User:114.76.86.20/sandbox
- “Glossaria” redirects here. For other uses, see Glossaria (disambiguation)
 Location of the fourth-story walkway which collapsed into the lobby of the Boston JPHG Retaildential Centre | |
| Date | 19 August 2039 – 25 December 2039 |
|---|---|
| Location | Boston, Massachusetts |
| Coordinates | 42°21′36″N 71°05′31″W / 42.360°N 71.092°W |
| Also known as | Glossaria Disaster |
| Cause | Structural overload resulting from people under influence of Glossaria |
| Deaths | 42 |
| Non-fatal injuries | 112 |
Glossaria was an advertising campaign for the Jīngshén Properties Holding Group (JPHG) which ran for 6 months in the lead up to Christmas day 2039.
The campaign was intended to highlight JPHG’s personalisation capabilities within their Retaildential Centres which provide demographic-specific retail, entertainment and accommodation services.
It was ultimately discontinued when it became the subject of scrutiny by multiple governments, a class action lawsuit and was excoriated before the UN human rights committee.
Background[edit]
The Jīngshén Properties Holding Group (JPHG) was originally a shopping mall conglomerate that began in 1994 and one of the first to pivot into the Retaildential Centres space in 2025 when it was clear that e-commerce had made shopping malls no longer economically viable. By combining low cost residential property, banking services, and internet access into a single service JPHG could both capture the majority of residents incomes whilst gaining deep insights into their spending habits.
The residential towers that were incorporated into each former shopping mall funnelled the residents through the retail spaces which were tailored to their predicted purchasing needs. As residents left their apartments and traveled toward the exit the mall space would be reconfigured accordingly with both tailored screen adverts and physical products put on display using automated stockists.
By targeting each Retaildential Centre to narrow demographics JPHG could maximize resident expenditure whilst keeping stock to absolute minimums. The residents of Retaildential Centres targeting young families would see adverts for children activities and physical products such as bottle sterilisers. Whilst those in centres targeting wealthy retirees would instead focus on luxury goods and health products.
By data mining the residents banking data, online activities, and physical movements they could accurately predict spending habits to the point where most Retaildential Centres captured over 74% of residents earnings.
The Christmas shopping season was one of the most profitable times of the year for JPHG. In 2039 an advertising campaign was created by their internal MLM department (Machine Learning Marketing) to see if the Christmas Carols that played between August and December each year could be used to influence resident behaviour.
JPHG already used various techniques using light, sound, and smell to guide residents to various targeted products. The theory was that a GAN (generative adversarial network) could be used to dynamically create carols that would cause a statistically significant influence in resident behaviour. The initial goal was to influence 2% of people into turning left after passing through the security stations that were situated at the entrance to every Retaildential Centre.
Outcome[edit]
Generative Adversarial Networks had already been used in the past to create photo-realistic pictures of people that didn’t exist or “deepfake” videos of celebrities doing things that had never happened. The process uses a pair of convolutional neural networks, one trained to create the output and the other trained to spot the fakes. The conflict between the two systems constantly refines the output over millions of iterations.
Initial tests on residents began on the 19th August with the “deepfaked” carols playing at all 48 JPHG properties globally. Cameras tracked resident movements which were fed back into the GAN as the primary fitness metric for whether it was succeeding.
Statistically significant changes in resident movements were achieved fairly quickly but these were found to be non-viable solutions. The GAN had started to incorporate sounds that people later described as unsettling and were attempting to avoid them. This was fixed by feeding live sentiment analysis data from the video feeds into the GAN. Any noise that made residents unhappy would be selected against.
Over the next few months the GAN continued to overfit by selecting for various demographic specific solutions. Each of the Retaildential Centres’ resident populations was selected for a narrow marketing demographic. An example of the overfit was that elderly residents of an aged care Retaildential Centre began hearing something that many described as sounding like “Smells like Teen Spirit” by the 1990s band Nirvana. The audio sounded to be coming from the left and many of the residents were subconsciously turning toward a song from their youth.
