Tower of Label

Tower of Label is a participatory audiovisual work. It invites exhibition visitors to speak to a system that evaluates the visitor's voice in the same way that digital platforms evaluate us all: based on commercial value and moderation risk. Which voices are amplified and which disappear is left to the markets to decide. Regardless of what was actually said.

• Scan the QR code next to the tower and use your phone to record a short voice message, preferably in English. Say something to the internet — it is listening.

Detailed Artwork Description

The internet was built on language — protocols, commands, declarations, and voices. However, not all voices are equal: algorithms decide which of them are heard. Attention became a resource to be harvested and resold long before the birth of the internet, and every new medium of communication found new ways to capture it (Wu, 2016).

Tower of Label is a participatory sound installation that makes this selection process visible and audible. Visitors are invited to record a short voice message for the future of the internet. Each voice connects to a living particle field, where one particle corresponds to each recording. The particles drift, group together, and sound from time to time according to the logic of algorithmic visibility: what is commercially valuable, what is flagged as problematic, what is amplified, and what disappears.

Three displays attached to the same rotating steel frame replicate the same view of the particle field. By scanning the QR code at the base of the tower, visitors can record a short voice message with their own phone. Their words are scored in real-time by the metrics of the attention economy — based on keyword value and moderation risk — and transformed into sound. The pitch of the audio snippets is modified according to musical intervals, and the snippets are cut and shaped based on how the system evaluates them. Hanging next to the QR code is a single long paper roll containing 9,799 words: all commercial keywords whose market price exceeds the set minimum threshold. Each of them has been read aloud and priced, and they live in the same particle field along with the visitors' own recordings. A price list for the words that the algorithm loves the most.

The work repeats in cycles of 12 minutes and 45 seconds. Most of the time, it follows the same commercial logic as the internet: quieter voices drift unheard, while the loudest, most valuable, or those classified as most problematic dominate the field. Four times during each cycle, this logic stops completely for 45 seconds. During these moments, only the voices of people who have actually spoken to the tower that day are heard. Each is heard in turn, one at a time; nothing is scored and nothing is bought. No visitor has to wait longer than about four minutes for such a moment. This time period is not a utopia. It is a glimpse. Viewed this way, attention resembles an ecosystem rather than private property — a shared environment that can become congested, depleted, or overrun by the fastest-spreading phenomena, just as an invasive species invades a wetland (Citton, 2017). What looks like a personal distraction of focus is often what philosopher James Williams calls a crisis of the will: the environment is designed to pull you away at every turn from what you would choose yourself (Williams, 2018).

At first, the tower speaks only to itself. Its source material consists of synthetic voices reading the keyword collection aloud — the algorithm's own language presenting itself even before the first visitor arrives. Over the three weeks of the exhibition, human voices gradually displace this synthetic source material, until in the final days, the field belongs almost entirely to the people who spoke into the work.

Tower of Label was created in collaboration with Anthropic's Claude AI. The entire software development of the work has been carried out through "vibe coding" — a conversational, AI-assisted development workflow instead of writing code hand by hand, line by line.

The concept of the work was developed together with Tom Kettunen. He was the first to bring up the idea that different keywords have different market values. This was the starting point for the operational logic based on the commercial value of the work.

Appendix: Calculation principles of the particle field – for visitors interested in the mechanics of the artwork

Each voice entering the field is scored as soon as it is converted into text, across three independent dimensions. The artist does not determine these values: each value is based on actual market data or automated analysis of the words and the audio itself.

Arousal level – determines if the voice is heard at all. Arousal describes how energetically something was said. It combines the characteristics of the voice — pitch, energy, and speech rate — with the emotional charge of the words used. It is kept separate from commercial value and toxicity so that even a calm voice can be heard and a loud voice is not favored solely for its volume.

