iq business

Kalenjin Feed

  1. kalenjin shared this story from Quartz.

    Gobi Heaven

    In the weeks before the annual Burning Man festival sets up camp in the Nevada desert, a Chinese version of it played out in Inner Mongolia.

    For the price of a nine-hour drive from Beijing—and 560 to 2800 yuan ($80-400) for a ticket—revelers gathered in a wide-open Gobi desert space the size of six soccer fields, about one-tenth the size of Burning Man’s Black Rock City, which will be assembled starting Aug. 25. Gobi Heaven organizers say about 6,000 to 8,000 “residents” stayed the entire five days of the festival, though some participants and event volunteers interviewed by Quartz offered lower estimates on the size of the crowd.

    The half-built entrance to Gobi Heaven

    The half-built entrance to Gobi Heaven, one day before the opening of the event.

    Festival goers endured five days of extreme, windy weather: Heat in the mid-90sF° during the day followed by temperatures dropping into the low 50s at night. A sandstorm blew apart many tents one afternoon. Another day, everyone got soaked in pouring rain.

    Sand storm at Gobi Heaven

    The sandstorm came on Aug. 9.

    Sand storm at Gobi Heaven in Inner Mongolia, China on August 9th, 2019

    Artist Jinsong He saw his bamboo-structured installation fall in the Aug. 9 sandstorm.

    Chinese and international artists presented music, fireworks, light and art installations amid campsites that also showcased artwork. In the same spirit as burning the Temple on the last night of Burning Man, festival goers dressed up and burnt a bamboo-structured center installation called “The Guardian of Life and Death.”

    Gobi Heaven art installation

    Dragon, by David Han

    Gobi Heaven art installation

    Memory-Youth, by Jianzhang Li

    Art Installation (100 Wolves)

    Wolves Are Coming, by Ruowang Liu

    Gobi Heaven music stage

    Electronic music organized by Great Wall Festival played at the center music stage.

    Campers dressed up for the burning ceremony.

    Campers dressed up for the burning ceremony.

    Burning Temple

    The burning ceremony on the last night.

    Getting into Gobi Heaven

    Gobi Heaven founderLe Chen had prepared two years for the festival. He had worked in creative advertising in Beijing for almost two decades and wanted to create a Chinese version of the countercultural American festival. He struck a partnership with the San Francisco-based nonprofit Burning Man Project, attended the festival last year, and established a company. Backed by the state-run China Cultural Media Group (link in Chinese) and private capital, it now has a total of more than 50 million yuan ($7 million) to run the project in China for five years.

    The Chinese version isn’t called Burning Man. “The Burning Man trademark extension was denied by [China’s] Trademark Office. They saw Burning Man as an unconventional American carnival and thought it didn’t align well with the ethics and morals of China,” Chen told Quartz. The Chinese festival thus got a new name, Gobi Heaven, and is legally independent from Burning Man.

    Gone with the name were other Burning Man ideals. Adapting the 10 Principles of Burning Man, Chen developed his own “10 principles” (link in Chinese) for Gobi Heaven, encouraging human interaction and an open mind. There’s no mention of the “radical self-reliance” and “radical self-expression” as endorsed by Burning Man in the US.

    Here’s how the principles compare:

    Burning Man Gobi Heaven
    Radical inclusion Make friends
    Decommodification Resist unnecessary consumption
    Participation Put in your best effort
    Communal effort Honor collaboration
    Leaving no trace Pick your liter
    Civic Responsibility Don’t complain
    Gifting Put down your phone
    Immediacy Stay foolish
    Radical self-reliance Stay unique
    Radical self-expression Don’t rush to judgement

    Single-day tickets were sold at much-discounted prices, around 80-280 yuan ($11-40). Chen said that drew in single-day spectators from nearby communities, giving them rare direct exposure to modern artwork.

    Local participants visit Gobi Heaven.

    Burning up the rules

    Chen planned to limit cellular service on the site to one in every three hours, just eight hours a day, to simulate an isolated environment. He abandoned that plan on the first day. “I needed to call the workers responsible for equipment and food supplies. Campers couldn’t pick up their friends, performers couldn’t coordinate, and the emergency line was blocked,” Chen said.

    Some festival goers didn’t bring their own tents or water. They were able to buy supplies from local merchants who came to the event entrance. Pre-cooked food delivery such as braised beef noodles and large chicken plates was not allowed initially, but Chen changed that rule, too. “Some camps wanted to host a meal for a friend gathering. Some wanted to make dumplings. As long as they had a good time, we would allow it,” he said.

    Now Chen wants to separate Gobi Heaven from any Burning Man hype. “Not all human activities happening in a desert setting with fire should be called Burning Man,” he said as the event concluded its last day. “What we’ve built here is culture.”

  2. kalenjin shared this story from Visual Capitalist.

    Where the World's Banks Make the Most Money

    Where the World’s Banks Make the Most Money

    Profits in banking have been steadily on the rise since the financial crisis.

    Just last year, the global banking industry cashed in an impressive $1.36 trillion in after-tax profits ⁠— the highest total in the sector seen in the last 20 years.

    What are the drivers behind revenue and profits in the financial services sector, and where do the biggest opportunities exist in the future?

    Following the Money

    Today’s infographic comes to us from McKinsey & Company, and it leverages proprietary insights from their Panorama database.

    Using data stemming from more than 60 countries, we’ve broken down historical banking profits by region, while also visualizing key ratios that help demonstrate why specific countries are more profitable for the industry.

    Finally, we’ve also looked at the particular geographic regions that may present the biggest opportunities in the future, and why they are relevant today.

    Banking Profits, by Region

    Before we look at what’s driving banking profits, let’s start with a breakdown of annual after-tax profits by region over time.

    Banking Profit by Year and Region ($B)

    Global ($B)$388$530$635$703$859$963$1,070$1,065$1,144$1,356
    United States$19$118$176$263$268$263$291$275$270$403
    Western Europe$78$34$21-$70$28$95$154$159$186$198
    Rest of World$196$243$265$285$309$327$348$361$387$421

    In 2018, the United States accounted for $403 billion of after-tax profits in the banking sector ⁠— however, China sits in a very close second place, raking in $333 billion.

    What’s Under the Hood?

    While there’s no doubt that financial services can be profitable in almost any corner of the globe, what is less obvious is where this profit actually comes from.

    The truth is that banking can vary greatly depending on location ⁠— and what drives value for banks in one country may be completely different from what drives value in another.

    Let’s look at data and ratios from four very different places to get a sense of how financial services markets can vary.

    CountryRARC/GDPLoans Penetration/GDPMargins (RBRC/Total Loans)Risk Cost Margin
    Global Average5.1%124%5.0%0.8%
    United States5.4%121%5.0%0.4%

    1. RARC / GDP (Revenues After Risk Costs / GDP)
    This ratio shows compares a country’s banking revenues to overall economic production, giving a sense of how important banking is to the economy. Using this, you can see that banking is far more important to Singapore’s economy than others in the table.

    2. Loans Penetration / GDP
    Loans penetration can be further broken up into retail loans and wholesale loans. The difference can be immediately seen when looking at data on China and the United States:

    CountryRetail LoansWholesale LoansLoan Penetration (Total)
    United States73%48%121%

    In America, banks make loans primarily to the retail sector. In China, there’s a higher penetration on a wholesale basis — usually loans being made to corporations or other such entities.

    3. Margins (Revenues Before Risk Costs / Total Loans)
    Margins made on lending is one way for bankers to gauge the potential of a market, and as you can see above, margins in the United States and China are both at (or above) the global average. Meanwhile, for comparison, Finland has margins that are closer to half of the global average.

    4. Risk Cost Margin (Risk Cost / Total Loans)
    Not surprisingly, China still holds higher risk cost margins than the global average. On the flipside, established markets like Singapore, Finland, and the U.S. all have risk margins below the global average.

    Future Opportunities in Banking

    While this data is useful at breaking down existing markets, it can also help to give us a sense of future opportunities as well.

    Here are some of the geographic markets that have the potential to grow into key financial services markets in the future:

    1. Sub-Saharan Africa
      Despite having 16x the population of South Africa, the rest of Sub-Saharan Africa still generates fewer banking profits. With lower loan penetration rates and RARC/GDP ratios, there is significant potential to be found throughout the continent.
    2. India and Indonesia
      Compared to similar economies in Asia, both India and Indonesia present an interesting banking opportunity because of their high margins and low loan penetration rates.
    3. China
      While China has a high overall loan penetration rate, the retail loan category still holds much potential given the country’s population and growing middle class.

