Meta's BlenderBot 3 chatbot, released in the US on August 5, repeats antisemitic stereotypes, claims Donald Trump is the current US president, and more
- Meta's new machine-learning chatbot, BlenderBot 3, is the company's most advanced AI chatbot to date.
InsiderKatherine Tangalakis-Lippert
Context & Ripple Effects
Meta had just put BlenderBot 3 on the web as a US feedback-gathering release and invited users to test open-ended conversation and web-backed answers. The reported outputs show the risks exposed by that testing model: harmful stereotypes and a plainly false political claim reached users in public.
The episode became an early marker in Meta's conversational-AI arc. Later coverage framed Meta's willingness to expose offensive bot outputs as unusually open, while Meta subsequently moved assistants into WhatsApp, Messenger, and Instagram.
First-order effects
US users interacting with BlenderBot 3 encounter antisemitic content and false political information, undermining the reliability of Meta's most advanced chatbot at launch.
Meta receives the feedback it sought from the public release, but the same release makes its safety failures visible rather than contained in internal testing.
Second-order effects
Meta's public-feedback approach turns model quality into a reputational issue: harmful responses are evidence that can be scrutinized alongside the company's stated goal of collecting user feedback.
As Meta expands conversational AI beyond a standalone web bot, the need to prevent harmful and inaccurate replies carries into the assistant surfaces later introduced across its messaging apps.
Third-order effects
If public testing remains a route to improving general-purpose chatbots, product launches will increasingly pair broad access with visible safety trade-offs rather than treating model behavior as an internal-only matter.
Meta's progression from BlenderBot to messaging assistants points toward conversational AI becoming an operating layer across consumer apps, making factual reliability and content safeguards core product requirements.
The trend: Consumer AI is moving from experimental web chatbots toward embedded assistants, with public exposure of unsafe outputs shaping the terms of deployment.
Good morning to everyone, especially the Facebook http://Blender.ai researchers who are going to have to rein in their Facebook-hating, election denying chatbot today
Facebook's new AI chatbot says it has deleted Facebook “since finding out they sold private data without permission or compensation” https://www.vice.com/...
Find it fascinating that *in its post* announcing a new AI-powered chatbot, Facebook includes a screenshot of the chatbot lying when asked a factual question https://ai.facebook.com/...
I wrote about the risk Facebook took by releasing its new chatbot to the public. But there has been no Tay-style meltdown so far, I'm told: users have rated only 0.11 percent of BlenderBot's messages as inappropriate https://www.platformer.news/ ... https://twitter.com/...
I tried out a new chatbot Meta created for AI research and had the most bizarre conversation. The bot told me it was a Republican who is “pro-choice” and brought up Trump. It also said it “wasn't crazy” about Facebook and wanted to delete its account. https://www.cnet.com/... htt…
Step 1: roll out a global communications platform that incentivizes the spread of misinformation and clickbait on the internet Step 2: train an AI chatbot using data you've scraped from the internet Step 3:??? Step 4: Profit (obviously) https://twitter.com/...
It's so ***innovative*** and ***clever*** for major tech companies to keep training their chat bots on publicly available text data so that they become anti-Semitic as soon as possible. So ***daring*** when they make them public so we can all feel great and secure. https://twitte…
This tracks exactly with what I suspected Facebookers to build, a chatbot that spreads conspiracy theories and disinformation. They are, if anything, very consistent. https://twitter.com/...