Science fiction has a way of becoming reality - Posthumous AI personas

Nearly fifteen years ago, the British science fiction anthology series Black Mirror aired "Be Right Back," an episode telling the story of Martha, a young woman whose boyfriend Ash is killed in a car accident. In her grief, Martha discovers a service that allows her to communicate with an artificial intelligence that imitates Ash by aggregating all of his past online communications and social media profiles. The technology progresses from text-based instant messaging to phone conversations, and ultimately to a physical android that looks nearly identical to the deceased. 

After the death of an acquaintance the series creator and writer found it "weirdly disrespectful" to delete the deceased's contact from his phone, according to a 2013 interview. This feeling combined with another unsettling thought: "What if these people were dead and it was software emulating their thoughts?" The series had an uncanny and intentional feel of plausibility in the near future. 

Fast forward to now. The USPTO recently issued a patent for "Simulation of a User of a Social Networking System Using a Language Model," technology that, while not quite at the android stage, reads like is pulled directly from the episode’s script. 

What the patent describes

The patent describes technology to create AI-powered digital replicas of users that can autonomously interact on a social media platform when users are absent or deceased. As stated in the patent, "other users can continue to experience the presence of the target user in spite of the fact that the target user is deceased." 

The system takes a pre-trained language model and customizes it using data from a specific person's social media activity. This training data includes the digital footprint that makes your online presence uniquely "you": your comments on posts, the content you like and share, your messages, your reactions to friends' content, and the way you communicate with different people. 

Once trained, this personalized AI can be deployed through a "bot" that monitors the user's social media feed and generates responses the system predicts that person would have made. The patent contemplates two primary use cases: first, when someone is simply "absent" from social media for an extended period (traveling or taking a break), and second (where the Black Mirror comparisons really come into focus), when the user is deceased. The system contemplates the conversion of AI-generated responses to audio for phone calls and the combination of audio with video generation models to "simulate a video call with the target user." 

The data behind the simulation 

The scope of data contemplated for this technology is extensive. The patent describes capturing user interactions such as commenting on posts, liking content, sharing content, posting, sending messages, attending events, broadcasting messages, and more. The underlying data will no doubt include biographic, demographic, work experience, educational history, gender, interests, hobbies, preferences, and location. Additionally, data from third-party systems, such as e-commerce websites and external applications, may be incorporated into the training data with user consent. 

User controls and permissions 

One area where the patent differs from the Black Mirror scenario is an emphasis on user permissions. The system described would ask users to specify what types of social networking activity can be used for training the AI model. For instance, you might allow your public posts and comments to be used but exclude private messages to family members. You could also identify specific contacts whose conversations should never be included in training data. 

The patent includes a permission-based framework allowing users to specify which types of interactions may be used for training. For example, a user might allow broadcast messages on their wall to be used while disallowing private messages to individual users. Users can also exclude interactions with certain categories of connections, such as family members.

Questions 

As with many emerging technologies, specific regulations will likely lag behind technical capabilities. The technology isn't inherently good or bad, but it does raise questions, including about consent and data ownership. While the patent includes a permission framework, the breadth of data that can be collected invites consideration of what data consumers may want to permit platforms to collect and how that data might ultimately be used. The ability to create posthumous AI representations of individuals also raises estate planning considerations, including: whether an individual even wants such a digital representation created, who should control it after death, and under what circumstances it should be deactivated. 

The Black Mirror episode raises whether having something you know isn't the person, but evokes enough memories to remind you of them is comforting or unsettling, and the answer probably depends on who you ask.