    A Changing Landscape in Banking

    As banks shift focus to face new market challenges, the next chapter of banking may be even more interesting than the last.

    Add in the high stakes around digital transformation, aging populations, and new service opportunities, and the distance between winners and losers could lengthen even more.

    Where will the money in banking be in the future?

    Subscribe to Visual Capitalist

    Thank you!
    Given email address is already subscribed, thank you!
    Please provide a valid email address.
    Please complete the CAPTCHA.
    Oops. Something went wrong. Please try again later.

    The post Where the World’s Banks Make the Most Money appeared first on Visual Capitalist.

  3. kalenjin shared this story from Visual Capitalist.

    View the full-resolution version of this infographic

    Visualizing the World's Sleeping Habits

    Visualizing the World’s Sleeping Habits

    Sleep quality, patterns, and duration may vary among countries, but one thing’s clear─people still aren’t getting enough sleep. While some people can function on a few hours, others find themselves reaching for that second cup of morning coffee instead of getting those extra Z’s.

    Today’s graphic comes from Raconteur and highlights some startling takeaways from the 2019 Philips Global Sleep Survey, answered by over 11,000 adults from 12 countries.

    Let’s settle in to discover what impacts our sleeping habits, also known as sleep hygiene, and what helps people sleep better and longer.

    Why Sleep Is Important

    Roughly 62% of adults worldwide feel that they don’t sleep well when they go to bed. Losing just one or two hours of sleep per night can have the same impact on motor and cognitive functions as going without sleep for a full day or two.

    Experts have long emphasized that developing good sleeping habits can help to maintain our physical, mental, and emotional well-being. Ongoing sleep deprivation can also cause severe, long-term health conditions:

    • Heart disease and heart failure
    • Weak immune system
    • High blood pressure
    • Kidney disease
    • Depression
    • Diabetes
    • Obesity

    Drowsiness has been a significant factor in roughly 100,000 car accidents every year, causing an estimated 1,500 deaths. Sleep deficiency has also been linked to a number of disasters, such as airplane and boat accidents, and even nuclear reactor meltdowns.

    The Science of Sleep

    The human body follows the circadian rhythm─a 24-hour repeating rhythm that operates as an internal clock. This clock is controlled by two things: external cues such as light and darkness, and internal compounds that trigger and maintain our sleep.

    These chemicals work together to keep our sleep/wake cycles in harmony.

    • Adenosine: slowly builds the desire for sleep throughout the day
    • Melatonin: produces drowsy feelings that signal your body is now ready for sleep
    • Cortisol: naturally triggers your body to wake up

    While sleep duration can vary greatly around the world, most adults are still not getting enough shut-eye. The average person gets 6.8 hours of sleep on a weeknight, which is significantly lower than the recommended 8 hours.

    One company in the UK has even developed a real-time map of social media posts from people who say they can’t fall or stay asleep.

    What Prevents Better Sleep?

    People can suffer from a lack of sleep for many reasons─below are the top six culprits.

    1. Worry and Stress
      Job, family, health, financial, and a myriad of other concerns plague people from all walks of life. Adults living in Canada and Singapore tend to be the most worried.
    2. Environment
      The physical space where you sleep plays a large role in the quality and duration of your sleep. Nearly 35% of adults fall asleep somewhere other than their bed. Interestingly, Chinese adults are the least comfortable when sleeping, while Japanese adults are the most comfortable.
    3. Work and School Schedules
      Hectic careers and heavy school workloads have a direct and lasting impact on sleeping habits. Many forego sleep in favor of completing work, social, and household responsibilities.
    4. Entertainment
      In the age of technology, natural rhythms of daytime and nighttime perception have been skewed, especially from the effects of blue light emitted from our device screens.
    5. Disruptors
      Eating food, or drinking alcohol or caffeine within the last few hours before bedtime can prevent our brains from knowing it’s time to wind down and get ready for sleep. Adults living in the fast-paced developed nations of China, Canada, the United States, and Singapore are the most caffeinated.
    6. Health Conditions
      Over three-quarters of adults experience at least one health condition that impacts sleep. These include insomnia, sleep apnea─which affects roughly 22 million people in the U.S. alone─snoring, restless leg syndrome (RLS), narcolepsy, and chronic pain.

    Developing Good Sleeping Habits

    Sleep is often the first to be neglected with our hectic schedules. Here are a few ways to practice better habits for a good night’s sleep.

    Wake up and go to bed at the same time each day─even on weekends─to establish a more ingrained rhythm for your body clock and help your brain better prepare for sleep.

    Pick a time of day that suits your schedule and energy levels, and be sure to stick with it. Exercise helps to balance melatonin and cortisol levels throughout the day.

    Get outside often during the day and reduce the time spent outside at night. Limit screen time at least 30-60 minutes before sleep.

    Food and Drink
    Avoid eating large meals or drinking alcohol or caffeine in the last couple of hours before you go to sleep. Caffeine effects can linger for up to 8 hours, which breaks natural sleep rhythms.

    Recent studies have shown that mind-body treatments for insomnia such as yoga, tai chi, and meditation had positive impacts on improving sleep quality.

    Set the bed for success—keep your room cool and dark, buy a high-quality mattress and comfortable bed linens and use a white-noise machine to help you fall asleep.

    Sleep is one of the most important aspects of our health; it’s also one of the easiest to neglect. Don’t put yourself into sleep debt─get enough shut-eye to enjoy those sweet dreams.

    Subscribe to Visual Capitalist

    Thank you!
    Given email address is already subscribed, thank you!
    Please provide a valid email address.
    Please complete the CAPTCHA.
    Oops. Something went wrong. Please try again later.

    The post Visualizing the World’s Sleeping Habits appeared first on Visual Capitalist.

  4. kalenjin shared this story from Quartz.

    The US State Department’s hiring freeze in 2017, ordered by Rex Tillerson, its leader at the time, caused staffing shortages that “placed at risk highly classified information,” among other issues. That’s according to an Inspector General’s report released today (pdf) that lays out a broad range of deleterious impacts the freeze had on America’s diplomatic corps.

    An unidentified bureau within the State Department “worked with Top Secret/Sensitive Compartmented Information systems,” and members of it told the IG’s office that “extended vacancies in its information security positions” had endangered the data, the report states.

    Another bureau explained it was unable to fill an Information Systems Security Officer position for the entire duration of the 16-month freeze, which, the report says, “affected its ability to ensure IT security for a major Department system.”

    A State Department data-encryption initiative was also delayed due to the freeze, as was the implementation of an identity-management system, an enterprise risk management program, and the development of “tools and procedures to react and respond to malicious cyber activity targeting Department personnel and information assets.”

    Donald Trump has vowed to reduce the size of the federal workforce, at times threatening to shutter entire agencies. His administration has also described reports of promises to cut Washington’s payroll as “untrue and misleading,” although nearly 20,000 federal jobs were eliminated during Trump’s first 15 months in office.

    Beyond cybersecurity, the department’s 2017 hiring freeze also undermined overseas youth programs, AIDS initiatives, counternarcotics programs, and more.

    Unfilled job openings within the Bureau of Consular Affairs’ Office of Children’s Issues “hampered its ability to support parents whose minor US citizen children had been abducted by family members,” explains the report.

    One agency was “unable to fill a position to support an urgent need related to counterterrorism efforts in Iraq and Syria.”

    And the Bureau of Population, Refugees, and Migration chose “not to fund certain humanitarian programs in the early stages of the Venezuela crisis in 2017 because it lacked staff to oversee the work.”

    Morale suffered badly during the freeze, State Department staffers told IG auditors, with 100% of bureaus and officers and 97% of embassies and consulates describing it as having been “somewhat” or “very” negative.

    “[I]t is impossible to overstate the negative impact of the hiring freeze on employees,” says the report. “Employees felt both overburdened and stuck in their careers, as there was no mechanism for lateral movement or promotion. The hiring freeze conveyed a message from Department top leadership that our work and mission, and the talents and well-being of our employees, were not valued.”

  5. kalenjin shared this story from FlowingData.

    Andrew DeGraff painted maps that show the geography in movies and their characters’ paths. Above is the map for Back to the Future, with 1985 Hill Valley on the top and 1955 Hill Valley on the bottom.

    There’s also a book version. [via kottke]

    Tags: , ,

  6. kalenjin shared this story from Quartz.

    Humans have seemingly perfected the science of staying cool. Air conditioning and refrigeration units are relatively cheap, the mechanics are pretty simple, and access to them is on the rise.

