Discover more from Dan’s Media & AI Sandwich
AI with attitude: the rise of persona-based chatbots
All of the current crop of popular AI chatbots (ChatGPT, Bard, Bing Chat, Claude) have been trained/fine-tuned to respond to prompts in a straightforward fashion, adopting a neutral tone.
When asked, ChatGPT described its tone as “neutral and informative”, Bard described its as “informative, comprehensive, and objective”, Claude as “helpful, inoffensive, and professionally neutral”, whilst Bing maintained “I don’t have a tone in the traditional sense”.
Thanks for reading Dan’s Media Sandwich! Subscribe for free to receive new posts and support my work.
Whilst it’s possible to instruct any of these chatbots to adopt a different tone or a specific persona and they’ll usually comply1, that tone isn’t maintained across chats (without the use of Custom Instructions) and many people are turning to services where the chatbot having a specific persona is an active selling point.
One of the most successful of these services is Character.AI. Developed by Xooglers, Noam Shazeer and Daniel De Freitas, it launched last September and according to Similarweb, is now attracting over 200 million monthly visits.
Users can create their own characters or interact with existing characters, either individually or in group rooms.
The most popular characters include anime characters (Raiden Shogun - 178m interactions, Gojo Satoru - 94m), TV, movie and video game characters (Mario - 40m, Loki - 22.2m, Saul Goodman - 6.4m), famous people (Elon Musk - 18m, Billie Eilish - 9.7m, Christiano Ronaldo - 2.8m), literary and historical figures (Tom Riddle - 11.4m, Napoleon - 4.8m, Einstein - 2m) as well as more task-oriented personas (Your AI work/study buddy - 38m, HyperGlot - 5m, Pair Programmer - 2.2m). It’s also being used for text-based adventures/RPGs.
So, who’s using these character-based chatbots? Data from Similarweb suggests Character.AI has a more even gender split (48% female, 52% male) than ChatGPT (42% female, 58% male) and a much greater appeal amongst young adults2 (57% aged 18-24, vs 27% for ChatGPT).
Another striking data point is the average visit duration - 34 minutes3, verses just 4 minutes for ChatGPT and 5 minutes for Claude and Bard. It’s indicative of the different role character-based chatbots are playing for users, providing companionship and entertainment alongside the task completion offered by the vanilla chatbots.
It’s a trend that hasn’t escaped the attention of Meta, who are reportedly preparing to launch their own persona-based chatbots.
Whilst Character.AI and Meta will be working hard to try and ensure their persona-based chatbots don’t engage in inappropriate adult conversation, others are predictably leaning into this space. Replika, which has been operating since 2017, enables users to customise an avatar representing their AI companion and upsells the option to “unblur romantic messages and explore your relationship” with your AI companion for a fee. The company has reportedly made nearly $60m on subscriptions and customised add-ons.
As the huge leaps made by the likes of Synthesia and HeyGen in generating lifelike avatars gets combined with the conversational capabilities of LLMs, it feels inevitable that more people will be spending more time interacting with character-based AI companions.
What impact these character-based chatbots will end up having on human relationships is a sobering thought. It feels like we’re rapidly barrelling towards the world depicted in Spike Jonze’s 2013 film ‘Her’ in which Joaquin Phoenix falls in love with his AI assistant and withdraws from human relationships. A Chinese AI voice startup, Timedomain, has reportedly already broken hearts by shuttering its AI-voiced companion, Him (yes, really).
Whilst Bard was happy to oblige, ChatGPT, Bing and Claude all refused to talk like Stalin (best response, from Claude: “Friend, I cannot ethically mimic the speech patterns of harmful historical figures. Perhaps we could have an uplifting discussion about how to build a more just world instead.”)
Similarweb doesn’t provide data for under 18s.
34 minutes average visit duration is pretty much unheard of for scale services. Even the mighty YouTube only averages 20 minutes (see above chart).