Artwork: Francesco Fabris, Below the Haze Line: An Extinct Gaze (AI International Biennale, Kraków). Star positions, magnitudes, colours and distances: HYG Database v4.1 by astronexus (CC BY-SA 4.0). Constellation figures: d3-celestial by Olaf Frohn. Positions are fitted to the drawn figures, so the drawing stays as it was made.
Below the Haze Line: An Extinct Gaze
Francesco Fabris · installation, AI International Biennale, Pałac Potockich, Kraków · curated by Patrycja Maksylewicz
The installation
Below The Haze Line: An Extinct Gaze, AI International Biennale, Palac Potockich, Krakow, PL
‘Below the Haze Line: An Extinct Gaze’ is an installation that reimagines our relationship with the cosmos by exploring the tension between human imagination and machine cognition. The project speculates on a future where humanity has lost the ability to directly observe the night sky, a starkly imminent reality exacerbated by environmental issues like light pollution and atmospheric haze. This erosion leaves us to rely upon raw archival data, ancient myths, and artificial intelligence to serve as our only telescope.
The sinusoidal textures originate from sonifications of scientific data from the stars within the zodiac. Here, an AI is tasked with interpreting these frequencies to retrieve a lost past, calling back to an era when our visual communion with the cosmos dictated and empowered the very rhythms of existence.
As the installation glows, the swells of brightness demonstrate light pollution’s 10% annual increase, taken in accordance with 20 years of data. This growing radiance controls an AI model trained on the constellation microtones and fed by the invisible humming of electromagnetic field recordings of global human-made light sources.
What emerges is a widening chasm of disconnection mediated by the machine. Technology fails to restore the purity of nature, hindered by the restless and omnipresent anthropocene transmissions that have saturated the world.
A fragile hope remains in the potential for human healing, currently suspended in the limbo of myth and pseudoscience. We are left to wonder: are we still in control and capable of restoring what we destroyed, or is that capability merely an illusion, realized too late to discern?
This web version
This is a reworked web version of the installation. The zodiac and the oldest written tuning system both come from Mesopotamia. Here a machine listens to the stars of the zodiac and remembers them in one of those tunings, nīd qablim, the tuning of the Hurrian Hymn. The extinct gaze becomes cultural as well as optical: what disappears is not only the dark sky but a whole way of relating to it. As the haze grows, the hum of the light that hides the stars is drawn into the same ancient scale.
It is not a reconstruction of how the sky once sounded but a fiction the machine tells: an imagined memory, not a recovered one.
The real sky and the machine do not move together. The stars disappear year by year, while the machine stays fixed in the one way it learned to hear. It never sees the data, only hears the stars' sound, and answers in the tuning it was taught. As light pollution rises, its answer comes less from the stars and more from the hum, until it hears stars that are no longer there. That widening gap, between the sky that is lost and the sky the machine remembers, is the subject of the piece.
How it works
The wheel is the installation's zodiac surface. Each medallion holds real stars from the HYG catalogue, placed on the printed dots of the drawn figures. The other stars that belong to each constellation, down to magnitude 6.2, fill its slice of the wheel around the medallion as smaller points: 1,234 stars in all. They keep their directions and the order of their distances from the centre of the constellation, spread out to fill the slice. Only the stars on screen are heard: choose a constellation in the Listen to menu, tap a medallion, or zoom in by hand (pinch, Ctrl/⌘ + scroll, or the + and − buttons). Tap a star to hear it on its own.
The timeline runs from 1965, the first year light pollution was measured, to , the year the last star of the zodiac disappears, and stops there. With Loop on, it holds a moment at the end and starts again from 1965. Play and Pause control it (Play at the end starts again from 1965), the speed menu sets how long one run takes (5 minutes by default), and a click or drag on the chart jumps to any year.