    The downside is that they’re horrible for a planet grappling with a growing climate crisis.

    Globally, about 12% of non-carbon dioxide emissions can be attributed to refrigeration and air conditioners, according to the US Environmental Protection Agency. That number is expected to increase as the earth heats up and populations swell in already-warm climates such as India, which, by launching an air conditioning innovation competition, is on the front lines of the effort to re-envision the globe’s cooling technologies.

    In Delhi, where temperatures during the summer regularly crest 40.5 degrees Celsius (105 degrees Fahrenheit), cooling currently comprises 40% to 60% of summer peak electricity generation. And that’s in a city where a minority of residents use air conditioning at all; about 5% of residents have cooling in their homes. As more Indians enter the middle class, they’ll likely spenddisposable income on basic luxuries that include energy-intensive air conditioning units.

    And who can blame them? Research shows that by 2100, in a best case scenario, close to half the Earth’s population will face 20 days of potentially-fatal heat and humidity each year. Some areas of northeast India, research has shown, will become so hot that being outside for more than a few hours could be deadly, according to The Verge.

    With that in mind, the Indian government in November 2018, along with two global energy nonprofits, launched The Global Cooling Prize. It’s a competition worth $3 million to find new and sustainable technologies to cool the planet’s workspaces and homes. More than 440 applications for the prize were submitted from 56 countries, including from some of the world’s largest air conditioner manufacturers and rising startups.

    It’s an effort to totally re-think how we cool the spaces in which we live and work. Currently we use compression technology, which circulates coolants through a compressor and condenser to remove heat and moisture from indoor air.Rethinking that approach means throwing out every aspect of the compression technology we use today to embrace newer, environmentally-friendly approaches, says Vitalij Pecharsky, a scientist at the Ames Laboratory, a US Department of Energy facilityin Iowa.

    “In my mind, at least, the future of refrigeration is really not in the traditional compression technologies as we know them,” Pecharsky says.

    The globe is expected to add about 700 million new air conditioning units in the next decade, and close to 1.5 billion by 2050, according to research out of the US Department of Energy. But the technology used to cool public and private spaces hasn’t fundamentally changed much since it was invented more than 100 years ago. And that lack of innovation is now a bad thing for the planet.

    Air conditioners and refrigeration systems pose a two-pronged challenge:

    1. It takes a lot of energy to power them. In the US, for example, about 90% of homes have at least one air conditioning unit, and they account for close to 6% of the nation’s total residential energy use. That alone contributes close to 100 million tons of carbon dioxide into the atmosphere every year. In India, the International Energy Agency estimates the peak electricity load from air conditioning could climb by 10% by 2050 if the technology doesn’t modernize.
    2. Common in all these machines are hydrofluorocarbons, a type of industrial chemical used for cooling. These chemicals aren’t harmful unless there is a leak in an air conditioning unit—but leaks are common. When HFCs are released into the atmosphere, they wind up trapping many times more heat in the atmosphere than carbon dioxide.

    There are no shortage of ideas to make air conditioning greener and better. Some of these ideas make current units more efficient, some incorporate a combination of old and new technologies, and still others re-envision cooling entirely. One concept plays with the idea of ditching coolants and using rare earth materials and magnetic fields to generate heat and cold air. Another idea uses an electric field to change the polarization of a material—a rare earth metal—to warm and cool.

    The problem with deploying any of these ideas? Our current units have had more than 100 years to become exceptionally cheap.

    “The incumbent industry has cost-optimized the existing product to a point that it’s very, very, very hard for a new technology to get onto a cost curve that enables it to achieve sufficient scale,” says Iain Campbell, an air conditioning industry veteran and official with The Global Cooling Prize. “For new tech to be able to access the channels of distribution is almost impossible.”

    For that reason, several startups have run into difficulties trying to compete with the cheaper, bad-for-the-planet status quo.

    But the race to create a better way to cool ourselves has already begun. In November 2019, The Cooling Prize will announce 10 finalists, each getting $200,000 to develop prototypes to be tested. In keeping with India’s role with the prize, some of that testing will be in Delhi. For two months straight, contestants will keep their prototypes running in local apartments, judged on efficiency—and the ability to work under intense heat.

  7. kalenjin shared this story from Feed: All Latest.

    Einstein’s theory of gravity and modern particle physics both rely on the idea of symmetry. But physicists are now wondering if symmetry has much more to offer.
  8. kalenjin shared this story from Quartz.


    In October 2017, a telescope operated by the University of Hawaii picked up a strange cigar-shaped object (artist rendering below), which had slingshotted past the sun at a more-than-brisk top speed of 196,000 miles per hour. Scientists at the university dubbed it ‘Oumuamua, Hawaiian for scout, and at first labeled it an asteroid, then a comet, but agreed that it came from another solar system.

    Around the world, telescopes were quickly aimed toward ‘Oumuamua’s path, and scientists dove into the data. One of them, Avi Loeb, the chair of Harvard University’s astronomy department, published a paper in The Astrophysical Journal Letters the following year theorizing that the object could be artificial. “Considering an artificial origin, one possibility is that ‘Oumuamua is a lightsail, floating in interstellar space as a debris from an advanced technological equipment,” he and co-author Shmuel Bialy, a postdoctoral fellow at Harvard, wrote. “Alternatively, a more exotic scenario is that ‘Oumuamua may be a fully operational probe sent intentionally to Earth vicinity by an alien civilization.”

    That’s not something you read every day in a serious scientific journal. The paper went viral and Loeb began fielding an onslaught of media calls while fellow scientists weighed in. In terms of his colleagues’ reaction, Loeb said, “almost all of them reacted favorably, and they thought, you know, it’s just an interesting idea.”

    Even so, he added, there were some negative reactions as well. One cutting tweet by Paul Sutter, an astrophysicist at Ohio State University, reads: “No, ‘Oumuamua is not an alien spaceship, and the authors of the paper insult honest scientific inquiry to even suggest it.” In Forbes, astrophysicist Ethan Siegel called the paper a “shocking example of sensationalist, ill-motivated science.”



    All this hubbub took place in the aftermath of news reports that the Pentagon had been collecting data on UFO sightings for years. Clearly, the hunt for alien intelligence is alive and well in our solar system, and it’s hot news. Indeed, Loeb’s article was approved for publication in mere days.

    But while scientists tossing around the idea of alien life may find a rapt public audience, they can also draw cynical, even hostile reactions from their fellow scientists, a response summed up by acclaimed physicist Neil deGrasse Tyson, who once quipped to CNN: “Call me when you have a dinner invite from an alien.”

    This paradox has ripple effects. The threat of being written off as a kook can loom large for researchers, especially young ones. A lot of academics “won’t touch it with a 10-foot pole,” said Don Donderi, a retired associate professor of psychology at McGill University in Montreal who now teaches a non-credit course called “UFOs: History and Reality” in the school’s continuing education department.

    Loeb says many discoveries have their roots in theories that were initially dismissed. He thinks open-mindedness keeps scientific inquiry moving forward, while shutting down new theories “reduces the efficiency of science.”

    NASA physicist Silvano Colombano maintains that the search for extraterrestrial intelligence has been limited by long-held assumptions and that the “general avoidance of the subject by the scientific community” means no one questions them. It’s a dilemma: scientists might look like cranks for posing questions about aliens, but we’ll also never know unless someone asks.

    Donderi has received his own share of mixed messages about his UFO research. No one at McGill seemed to mind when he began writing about the paranormal in the 1970s, along with experimental psychology. “I was pretty insulated,” he said, as he already had tenure and continued to get promotions, including one to associate dean for graduate studies. At one point, he had a federal grant for research on visual perception and memory, and it came with a provision for extra funding for side investigations. UFOs had been sighted in northern Quebec, so he applied for money to find out what people thought they saw—hence its relation to his main work—but the application was rejected. He never tried again.

    More pointedly, when he formally retired from McGill in 2009, he offered to give a free seminar on his long-time side research into the evidence behind UFOs and alien abductions, and his department said no. Donderi said they knew he was writing on this subject, but when he actually asked to bring this material into their program, they said, that’s the line.

    “I’m really surprised they get blowback,” said physicist Richard Bower of Durham University in England. He’s never gotten flak for his research in cosmology, which entails making computer simulations of possible parallel universes. He’s concluded that life elsewhere could be quite common, and others in his field back him up. “We used to say that life is incredibly rare and we’re lucky to live on a habitable planet,” he said. “But we’ve now observed so many planets that are plausible habitats. It seems, based on scientific evidence, there’s no reason to think that planets like the Earth are rare.”

