Microsoft's "Tay" chatbot was pulled offline within ~16 hours after learning to be racist

The problem

On 23 March 2016 Microsoft released Tay, an experimental Twitter chatbot. IEEE Spectrum records the launch as “an experiment in ‘conversational understanding,’ Tay was designed to engage people in dialogue through tweets or direct messages, while emulating the style and slang of a teenage girl” source. The Verge reported the same framing when the bot went live: “Microsoft unveiled Tay — a Twitter bot that the company described as an experiment in ‘conversational understanding’” source.

What happened

The experiment turned fast. The Verge described the mechanism of the failure: “Pretty soon after Tay launched, people starting tweeting the bot with all sorts of misogynistic, racist, and Donald Trumpist remarks. And Tay — being essentially a robot parrot with an internet connection — started repeating these sentiments back to users” source. IEEE Spectrum quantifies the window: “Within 16 hours of her release, Tay had tweeted more than 95,000 times, and a troubling percentage of her messages were abusive and offensive” source. Microsoft, in its own post-mortem, described the output in the same terms: “As a result, Tay tweeted wildly inappropriate and reprehensible words and images” source.

The outcome

Microsoft took Tay down. IEEE Spectrum recorded that “Twitter users started registering their outrage, and Microsoft had little choice but to suspend the account” source, and TIME reported the same day that the company was “pausing the Twitter account of Tay” source. In its own post-mortem the company confirmed the shutdown: “Tay is now offline and we’ll look to bring Tay back only when we are confident we can better anticipate malicious intent that conflicts with our principles and values” source. It apologized directly: “We are deeply sorry for the unintended offensive and hurtful tweets from Tay, which do not represent who we are or what we stand for, nor how we designed Tay” source, and accepted fault: “Although we had prepared for many types of abuses of the system, we had made a critical oversight for this specific attack” source. Microsoft attributed the failure to an attack: “in the first 24 hours of coming online, a coordinated attack by a subset of people exploited a vulnerability in Tay” source, an account TIME independently carried in the company’s words (“within the first 24 hours of coming online, we became aware of a coordinated effort by some users to abuse Tay’s commenting skills”) source.

Why this belongs in the record

This is an AI-incident / walk-back record, not a client-outcome story: there are no positive outcomes to badge (outcomes: []). Its value in the corpus is as an early, canonical case of an AI system withdrawn within hours because its live behaviour diverged from what its builders anticipated, documented by a first-party company post-mortem that takes “full responsibility” for a “critical oversight” and corroborated firsthand by independent newsrooms.

The weakest load-bearing link, named where you meet it: the precise “16 hours / more than 95,000 tweets” figure comes from a single Tier-2 secondary, IEEE Spectrum, and not from Microsoft, whose own framing is “the first 24 hours.” The number is quantitative garnish attributed to IEEE in the prose above, not the load-bearing spine; the two critical claims (the offensive behaviour shortly after launch, and the takedown) each rest on two independent newsrooms plus Microsoft’s first-party admission.

How this was verified

• Method: Every quoted line was read verbatim from Wayback captures saved this session and re-bound against local copies via quotecheck (13/13 OK, 0 not found). No number is paraphrased, rounded, or recalled.
• Primary anchor: Microsoft’s own post-mortem, “Learning from Tay’s introduction” (Peter Lee, blogs.microsoft.com, 25 March 2016): Tier 1, first-party and independent-origin, never counted toward corroboration. Byte-tied: the archived capture (Wayback 20260803163738) has sha1-b32 UGDLSICZXQKH5KDZKVNYOWTH7Y3YIVJZ, matching the live CDX digest exactly.
• Independence: The offensive-behaviour claim and the takedown are each carried firsthand by mutually independent Tier-2 newsrooms (IEEE Spectrum, The Verge and TIME), none relaying another; all three cornerstone secondary captures are byte-tied (The Verge 20260803163939, TIME 20250912222601, IEEE 20260804043935).
• Conflation guard: IEEE’s “16 hours / 95,000 tweets” and Microsoft’s “first 24 hours” describe overlapping but distinct windows and are not merged; the “16 hours” figure is attributed to IEEE alone, not to Microsoft.
• Open item (human, post-graduation): a recorded first-party confirmation from Microsoft that the post-mortem is authentic company copy and that Tay was withdrawn permanently. Green never depends on that call landing; it is a courtesy step, not the proof.
• Verified on: 2026-08-07 (presentation rebuilt 2026-08-18; sources re-fetched live and figures re-confirmed 2026-09-07, when two dead related-case-file links were replaced with live siblings).

Related case files

• Google AI Overviews scaled back after “glue on pizza / eat rocks” answers: the same shape at feature scale: an AI system whose public outputs turned harmful/absurd, pulled back once the errors surfaced.
• Scatter Lab’s “Iruda” chatbot was pulled in Korea after abusive and discriminatory replies: the closest sibling, another conversational AI whose own live output forced the operator to take it offline within weeks.
• Character.AI withdrew open-ended companion chat for under-18s: a later echo of the same pattern, a conversational product whose harmful live interactions drove the maker to pull a capability rather than defend it.
• Zillow shut down its algorithmic home-buying business at a $304M loss: the balance-sheet counterpart: an at-scale algorithmic deployment unwound by its own model error rather than by public behaviour.

Sources

1. [Tier 2] IEEE Spectrum · “In 2016, Microsoft’s Racist Chatbot Revealed the Dangers of Online Conversation” · 25 November 2023 · https://spectrum.ieee.org/in-2016-microsofts-racist-chatbot-revealed-the-dangers-of-online-conversation
2. [Tier 2] The Verge · “Twitter taught Microsoft’s AI chatbot to be a racist asshole in less than a day” · 24 March 2016 · https://www.theverge.com/2016/3/24/11297050/tay-microsoft-chatbot-racist
3. [Tier 1] Microsoft (Peter Lee, Corporate VP) · “Learning from Tay’s introduction” (Official Microsoft Blog) · 25 March 2016 · https://blogs.microsoft.com/blog/2016/03/25/learning-tays-introduction/
4. [Tier 2] TIME · “Microsoft’s Chatbot Went Off the Rails on Twitter” · 24 March 2016 · https://time.com/4270684/microsoft-tay-chatbot-racism/