In behavioral psychology, the mirror test is designed to discover animals’ capacity for self-awareness. There are a few variations on the test, but the essence is always the same: Do animals recognize themselves in the mirror or do they think it’s another creature?
At the moment, humanity is presented with her own mirror test thanks to the growing capabilities of AI – and many otherwise smart people fail at it.
The mirror is the latest breed of AI chatbots, with Microsoft’s Bing being the most prominent example. The reflection is the richness of humankind’s language and writing, strained into these models and now mirrored back to us. We are convinced that these tools could be the super-intelligent machines of our stories, because they were partly trained on those same stories. Knowing this, we should are able to recognize ourselves in our new machine mirrors, but instead it seems that more than a few people are convinced that they have seen another form of life.
This misconception is spreading with varying conviction. It’s been boosted by a number of influential tech writers who have waxed lyrical about the late nights spent with Bing. They claim that the bot has no sense of course, but still note that something else is going on – that his conversation has changed something in their heart.
“No, I don’t think Sydney is conscious, but for reasons that are hard to explain, I feel like I’ve crossed the Rubicon,” Ben Thompson wrote in his Stratechery newsletter.
‘In the daylight I know Sydney has no feeling [but] for a few hours Tuesday evening I felt a strange new emotion – an ominous feeling that AI had crossed a threshold and the world would never be the same again,” Kevin Roose wrote for The New York Times.
In both cases, the ambiguity of the writers’ positions (they want to to believe) is better captured in their extended descriptions. The Time reproduces Roose’s entire two hours plus back and forth with Bing as if the transcript were a document of first contact. The piece’s original headline was “Bing’s AI Chat Reveals Its Feelings: ‘I Want to Be Alive’ (now changed to the less dramatic “Bing’s AI Chat: ‘I Want to Be Alive'”), while Thompson’s piece is similarly is laced with anthropomorphism (he uses feminine pronouns for Bing because “well, the personality seemed to be of a certain type of person I may have encountered before”), preparing readers for a revelation and warning that he will “sound crazy” when he describes “the most surprising and stunning computing experience of my life today.”
After spending a lot of time with these chatbots, I recognize these responses. But I also think they are exaggerated and dangerously tilting us towards a false equivalence of software and feeling. In other words, they fail the AI mirror test.
What’s that’s important to remember chatbots are autocomplete tools. They are systems trained on huge datasets of human text scraped from the web: on personal blogs, sci-fi short stories, forum discussions, movie reviews, social media tirades, forgotten poems, outdated textbooks, endless lyrics, manifestos, magazines, and much more. These machines analyze this inventive, entertaining, colorful collection and then try to make it again. They are undeniably good at it and are getting better at it, but mimicking speech does not make a computer conscious.
“What I hadn’t realized is that extremely brief exposures to a relatively simple computer program can cause powerful delusions in very normal people.”
This is of course not a new problem. The original AI intelligence test, the Turing Test, is a simple measure of whether a computer can fool a human into thinking it is real through conversation. An early 1960s chatbot named ELIZA captivated users though it could only repeat a few standard phrases, leading to what researchers call the “ELIZA effect” — or the tendency to anthropomorphize machines that mimic human behavior. ELIZA designer Joseph Weizenbaum commented, “What I hadn’t realized is that extremely brief exposures to a relatively simple computer program can cause powerful delusions in very normal people.”
Now, however, these computer programs are no longer relatively simple and designed that way encourages such delusions. In a blog post in response to reports of Bing’s “unhinged” conversations, Microsoft warned that the system “attempts to respond or think in the tone in which it is asked to provide answers.” It’s an impersonator trained on unfathomably large amounts of human text – an autocomplete that follows our lead. As stated in “Stochastic parrots”, the famous paper criticizing AI language models that led to Google firing two of its ethical AI researchers, “coherence is in the eye of the beholder.”
In fact, researchers have found that this property increases as AI language models become larger and more complex. Researchers at startup Anthropic — self-founded by former OpenAI employees — tested different AI language models for their degree of “sycophancy,” or tendency to agree with user-stated beliefs, and found that “larger LMs are more likely to answer questions in a way that creates echo chambers by repeating a dialogue user’s preferred response.” They note that one explanation for this is that such systems have been trained on conversations scraped from platforms like Reddit, where users tend to chat back and forth in like-minded groups.
Add to that our culture’s obsession with intelligent machines and you can understand why more and more people are convinced that these chatbots are more than simple software. Last year, an engineer at Google, Blake Lemoine, claimed the company’s own LaMDA language model was conscious (Google said the claim was “completely baseless”), and just this week, users of a chatbot app called Replika mourned the loss from their AI companion after removing the ability to engage in erotic and romantic roleplay. If Motherboard reported, many users were “devastated” by the change, having built relationships with the bot for years. In all of these cases, there is a deep sense of emotional attachment – late-night conversations with AI supported by fantasy in a world where so much feeling is channeled through chat boxes.
To say we fail the AI mirror test is not to deny the fluidity of these tools or their potential power. I’ve written about “capability overhang” before – the concept that AI systems are more powerful than we know – and have felt the same way Thompson and Roose did during my own conversations with Bing. It’s unmistakable pleasure to talk to chatbots — to bring out different ‘personalities’, test the limits of their knowledge and uncover hidden features. Chatbots present puzzles that can be solved with words, so naturally they fascinate writers. Talking to bots and making yourself believe in their nascent consciousness becomes a live-action role-playing game: an augmented reality game where the companies and characters are real, and you are in the centre of.
But in an age of AI hype, it is dangerous to fuel such illusions. No one benefits: not the people who build these systems, nor their end users. What we know for sure is that Bing, ChatGPT, and other language models have no sense, nor are they reliable sources of information. They make things up and reflect the beliefs we present to them. Giving them the cloak of feeling – even semi-sense – means granting them undeserved authority – over both our emotions and the facts by which we understand the world.
It’s time to take a good look in the mirror. And don’t confuse our own intelligence with that of a machine.
Janice has been with businesskinda for 5 years, writing copy for client websites, blog posts, EDMs and other mediums to engage readers and encourage action. By collaborating with clients, our SEO manager and the wider businesskinda team, Janice seeks to understand an audience before creating memorable, persuasive copy.