Each star has a data tone. It is made from the star's own data, with the same sonification as STELLAR DATA FIELDS: colour index sets the pitch (hotter stars sound higher), absolute magnitude sets the level (more luminous stars are louder) and distance sets the length (farther stars last longer, from 0.25 seconds for the nearest star to 5 seconds for the farthest, on a logarithmic scale). The pitch is set from the colour index directly, between 20 Hz and 5 kHz, spread evenly on the ear's ERB-rate scale (Glasberg & Moore), and the level is corrected by the ear's equal-loudness curve (ISO 226). These data tones are not heard: they are what the AI listens to. Stars play all the time, each one at most every 2.5 seconds; as the haze hides them, the ones left play more often, so the AI hears stars until the last one disappears.
The AI is what you hear. A small neural network listens to the data tones and answers in nīd qablim, one of the seven Babylonian tunings named on cuneiform tablets (2nd millennium BCE) and the tuning of the Hurrian Hymn from Ugarit (about 1400 BCE), the oldest piece of music written down well enough to be reconstructed. The scale is read here as Pythagorean, with the step pattern of the Lydian mode (the exact intervals are still debated), over six octaves from 55 Hz to 3.5 kHz: 43 notes, played as sine tones with the same loudness correction. Which star becomes which note is this work's own choice. The network was trained on the app's own data tones paired with the nīd qablim notes they become, with the field recordings below mixed in at random levels, so it learnt to hear the stars through the hum. It hears about 0.1 seconds of sound at a time and knows nothing of the data itself. Each of its notes is placed in the stereo field where the stars that feed it are on the wheel; a note no star is feeding drifts slowly to the centre.
The stars disappear. When the sky becomes too bright to see a star with the naked eye, it fades from the wheel, the printed dots of the drawn figures included, and its data tone stops.
Light pollution follows the sky of a typical European. In 2014, 60% of people in the EU lived where artificial skyglow is about four times the natural sky, the point where the Milky Way disappears (Falchi et al., Science Advances, 2016, from 2014 satellite data). Skyglow has grown about 10% a year whenever it has been measured: from 1965 to 1996 in archive measurements at the Asiago and Ekar observatories in Italy (Cinzano), and 9.6% a year worldwide from 2011 to 2022 (Globe at Night, Kyba et al., Science, 2023). The timeline starts in 1965, the first measured year, and runs forward at that rate as a projection until the last star of the zodiac disappears. The measured years are shaded on the timeline. One curve stands for many places; real skies range from pristine to city centres. The sky brightness (mag/arcsec²) is converted into the faintest star the naked eye can see, and each star disappears in the year that limit passes its magnitude.
The haze line is the flickering grey ring. It closes in at one steady pace, from the rim of the wheel in 1965 to the edge of the Dendera disc in , where it stops. As it closes, its flicker grows faster and brighter, reaching its brightest in the final year.
Behind the haze line, the light that hides the stars is heard. The stars the line has passed over sound through field recordings of light-pollution sources: street lights, light bulbs, shop lights, neon and a light sign, recorded with an electromagnetic-field receiver for the Outer Fields library. The recordings are cut into 432 fragments of 2.5 seconds. For each star, one of the fragments closest to its pitch and colour is chosen (bright, high fragments for hot stars, dark hum for cool ones) and played as recorded, with no tuning and no filter, placed where the star is on the wheel. Most of these sources hum at 100 Hz and its harmonics, the flicker of Europe's 50 Hz mains. The recordings are heard as they are, and they are also fed to the AI. They start silent in 1965 and grow louder as the sky brightens and the haze line takes more stars: as they cover the stars, the AI's answer drifts, and once the hum takes over, it hears stars in the noise that are not there.
The sky panel shows the year, the sky brightness, the faintest visible star, how many of the stars on screen can still be seen and the level of light pollution. Below it is the last star heard, or the one you tapped: its magnitudes, colour and temperature, distance, the year its light left it, its data tone (frequency and length), its data level in dB and the year it disappears into the haze.