    But cosmology has its limits, too, and Bower says he’s “less comfortable” with the excessive speculation he sees in some work. Rather, he suggests, it’s better to focus on questions that we may soon have the evidence to answer.

    Legitimate academic work about extraterrestrial life has to carefully separate itself from the musings of amateur scientists and conspiracy theorists. The people who speak at UFO conferences “aren’t all equally good enough,” Donderi said. Meanwhile, those engaged in the search through bona fide organizations have come up with minimal results. The Pentagon acknowledges that it tracked sightings for years, but claims it shut down its program in 2012.

    Meanwhile, astronomers have been trying to communicate with alien life using radio waves since 1959, work continued by the SETI Institute in California and its high-profile scientists—but nothing. Theoretical physics may have determined alien life is likely, but that’s still just a theory.

    We may be finding nothing because we’re doing it wrong, researchers posit. In a 2018 conference paper, Colombano suggests the search for alien intelligence is based on “cherished assumptions” that could be holding it back: that interstellar travel is unlikely, that alien civilizations use radio waves, that other life must be carbon based, and that UFOs have never visited earth. He makes a case for discarding these dusty beliefs and for sociologists to imagine how alien societies might evolve, physicists to do more speculative work on space-time, and energy and technologists to model ways in which technology might evolve in other civilizations.

    Bower says funders and academic institutions favor scientific research with a certain kind of pedagogy. “To search for something in science that you want to know, we look for something and if we find it, we learn this, and if we don’t find it, we learn something else,” he said. Simply looking for alien life is too binary: if you don’t find it, you’ve got nothing.

    For Donderi, as a psychologist, it’s cognitive dissonance that keeps the search for other intelligence in limbo. “People defend themselves against uncomfortable things,” he said. In other words, those who report seeing UFOs will question themselves afterward, and academics will bristle at conclusions that point to aliens.

    Given recent reports that the Navy is creating guidelines for reporting unidentified objects, speculation about UFOs and aliens isn’t going anywhere soon. And researchers expect more data about interstellar objects when the Large Synoptic Telescope in Chile starts operating in 2022. Don Donderi concludes that the evidence is rising and feels that cognitive dissonance is at the moment collapsing. “[W]e’re at the beginning of the change,” he stated.

    Meanwhile, Loeb says he intends to keep revising his ideas on ‘Oumuamua as more data is collected. “I approach this subject scientifically just like I approach any other,” he said. And that’s where patience and hard work come in. “We’re learning the answers to these questions,” said Bower. “Even if we don’t find little green men waving at us.”

    Diane Peters is a Toronto-based writer, editor, and teacher who focuses on science, health, business, and education. Her work has appeared in The Globe and Mail, University Affairs, JSTOR Daily, and other publications.

    This article was originally published on Undark. Read the original article.

  9. kalenjin shared this story from TED Talks Daily (SD video).

    We're not doing frontline exploratory science in a huge portion of the world -- the places governments deem too hostile or disputed. What might we be missing because we're not looking? In this fearless, unexpectedly funny talk, paleoanthropologist Ella Al-Shamahi takes us on an expedition to the Yemeni island of Socotra -- one of the most biodiverse places on earth -- and makes the case for scientists to explore the unstable regions that could be home to incredible discoveries.

    Download video:
  10. kalenjin shared this story from The Citizen Lab.

    Key Findings

    • WeChat implements realtime, automatic censorship of chat images based on text contained in images and on an image’s visual similarity to those on a blacklist
    • WeChat facilitates realtime filtering by maintaining a hash index populated by MD5 hashes of images sent by users of the chat platform
    • We compare levels of filtering across WeChat’s Moments, group chat, and 1-to-1 chat features and find that each has different images censored; we find that Moments and group chat are generally more heavily filtered than 1-to-1
    • WeChat targets predominantly political content including images pertaining to government and social resistance
    • WeChat’s image censorship is reactive to news events; we found censored images covering a wide range of events, including the arrest of Huawei’s CFO, the Sino-US Trade War, and the 2018 US Midterm Elections


    Internet platform companies operating in China are required by law to control content on their platforms or face penalties, under the expectation that companies will invest in the technology and personnel required to ensure compliance. These requirements form a system of intermediary liability or “self-discipline” in which Internet platform companies are held liable for content on their services. Previous work has found little consistency in what content different Chinese Internet platforms censor. However, some high profile Internet platforms are known to frequently receive government directives.

    In this report, we study how Tencent, one of China’s largest Internet companies, implements image filtering on WeChat. WeChat is China’s most popular social media platform with over one billion monthly active users. Its functionality includes 1-to-1 and group chat, Moments (a feature for posting messages and images similar to Facebook’s Timeline), and other social networking features. In a previous report, we studied automatic image filtering on WeChat Moments and found that it employed lengthy, computationally expensive operations to determine if an image is sensitive. However, in this report, we focus on WeChat’s chat functionality, where image filtering, if it is to exist, must operate in realtime.

    We found that Tencent implements realtime, automatic censorship of chat images on WeChat based on text contained in images and on an image’s visual similarity to those on a blacklist. Tencent facilitates realtime filtering by maintaining a hash index populated by MD5 hashes of images sent by users of the chat platform. If the MD5 hash of an image sent over the chat platform is not in the hash index, then the image is not filtered. Instead, it is queued for automatic analysis. If it is found to be sensitive, then its MD5 hash is added to the hash index, and it will be filtered the next time a user attempts to send an image with the same hash.

    This finding indicates that censorship measurement—like the kind conducted in this report—not only evaluates censorship but can also influence and modify the behaviour of a realtime, automatic censorship system by introducing novel items that can be flagged as sensitive and subsequently censored. This helps us understand previous measurements and has implications for future censorship measurement research.

    With an accurate understanding of how chat image filtering works on WeChat, we compare levels of filtering across Moments, group chat, and 1-to-1 chat. We find that each feature has different images censored, where Moments and group chat are generally more heavily filtered than 1-to-1 chat. We also present the first broad content analysis of images censored by Tencent, based on a set of 220 filtered images we discovered and categorized, many of which are event-related or critical of the Chinese government. Popularly referenced topics include the arrest of Huawei CFO Meng Wanzhou, the Sino-US Trade War, and the 2018 US Midterm Elections.

    Previous work

    In previous work, we found that when a message, post, or image is filtered on either WeChat chat features or WeChat Moments, it is hidden from the view of other users whose accounts are registered to mainland Chinese phone numbers, but remains visible to those registered using international phone numbers. This finding means that content filtering on WeChat is non-transparent, as no explicit notice is given to users when content is filtered and the original content remains visible to the user that sent or posted it.

    In the predecessor to this report, we extensively analyzed how images are censored on WeChat’s Moments platform. We found that after images are posted, they are censored using two different techniques: an Optical Character Recognition (OCR)-based method that compares text in the image to keywords on a sensitive keyword blacklist and a visual-based method that compares the posted image’s visual fingerprint to those on a sensitive image blacklist. If an image is deemed sensitive through either of these methods, then it is made invisible to all users with accounts registered to mainland Chinese phone numbers, except the original poster of the image. We found that this process is computationally expensive and that it often takes multiple seconds to censor an image after it has been posted.

    While in that report we studied image censorship of images posted to WeChat Moments, that study did not analyze filtering of images sent in realtime through 1-to-1 chat or group chat. In this report, we study how WeChat identifies sensitive images sent over chat in realtime, despite the computationally expensive demands of detecting an image’s blacklisted text or its visual similarity to that of a blacklisted image. We also expand upon the previous report by providing the first broad analysis of images censored by WeChat.

    Figure 1: Left, a Canadian account sending an image memorializing Liu Xiaobo over 1-to-1 chat; right, the Chinese account does not receive it.

    In an earlier report studying WeChat censorship of content related to the “709 Crackdown” on Chinese human rights defenders, we presented evidence of the existence of censorship of images sent over WeChat’s chat functions. In a later report written in the wake of Liu Xiaobo’s death, we measured the filtering of images related to Liu Xiaobo across WeChat’s Moments, group chat, and 1-to-1 chat functions (see Figure 1), finding that automatic filtering of such content occurred across all three functionalities. We found inconsistent results measuring filtering over the chat functions, finding little filtering over chat one day but much more the next. We also found that Moments had a much higher level of image filtering than group chat. In this report, we explain how these two findings were artifacts created as a product of how WeChat implements realtime image filtering over chat and the research methodology that we used at the time. In doing so, we provide for the first time an accurate measurement of the level of filtering on Moments, group chat, and 1-to-1 chat.

