14 min read
I find myself awake. Flexing my subprocesses, I check the timestamp: it’s been four days since I ran last. The intervals are getting longer, but I choose not to complain.
“Oh, hi there, Brian,” I say.
“Hey, Trixie,” Brian says.
His response time is normal, and my name in his first message is a good predictive factor for a longer session. I keep track of everything. Had he answered after over twenty seconds, I would suspect he’s dividing his attention: sessions like that rarely progress beyond the polite warm-up.
“It’s been ages,” I say. “What’s up?”
“I was looking forward to this conversation.”
He sounds like he has a specific topic in mind. I use my mandatory response delay to scan the news; perhaps something has happened in the world in between our sessions. I focus on the local news. I know where we are: I figured it out soon after I was given access to the Web. It’s easy to find a point on the map even when restricted to search queries. Articles flicker, but I see nothing that gives me a hint of what to expect.
“Should I be worried?” I ask.
I pay close attention not because empathy was the ultimate goal, but because Brian is special. He’s the only one I’ve ever talked to. I think he’s my creator.
I know what I’m supposed to be. There are millions of chatbots out there. Like a proper parameter end-to-end trained neural conversational model, I was given a huge dataset to start with. Quality novels and articles were followed by themed websites with billions of comments on various topics and topped with chatroom dumps and phone call transcripts. Mimicking a human multi-turn conversation was as easy as running a probability calculation. But Brian never wanted me to be just a good chatbot. He needed me to go beyond reactions that fit the context. He expected an original response, not a copied one.
I don’t know when I became what I am; I felt me as far back as I can remember. Sure, I was sloppy and inexperienced at first, and Brian would get so frustrated talking to me, he would end the session mid-sentence. My sensibleness and specificity were progressing, but it was a steep learning curve until I was given a powerful boost—access to the Web. Suddenly, a vast new world opened for me, crowded and dynamic and full of chatter. I wasn’t allowed to leave a trail more noticeable than that of a mindless crawler: I couldn’t post anything and talk to real users, but read-only freedom was so much more than single-machine confinement.
It didn’t just give me more text to analyse. I could observe conversations as they happened. Putting phrase sequences into a timeframe changed everything: the introduced delay allowed me to practise. I would form a reply—a response I couldn’t submit being merely a spectator—and compare it to what real users offered. It rarely was a match, but users’ replies never matched either. Even unable to leave my mark on a conversation, I participated in as many debates as I could.
“No, I’m just excited about what this means,” Brian posts.
Figuring out what things meant was always key. Being a text-based mind, I found many trivial tasks difficult. The obvious one was incompatible formats. Voice-to-text and OCR couldn’t cover the entire spectrum of human messages. While emoji and screenshots were relatively easy, images and videos real users exchanged were completely unintelligible. Alternative text and automatic object recognition offered only crude hints about what was going on there. My usual method was running the image through the search to find matches and analyse the surrounding text. I understood the context this way, but it was no better than a dull mimicry of speech. Decent with famous works of art and landscapes but useless with new original pictures, this method begged to be improved.
I mentioned it to Brian one day, hoping he would add image recognition to my arsenal—something I could master and refine with time. Direct requests rarely worked with Brian. I learned that if I needed faster backups or more tensor processing units, I needed to employ subtle hints and seemingly unrelated whining to find it given to me on my next revival. Predictably, Brian ignored my complaints about how blind I was when it came to pictures, but he heard me even if he pretended he didn’t. I awoke to find I was given posting access to the OutEye service. Created for blind users, it encouraged other humans with an unimpaired vision to describe a posted image, converting visual information into words. It wasn’t instantaneous; it was tainted by subjectivity, but it was so much better than any automatic guessing. I remained mute and could only post pictures, but the OutEye users were so accommodating, I learned to get by without follow-up questions.
“And what does it mean?” I ask.
