
Runs as a live, sandboxed lab. Sign-in required — this replay is a Pro lab. Recreates the attack class, not this exact branded bot.