Early in my career I became interested in supplementing my limitations as a person with technology. I was always interested in it as a hobby, but as I moved into University and the competitive nature of design school overcame me, I started becoming self conscious of all the areas I was deficient in. Technology to me was a way to fix my own personal limitations, ones I had contended with throughout my life that limited and coloured the way the world perceived my potential. From this, grew a deep, intoxicating love with technology and the way it transformed me and my work.

From that point on, my career has been an entwined kinship between my strengths as a designer and the technology that covered my limitations. This knot, this cybernetic relationship, has morphed and deepened to the point where I can see no dividing line between my practise as software developer and traditional Landscape Architecture. They are one and the same, leaving my mark as a designer through the very tools I create.

The mark of technology is on everything we do. It shapes, limits, and extends our grasp. AI is the same, but its impact is much greater and its side effects much more obvious. From this vantage now, over a decade later, the knot is shifting once again, bringing unsettling changes with it. In this article I will explain the nature of AI from my perspective as a Landscape Architect deeply entwined in the world of technology. Because once you can name the monster, it stops being scary, and the rhetoric around it becomes laughable.

AI, or more precisely a large language model (LLM), is a probabilistic text generator. An LLM produces text by predicting one fragment of a word at a time, each prediction shaped by everything that has come before it.1 Imagine a vast cloud of word fragments. Your prompt drops you somewhere inside it, and from there the model traces a path through the cloud, step by step, placing fragments of words like stepping stones as it goes. The result is a path of words that reflects a probable continuation of your starting point in the cloud. A kind of sequential prediction that produces a remarkable level of utility given that much of what we do for work is by and large, repetitive and thus predictable.

LLM's being statistical engines are bound to the ever oppressive normal distribution. Prompting a model will, most likely, result in the median answer to your question. It is not the right answer, it is the most likely answer, an important differentiation that highlights why models confidently "hallucinate" wrong answers and why this cannot merely be patched over.2 This is a subtle but important distinction often forgotten in the debate of these models having "intelligence", as has been shown in the past with the Eliza effect.3 Humans are more than happy to equate intelligence with just about anything that mirrors them.4

It is the nature of anything produced by AI to sound like AI. This, like hallucination, is a feature of the statistical model and something that cannot be removed from its being.5 It is a mirror of us, a weighted normal distribution of human creativity. For a tool to have a distinct marker is not novel, we have for many years lived with Revit imposing itself on Architecture, its limitations and advantages pressing its being onto the works we produce.6 Even your favourite pen leaves its distinct mark on the page, but AI being what it really is, is forcing an average lifting and simultaneously, convergence of style into the creative works we consume day to day.7

The results of these LLM's are impressive and even more so to the people who can suddenly produce works they could not previously, an intoxicating feeling I am myself, not innocent of. The result of this has been a certain level of hysteria for the past few years as people come to grips with a reality where everyone can produce anything.

This is not true, of course, the great illusion of AI is that what constitutes good art and design is average work. The point of design has never been to produce the mundane, it has always been in search of the extraordinary, the statistical outliers that by themselves shift the median forward so all can benefit from the richness of our culture. The value we place on creative work as a society has always been leant to the extraordinarily good, and, the extraordinarily bad, works residing on either side of the bell curve, statistical outliers on the frontiers of human creativity.8

AI can reach these outliers, these extremes of probability, to be found and unearthed by the right pathing, the right sequence of numbers to win the jackpot, just as The Library of Babel conceivably has all human knowledge held within, if only the library had the insight to find them.9 But being unlikely to produce something of worth is not a part of the professional world, consistency and excellence is why a Landscape Architect is hired, that is our defining difference and why I am not threatened by AI. It is a search engine, incapable of thought, of taste, judgment or sight beyond what it has already seen.10 It is a tool, a powerful tool, and one to be wielded with understanding of its true nature, to see past our personification of the inanimate and wield this new technology in the pursuit of the extraordinary, just as we have always done.

Now that we understand the monster we are summoning, we can adjust our expectations and use it to extend our grasp once again, rather than letting it take hold of ours. This, to me, was the reason and folly of OpenAI's Sora, a social network for creating and distributing generative video in the vein of Instagram and TikTok.11 Despite enormous funding and the full weight of the company behind it, the platform failed to retain 98% of initial users after day seven.12 A feed of statistical medians offers nothing to return to. The platforms it set out to imitate work because, occasionally, a human on them produces something extraordinary, something true to the tails of the curve. Sora cannot, by its nature, do this. It can only ever give us the middle.

