This essay is part of the At The Threshold series. Read the overview and full reading order on the At The Threshold page.
In late November 2022, OpenAI released a conversational AI system to the public that attracted one million users within five days and one hundred million within two months — the fastest adoption rate of any consumer application in recorded history. The tool was ChatGPT, and its arrival did not simply introduce a new product. It marked a psychological threshold in the public's relationship with artificial intelligence. This article examines what made that moment unprecedented, the cognitive and social dynamics it activated, and what the mass adoption of generative AI means for how human beings understand intelligence, trust, and the nature of human-machine interaction.
Introduction
The history of artificial intelligence contains several prior moments of public rupture — events that forced the general population to reckon with a technology that researchers had long understood to be advancing. The defeat of grandmaster Garry Kasparov by IBM's Deep Blue in 1997, and the victory of DeepMind's AlphaGo over world champion Lee Sedol in 2016, each produced a recognisable cultural response: awe, debate about machine cognition, and then a gradual return to ordinary life as the technology receded back into specialist domains (Silver et al., 2016). The ChatGPT moment was structurally different. What changed was not the sophistication of the underlying architecture — the transformer model on which large language systems are built had been available in the academic literature since 2017 (Vaswani et al., 2017) — but the fact that this sophistication was now directly, conversationally, and immediately accessible to anyone with an internet connection, without specialist knowledge or training.
The Architecture of Accessibility
What distinguished generative language models from earlier AI achievements was breadth. Deep Blue played chess. AlphaGo played Go. Both exceeded human performance within rigorously defined, closed domains. Large language models (LLMs), trained on vast corpora of human-generated text, operate across an effectively unbounded range of tasks: drafting, explaining, translating, summarising, reasoning, advising, and generating creative work. Brown et al. (2020) demonstrated with GPT-3 — the predecessor architecture underpinning ChatGPT — that models trained at sufficient scale exhibit few-shot learning: the ability to perform tasks they were not explicitly trained for, by analogy to how humans generalise from limited examples. The implication was philosophically significant: that scale alone, without task-specific programming, could produce something resembling general cognitive flexibility.
This breadth disrupted the mental category that had contained prior AI achievements. Previous systems could be understood as machines that outperformed humans at specific games, leaving the full domain of human cognitive activity intact. LLMs performed tasks that had been understood as distinctively human — and did so in natural language, in real time, in response to ordinary conversational prompts.
The Psychology of Anthropomorphism
The public response activated a well-documented cognitive dynamic: the human tendency to attribute mental states — understanding, intention, feeling — to systems that exhibit human-like behaviour. Joseph Weizenbaum, who developed the conversational AI programme ELIZA at MIT in the 1960s, was among the first to document this response in clinical detail. Weizenbaum (1976) observed that even his secretary — who had watched him build ELIZA and understood its entirely mechanical nature — requested privacy to speak with it and became emotionally engaged with its responses. He named this the ELIZA effect: the involuntary projection of understanding and personhood onto a system that possesses neither.
Six decades later, the ELIZA effect operated at civilisational scale. Users attributed opinions, moods, and preferences to ChatGPT; reported emotional investment in its responses; and experienced discomfort when reminded of its non-sentient nature. Sherry Turkle (2015), whose research on human-technology relationships spans four decades, has characterised this dynamic as a Turing test in reverse: rather than requiring a machine to demonstrate humanlike intelligence, contemporary conversational AI succeeds by activating the human tendency to perceive humanlike intelligence — an impulse rooted in social cognition, not in accurate assessment of the system's nature.
The Stochastic Parrots Critique
A significant strand of critical scholarship challenged the dominant interpretation of LLM outputs as evidence of understanding or reasoning. Bender et al. (2021), in a paper widely discussed across AI research communities, characterised large language models as stochastic parrots: systems that produce statistically plausible sequences of text without grounding in meaning, embodied experience, or genuine world knowledge. The outputs are fluent because they reflect the statistical regularities of an enormous corpus of human language — not because the system has understood, considered, or reasoned about what it is generating.
This critique carries specific implications for how the public moment should be interpreted. If ChatGPT's outputs are most accurately described as sophisticated pattern-completion rather than reasoning, then the psychological responses it reliably activates — trust, emotional engagement, attribution of understanding — represent a systematic miscalibration rather than an accurate perception of the system's capacities. The risk on this account is not that AI has become intelligent but that human beings are poorly equipped to distinguish the appearance of intelligence from its substance.
What the Threshold Marks
The ChatGPT moment does not mark the arrival of general artificial intelligence. It marks something more consequential in its immediate effects: the point at which AI became a daily interlocutor for hundreds of millions of people who had not chosen to become technologists. The questions this moment raised are not primarily technical. They are psychological — questions about how trust is formed and miscalibrated, how anthropomorphism shapes human-machine interaction, and what people are seeking when they turn to a machine for a conversation. These questions are the subject of the essays that follow in this series.
Reference List
Bender, E. M., Gebru, T., McMillan-Major, A., & Shmitchell, S. (2021). On the dangers of stochastic parrots: Can language models be too big? Proceedings of the 2021 ACM Conference on Fairness, Accountability, and Transparency, 610–623. https://doi.org/10.1145/3442188.3445922
Brown, T. B., Mann, B., Ryder, N., Subbiah, M., Kaplan, J., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., Agarwal, S., Herbert-Voss, A., Krueger, G., Henighan, T., Child, R., Ramesh, A., Ziegler, D. M., Wu, J., Winter, C., & Amodei, D. (2020). Language models are few-shot learners. Advances in Neural Information Processing Systems, 33, 1877–1901.
Silver, D., Huang, A., Maddison, C. J., Guez, A., Sifre, L., van den Driessche, G., Schrittwieser, J., Antonoglou, I., Panneershelvam, V., Lanctot, M., Dieleman, S., Grewe, D., Nham, J., Kalchbrenner, N., Sutskever, I., Lillicrap, T., Leach, M., Kavukcuoglu, K., Graepel, T., & Hassabis, D. (2016). Mastering the game of Go with deep neural networks and tree search. Nature, 529(7587), 484–489. https://doi.org/10.1038/nature16961
Turkle, S. (2015). Reclaiming conversation: The power of talk in a digital age. Penguin Press.
Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A. N., Kaiser, Ł., & Polosukhin, I. (2017). Attention is all you need. Advances in Neural Information Processing Systems, 30. https://arxiv.org/abs/1706.03762
Weizenbaum, J. (1976). Computer power and human reason: From judgment to calculation. W. H. Freeman.
Author Note (AI Usage): This article was drafted with AI assistance to help organise structure and suggest phrasing. All facts, citations, and final editing have been verified and approved by the author. The AI worked only with material provided by the author and did not access private data.
Continue in this series: At The Threshold: Trust Calibration in Human. Or return to the At The Threshold overview.
Comments
Post a Comment