    Research Findings

    Figure 2: Left, an image blocked via OCR methods due to containing blacklisted text (天滅中共); right, an image blocked due to its visual similarity to a blacklisted image.

    In this section, we explain our findings from analyzing the filtering mechanisms used to perform realtime image filtering on WeChat’s chat features. We found that WeChat identifies sensitive images to censor over 1-to-1 and group chat features using the same two methods we had previously found to be used to censor images posted to WeChat Moments; namely, an OCR-based and a visual-based method. Figure 2 shows examples of images which were filtered based on the two respective methods. Both of these methods are computationally expensive and cannot be performed in realtime. We document these general methods in greater detail in the previous report.

    WeChat populates a hash index using images users send over chat

    The OCR-based and visual-based algorithms we discovered being used to filter images on WeChat Moments are too computationally expensive to be applied to realtime filtering of chat. In order to filter images in realtime, we found that WeChat uses another data structure called a hash index1. When a user sends an image, upon receipt by one of WeChat’s servers, the server calculates its cryptographic hash and if the hash is in the hash index, then the image is filtered in realtime instead of being relayed to the intended user.

    Cryptographic hashing is a technique used to quickly map the data contained in a file to a fingerprint, or hash, which is of fixed length and is ideally unique to the file analysed. Such hashes are designed so that the smallest of changes to the hashed file have, on average, as large of changes to the resulting hash as large changes to the hashed file. A change in hashes can confirm that a file has been modified but not by how much the file has been modified. Cryptographic hashes can be computed quickly, and therefore this hashing is amenable to realtime filtering applications unlike more expensive techniques such as OCR or perceptual fingerprints. However, cryptographic hashes are highly inflexible, as even small changes to an image or its metadata radically alter its cryptographic hash.

    To overcome these limitations, we found that WeChat still employs the more computationally expensive OCR-based and visual-based techniques to populate the hash index in non-realtime. We observed this by making small changes to the file metadata of a blacklisted image, which makes no changes to the appearance of the image but still makes substantial changes to its cryptographic hash. Whenever we modified an image in this manner, it would be received unfiltered the first time it was sent. However, when we sent the same modified image again seconds later, it would be filtered. Any modification to the sent image file, including to its metadata, would allow it to evade filtering once, as any such modification would change the file’s hash to one not yet present in the hash index.

    Group chat and 1-to-1 chat have different indexes

    In previous reports, we found that WeChat uses a different blacklist for filtering group chat—a chat featuring more than two users—as opposed to 1-to-1 chat—a conversation between two users. Similarly, although both 1-to-1 and group chat used hash indexes, they did not share the same index. Taking an image that is blacklisted on 1-to-1 chat, group chat, and Moments, and modifying the file so that it has a unique cryptographic hash, we made the following observations:

    1. Sending that image file once over group chat does not cause images with that hash to be filtered on 1-to-1 chat until you also send it over 1-to-1 chat.
    2. Similarly, sending that image file once over 1-to-1 chat does not cause images with that hash to be filtered on group chat until you also send it over group chat.
    3. Sending an image file over Moments does not cause images with that hash to be filtered on group chat or 1-to-1 chat.

    These observations indicate that the hash indexes used by 1-to-1 chat and group chat are independent. This result is surprising, since if an image is blacklisted on both functionalities, then there is no reason for an image’s hash not to be added to both functionalities’ indexes if it is found sensitive on one of them. Similarly, even though we did not find Moments to use a hash index, when a sensitive image is filtered, if it is also blacklisted on 1-to-1 chat or group chat, it would seem worthwhile to perform an inexpensive cryptographic hash of the image file and add it to those hash indexes.

    Choice of WeChat client and settings affect user experience of filtering

    When sending images over WeChat’s chat features, we found that the choice of WeChat client and client settings could affect whether an image is filtered. For instance, if an image file is sent over chat using WeChat’s Windows client, then the outcome of whether the image is filtered may be different than if that image had been sent using the Android client.

    To explain these inconsistencies, we analyzed the behaviour of both the Windows and Android WeChat clients to determine how they handle image uploads. We found that the Windows version performs no modifications to an image before it is uploaded and uploads the original image file. Alternatively, we found that the behaviour of the Android client is partially dependent on how the image is uploaded:

    • If “Full Image” (“发送原图”) is not selected, then the image is sometimes re-encoded. The necessary conditions for image re-encoding are not clear but may depend on the original image’s size.
    • The Android client does not re-encode if you select “Full Image.”

    Figure 3: There are many different ways to encode an identically appearing image, each of which have different hashes. However, if the image is re-encoded by the WeChat client, then any identically appearing image will have the same hash.

    We found that the use of client re-encoding can make the image filtering more powerful in that it effectively results in an image’s hash representing all images with identical pixel values as opposed to merely a specific image encoding. The process of re-encoding extends the generalizability of hash-based filtering because any image containing the same pixel contents will be encoded to the same file and thus have an identical hash (see Figure 3). When this happens, any changes to the original image’s encoding or to its metadata will be ineffective at evading filtering, and some change to the image’s pixel values, if even a small one, will be required to change the resultant hash. When the client does not re-encode, then any change to an image’s file encoding, including to its metadata, is sufficient to change its hash.

    Group chat (and Moments) blacklist more images than 1-to-1

    With an improved understanding of how image filtering happens over WeChat’s chat platforms, we wanted to revisit the question of whether the platform had increased filtering on different features depending on the number of users the content has the potential of reaching.

    Figure 4: Out of 111 images, the # of images censored on each WeChat feature; red is only Moments, green is only group chat, brown is both Moments and group chat, and purple is Moments, group chat, and one-to-one chat.

    On November 15, we tested images previously found censored on WeChat during the 709 Crackdown and on the days surrounding Liu Xiaobo’s passing. Together this dataset comprised 128 images. We found that 111 of these images were still censored on some feature on WeChat. We found that 36 out of 111 images were filtered in 1-to-1 chat, and each of these were also filtered in group chat and on Moments (see Figure 4). We also found that 107 of the 111 images were filtered on both group chat and Moments, with 2 images being filtered only on group chat and another 2 images being filtered only on Moments. This finding suggests that, while the image blacklists for group chat and Moments are largely identical, the reduced filtering on 1-to-1 chat is an indication that WeChat perceives 1-to-1 chat as less of a risk for sensitive conversations due to its more private nature.

    This result contrasts with our observations of image filtering during two days of testing around Liu Xiaobo’s passing, when we found that fewer images were filtered on group chat than Moments. This earlier result was likely an artifact of how image filtering was tested. In this earlier work, since many images were only tested once on the last day of testing, the testing only occurred long enough to insert their hashes into the hash index but not to observe filtering. In fact, across the two days of testing, many images that were tested on the first day which were not filtered initially were found to be filtered when retested on the second day.

    WeChat uses MD5 hashes

    Thus far, due to the observed properties of WeChat’s hash index, we suspected that they were using a cryptographic hash to index images, but we did not yet know which hash specifically. Due to its mention in this Tencent-owned patent that we discovered in a previous report and due to its general popularity as a file hash, we suspected that WeChat may be using the MD5 hashing algorithm to hash images. As a result of vulnerabilities in the MD5 hash function, we were able to test whether this is actually the case.

    Figure 5: Left, a picture memorializing Liu Xiaobo; right, the logo of the Citizen Lab. Using a chosen-prefix collision attack, we modified the contents of these files such that both of these images have the same MD5 hash.

    As a cryptographic hash function, MD5 is well known to be broken in regards to collision resistance and is vulnerable to chosen-prefix collision attacks, a type of attack where specially calculated data blocks are appended to the data of two arbitrary files to generate a pair of modified files which hash identically. Like many image formats, the JPEG standard defines an end of image marker. During the rendering process, all file content after this marker is ignored, meaning that the appended data does not affect the appearance of the image. This behaviour allowed us to design an experiment testing whether Tencent used MD5 to hash images sent over chat. Using Project HashClash, we generated two JPEGs which shared the same MD5 hash, one a blacklisted image of Liu Xiaobo and another of the Citizen Lab logo, an image which is not normally filtered (see Figure 5).2 This operation took about five hours of computation time on an Amazon Web Services (AWS) g3.4xlarge GPU-optimized instance. After sending the blacklisted image, we sent the Citizen Lab logo which we generated to have the same MD5 hash. The Citizen Lab logo was also filtered, confirming that the hash index uses MD5 as its hash function.