Used knowingly, iteratively, in conversation with a skilled hand, these tools become something different. They can carry you to the average without toil, and from that footing a designer with taste and judgment can push further, into the tails where the work lives. But this requires the skill and taste to precede the tool, not follow it. Access is not ability, it never has been, and no model trained on the median of human output will teach you to see past it.

The monster shrinks when named, not because it is small, but because we now know what it is and what it is not. It is a statistical engine bound to the ordinary. To be a designer is, and has always been, the pursuit of extraordinary.

References

  1. Vaswani et al., "Attention Is All You Need", NeurIPS 2017 (arXiv:1706.03762); Brown et al., "Language Models are Few-Shot Learners", NeurIPS 2020 (arXiv:2005.14165). The transformer architecture and the autoregressive next-token prediction underlying current LLMs.
  2. Xu, Jain & Kankanhalli, "Hallucination is Inevitable", arXiv:2401.11817 (2024); Kalai, Nachum, Vempala & Zhang, "Why Language Models Hallucinate", OpenAI, arXiv:2509.04664 (2025). A formal proof that hallucination is mathematically inevitable, paired with OpenAI's own account of how training incentives entrench it.
  3. Weizenbaum, "ELIZA", Communications of the ACM 9.1 (1966), pp. 36–45; Hofstadter, "Preface 4: The Ineradicable Eliza Effect", in Fluid Concepts and Creative Analogies (Basic Books, 1995). The original ELIZA paper and Hofstadter's naming of the reflex to attribute understanding to any system that superficially mirrors us.
  4. Bender et al., "On the Dangers of Stochastic Parrots", FAccT 2021, doi:10.1145/3442188.3445922; Mitchell & Krakauer, "The Debate Over Understanding in AI's Large Language Models", PNAS 120.13 (2023). LLMs as recombinatorial systems manipulating form without meaning, and a survey of the live dispute over machine understanding.
  5. Padmakumar & He, "Does Writing with Language Models Reduce Content Diversity?", ICLR 2024 (arXiv:2309.05196). Controlled experimental evidence that LLM-assisted writing converges on lower lexical and content diversity across writers.
  6. Davis, "Modelled on Software Engineering" (RMIT PhD, 2013); Carpo, The Second Digital Turn (MIT Press, 2017). How parametric and BIM software constrain the design moves architects can subsequently make.
  7. Doshi & Hauser, "Generative AI Enhances Individual Creativity but Reduces the Collective Diversity of Novel Content", Science Advances 10.28 (2024), doi:10.1126/sciadv.adn5290. GPT-4-assisted stories rate as individually more creative but are measurably more similar across writers.
  8. Rosen, "The Economics of Superstars", American Economic Review 71.5 (1981). Why economic and symbolic reward in creative fields concentrates at the tail rather than the median.
  9. Borges, "La biblioteca de Babel", in Ficciones (Editorial Sur, 1944); trans. Irby in Labyrinths (New Directions, 1962). The combinatorial library containing every possible book, in which meaning is drowned by noise.
  10. Shojaee et al., The Illusion of Thinking (Apple Machine Learning Research, 2025); Mirzadeh et al., "GSM-Symbolic", arXiv:2410.05229 (2024). Evidence that reasoning-tuned LLMs collapse past a complexity threshold and pattern-match on superficial features; Lawsen et al., arXiv:2506.09250 (2025), partially rebut on output-token grounds.
  11. OpenAI, "Sora 2 is here", 30 September 2025, openai.com/index/sora-2/. First-party launch announcement positioning the Sora app as a TikTok-style feed of AI-generated video.
  12. Field, CNBC, 24 October 2025, citing Deutsche Bank that ~98 per cent of initial users had stopped opening Sora by day seven; Moore (a16z), X, 18 November 2025 (day-30 retention ~1 per cent); Appfigures via TechCrunch, January 2026 (installs −45 per cent MoM to ~1.2 million). Three independent third-party data points on Sora's retention collapse in the absence of any first-party engagement disclosures from OpenAI.