Augmented Intelligence
From a Website Redesign to AI Displacement
Back in 2019, I worked on a complete redesign of my personal website. Coming up with the design language and my personal brand was a lot of fun. Coding it, however, was a mixed bag. There were great moments (I have always loved writing React), but also tedious chores like wrangling build configurations to deploy to Netlify. All in all, it took a few weeks of after-hours work.
Fast forward to last weekend: I decided to refresh the site, refactor the codebase, add a newsletter, launch a blog with a custom admin panel, and build a dedicated landing page for my upcoming book. The entire process took less than two hours. Between visual changes steered via Claude Design, and execution and deployment handled by Claude Code, it felt effortless. I was freed up to focus entirely on quality, direction, and steering the features I needed.
Weeks of work compressed into hours. It felt incredible, and a little unnerving.
Like many of us in tech, I occasionally worry about AI displacing human labor. After all, tasks I used to perform manually are now executed by a machine in seconds. But that perspective misses the point. My entire job was not replaced; specific tasks were automated. In reality, I became augmented by the machine. That realization pushed me to look past the hype and dig into the history of human-computer interaction, as well as the current data on AI job displacement.
A Time Study from 1957
In the spring and summer of 1957, J.C.R. Licklider kept a meticulous log of his own workday. As a psychoacoustician managing research at Bolt Beranek and Newman, he wanted to see where his intellectual energy actually went. Lacking a better subject, he chose himself.
To his surprise, roughly 85 percent of what he considered thinking time was spent simply getting into a position to think: tracking down references, plotting graphs, explaining to assistants how to plot them, or spending hours reconciling experiments measured on different scales. Once the numbers were standardized, arriving at the actual insight took seconds.
His most striking observation is one I keep returning to. Licklider admitted that his choices about which scientific problems to pursue were shaped to an embarrassing degree by what was clerically feasible, rather than what he was intellectually capable of solving.
Something similar happened to me over the weekend. The work went faster, but more importantly, I built things I would not have attempted in 2019. Back then, a custom admin panel and a bespoke blog engine sat on the far side of an invisible line, a line where a good idea ceases to be worth the evenings it costs to build. Figuring out Netlify build configurations was my modern equivalent of plotting graphs by hand. The hours it consumed mattered less than the ambitious ideas it quietly disqualified.
Licklider published his findings in his landmark 1960 paper, "Man-Computer Symbiosis," proposing machines that would take on clerical preparation so humans could spend their days on the parts that require human cognition. Two years later, Douglas Engelbart formalized this into a research agenda titled "Augmenting Human Intellect", an effort that gave us the computer mouse, bitmapped displays, and hypertext. Making humans more capable was one of the two founding ambitions of computer science, alongside building autonomous intelligence.
I heard Nvidia CEO Jensen Huang reach for this exact intuition in a recent interview while describing how his engineers work. He called it "augmented development," grabbing the ordinary phrase that best described what he observed across his company's desks. I found the term accurate and delicate. Instead of chasing pure automation, his organization focuses on human augmentation.
Is this emphasis on symbiosis representative of how the broader tech industry views generative AI, or is it an outlier? I turned to economic data to find out.
The Reality "Right Here, Right Now"
In a working paper for the Stanford Digital Economy Lab, Erik Brynjolfsson, Bharat Chandar, and Ruyu Chen analyzed ADP payroll records covering millions of American workers. They found that early-career workers aged 22 to 25 in the most AI-exposed occupations experienced a 16 percent relative decline in employment following the widespread adoption of generative AI. Crucially, this adjustment occurred not through active layoffs, but through reduced hiring. Entry-level roles are simply being left unfilled.
If you work in a high-exposure knowledge sector, the ladder is missing its bottom rungs.
Is this shift driving clear business gains? Are companies cutting entry-level roles because AI has made their operations drastically more productive? The data does not support that assumption. Research by Fiona Chen and James Stratton at Harvard examined roughly 200 million work events across 500 engineering organizations. They observed an 8.5 percent increase in coding activity and an 8.7 percent faster task completion time among engineers using tools like GitHub Copilot and Cursor. However, they found no statistically significant impact on total business output or revenue.
Individual tasks get completed faster, but overall organizational output remains flat.
We are currently navigating the fog of war in the early days of an AI-driven economic transition. Business leaders are reacting to a narrative rather than reality. Influenced by headlines predicting full autonomy, companies are cutting entry-level headcount before seeing true productivity breakthroughs. In turn, product teams build tools designed for total task delegation rather than human augmentation. It threatens to become a self-fulfilling prophecy.
