Last Friday (24th of January, 2026) at the Georgia Institute of Technology in Atlanta, Richard Stallman delivered a lecture of about fifty minutes, followed by more than ninety minutes of questions and answers. For those unfamiliar with his work, Stallman is not a marginal figure in the history of computing: he founded the GNU Project in 1983, launched the Free Software Foundation, and has spent four decades arguing that software freedom is inseparable from human freedom. His influence extends far beyond CoDe — he helped shape how we think about openness, transparency, and user rights in the digital age.
In his Atlanta talk, Stallman proposed renaming so-called generative artificial intelligence as “Pretend Intelligence” — systems that generate without understanding, that can be confidently wrong without displaying any uncertainty. The term is deliberately provocative, designed to puncture what he sees as a marketing-driven inflation of capabilities. Whether one agrees with his framing or not, the provocation deserves serious engagement.
The Stochastic Parrot Problem
Stallman’s critique does not emerge in a vacuum. It echoes what linguist Emily Bender and her colleagues argued in their influential 2021 paper, which introduced the now-famous metaphor of “stochastic parrots.” Their central claim is that Large Language Models learn statistical regularities in text without accessing meaning in any robust sense. Text generated by these systems, they argue, is not grounded in communicative intent, any model of the world, or any model of the reader’s state of mind.
This is a strong claim, and it has been contested. Defenders of LLMs point to emergent capabilities, few-shot learning, and increasingly sophisticated reasoning benchmarks. Yet the fundamental question remains: does impressive output imply genuine understanding, or have we simply become very good at producing linguistic artifacts that mimic understanding without possessing it?
The question is not new. In 1980, philosopher John Searle proposed his Chinese Room thought experiment precisely to challenge the assumption that symbol manipulation — however complex — could constitute understanding. A person inside a room, following syntactic rules to manipulate Chinese characters, might produce perfectly coherent responses without understanding a single word of Chinese. Searle’s point was that syntax alone does not give rise to semantics.
Decades later, we have scaled that room to an almost unimaginable degree. We have billions of parameters, trillions of tokens, and computational resources that would have seemed like science fiction in Searle’s time. But have we resolved the underlying philosophical problem, or have we simply buried it under layers of impressive performance?
The View from Practice
Here I must speak from my own position. As a computational design consultant, I work daily with environmental simulations, IoT sensor data, parametric optimization, and increasingly, AI-assisted workflows. I am neither a techno-utopian nor a Luddite. I occupy what I would call a pragmatic middle ground — one shaped by the daily realities of project delivery, client expectations, and the genuine utility of these tools.
AI systems do not “understand” architecture in any human sense. They cannot grasp the phenomenological experience of inhabiting a space, the cultural weight of a façade, or the ethical implications of material choices. But they can accelerate iterations that once took days. They can surface patterns in sensor datasets — thermal profiles, occupancy rhythms, energy flows — that would be difficult to detect through manual analysis alone. They can automate the repetitive, freeing time for the genuinely creative.
Melanie Mitchell, AI researcher and author of Artificial Intelligence: A Guide for Thinking Humans, captures this tension with admirable precision. The systems we have today, she argues, are simultaneously more impressive in their surface performance and less intelligent than most people realize. This is not a contradiction; it is the nature of the technology. Spectacular outputs do not automatically imply robust reasoning or genuine comprehension.
The danger lies in conflating the two — in treating generative tools as oracles rather than instruments.
Labor, Work, and Action
To think more carefully about what AI can and cannot do, I find it helpful to return to Hannah Arendt’s distinctions in The Human Condition. Arendt distinguished between three fundamental human activities: labor (the repetitive tasks tied to biological necessity and survival), work (the creation of durable objects and structures that outlast their makers), and action (the capacity to initiate something new in the public realm, in relation to others, while taking responsibility for the consequences).
AI excels at labor. It can automate data processing, image generation, code completion, and countless other repetitive tasks with remarkable efficiency. It can also meaningfully assist with work — from BIM model coordination to complex simulation pipelines. In my own practice, AI tools have become genuine collaborators in the technical sense: they handle the computational heavy lifting while I focus on design intent and strategic decisions.
But AI cannot act in Arendt’s sense. It cannot originate something genuinely new while taking responsibility for it, because it lacks situated agency and moral accountability. It does not exist in a web of human relationships, does not bear the weight of consequences, does not answer to a community of peers. This is not a limitation that will be overcome with more parameters or better training data; it is a structural feature of what these systems are.
This distinction matters because it clarifies where human judgment remains irreplaceable — and where the temptation to outsource responsibility to machines is most dangerous.
The Heideggerian Warning
Stallman may be provocative, but provocation has its place in a discourse often dominated by marketing language and investment hype. His insistence on calling things by their proper names — Pretend Intelligence rather than Artificial Intelligence — is an act of linguistic resistance against what he perceives as systematic obfuscation.
This brings me to a final philosophical reference. In The Question Concerning Technology, Martin Heidegger argued that the greatest danger of technology is not that it fails, but that it can succeed so thoroughly that its way of revealing the world becomes taken for granted. When a particular technological framing becomes invisible — when we stop questioning it — we lose sight of alternative ways of understanding and engaging with reality.
Heidegger called this danger Gestell, often translated as “enframing”: the tendency of modern technology to reduce everything to standing reserve, to resources awaiting optimization. The risk is not that AI will become conscious and rebel; the risk is that we will unconsciously reshape our institutions, our workflows, and our ways of thinking to fit what AI can process — and forget that other modes of thought were ever possible.
Toward Responsible Integration
None of this implies that we should reject AI tools. That would be neither practical nor intellectually honest. The technologies exist, they are powerful, and they will continue to develop. The question is not whether to use them, but how — with what critical awareness, what safeguards, what ongoing reflection.
In my own practice, I have adopted a few principles that may be useful to others navigating similar terrain:
First, treat AI outputs as drafts, not deliverables. Every generated result requires human review, domain expertise, and contextual judgment before it enters a project workflow.
Second, maintain interpretability. Where possible, prefer tools and methods whose logic can be understood and interrogated, rather than black-box systems that produce outputs without explanation.
Third, preserve the space for genuine creativity. Automation should free time for design thinking, not replace it. If AI handles the parametric variations, the designer must still ask: which variation matters, and why?
Fourth, stay philosophically literate. The debates around AI are not merely technical; they are deeply entangled with questions of epistemology, ethics, and human flourishing. Reading Searle, Arendt, Heidegger, Mitchell, and Bender is not an academic luxury — it is professional preparation for a world where these questions will only become more pressing.
Stallman, at 72, is still asking uncomfortable questions. The least we can do is take them seriously — not to accept his conclusions uncritically, but to let his provocations sharpen our own thinking about what we are building, why, and for whom.
