The big fashion for “chatbots” in the mid-2010s seemed to have passed. But on Friday August 5, Meta recalled that work on this technology was continuing by presenting BlenderBot 3, its new “state-of-the-art chatbot“. According to the company, this text-based robot can “converse naturally with people” on “almost any subject“, a promise repeatedly made by the creators of chatbots but never realized.
Still in the prototype state, BlenderBot 3 is accessible free of charge (only in the United States for the moment), so that a large number of voluntary testers can make it progress through a system of evaluation of the discussion. It has therefore been questioned at length by the media and other curious people since it was put online, and the first assessment looks like a sad refrain: BlenderBot 3 quickly plagues Facebook, criticizes Zuckerberg’s style of dress, then spins with conspiratorial remarks, even anti-Semites. Just before launching the tool, Meta warns users that the chatbot “is likely to make false or offensive statements“. But in his press release, he specifies that he has put in place guarantees to filter out the worst of them…
Meta’s chatbot, Meta’s first critic
BlenderBot’s goal is long term. The researchers don’t want to create a functional and marketable tool in the short term, they just want to improve the state of the art of chatbots. In concrete terms, their tool aims to integrate human conversational qualities (such as personality traits) into its responses. With a long-term memory, it must be able to adapt to the user as the exchanges progress. In their statement, the researchers specify that BlenderBot must advance the conversational skills of chatbots “avoiding learning unnecessary or dangerous answers“.
The problem, as always, is that the chatbot will search the internet for information to fuel the conversation. Except that it doesn’t do enough sorting. Asked about leader Mark Zuckerberg, he can answer, “he is a competent businessman, but his practices are not always ethical. It’s funny that he has all that money but still wears the same clothes!“, reports Business Insider. He does not hesitate to recall the myriad of scandals that have marred Facebook (and partly justified its change of identity) when it comes to its parent company. Or, he says that his life is much better since he deleted Facebook.
If the bot is so negative towards Meta, it is simply because it will tap into the most popular search results on Facebook, which relate the history of its setbacks. By this operation, it maintains a bias, which turns out to be to the disadvantage of its own creator. But these drifts are not confined to fun projections, which poses a problem. To a journalist from the wall street journalBlenderBot claimed that Donald Trump was still president, and “would still be with his second term ending in 2024“. It thus relays a conspiracy theory. To top it off, Vice indicates that BlenderBot’s answers are only “generally neither realistic nor good“and that he”frequently changes the subject“brutally.
The story repeats itself
These skids from the amusing to the dangerous have the air of deja vu. In 2016, Microsoft launched the Tay chatbot on Twitter, which was supposed to learn in real time from discussions with users. Failed: after a few hours, the text robot relayed conspiracy theories as well as racist and sexist remarks. Less than 24 hours later, Microsoft unplugged Tay and apologized profusely for the fiasco.
Meta has therefore attempted a similar approach, relying on a massive learning model with more than 175 billion parameters. This algorithm was then trained on gigantic text databases (mostly publicly accessible), with the aim of extracting an understanding of language in mathematical form. For example, one of the datasets created by the researchers contained 20,000 conversations on over 1,000 different topics.
The problem with these large models is that they reproduce the biases of the data to which they have been fed, most often with a magnifying effect. And Meta was aware of these limitations: “Since all AI-powered conversational chatbots are known to sometimes mimic and generate dangerous, biased, or offensive remarks, we conducted large-scale studies, co-hosted workshops, and developed new techniques to create safeguards for BlenderBot 3. Despite this work, BlenderBot may still make rude or offensive comments, which is why we collect feedback.” Obviously, the additional guarantees do not have the desired effect.
Faced with repeated failures of major language models, and a good number of abandoned projects, the industry has returned to less ambitious but more effective chatbots. Thus, the majority of customer assistance robots today follow a predefined decision tree without ever leaving it, even if it means telling the customer that they do not have the answer or directing them to a human operator. The technical challenge then becomes to understand the questions asked by users, and to bring up the questions left without the most relevant answers.
Meta is transparent
While BlenderBot3’s success is more than questionable, Meta at least demonstrates a rare transparency, a quality usually lacking in AI-powered tools. The user can click on the answers of the chatbot in order to obtain the sources (in a more or less detailed way) on the origin of the information. Additionally, researchers share their code, data, and model used to power the chatbot.
At Guardiana spokesperson for Meta also clarifies that “youAnyone using Blender Bot is required to acknowledge that they understand that the discussion is for research and entertainment purposes only, that the bot may make false or offensive statements, and that they agree not to intentionally incite the bot to make offensive statements.“
In other words, BlenderBot recalled that the ideal of sensitive chatbots capable of expressing themselves like humans is still far away, and that there are still many technical barriers to overcome. But Meta has taken enough precautions in her approach so that this time the story does not turn into a scandal.