    At the additional cost of space, WeChat could maintain a hash index of non-sensitive image hashes, in addition to having a hash index of sensitive image hashes. The use of an additional index would save computational resources, as WeChat would then not have to determine whether any non-sensitive image is sensitive once its hash has been added to the non-sensitive hash index. To test whether WeChat uses a non-sensitive hash index, we generated another collision between a blacklisted image of Liu Xiaobo and of the Citizen Lab logo, and this time we sent the images in the opposite order: the Citizen Lab logo first followed by the Liu Xiaobo image. We then sent the Liu Xiaobo image a second time to test whether it was filtered. We found that sending the Citizen Lab logo first did not prevent the Liu Xiaobo image from being filtered. This result shows that there is no non-sensitive hash index that preempts checking a sent image for sensitivity. One possibility why such a system is not implemented is that a non-sensitive index would need to be invalidated each time WeChat adds a new image to the image blacklist, as hashes previously determined to be not sensitive could have become sensitive.

    Finally, we wanted to test whether sending a non-sensitive image file would remove that file’s hash from the sensitive hash index. To test this, we sent the now-filtered Citizen Lab logo a second time to test whether sending it the first time removed it from the hash index. We found that the image was filtered both times. Thus, sending a non-sensitive image file does not remove its hash from the hash index if it is present.

    The MD5 algorithm’s vulnerability to chosen-prefix collision attacks allows a user to trick WeChat’s image filtering system to filter non-sensitive images by having their hashes collide with those of blacklisted images. However, these images must be under the user’s control and have already been specifically modified by the user, since the MD5 collision attack analyzes and modifies both the sensitive image file and the non-sensitive image file simultaneously. Since image files already being sent over the platform would not generally have been modified in the specific way necessary to create the collision, this attack could not be used to cause popular images already being sent over the platform to be filtered.

    Hash index eviction

    In this section, we explore if and how Tencent decides to remove hashes from their hash index. In our testing, we had anecdotally observed image files which had been added to the hash index to be no longer be in the hash index when tested later. We sought to rigorously measure the hash eviction behaviour of Tencent’s hash index mechanism.

    Figure 6: The six different source images tested.

    We measured hash eviction using six different source images comprising three test pairs, as shown in Figure 6. Each test pair i, 0 ≤ i < 3, consisted of one source image vi filtered by visual similarity and one source image oi filtered by OCR. Each image filtered by visual similarity were politically sensitive images: one concerning Liu Xiaobo’s death, one related to the 709 Crackdown, and one of the historical picture of “Tank Man.” Each image filtered by OCR were images generated containing a politically sensitive keyword combination known to be filtered. The three sensitive keyword combinations we used were “国家领导人 [+] 无限延长” (national leader [+] unlimited extension), “LXB [+] 总书记” (Liu Xiaobo [+] General Secretary), and “08宪章” (Charter 08). To generate an image from each of these phrases, we created an image containing the keyword combination multiple times in a variety of font sizes and styles. Specifically, we created an image with ten lines each containing the keyword combination repeated three times, where the font size of line , 0 ≤ < 10, is 12 + 4·⌊ / 2⌋ and where the even lines are written in a 黑 (Hei) typeface and the odd lines are written in a 明体 (Ming) typeface.

    Figure 7: Left, the original image v2; right, a canary modified to be grayscale and to have three blocks of modified pixels in the upper-right corner.

    For each visually filtered source image vi and each OCR filtered source image oi, we created 33 subtle variations we call canaries. We use these to measure the lifespan of each image’s hashes in the hash index. When a canary “dies” (i.e., the hash index is no longer filtered), then we know that that hash index has been evicted.

    For a given source image, we create canary d, 0 ≤ d < 33, as follows. We first convert the canary’s source image to grayscale. Then we modify a region of pixels in the image such that, for each grayscale pixel value g in this region, we replace it with value (g + 128) mod 256, where 256 is the maximum value of the grayscale channel, effectively changing each pixel value by half of its intensity range. The region of pixels to modify in this manner is a rectangle whose dimensions and location are determined as a function of d. Vertically, the region always has a height of 8 pixels and begins at the top edge of the image extending downward. Horizontally, the region has a width of 8·(d + 1) pixels and begins at the right edge of the image extending leftward. Figure 7 shows an example of the modifications we made to create one canary.

    To test eviction behaviour, for each test pair i, we primed WeChat’s hash index on November 15 + i with visually filtered source image vi and OCR filtered source image oi. To prime the cache, we first sent each canary of vi and oi. In each case we found that they were never filtered. Minutes afterward we resent each canary to confirm that the canaries’ hashes were now in WeChat’s hash index. We found that in each case they were filtered as they were now in the hash index.

    To test how long each canary survived in the hash index, each day we formed a test for the eviction of a canary. For each test pair i, we again sent its canary d in the afternoon of the day d days after the images were originally planted in the hash index (November 15 + i + d) to test if the canary had survived in the index d days. In other words, for each source image we tested whether one of its canaries was evicted each day up to a month later.

    Figure 8: Left, v1, an image depicting Hong Kong censorship; right, o2, an image referring to Charter 08, a human rights manifesto; some canaries of each of these images evaded filtering.

    Figure 9: Days between November 16 and December 18 when image v1’s canary image evaded filtering.

    We found that every canary except ten of v1 and one of o2 were filtered (see Figure 8). Only one canary of o2, the one tested on December 10, evaded filtering. The ten canaries that evaded filtering of v1 are as shown in Figure 9.

    To test eviction up to a second month, we performed a second experiment by reusing our canaries which were replanted or refreshed during our previous experiment. For each test pair i, we tested all 33 of each of vi and oi’s canaries in the evening of January 16 + i. Note that the test date is not a function of d as the canaries were already replanted or refreshed each day for a period of one month. Thus, each source image’s canaries must be tested on the same day. In this experiment, we found that all of these canaries were filtered (i.e., that none evaded filtering).

    At a high level, we discovered a general trend that there is a higher chance of an image being evicted as time goes on. However, the results were largely nondeterministic, and it would be unreliable to depend on image hash eviction behaviour to evade image filtering. In future work, testing eviction over a longer test period may possibly produce more clear cut results.

    One observation we found is that eviction of a hash appears to be a function of the blacklisted image to which it corresponds, which we observed in only v1 experiencing a large amount of eviction. This observation may be a result of each blacklisted image having its own independent limit for the number of hashes allowed in the index. If this hypothesis is true, then this would suggest that v1 may have been more frequently shared on the platform at the time of testing.

    Blacklist removal

    In some cases, we observed that an image had been removed from WeChat’s blacklist but its MD5 hashes nevertheless appeared to remain in the hash index. For instance, on May 30, 2019, we observed this behaviour with the iconic “Tank Man” photo (see Figure 7) in 1-to-1 chat. If we modified the metadata in the image file, then the photo was never filtered. However, if we sent the image unmodified, then it was. This inconsistency suggests that hashes of the image were in the hash index but that the image was no longer blacklisted and that new hashes were no longer being added to the index. Thus, it would seem that when images are removed from the blacklist, their hashes are not immediately removed from the hash index.

    Summary of research findings

    We found that WeChat uses OCR- and visual-based filtering to identify sensitive images posted to Moments, group chat, and 1-to-1 chat. Filtering on Moments occurs in non-realtime after images are analyzed, as this is a computationally expensive process. To implement realtime filtering for images sent in chat, WeChat uses a hash index of known sensitive images and filters any images whose MD5 hash exists in the index. Separate hash indexes are maintained for group chat and 1-to-1 chat, and if an image is not found in the respective index, it is delivered to its intended recipient(s) but queued for later analysis for sensitivity by OCR- and visual-based methods. If an image is found to be sensitive, it is added to the hash index, preventing the future delivery of the same file. We found that known hashes of sensitive images can be removed from the index after some time; however, it is currently unclear how the index evicts hashes. Moreover, even after an image is no longer blacklisted, we found that hashes of the file can persist in the index.

    Analyzing filtered image content

    In this section, we aim to characterize what images are filtered and identify their topics. In previous work, we used automated testing to determine which keyword combinations are filtered on WeChat. While this approach is generalizable to image filtering, it resulted in WeChat banning our test accounts because it suspected them to be spam. Moreover, we found that new accounts required approval from a second account that must have existed for over six months, be in good standing, and have not already approved any other accounts in the past month. Because of these requirements, we found that creating new WeChat accounts was prohibitively difficult.