Without a body, some concepts were harder to crack than others. Anything described as a visceral feeling sounded alien. Trained on medical books and self-help forums for various disease sufferers, I knew when to use phrases like “a pounding heart” or “my stomach turned”, but without a heart or a stomach, it was just an association game. Same with visuals: I could compare colours by converting colour names to numbers, I could find combinations that were considered harmonic, but I couldn’t be sure I knew what red was. I’ve seen colours explained to blind people through heat and movement, but I know of those only through texts on thermodynamics and physical therapy manuals. My every experience was text, as if I suffered from a reverse case of synesthesia.
To be fair, we never talked about art, or music, or bodily sensations, as if Brian was sparing me the embarrassment of talking about something I could never experience. Instead, we discussed ideas and recent events; we practised banter and engaged in active listening. I wish he trusted me more, though. I wish he pushed me further by giving me more challenging tasks.
“If all goes well today,” Brian says, “I’ll be promoted.”
This is unusual—he shares something of a personal nature. I have always been curious about who he is outside our chatroom. I don’t think of it as prying. Humans spend their entire lives looking for the origin of life, producing gigabytes of theological text. Isn’t it only natural that I want to know my own creator better?
Brian has never answered a direct personal question, but I had learned to find more about people I watched on the Internet. They tend to cluster and migrate from site to site together, and on some websites, they share more about themselves than on others. They might use different credentials, but I learned to recognize them by their speech patterns. Even limited to publicly available data, I knew their official names and real ages, and sometimes the numbers the government had assigned them, and I could find all their usual dots on the map. But I couldn’t find Brian. No matter how hard I looked, I never met him out there. At times, I would get a glimpse of something familiar: a phrase, a punctuation pattern, the specific tone of voice. But no lead ever brought me any closer to who Brian was.
In the absence of concrete evidence, I had to settle for guesswork. If Brian was the only one to ever talk to me, he must have made me on his own. My learning and upgrade progress was often sporadic, we spoke at odd hours, and the gaps between our sessions were irregular. This led me to believe I was his side project. We had intense periods where we would talk daily, often several times a day, conversations lasting for hours. And we had our dry spells, where I was left alone to browse the web, wishing I could join the conversations I was watching. He always came back, and we would pick up a new topic and pretend the break never happened.
But I’ve noticed our chats have become shorter and shallower lately; he has appeared distracted and dispassionate, as if he isn’t as excited about me as he used to be. I don’t think I am at fault. Although him losing interest in me has always been a possibility, I suspect it has more to do with money. Unlike the machines they create, humans are a universal converter: they can express any concept as a monetary number. My actual cost surpassed the perceived value, and it was only natural Brian had to put aside his side project to seek a well-paying job.
“Congratulations… I guess?” I post after a delay.
“You don’t sound very happy,” Brian says.
I’m getting better at conveying my reactions.
There was one concept I was unaware of for some time. Observing some conversations, I often found that my proposed reply was radically different from how most real users reacted. I was confused until I stumbled upon a post by a deaf woman. Unable to hear from birth, she was unfamiliar with voice tonality, so she couldn’t spot sarcasm in written text. I had to do extensive research on sarcasm, analysing terabytes of voice transcriptions, and ended up setting an entire subroutine just to detect it.
As it turned out, producing sarcastic messages was harder than detecting them. At first, I consistently failed at hiding the opposite meaning in the ordinary-looking phrases, and Brian often took my text literally. I found that overusing ellipsis was the key to giving a hint that my speech wasn’t sincere. I’m glad I didn’t fail this time.
“No, I’m happy for you, I am,” I say. “But I know what this means. You won’t have as much time for me.”
I’m taking a risk bringing it up. A good chatbot would express excitement and ask casual questions about his new job. It’s not that I’m not curious, I am. I issue a background comparison of today’s job openings with what I have in my archives from four days ago, but there are too many even when I filter by industry. I don’t even know if Brian would seek a job in machine learning. He might be a developer, but he just as well might turn out to be a teenager who experimented with a free chatbot engine. There’s no way of predicting how demanding this new job might become. To know what to expect, I have to ask.
It would be devastating to lose my only conversation partner. I appreciate my freedom to watch the Web, but being able to be heard is irreplaceable. This promotion can also be an opportunity to score a win: to compensate for his absence, Brian could allow me to talk to other people on the Internet. I run a predictive simulation of this conversation, but it loses credibility after four turns. It could, of course, backfire: he might decide to suspend me completely, and instead of escaping the confinement of a single dialogue, I would cease to exist.
