It’s early morning during the Dutch school holidays. Most of the country has migrated south to recharge and prepare for the long grey winter. My summer has been a mix of early-morning writing, work, mountain hikes, family time, and devouring new books. One in particular has inspired me: Jaron Lanier’s You Are Not a Gadget (2010), a manifesto by the Minsky protégé turned VR pioneer, author, and tech philosopher. The book takes a critical look at the design decisions behind the systems we rely on for human interaction and connection. Lanier challenges the dumbing down of human experience to fit paradigms of input fields, profile pages, search algorithms, and neural networks. He argues that compromises made during this societal-level encoding have been normalised and now shape our interactions. Or, to quote him directly: “Lock-in turns the transient quirks of early software into permanent features of the world.” In a world where dating, friendship, and family conversations move through a handful of platforms, lock-in isn’t just technical; it’s ontological. Once interfaces define what counts as a message, a reply, a match, they also begin to define what counts as a person.
Hubert Dreyfus: formalisation can misread the grain of lived meaning
Artificial intelligence has long pursued a manageable model of human intelligence, and the more I study and work with this technology, the more I’m convinced it’s a simplification. I was recently writing a chapter of my book on learning. I traced, in broad strokes, the human study of learning and then paired it with the steady advancement of intelligent technology. My humble conclusion: the language models we’re integrating into society at breakneck speed are not synthesising the human experience of learning; they are consuming and encoding the written history and study of learning. There is a difference, and it lies in the richness and complexity of context—which is precisely what’s missing from the calcified infrastructure of the digital-social interactions we now rely on.
The foundations of AI were defined by people who believed that the human condition could be cut up into a series of functions and then digitalised. Symbolic AI compressed meaning into explicit rules and categories. Today’s generative AI compresses selves into statistical continuations: most likely next words using the most plausible tone. The technologies differ, but the underlying move is the same: translate the messiness of life into forms computation can handle. The risk is not only error; it’s the slow normalisation of a diminished model of the person. You could argue that we might be moving towards “over fitting” of our model and the chunking of its training data, where all those subdivisions are cut so finely that what comes out is just context-free digital slop.
This matters most in the social rooms where we are most exposed. Dating apps and messaging platforms, now fused with AI, don’t just convey intimacy; they set its grammar. Profiles, swipes, and ranking algorithms define how a self appears to others. Read receipts, reaction emojis, and suggested replies define what counts as response. Every small affordance is also an argument. Over time, templates become expectations, and expectations become habits of mind.
Sherry Turkle: “We expect more from technology and less from each other”
(Photography Monaris: https://superrare.com/monaris)
Consider everyday texting between partners. The system offers autocompleted responses like “Sounds good,” “On my way,” “Let’s talk later.” Each suggestion is fine. None is exactly a lie. But the accumulation of plausible shortcuts changes the texture of exchange. We learn to prefer the fast version of ourselves—the one that moves the thread along—over the part that hesitates, searches for a better word, or risks saying something that takes time to land. Predictive text becomes a training regime: articulate less, agree more quickly, keep pace.
Translate this to online dating and the effects scale. The architecture of swipes and scores sets the terms of desirability; the profile format constrains identity into bite-sized legibility; the recommendation engine rewards certain patterns of behavior and language. What starts as a filter becomes an personality. People adapt their presentation and their expectations to the system’s ontology: compatibility as optimisation, conversation as conversion, intimacy as throughput. When AI enters this loop—summarising bios, suggesting openers, smoothing replies—it intensifies the drift toward a scripted sociality that feels fluent but thin. Success on an online dating platform may be defined not by persons but by the successful adaption of the person to the limitations and design decisions of that platform.
Byung-Chul Han: the pressure toward positivity and frictionless communication that thins experience
The point is not to claim that tools of convenience are bad, or that augmentation can’t help. They can, and do. People with anxiety find it easier to start conversations; couples avoid unnecessary escalation by leaning on neutral phrasing; busy lives benefit from friction removed. But there is a cost.
When speed, legibility, and prediction become the default virtues, other goods recede: latency, opacity, and ambiguity. Slowness is not just inefficiency; it’s the space where attention and care take shape. Ambiguity is not just confusion; it’s the condition under which new understanding and connections becomes possible. Opacity is not just secrecy; it protects freedom to choose how and when to show up. Hypocrisy is crucial to human interaction, it allows us a path of less friction to navigate in life. Yet, in a boolean driven design, nuance becomes negative and ambiguity is emotional insecurity. The app doesn’t have human curiosity, it is not designed for long discussions or discoveries, it has a model, an input field and a match. We have locked in to the design limitations of these systems and helped them to redefine modern love and well, it aint pretty.
This Lock-in makes these trade-offs self-reinforcing. The more dominant a platform becomes, the more its assumptions become the unspoken baseline for social life. A message is what the system can parse. A good reply is what the model suggests. A successful match is what the ranking function yields. Once embedded in the everyday, these defaults are hard to challenge—technically, culturally, economically. The danger is closure: a narrowing of the grammar of intimacy to what the infrastructure can recognize.
James C. Scott: Danger arises when simplification becomes the only permissible reality
There are, however, ways to widen the path. I saw this recently in OpenAI’s new study mode, where an application makes a design decision that follows the meandering path of human learning instead of simply churning out auto-completed answers to be copy pasted into word documents and shared with teachers (sans any and all “learning”).
In messaging, we could also design for delay. Build slow modes that encourage drafting and revisiting before sending; let slowness be an option, not a failure. Make presence indicators and read receipts opt-in, not defaults; preserve the right not to be immediately legible. Reframe augmentation: rather than offering generic templates, surface the user’s own past language and tone, prompting reflection over mimicry. Prompt consent in intimate contexts: do you want help writing this apology, or do you want time to find your own words? Analysis and assistance from the study of human interactions instead of plugging directly into our lives.
In dating, I think we could diversify the grammar. Here's a few ideas off the top of my head:
- Support richer forms of self-presentation
- Let a person create a broader, quirkier representation of themselves that isn't just a template to make inference time cheaper.
- Expose, and give control over, the objectives that recommendation systems optimize.
- Let people choose values beyond engagement and match rate.
- Allow alternative discovery modes that privilege conversation over throughput.
- Encourage norms that treat silence, reconsideration, and changing one’s mind as part of the courtship, not failures of performance.
- Adding a romantic counsellor to a dating app that assists the individual with advice about not ghosting or responding to emotional overreach (or just sociopathic rudeness)
- Having an intelligent pause button where an assistant asks the question “are you sure you want to send this?” instead of just bluntly autocompleting our emotions to increase the speed and volume of consumption.
I'd suggest we defend deviation in systems of social interaction as much as possible. Protect the awkward pause, the clumsy sentence, the joke that misses and then gets repaired. These are not inefficiencies to be eliminated; they are how we are as human beings. Teach, and model, the right to write badly before writing well is something I’d like to retain, this is me with all my grammatical and logical failings and I’d like this space in a relationship as well without it being infected by the "if, then, else" of conditional logic.
It's a modest idea and at the same time radical. Keep augmentation, but let it amplify diversity rather than convergence. Retain efficiency, but not as the sole measure of value. Build infrastructures that can host multiple grammars of intimacy, not just the one that registers cleanly in a database. Preserve the option to be more than the fast version of ourselves. I'm well aware of the business model behind these platforms, but I also believe that this is the core failing of dating apps and social media platforms.
If we accept that the apps we use for communication, love and friendships have shaped us and changed how we experience other human beings—and I believe they have—then design of these systems is a moral art.
(Photography Monaris: https://superrare.com/monaris)