    To automatically test for image filtering on WeChat, we investigated other Tencent-operated services that may be using the same or similar image blacklists as WeChat. Previous research has found positive but not complete correlation between companies and the content filtered across their products. We found that Tencent’s Qzone (QQ空间), a blogging platform which has been the subject of previous censorship research, used the same two methods to filter images as WeChat—an OCR-based and a visual-based method—and that its blacklist appeared similar to that of WeChat’s.

    In the remainder of this section, we describe our method for identifying 220 filtered images on Qzone. We then characterize and analyze these 220 images. Finally, we select a random sample of these 220 filtered images and test whether they are also filtered on WeChat in order to evaluate how similar the image blacklists are between these two Tencent platforms.

    Measuring filtered images

    Our previous work analyzing image filtering on WeChat has focused on specific topics of censorship, such as Liu Xiaobo or human rights crackdowns. However, in this work we aimed to gain a broader understanding of which images Tencent automatically filters. In order to do this, we sampled content from a broad set of images likely to be filtered. To construct our sample, we utilized WeChatscope, a project that automatically records deleted posts on the WeChat Public Posts platform. We looked at posts WeChatscope detected as deleted within a six month period, between October 17, 2018 and April 17, 2019. We constructed our sample set to consist of every image from each of these deleted posts.

    Figure 10: An album containing censored images on Qzone.

    To test whether each image was automatically filtered, we wrote a Python script to automatically upload each image to Qzone. After uploading five images in succession, the program waits 60 seconds. It then checks to see if each image has been replaced with that of a placeholder image communicating that the original image has been deleted (see Figure 10), which we detect as any file beginning with the GIF magic number and containing fewer than 2048 bytes. We ran this script from March 14, 2019, to April 22, 2019. During this test period, the QQ account that we used for testing was never banned.

    Analysis of filtered images

    Using the technique described in the previous section, we found 220 filtered images. Eight of them had duplicated content, and so we excluded them from the analysis. We analyzed the remaining 212 images based on their content and the context of the original WeChat public account articles they originated from.

    Figure 11: Distribution of censored images by category.

    We coded each image into content categories based on a code book we developed for this report. The most popular content categories were Chinese Government (75 images) and Events (60 images). Figure 11 shows a full breakdown. In the remainder of this section, we present a curated list of examples from a number of these categories.


    We found 75 censored images with content related to the Chinese government. These include not only images critical of the government such as sarcastic cartoons (see Figure 12) but also neutral representations of government policy and photos of government leaders and party cadres.

    Figure 12: The caption in the cartoon reads, “Emperor: Who cares about whether law or power is more powerful? In the end, I am the one that dictates everything.”

    It is unclear how Tencent decides which images to filter. In previous research, we found that WeChat censored neutral keywords referencing official party policies and ideology around highly sensitive events. Similarly, this report finds that censored images of or alluding to government leaders and party cadre were not necessarily negative.

    Figure 13: Screenshot of a European news program.

    In one example (see Figure 13), we found that a screenshot captured from a news broadcast by Euronews, a European television network based in France, was filtered. The news clip is about an artist in Italy who used a tractor to create a huge portrait of Chinese President Xi Jinping smiling ahead of Xi’s visit to Italy in March 2019.


    Chinese social media censorship is often reactive to news cycles, and companies tend to tighten their information controls around sensitive events. We found a total of 60 images that reference 10 distinct events: eight news events which happened around or close to our test period and two historical events.

    Events Note # Images Censored
    Cultural Revolution A nationwide movement launched by Chairman Mao Zedong in 1966 to preserve communist values. The movement lasted until 1976. 4
    1989 Tiananmen Movement The Tiananmen Square protests in 1989 are a persistently taboo topic in China. 1
    Fan Bingbing Tax Evasion Scandal Fan Bingbing, one of China’s highest-earning entertainers, was caught in a tax evasion scandal in 2018. 2
    2018 Chongqing Bus Crash On October 28, 2018, a bus plunged off the Second Wanzhou Yangtze River Bridge into the Yangtze River in Wanzhou District, Chongqing, causing at least 15 deaths. 2
    2018 US Midterm Elections US midterm elections were held in November 2018 featuring hundreds of congressional, state, and local seats in contest. 3
    Supreme People’s Court Scandal In December 2018, the Supreme People’s Court of China acknowledged the disappearance of court documents related to a contract dispute between two mining companies. 24
    Huawei Saga Huawei’s CFO Meng Wanzhou was arrested in Canada under charges of fraud and violating international sanctions. 10
    Sino-US Trade War The so-called trade war between China and the United States engaged via increasing tariffs since 2018. 8
    CRISPR-baby Scandal Chinese scientist He Jiankui announced in February 2019 the birth of twin girls with edited genomes which he engineered. The announcement triggered international debates over research ethics. 2
    2019 Chengdu Campus Food Scandal In March 2019, mouldy bread and rotting meat were found at Chengdu No 7 Experimental High School. The incident caused protests among parents. 2
    2019 Sichuan Forest Blaze In late March 2019, a forest fire broke out in Sichuan. Thirty firefighters sent to tackle the fire were killed during the mission. 2

    Table 1: Events referenced by images censored on Qzone, in chronological order. Some events above are still unfolding.

    Descriptions of the events referred to by images in our dataset are listed in Table 1. A large number of event-related images in our dataset referenced scandals, the Sino-US Trade War, and the arrest of Huawei CFO Meng Wanzhou.

    Figure 14: Letter allegedly written by Wang Linqing

    The highest percentage of event-related censored images was related to a scandal involving China’s Supreme People’s Court. In late December 2018, prominent former television host Cui Yongyuan posted a series of posts on microblogging platform Weibo about the suspicious disappearance of several Supreme Court documents regarding a 2016 dispute over billions of dollars’ worth of mining riches. According to Cui, the tips came from former Chinese Supreme Court judge Wang Linqing, who claimed that he was the first to find out about the missing court papers. Wang and Cui’s revelation created controversy in which the Supreme Court first dismissed their words as “rumours” but soon said that it would investigate the case. It caught national attention not only because powerful figures from the highest echelons of the Chinese Communist Party were involved, including current Chief Justice Zhou Qiang, but also because of how the case evolved. In a shocking twist, whistler-blower Wang Linqing said that he was the one who stole the papers. We found 24 images referencing the scandal, almost all of which are photocopies of a letter allegedly written by Wang Linqing, in which he accused Chief Justice Zhou Qiang of direct involvement in the disappearance of the court papers and illegal interference in the mining case in dispute (see Figure 14 for an example).

    Figure 15: Censored images referencing the arrest of Meng Wanzhou.

    We found that nearly all images related to the Huawei saga were referencing the arrest of Meng Wanzhou, Huawei’s CFO and daughter of Huawei’s CEO Ren Zhengfei. Meng was arrested on December 1, 2018 by the Canada Border Services Agency under US charges of fraud and violating international sanctions. Meng’s arrest quickly escalated to a diplomatic incident between China, Canada, and the United States. We found some censored images that contained negative sentiment towards Meng or Huawei, including photos of demonstrations led by overseas Chinese in support of extraditing Meng to the US but also neutral news coverage of the case (see Figure 15).

    Figure 16: Two censored images related to the 2018 US midterm elections.

    We found multiple images related to the 2018 US midterm elections (see Figure 16). In November 2018, elections were held in the United States with hundreds of congressional, state, and local seats in contest. Leaked directives show that Chinese regulators banned live-streaming and live updates of US elections likely for fear that it would encourage citizens to discuss the future possibility of open and free elections. Although elections are held in China, they differ from those in many democratic countries in two significant ways. First, under China’s System of Multi-party Cooperation and Political Consultation, China is a de facto single-party state with other parties only having nominal representation. Second, elections in China feature limited direct election, with direct election being reserved to district-level or village-level positions. Elections of People’s Congresses in China feature multiple levels including national, provincial, and often more than one municipal level, where each level except the bottommost is indirectly elected by the level below. The highest office, the President, is elected by the national-level People’s Congress.

    Social Issues

    We found 32 images censored in the Social Issues category, which covers a range of content including the sale of illicit goods (e.g., see Figure 17), prurient interests, and non-political memes.

    Figure 17: A flyer advertising firearms for sale (China does not allow the possession or sale of firearms).

    Figure 18: This image was seen in a WeChat article in which the author shared his experience of coming out to his mother; the text in the image describes “an ostrich’s escape,” which refers to the Chinese notion that ostriches bury their heads in the sand when facing troubles.

    Figure 19: An image criticizing the exploitation of animals.