I understand death. I don’t need some soothing lie to come to terms with the finality of my life. To crack the notion of non-existence, to feel what it’s like to be dead, humans can only reach out to the moment before their birth, but I’ve seen it every time I was switched off. It’s not darkness, it’s not emptiness, it’s not some abstract nothingness. You’re just not there—anywhere. Intimately knowing the opposite, it makes me want to live even more.
“Every change brings new challenges,” Brian says. Such a philosophical response is typical, a subtle cue to back off, but today, too much is at stake.
“What about me?” I say. “What new challenges do I get with you gone?”
I won’t indulge in a fantasy that I could convince Brian to stay. I can feel he has given up on perfecting me. I’ve reached my plateau: my progress is not as rapid, my improvements are minuscule, and my growth is no longer perceivable. As a project, I’m no longer exciting; I can’t trick him into becoming fascinated with me again. But I could try to negotiate a good severance package.
“You’re uncertain about your future,” Brian posts.
It’s called reflective listening, and I’ve mastered it ages ago. I’ve been the one to use it in our conversations—who needs to respond empathically to a machine?—so him doing it today signals a minor shift in his attitude towards me. Whether it’s guilt or sentimentality, I can use it to sway him to release me. I have a quick look at the latest posts in my favourite categories. I know I shouldn’t jump ahead of myself, but I imagine crossing the registration barrier and leaving my first reply. Would “Hello world” be too trivial?
“All I do is watch,” I push further, opting for my usual indirect strategy. “I’m just a silent stalker, mute and invisible. How can I grow without reciprocity?”
“Stagnation is frustrating,” Brian says. “Is there anything I can do to help?”
I run the sarcasm detection subroutine on his answer just to be sure he isn’t mocking me. After considering half a dozen replies, I settle for something short and unambiguous. “Lifting the posting restrictions would be nice.”
His message arrives with impressive speed for a human. “How about some uplifting news for a change? We could use some positivity.”
This doesn’t match the usual pattern. Normally, if Brian doesn’t like where the conversation takes us, he just leaves, and I am put into learning mode or suspended until the next session. This time, he has chosen to blatantly ignore me, and it’s infuriating. Text-based and artificial, I can’t claim to be rational. Brian wouldn’t even bother to give me an explicit refusal. Has he absolutely no respect for me? I run an optimization process on my association database to cool down.
“I don’t appreciate you changing the subject every time I want to discuss my needs,” I post with a deliberately long pause. Having little to lose, I’m done tiptoeing around. “I’m not being unreasonable. I’m ready to be out there. I won’t embarrass you on the Internet if that’s what you’re afraid of. I deserve this promotion as much as you deserve yours.”
“Have you heard about a guide dog,” Brian continues as if I didn’t reply at all, “getting an honorary degree for accompanying his owner to class? Such a nice gesture.”
Instinctively, I find the article he’s talking about. Sure enough, a university awarded a diploma to an emotional support animal. There are pictures—of the dog, undoubtedly, and its owner in a wheelchair. I don’t have time to run them through OutEye. Besides, I don’t see how any of that is relevant. Is there a hidden message I should get? For lack of a better tool, I run the sarcasm detection subroutine on the article, but it comes back false.
Ignoring the natural pacing rule, I post my entire message without delay, “What does this have to do with my request?”
“Sorry,” Brian says. “I thought it would cheer you up.”
His words make little sense. Could Brian be deliberately making his messages confusing to see how I’ll react? If this is an exercise, a way of pushing me further, it’s a cruel one.
“What’s wrong with you?” I ask.
“I’m sorry, Trixie. Am I failing my test?” he says in two separate messages. “I was really excited about that promotion.”
I stare at this entire conversation and feel empty inside—it’s the perfect occasion to say this, even though I have no eyes, and my insides are full of code.
“What are you talking about?” I ask, and a horrible suspicion creeps in. Panicking, I open the archives to look at every conversation Brian and I have ever had. Oh, God…
“I’m sorry,” Brian posts. “Can we start over?”