The Shape of Interaction is a Design Choice
In mid-2026, Anthropic published a comparison from their Economic Index report that I have not stopped thinking about. Looking at how users completed similar writing tasks across two different surfaces powered by the same underlying model, they found a stark contrast:
- The median chat session producing a blog post involved 13 back-and-forth turns between human and model.
- The median agentic environment (Claude Code) producing a text artifact contained a single initial prompt.
The underlying model was identical; the interaction surface determined the nature of the work.
When I refactored my website, I used an automated agentic surface, yet I experienced it as an augmented dialogue. I constantly intervened, steered, rejected outputs, and treated every generated draft as a starting point rather than a final product. The software leaned toward total automation, but I leaned back toward human oversight.
We cannot rely on users to fight their tools. Interfaces set default behaviors. A system that delivers a finished artifact encourages passive acceptance. A system that presents malleable, inspectable work invites critique and iteration. The distance between those two experiences is not determined by breakthroughs in machine learning. It is determined by UX designers and Product Managers deciding how the interface tilts the floor.
Designing for Symbiosis
As product leaders, designers, and engineers we really should stop building interfaces that treat human beings as passive consumers of AI output, and start building products engineered for true human-machine symbiosis.
If we default to building tools that turn humans into button-pushers who blindly accept generated outputs, we will erode domain expertise, degrade work quality, and shrink human agency. If we build for symbiosis, we remove the clerical drag on human creativity, unlocking problems we previously lacked the bandwidth to attempt.
To build products that elevate human capability rather than replace it, we can anchor our work on four core design principles:
1. Design for Steering, Not Just Prompting
Prompting is an instruction given at the start. Steering is continuous control applied throughout the process. Symbiotic products provide real-time controls, inline branch points, and mid-flight interventions that allow users to shape the outcome as it forms.
2. Expose the Scaffolding of Thought
Never present complex AI outputs as an unexplainable fait accompli. Make the underlying reasoning, sources, assumptions, and trade-offs visible and editable. When users can see how a machine arrived at a proposal, they can critique, refine, and learn from it.
3. Keep the Feedback Loop Frictionless
The moment an interface makes it difficult to challenge, adjust, or tweak an AI output, users default to passive approval. Treat every generated output as a draft. Provide lightweight micro-interactions, such as single-click alternatives, granular inline edits, or targeted re-prompts, to make iteration effortless.
4. Protect the Thinking Time
Automation should eliminate clerical setup (the modern equivalent of plotting graphs by hand) so humans can focus on evaluation, strategy, and domain judgment. Measure your product's success not by how long it operates without human involvement, but by how effectively it equips the user to make high-leverage decisions.
There's lots of work for us left to do to make AI perform well in the symbiotic relationship with people.
ps. Premium subscribers to my Substack Newsleter: Augmentation: UX, Product, AI will receive a full prompt by prompt history of my website redesign. You'll be able to see how I steered Claude towards the right results.
References
Autor, D. H. (2015). Why are there still so many jobs? The history and future of workplace automation. Journal of Economic Perspectives, 29(3), 3–30. https://doi.org/10.1257/jep.29.3.3
Brynjolfsson, E., Chandar, B., & Chen, R. (2025). Canaries in the coal mine? Six facts about the recent employment effects of artificial intelligence (revised 13 November 2025). Stanford Digital Economy Lab. https://digitaleconomy.stanford.edu/publications/canaries-in-the-coal-mine
Chen, F., & Stratton, J. (2026). Artificial Intelligence in the Firm. Harvard University. https://fion.ac/jellyfish.pdf
Engelbart, D. C. (1962). Augmenting human intellect: A conceptual framework (Summary Report AFOSR-3233, SRI Project 3578). Stanford Research Institute. https://dougengelbart.org/content/view/138/
Licklider, J. C. R. (1960). Man-computer symbiosis. IRE Transactions on Human Factors in Electronics, HFE-1, 4–11. https://worrydream.com/refs/Licklider_1960_-_Man-Computer_Symbiosis.pdf
Massenkoff, M., Lyubich, E., Sacher, S., Hitzig, Z., Zhang, S., Heller, R., & McCrory, P. (2026). Anthropic Economic Index report: Cadences. Anthropic. https://www.anthropic.com/research/economic-index-june-2026-report
U.S. Bureau of Labor Statistics. (2025). Occupational outlook handbook: Tellers. https://www.bls.gov/ooh/office-and-administrative-support/tellers.htm