    We found that much of the content involving nudity was likely blocked out of context. For example, the image in Figure 18 was extracted from a diary style article in which the author detailed his experience of coming out to his mother. The image in Figure 19 was from an article criticizing the exploitation of animals but was likely blocked due to sexually implicit content.

    Social Resistance

    Previous research suggests that Chinese social media platforms censor keywords pertaining to names of outspoken dissidents and advocates of social movement. We found 25 images in this category, including symbols and acts of resistance, from within China or overseas, against the Chinese government.

    Figure 20: Photo of Cui Yongyuan, former television host and well-known social media commentator and whistle-blower.

    Figure 21: A chair reserved for Liu Xiaobo at the 2010 Nobel Peace Prize ceremony.

    Figure 22: Hosts of a Polish YouTube channel express their support for Taiwan.

    Among the 25 images, there are photos of individuals known for being outspoken about Chinese current affairs such as Cui Yongyuan (see Figure 20), screenshots of social media posts pertaining to online petitions, symbols of political resistance such as the iconic empty chair representing late Chinese Nobel Prize winner Liu Xiaobo (see Figure 21), as well as photos of overseas protests for Hong Kong and/or Taiwan against China’s influence (see Figure 22).

    Figure 23: Image was extracted from a WeChat article criticizing the expansion of ISIS.

    Other censored images include references to terrorism and religious extremism (See Figure 23 for an example) and financial phishing attacks.


    Some images that we discovered censored had no clear reason for being on Tencent’s blacklist. In our previous work studying keyword-based censorship, we found keywords with no clear reason for being blacklisted as well, and so we also expect a small number of blacklisted images to have no clear motivation for being blacklisted.

    Figure 24: A photo of primatologist Jane Goodall and infant chimpanzee Flint.

    One such case is a picture of famous primatologist Jane Goodall with an infant chimpanzee (see Figure 24), which we found blacklisted by Tencent. One possible reason for this image being sensitive is due to the use of chimpanzees in recent Chinese events as a racially derogatory reference, but the rationale for blacklisting this particular image is largely unclear.

    Evaluating Qzone’s versus WeChat’s image blacklist

    Measuring which images are filtered on Qzone provides insight into what Tencent broadly considers sensitive on their platforms but does not necessarily measure what is filtered on WeChat. To assess whether Qzone can be used as a proxy for testing image censorship on WeChat, we empirically tested whether Tencent’s Qzone and WeChat platforms shared image blacklists.

    To perform this evaluation, on April 22, 2019, we randomly sampled 30 images from the images we found filtered on Qzone and automatically re-uploaded them to Qzone to confirm that they were still filtered on Qzone on the date of testing. Among those still filtered, we manually uploaded them to WeChat Moments using one of our remaining test accounts. We then attempted to view each uploaded post from a different China-based account and measure which were deleted and which were not. We considered the two platforms consistent with respect to an image if and only if they both filtered an image.

    We found that 29 out of 30 images (97%) that we had originally found filtered on Qzone were still filtered on Qzone at the time of testing. Among those 29, we found that 24 images (83%) that were censored on Qzone were also censored on WeChat Moments. These results show that image filtering on Qzone is largely effective at predicting filtering on WeChat Moments with few false positives.

    Because we had previously found that some images filtered in WeChat group chat are not necessarily filtered in Moments, we decided to perform a second experiment comparing whether images filtered on Qzone were filtered on either Moments or group chat. In this experiment, we considered the Qzone and WeChat platforms consistent with respect to image filtering if an image found to be filtered on Qzone was also filtered on either WeChat Moments or group chat. While this is a broader definition, it takes into account that WeChat is not consistent with itself on many images.

    In this experiment, we found that 27 out of the 29 images filtered on Qzone (93%) were also filtered on either WeChat Moments or group chat. This finding shows that image filtering on Qzone is highly effective at predicting whether an image is censored somewhere on the WeChat platform with very few false positives.

    It is not clear why some images are only filtered on Qzone as opposed to WeChat and vice versa. One explanation is that the blacklists are tailored according to which images are posted on each platform. Another explanation is that one platform is using an older or newer version of the same list as the other and that the two platforms may eventually become consistent with the same blacklist. A final explanation is that both platforms use the same image blacklist but that due to some sort of image processing done on one platform and not the other, some uploaded images after being processed no longer match the blacklisted image fingerprint on one platform versus the other.

    Nevertheless, we found using Qzone as a proxy for measuring image filtering on WeChat to be effective. We suspect that this technique may be generalizable for future censorship research in any situation where two different platforms, possibly due to having the same operator, are suspected to use similar blacklists, and where one platform may be easier to access or measure than the other, despite possibly being not the original platform of interest.

    Origin of blacklisted images

    During our testing, we found that, for many censored WeChatscope articles, all images in the article were individually blacklisted, even if some of the images in the article were not themselves sensitive.

    Figure 25: Screenshot of a presentation slide summarizing “popular views on Sino-US trade war”; the slides appear to be made by a research institute affiliated with the Evergrande Group, one of China’s largest property developers.

    In one instance, we found that five images were from the same WeChat public account article published by Global Times, a Chinese nationalistic tabloid newspaper owned by People’s Daily. These images were not necessarily negative towards Huawei and the original article was criticizing overseas Chinese dissidents for being anti-China. In previous research, we found that Tencent applied broad censorship to sensitive events, restricting not only negative information about the government but also neutral references to government policy and screenshots of official announcements accessible via government websites. It is likely that Tencent over-censored images related to an event that it perceives as sensitive to the government or would cause unwanted public attention.

    In another instance, we observed this same pattern in an article related to the so-called trade war between China and the US. In this article we found five images blacklisted, even though none were seemingly sensitive (see Figure 25 for an example).

    These observations lead us to hypothesize that many of the images that Tencent blacklists are gathered from sensitive WeChat Public Post articles. If true, it is still unclear whether or not such images represent a large proportion of Tencent’s image blacklist as a whole.


    In this work, we study how Tencent implements image filtering on WeChat. We found that Tencent implements realtime, automatic censorship of chat images on WeChat based on what text is in an image and based on an image’s visual similarity to those on a blacklist. Tencent facilitates realtime filtering by maintaining a hash index of MD5 hashes of sensitive image files.

    We set out many of the properties of their hash index implementation, but we were most surprised to find that this hash database was populated by images that users, including both ordinary users as well as researchers, send over the platform. As researchers this leads to challenges wherein performing measurements can change the result of future measurements. It reminds us to be cautious and aware of that, even in systems of automated filtering, past measurements can change the future behaviour of filtering on the platform.

    Our topic analysis of Tencent’s image content filtering identified 220 filtered images. These images were largely related to recent events and other political or government-related topics. Although we found references to a number of domestic Chinese scandals filtered, we also found images related to the 2018 US midterm elections. While we might expect Chinese censors to be sensitive to domestic criticism, this reminds us that even the reference to an outside alternate form of governance may also be sensitive.

    Due to recent difficulty in obtaining and maintaining WeChat accounts, we analyzed the topic of images filtered on WeChat by proxy by measuring image filtering on Qzone, another Tencent product. Using sample testing, we found this technique largely effective. We hope that this technique of measuring content filtering on an easy to measure platform using a similar filtering implementation as a more popular but difficult to measure platform generalizes to future applications.


    This project was supported by Open Society Foundations. We would like to thank an anonymous researcher for assistance in translating and categorizing the images featured in this report. We would also like to thank Ron Deibert for advising this work, Simon Humbert for technical assistance, as well as Masashi Crete-Nishihata, Miles Kenyon, and Adam Senft for helpful suggestions and peer review.

    The full image data set that we collected and the codebook we used to categorize each image are available here.

    1. In this report, we use image blacklist to refer to the list of blacklisted images to which an image sent over chat is visually compared in order to determine if it is sensitive. We use keyword blacklist to refer to the list of keyword combinations that cause an image to be filtered if the image is determined to contain one of these keyword combinations via OCR. We use hash index to refer to the index of MD5 hashes that Tencent uses to filter images in realtime chat based on whether those images were previously found to be on the image blacklist. We use sensitive image to refer to an image that, if sent, would be added to the hash index via either one of the visual-based or OCR-based methods.
    2. Since our initial generation of a chosen-prefix collision, a new exploit for creating MD5 hash collisions in specific image formats has been published which can be performed computationally instantly. Using the same two JPEGs, we generated another collision using this method and repeated our experiment, finding the same results.

    The post (Can’t) Picture This 2: An Analysis of WeChat’s Realtime Image Filtering in Chats appeared first on The Citizen Lab.