Commercial value (price) – determines how loudly the voice sounds, how large its particle grows, and whether the particle takes the shape of an image representing a harmful invasive species. The value is calculated by comparing the text generated from the speech to actual advertising market prices for the 9,799 commercial keywords printed on the roll of paper next to the QR code. All of their prices exceed the set minimum threshold, all have been read aloud, and they live in the particle field even before the first visitor arrives. The artist does not choose which words are valuable; the market has already made the choice.

Toxicity – determines how long the voice sounds and whether the particle takes the shape of an image representing a poisonous plant. The more toxic the voice is, the longer it sounds. This is an intentional amplification of audience reaction, as anger is what spreads. Toxicity is measured using the same content moderation model for all voices in the field, both visitor recordings and synthetic voices. Thus, no voice is evaluated on different grounds than others.

When neither commercial value nor toxicity dominates, the particle takes its shape from a historical image depicting biodiversity or a rare plant specimen. These particles are light and remain on the outskirts of the field. Numerically, the majority of voices end up in this group.

Pitch is not calculated based on any score. It follows a pre-determined melody playing continuously in the background — the arpeggiated notes of Grouper's "Poison Tree." Each voice receives the note that the melody is on when the voice starts.

Two voices are in use, one of which is heard less frequently. Both the source material and the visitors' recordings utilize two voice banks of nearly equal size and nearly identical average commercial value. However, the female voice is heard less frequently than the male voice: for every one hundred occurrences of the male voice, the female voice is heard approximately 84 times. The ratio is not arbitrary. It is based on the 2025 US Census Bureau figure representing the earnings of full-time working women relative to men's earnings. Here, the ratio is expressed by how often the voice is heard, not by how much it is paid (Kollar & Scherer, 2026). The voice used for each keyword is permanently determined by a hash value calculated from the word. This ensures that no topic or register of language belongs to only one voice — the difference is solely in how often each is heard, never in what they say.

The repeating cycle of the work and the human turn. The field goes through five different phases — vortex, column, tornado, spiral, and a slowly rotating music box — as well as two shorter, repeating views of the installation. The entire sequence rotates in a cycle of 12 minutes and 45 seconds. Most of the cycle follows the logic described above: the loudest, most valuable, and those classified as most problematic voices dominate the field, reflecting how attention online is actually distributed. Four times during each cycle, the scoring ceases to have an effect for 45 seconds. Next, the one whose previous turn has passed the longest time ago is heard, for the same duration and at the same volume as everyone else. Only real voices recorded by visitors are included. The synthetic source material remains visible but goes silent for these 45 seconds. The only exception is voices that exceed the threshold set for hate speech: they dominate in other parts of the cycle, but for these moments, they are silenced completely. No visitor has to wait longer than about four minutes for such a moment.

At first, the field speaks only to itself. Its source material consists entirely of synthetic voices reading the keyword collection aloud. The algorithm speaks its own language before the first visitor arrives. Over the three weeks of the exhibition, genuine recordings gradually replace these original voices, until in the final days, the field belongs almost entirely to the people who spoke into the work.

The plant illustrations are found, not made. Each image visible in a particle is a copyright-free botanical print from the Biodiversity Heritage Library collection (biodiversitylibrary.org). This is an international consortium of natural history museums and libraries that has digitized millions of pages of historic scientific illustration. No images have been drawn or generated for the work: each illustration plate is an authentic 19th-century print exactly as it was found.

Images representing harmful invasive species — plants introduced as ornamentals that have since spread uncontrollably: kudzu (Pueraria montana), water hyacinth (Eichhornia crassipes), Japanese knotweed (Reynoutria japonica), and purple loosestrife (Lythrum salicaria).

Images representing poisonous plants — toxic plants, several of which have historically also been used as medicinal plants: deadly nightshade (Atropa belladonna), common foxglove (Digitalis purpurea), oleander (Nerium oleander), and monkshood (Aconitum napellus). Many of these plates are from Köhler’s Medicinal Plants (1887), an atlas originally illustrated to help identify dangerous species.

Images representing biodiversity — orchids, ferns, cross-sections of seeds and fruits, and plant specimens cataloged for the first time during 19th-century scientific expeditions.