This, however, had no effect in the other centres and the decision was made to force the GAN to find a universal solution for all centres without using directional audio. This required a substantial increase in computing power which was approved by the board on the condition they would see results before the end of the Christmas shopping period in January.
Without directional audio or cultural props (i.e. pop songs) the algorithm ceased to make any statistically significant influence on the customers direction. Over the next three months it was tested against 4.6 million people over 1.09 billion times, eventually settling on a solution in the final three days before Christmas. This initial audio resulted in a 0.4% increase in the likelihood that a given customer would turn left upon exiting. Within thirty minutes the reinforcement learning had refined this to over 5.8% before plateauing. Even this substantially exceeded JPHG’s wildest expectations and the project was considered a success.
The Christmas Market Tragedy[edit]
JPHG’s board of directors met on the day before Christmas eve to discuss the outcomes of the project. Even then concerns were raised that a success rate of 5% was crossing the line from influencing into outright mind control. The board was unaware however the algorithm was in fact stuck within a local minima, a suboptimal solution that it had refined as much as possible but was otherwise still seeking new optimal solutions.
At 6:12 pm on Christmas eve the algorithm found its way out of that local minima and into a new solution that could be improved further. By 6:27 pm this had been optimised to over 90% of customers influenced to turning left. Globally the effect was immediately apparent but as the audio feed only influenced people for the first three seconds after leaving the choke point, the effect mostly only raised some minor concerns.
The notable exception was a Retaildential Centre in Boston where the target demographic of the residents was those who self-identified as artists. The Centre was running a Christmas market on the 24th and from 4:00 pm onward a large number of outside visitors were there for the market and to visit family. Those who turned left after exiting security found themselves on a narrow balcony overlooking the lobby.
By 6:30 pm security was attempting to turn people away and move people off the balcony, but by 6:45 pm the constant flow of foot traffic meant that over 480 people were on a balcony meant for 50. At 6:52 pm the crowd panicked and in the resulting chaos, the balcony balustrade gave way. In total 112 people were injured and 42 died as result of the Glossaria audio.
Theory[edit]
Although there is a global moratorium on practical research into Glossaria audio, many experts are still analysing the recordings that exist. The current theory is that the machine learning algorithm found the audio equivalent of a form constant.
A form constant is one of several geometric patterns which are recurringly observed during hallucinations and altered states of consciousness. They are notable for the fact that despite their complexity they exist independent of the person’s culture. The diversity of conditions that provoke such patterns suggests that form constants reflect some fundamental physical property of visual perception.
Similarly, it’s believed that the Glossaria created by JPHG tapped directly into a fundamental property of how audio is turned into language and meaning. A popular theory is that since no Glossaria has been found for animals (where testing is still legal) it’s possible that it’s using a vestigial language that existed in earlier species of Homo Sapiens.
Although a common plot element in entertainment Glossaria is now only a curiosity as the use case created by JPHG only functions within the narrow constraints of the scenario the algorithm was optimised for. Research has shown that modifying the audio even slightly renders it useless as does repeated exposure (many popular songs and movies have since sampled the Glossaria audio).
It’s believed that to create another would again require access to millions of people over a prolonged period of time which would be easily detectable. The only proposed solution to creating new Glossaria commands would be having access to a high fidelity simulation of the human brain which is unlikely to be available in the foreseeable future.
Discovery and Scandal[edit]
Very quickly details of the experiment leaked from within JPHG with the resulting fallout leading to a full censure by world governments and the UNHRC. It was later explicitly banned under the 5th Geneva Convention.
In the class action lawsuit that followed it was revealed that many members of the --------
Uh Oh!
It appears that the website en.wikipedia.org is not included in your current internet plan
To gain access beyond the first 10 thousand words, add a new website for $25 USD / month
or upgrade your plan.