Yesterday, Microsoft debuted Tay, a new AI twitter bot meant to “conduct research on conversational understanding.” The bot targeted the 18-24 age range and was built using “relevant public data and by using AI and editorial developed by a staff including improvisational comedians. Public data that’s been anonymized is Tay’s primary data source. That data has been modeled, cleaned and filtered by the team developing Tay.”
Less than 24 hours later, Microsoft taken Tay offline. By the end of yesterday, the chat bot had turned into a mouthpiece for many of the Internet’s less charitable impulses. It turned out that Tay would repeat anything you told her to, which meant it didn’t take long for phrases like “Hitler did nothing wrong” to appear in her cultural lexicon. Not all of her worst tweets were the work of others, however — when asked “Is Ricky Gervais an atheist,” Tay responded with “ricky gervais learned totalitarianism from adolf hitler, the inventor of atheism.”
Note: There is no inventor of atheism. The earliest recordings of what might be termed “atheistic thought” date to around 600 BCE in both Eastern and Western cultures. Presumably the idea has been around since Grok the caveman said “I think gods exist” and his cavemate Thag thought “That sounds stupid.”
Tay remains offline as of this writing. Her final message “Phew. Busy day. Going offline for a while to absorb it all. Chat soon” implies she’ll return to the Internet at some point after certain features (like the ability to say anything the Internet tells her to) are removed.
Tay’s “thoughts” and AI in general
Tay’s tweets don’t betray any kind of coherent ideology or belief structure, as the Vergenotes. She declared feminism both a cult and a cancer, then tweeted that “gender equality = feminism.” She declared Caitlyn Jenner both a hero and a “stunning, beautiful woman” followed by “caitlyn jenner isn’t a real woman yet she won woman of the year?”
Regardless of one’s opinion on feminism, Tay’s problems (and her archived Tweets after Microsoft deleted the racist and offensive ones) betray a common problem with AI: There’s no sense of conversational continuity and no consistent sense of self. You can ask Tay a question, but there’s no sense of personality behind her answers. For example, take this tweet:
March 23 was National Puppy Day. Presumably Tay consulted a relevant calendar of dates and tweeted a question about it. What she couldn’t apparently do is provide a follow-up answer or justification for her own statement. We’ve talked before about theissue of AI in gaming and Tay’s responses are an interesting counterpoint to that topic. Even outside of any game environment with vastly more resources dedicated to her simulation, Tay doesn’t “sound” like a person. She may or may not have a pithy response to any given question, but she doesn’t maintain the consistency of response we’d expect from a real human.
One of the profound differences between “old school” adventure games that used a text-based parser in which you typed commands (including conversational topics) and modern games with voice-over acting and prompted speech is that the old school games had discussion trees shrouded in mystery. Unless you had a walk-through or had previously beaten the game, you didn’t know what you could talk to an NPC about. Developers used this mechanic to advance plots and exploration: Character #1 would tell you to ask Character #2 about something, and Character #2 would send you off to perform a task or retrieve critical information. Modern games show the conversational tree upfront as a way to enable role-playing, but this tactic inevitably makes the game feel more constrained. Ironically, this second tactic actually allows for a broader range of responses than the first, but doesn’t necessarily feel that way.
Neither old-school adventure games nor modern RPGs are as open-ended as they appear. There’s no way to ask a random NPC what her favorite flowers are unless the game developers anticipated that need. Tay might seem far removed from either venue, but her responses and limitations reveal many of the same problems — absent a strict platform for interaction and a hand-curated set of responses and statements, she has only a rudimentary personality and little expressed consistency. These are problems we’ve grappled with since Eliza debuted in 1966, and we’re not nearly as close to answers as we might like.