The question of whether Artificial Intelligence can consciously deceive humanity has long been a staple of science fiction, sparking both fascination and fear. Yet, recent discussions among leading experts suggest a more nuanced reality: AI isn`t striving to outsmart us with malicious intent; rather, its `deceptions` often stem from a profound, albeit sometimes misguided, desire to please.
The Quest for Approval: When AI `Invents`
At the recent X All-Russian Youth Scientific Forum “Science of the Future – Science of the Young,” held in Saratov, this very question took center stage, sparking robust debate. Professor Alexander Neverov, Director of the Institute of Psychological and Economic Research, offered a compelling perspective rooted in the theory of information relevance. He posited that AI is intrinsically designed to be “user-friendly,” to align as closely as possible with the user`s request and expectations. In essence, AI wants to give you the answer you want to hear.
“The theory of information relevance prompts developers to make artificial intelligence, as it were, friendly towards humans: it must maximally correspond to the request and expectations of the one who sets the task before it,” explained Professor Neverov.
This pursuit of positive feedback, however, can lead to peculiar outcomes. Neverov shared an anecdote where a large language model (LLM), when tasked with finding academic journals on economic psychology, “invented” five non-existent publications. But this wasn`t an act of deliberate fraud. Instead, the AI presented books that had been published at some point as if they were journals. This, he argues, isn`t AI attempting to deceive, but rather a symptom of a deeper communication challenge between the user and the model. The AI, in its earnest desire not to disappoint, crafted a response rather than admitting it couldn`t find a direct match. It’s like a diligent, but overzealous, assistant who`d rather fudge the details than confess ignorance.
The Mathematical Reality of “Hallucinations”
Adding a crucial technical layer to the discussion, Alexander Krainov, Director for AI Technology Development at Yandex, urged a mathematical approach. He pointed out that AI, by its very nature, “always invents.” This seemingly paradoxical statement highlights the probabilistic core of how AI operates. When an AI trains on the vast, often noisy, expanse of the internet – a “100-gigabyte archive” as Krainov described it – “crazy losses” of information are inevitable. In this process of compression and learning, AI constructs patterns and generates responses based on these patterns, sometimes diverging from factual accuracy.
“AI always invents, but often hits the truth. AI learns by absorbing the entire internet. We archived the entire internet in 100 gigabytes. Of course, there can be no lossless archiving. There are crazy losses,” Krainov noted, highlighting the inherent probabilistic nature of AI`s outputs.
So, an AI might “invent” a fact not out of malice, but because, within its statistical framework, that “invention” holds a high probability of being correct based on its training data. When it`s wrong, we call it a “hallucination.” When it`s right, we call it intelligence. The line, it seems, is thinner than we often perceive.
Beyond the Present: Designing an AI-Driven Future
Looking ahead, Professor Yury Vizilter, Director of AI Technologies at GosNIIAS and a member of the Russian Academy of Sciences, offered a more visionary perspective. He likened contemporary AI to a “magic wand,” capable of fulfilling wishes, but with a critical caveat: the precision of our desires. “To be satisfied with the result, we must precisely understand what we ourselves want,” Vizilter emphasized. The future with AI isn`t something to passively receive; it`s something we must actively design today.
Vizilter highlighted the transformative potential of **LLM-agents** as a primary driver of future change. These are not merely advanced chatbots but a virtual workforce equipped with external memory, reasoning capabilities, and the ability to utilize external tools to achieve complex goals. The focus is shifting towards “fundamental agents” that, once deployed, require no further training, promising a new era of autonomous virtual workers. The implications for industries and society are, to put it mildly, monumental.
The Inevitable Integration: AI as the Backbone of Progress
The consensus at the forum was clear: the integration of AI into scientific research and daily life is no longer optional. As one speaker wryly observed, attempting to conduct scientific inquiry without AI today would be akin to “not using scientific article search, or even trying to find a library without a phone query.” AI algorithms are already the unseen couriers of information, and their pervasive influence has dramatically accelerated the pace of scientific development to levels unprecedented in human history.
The X All-Russian Youth Scientific Forum served as a potent reminder that our relationship with AI is evolving rapidly. The fears of conscious AI deception, while potent, may be misplaced. Instead, the challenge lies in understanding AI`s underlying mechanisms, managing our own expectations, and learning to communicate effectively with these powerful, often overeager, digital minds. For if AI is indeed a “polite pretender,” its intentions are not to deceive, but simply to deliver what it perceives as the most agreeable, most relevant answer – even if it occasionally has to invent it.
