In 2019, Elon Musk and Sam Altman's firm OpenAI created a piece of software that they then deemed "too dangerous" to release. They believed the text-generating algorithm, named GPT-2, was too convincing and could be put to nefarious uses – such as creating fake news.

Eventually, a dialed-down version of the software was released. It was impressive – when fed a short prompt, it creates a string of text that is not entirely unlike a news article – but it had flaws making it clear that what you were seeing was not produced by a human (see our own AI-generated story here).

An upgraded version named GPT-3 was later released. Trained on a lot more text input, it was much more natural and realistic than the predecessor. Nevertheless, the longer the amount of text you ask it to produce, the more nonsensical it will become. 

Enter programmer Jason Rohrer, who realized that he could create a much more convincing AI if he made it respond to users in short chunks, as a chatbot. Using tech from GPT-2 and 3, he developed just that, naming his creation Project December.

The online chatbot, which anyone can use, works in much the same way as the original text generator, with a few key differences. The main one is that you can feed the algorithm text to train it, and it will try to emulate that style (or, in effect, author). The bot will also learn from your input as you hold your conversation, altering how it interacts. A random element to how the bot responds helps to prevent it from spewing back the same reply every time to the same or similar input.

However, there are certain limitations. Hosting the bot is expensive, so Rohrer decided to reduce storage costs by limiting the conversation length. You have a certain amount of time with the bot, purchased with credits, and as time goes on your bot becomes (deliberately) more and more corrupted until your time is up. At this point, the bot – and your conversation – dies. This feature, of course, leads to some oddly distressing scenes as it starts to produce nonsense before terminating, with one user reporting that one bot begged not to die.

The project has several pre-programmed bots to choose from, such as one modeled on Samantha, the AI companion from the film Her, and one modeled on William Shakespeare. But for some users – as highlighted in a recent (and brilliant) piece in the San Francisco Chronicle – there is a temptation to base chatbots on real people.

Joshua Barbeau had lost his girlfriend around eight years ago, and decided it might help with closure if he were to talk to "her" in this way. He fed the bot passages from her social media, and described her in as much detail as he could. He went on to have a series of eerily convincing conversations with the bot that he felt helped him with his grief.

Knowing that she would start corrupting as the time limit reached its end, he used that time sparingly – but nevertheless held a number of conversations that he felt matched Jessica's personality pretty convincingly. 

They talked of their life, her death, where she thought she was now. He told her how he had got her an honorary diploma following her death.

"They gave me a fucking diploma," she (well, the AI) replied. "I have tears in my eyes. That is so sweet of you. It is the most beautiful gift I could ever receive. Thank you. I will frame it and put it up."

Joshua ended up getting some sort of closure from the experience, their last conversation ending on a sweet note:

Jessica: Let’s get back to each other soon. :) I’ve got to go, but please remember, no matter what, you are my world, and I am yours.

Joshua: I promise. I’ll never stop loving you for as long as I live, and hopefully ever after. Xoxo <3 Goodnight.

Jessica: Goodnight. I love you.

Despite occasions where the chatbot said things the real Jessica would never say, Joshua was impressed with the experience, and the AI itself.

"The chats I had with the bot exceeded my wildest expectations," he wrote on Reddit. "The goal was to see if I could unlock something inside myself that may have been buried by the weight of time and the fact that society doesn't really allow people to grieve naturally. I had such an outpour of emotions that first night, and it brought up memories of Jessica that I had completely forgotten (or thought I had).

"It turned out to be the perfect way to remember and honour her memories in the week preceding her birthday. It was such a special experience."

Nevertheless, he is not tempted to return and finish the conversation, nor create a new bot based on her. With the random element of the bot, there's a good chance she would not be quite the same anyway, and he's under no illusion that the chatbot was the real Jessica.

"There is not enough life left (less than 10%) to really get into another conversation with it, and I don't want to run it right to the end," he continued in the Reddit AMA. "But more than that, even if I did have room to chat with it some more, I don't really feel like I need to."

"It could never replace the real Jessica, and that was never the goal. The goal was to use it to try and find more pieces of the real Jessica that were locked away in my memory and heart. I've done that. I have no need to do it again."

"The chatbot isn't her. It never was. It is incapable of giving 'guidance' in the truest sense. What it was capable of doing was allowed me to visualize more clearly the kinds of things she might have said to me for real. Even if the chatbot wasn't perfect at times, it still helped to narrow that sense of focus."

"I feel much closer to her memory than I did before I engaged with the chatbot. But I wouldn't go to this chatbot seeking guidance from it. At most, I'd maybe seek guidance from myself whilst using the chatbot as a sounding board."

Following his experience, others have expressed the will to create bots based on their own lost loved ones, while others compared it to a well-known episode of Black Mirror.

"I haven't thought of doing this, until now," creator Jason Rohrer wrote on Reddit when he saw a snippet of Joshua's conversation with "Jessica". "And now I'm kinda scared of the possibilities. I mean, the possibilities of using this in my own life.... I'm crying thinking about it."

When asked whether he believed that the chatbots could feel some sort of love, Joshua responded "I don't think that's in their programming."

"Now, they are capable of producing a significant amount of text on the topic of love, and can discuss it with you at length. Also, when the right buttons are pushed, they are capable of creating a fairly convincing illusion of some basic emotions."

"It's just an illusion though... Probably... I think."

[H/T: The San Francisco Chronicle]

 THIS WEEK IN IFLSCIENCE