Designing Against the Algorithm

Designing Against the Algorithm

The next desirable interface might be the one that refuses to anticipate you. A phone without an infinite feed. A writing machine that cannot open another tab. A photograph altered to resist machine learning. As artificial intelligence spreads through everyday products, refusal is becoming a design proposition—and, inevitably, a sales pitch.

Dezeen’s September 2026 roundup of five anti-AI designs frames this emerging territory around giving people an element of control. That qualification matters. An element is not sovereignty. Beyond the roundup, a wider field of restrictive devices, privacy tools and defensive software exposes the central dispute: does designing against the algorithm transfer power to people, or merely give powerlessness a better-looking enclosure?

The answer depends less on whether a product contains AI than on who can set its limits. First, the case for refusal. Then, the case against selling refusal as liberation.

PRO: Limits Can Be an Interface for Freedom

Good design does not always remove friction. Sometimes it puts friction precisely where a business model would prefer none. The Light Phone II, created by the company founded by Joe Hollier and Kaiwei Tang, uses a small E Ink interface and a deliberately limited toolset to separate communication from the compulsive architecture of the smartphone. It predates today’s generative-AI boom; calling it an explicit response to that boom would distort its history. Nevertheless, it offers a useful precedent: capability can be reduced to protect attention rather than to force an upgrade.

Astrohaus’s Freewrite makes a related argument through the dedicated writing machine. A keyboard and focused drafting environment place composition ahead of browsing. Its cloud-connected features complicate any fantasy of total digital withdrawal, but its central intervention remains legible: the device cannot casually become every other device. The constraint is architectural, not motivational. It does not ask the writer to become more disciplined while surrounding them with temptations.

Applied to AI, this principle suggests products that do not automatically summarize private correspondence, propose replies or turn personal archives—including shared family health data—into recommendation engines. A genuine off switch should terminate a function, not hide its interface while processing continues elsewhere.

Chosen friction can protect an intention from being overwritten. The important word is chosen. A reversible limit selected by the user is a tool; a restriction imposed by the manufacturer is simply another form of control.

PRO: Privacy Can Become a Material Practice

The strongest anti-AI work addresses extraction, not just distraction. Glaze, developed by researchers at the University of Chicago, applies calculated changes to digital artworks intended to interfere with models learning an artist’s style. Nightshade, from the same research environment, takes a more confrontational approach: it is designed to make unauthorized training on treated images disruptive to the resulting model. Neither should be mistaken for universal protection. Their importance is also political: they treat artists as parties entitled to resist, rather than as a frictionless supply of training material.

Adam Harvey’s CV Dazzle likewise made machine vision a design problem. Its experimental hair and makeup strategies explored how to interfere with certain face-detection systems. It is not a dependable disguise against every contemporary recognition system, but it established a powerful reversal: instead of optimizing the human face for machine readability, redesign its presentation to contest that readability. A related ambition appears in clothing designed to resist machine vision.

These projects also illuminate an architectural question. A workplace can specify cameras and sensors as though observation were simply another utility. Alternatively, it can minimize collection, provide genuinely unmonitored areas and disclose where processing occurs. A physical lens shutter or electrically meaningful microphone disconnect makes a boundary understandable without requiring faith in a settings menu.

Privacy becomes agency when the boundary is both comprehensible and enforceable. Locally controlled systems, including appropriately configured Home Assistant installations, demonstrate that connected environments need not surrender every decision to a remote platform. The objective is not technological purity. It is the ability to inhabit a space without continuously feeding somebody else’s intelligence.

CONTRA: Refusal Is Easily Repackaged as Luxury

The market has an extraordinary ability to sell relief from conditions it helped normalize. First, the smartphone absorbs the camera, notebook, map and music player. Then a new generation of dedicated objects sells those functions back separately, accompanied by a moral language of focus. A pared-back phone or writing machine may genuinely help its owner. It may also become an expensive accessory to a primary smartphone and laptop, adding hardware without meaningfully reducing dependence.

This is where anti-AI design risks becoming an aesthetic category: matte surfaces, restrained typography, tactile buttons and photographs of uninterrupted concentration. None of these qualities establishes who controls the firmware, how long repairs remain possible or whether a companion account quietly collects usage data. A device can look reassuringly analogue while relying on thoroughly conventional platform infrastructure.

The Light Phone and Freewrite are useful tests precisely because their constraints are tangible. But no constrained product deserves exemption from ordinary questions about maintenance, accessibility, manufacturing and service dependence. If a supposedly liberating object becomes unusable when its vendor disappears, the freedom was conditional from the beginning.

There is also a class problem. A professional who can afford multiple devices and delegate administrative work can perform digital refusal far more easily than a delivery worker whose income depends on algorithmic dispatch. Agency cannot mean merely purchasing a quieter position inside the same system. When resistance is sold chiefly as personal lifestyle improvement, the infrastructure of surveillance remains intact—and gains a tasteful outlet for consumer dissatisfaction.

CONTRA: Making People Harder to Read Is Not Enough

Defensive design can also assign responsibility to the wrong party. CV Dazzle asks the observed person to alter their appearance. Glaze asks the artist to process their work before publishing it. These interventions can be valuable without being fair arrangements. The institution deploying surveillance and the company collecting training data retain the initiative; everyone else must spend time, knowledge and sometimes money making themselves less available.

Technical effectiveness is another fault line. An adversarial pattern can fail when a vision model changes, or when lighting, distance and camera angle differ from test conditions. A signal-blocking pouch can isolate a phone from radio connections when properly constructed and closed, but it does not erase information already held by platforms. The physical certainty of an object can encourage claims far broader than its actual protection.

Blanket rejection also mistakes automation for domination. Speech recognition, image description and predictive text can expand autonomy for disabled users. For someone with limited mobility, voice control may create access rather than diminish it. The relevant distinction is not human versus machine; it is assistance under meaningful user control versus compulsory mediation.

A credible anti-AI position therefore needs demands beyond products: limits on collection, defensible consent, procurement accountability and workable routes to contest automated decisions. Designers should publish what their interventions protect against, what they cannot prevent and how those limits were tested. A beautiful refusal that leaves institutional power untouched is not a counter-system. It is an accessory to one.

FAQ: What Counts as Meaningful Resistance?

What is anti-AI design?

Anti-AI design describes products and practices that resist unwanted automated inference, data extraction or algorithmic mediation. It includes defensive software and privacy hardware, but also deliberately limited interfaces. Not every minimalist or distraction-free device was designed specifically to oppose AI.

Do anti-AI products actually protect privacy?

Some protect against particular threats. A properly functioning camera shutter blocks visual capture through that lens; it does not stop other sensors or existing data processing. Evaluate the stated threat, independent testing, update requirements and failure conditions—not the product’s reassuring appearance.

Does rejecting automation increase human agency?

Only when rejection is a usable choice. Agency increases when people can understand, configure and override a system without losing essential access. Removing assistive functions or forcing laborious workarounds can diminish autonomy just as readily as compulsory automation can.

How can designers avoid turning resistance into an aesthetic?

Make boundaries verifiable. Disclose data flows, provide effective shutdown controls, support repair and explain dependencies on remote services. Measure success through reduced unwanted collection and greater user control, not through a visual language of analogue authenticity.

If freedom from algorithmic mediation becomes a premium feature, are we designing human agency—or designing who gets to afford it?