Blog: Dead but not gone
The time has come when no one will ever be really gone for good. A recent advancement in artificial intelligence (AI) could mean that when people message their deceased Facebook friends, the social network will be able to talk back to them like their friends used to.
If you’re a Black Mirror fan, the concept of chatting with dead people online probably doesn’t seem like such a strange idea. In the 2013 episode Be Right Back, a pregnant woman called Martha is devastated when her fiancée Ash dies in a car crash. Martha subscribes to a new online service that mimics his personality with alarming accuracy. The service goes through Ash’s digital remains by reading everything he ever said on the Internet — Facebook updates, tweets, anything he said in public and his private emails — to develop a digital replica that talks just like him. Martha finds comfort in communicating with the digital Ash and at one stage contacts him almost constantly, keeping him updated on her pregnancy.
Every month more than 30 million Facebook users visit memorialised pages to post stories, commemorate anniversaries and remember their departed loved ones. What does 30 million users look like you may ask? Well, it’s nearly the size of Australia’s population (25m) but slightly more than Ghana, Mozambique, Angola and Malaysia. Facebook can now detect if a user is deceased and, if it keeps expanding at its present rate, the number of dead people on the network will exceed 4.9 billion by 2100. Sheryl Sandberg, their Chief Operating Officer, said Facebook’s AI hides pages belonging to a person who have died that have not been memorialised, “… because we know memorializing a profile is a personal choice made by loved ones and may take time”.
Facebook have just released their Tributes section, a new page tab that encourages living users to communicate with memorialised ones, while preserving the original page as an online monument. Like a digital photo album that tells their life story. The network wants people to, “share stories, commemorate a special day or let family and friends know you’re thinking about [them]”, as they would in real life. Sandberg personally recommended the feature when they said, “After my husband Dave died, Facebook took on an even deeper meaning for me… I hope Facebook brings the same comfort to anyone who has lost someone they love”. According to research into online grieving on Facebook, almost half of people surveyed said they understood why someone would message deceased users, even though they would not do it themselves. In a study on the roles of social media communication for the bereaved, a participant said, ‘‘Even though she has died, I still post on her wall, mostly that I miss her and love her very much . . . . As odd as this sounds I feel she can see her Facebook page in Heaven’’.
Computers are yet to reach the same level of intuitive language comprehension as humans. Fear not, dear reader! When we read something in a blog post, we can understand what it means in our reality. We feel the emotions that arise from reading and often visualise what it would look like in real life. Computers cannot really understand what human language is trying to say — a computer can’t read between the lines, yet. However, when GPT-2 was evaluated using the Facebook AI Research Children’s Book Test which examines how well an AI system can predict the word missing from a given sentence in a story, the model achieved state-of-the-art scores. It was 93.3% accurate in predicting nouns (compared with humans’ 96%) and 89.05% accurate in guessing names (compared with humans’ 92%).
Industry experts consider GPT-2 to be hugely significant in the advancement of natural language processing (NLP) research and many agreed with OpenAI when they announced it would probably be dangerous if it fell into the wrong hands. For that reason, the organisation made the controversial decision to not release the complete model to the public. OpenAI anticipate GPT-2 could be used for malicious purposes such as writing misleading news articles or impersonating people online. Deceased people too, perhaps. Meanwhile, Facebook AI research are offering academics funding to develop NLP models that run on mobile devices, as 96% of users access the network on smartphones and tablets. Currently, only large computers can run NLP models, with slower processing speed than computers typically used to run applications in industry. According to Facebook, the biggest barrier to using models like GPT-2 in the real world is that you can’t talk to them like you do with Siri, Google Assistant and Alexa, yet.