# notes --- - [[data voids are gaps in information online]] - [[media manipulators exploit data voids by capitalizing on missing data and the logic of algorithms and search engines]] # executive summary --- The logic underpinning search engines is akin to a lesson from kindergarten: no question is a bad question. But what happens when innocuous questions produce very bad results for users? Data voids are one such way that search users can be led into disinformation or manipulated content. These voids occur when obscure search queries have few results associated with them, making them ripe for exploitation by media manipulators with ideological, economic, or political agendas. Search engines aren’t simply grappling with media manipulators using search engine optimization techniques to get their website ranked highly or to get their videos recommended; they’re also struggling with conspiracy theorists, white nationalists, and a range of other extremist groups who see search algorithms as a tool for exposing people to problematic content. Data voids are difficult to detect. Generally speaking, data voids are not a liability until something happens that results in an increase of searches on a term. Some are created by media manipulators, and escape notice for long periods of time. Others are the sudden products of a news spike, as millions are prompted to search names or terms for the first time, and misleading or hateful content is created to meet demand. Search-adjacent recommendation systems, like search bar auto-suggestions, further complicate the data voids problem by providing auto-suggestions that can send people down deeply disturbing paths. Search engine creators want to provide high quality, relevant, informative, and useful information to their users, but they face an arms race with media manipulators. In this report, we focus on five types of data voids that are currently being corrupted by those spreading conspiracies or hate: Breaking News: The production of problematic content optimized to terms that are suddenly spiking due to a breaking news situation; these voids will eventually be filled by legitimate news content, but are abused before such content exists. Strategic New Terms: Manipulators create new terms and build a strategically optimized information ecosystem around them before amplifying those terms into the mainstream, often through news media, in order to introduce newcomers to problematic content and frames. Outdated Terms: When terms go out of date, content creators stop producing content associated with these terms long before searchers stop seeking out content. This creates an opening for manipulators to produce content that exploits search engines’ dependence on freshness. Fragmented Concepts: By breaking connections between related ideas and creating distinct clusters of information that refer to different political frames, manipulators can segment searchers into different information worlds. Problematic Queries: Search results for disturbing or fraught terms that have historically returned problematic results continue to do so unless high quality content is introduced to contextualize or outrank such problematic content. Data voids raise questions about what role search engines can and should play in diverting their users from disturbing search results. We argue that there is no “fix” for data voids. Search engines and content creators must work together to identify these vulnerabilities, iteratively respond to attacks, and produce the high-quality content that is needed to fill these data voids. # highlights --- ## introduction >[!highlight|#ffd400]+ pg. [5](zotero://open-pdf/library/items/L2K3PJR7?page=5&annotation=GLV5CELL) > >While these are frequently designed to help increase clarity for the search engine, they may also invite users to traverse a network of information into areas that the searcher never previously considered. > > > > > > >these = recommended searches >[!highlight|#ffd400]+ pg. [6](zotero://open-pdf/library/items/L2K3PJR7?page=6&annotation=TFAUNCR8) > >There are many search terms for which the available relevant data is limited, nonexistent, or deeply problematic. Recommender systems also struggle when there’s little available data to recommend. We call these low-quality data situations “data voids.” Data voids lead to low quality or low authority content because that’s the only content available. They come about both naturally and through manipulation. >[!highlight|#ffd400]+ pg. [6](zotero://open-pdf/library/items/L2K3PJR7?page=6&annotation=2MIRATD6) > >Some of these data voids are intentionally exploited to introduce disturbing content, while others are created to promote political propaganda. Moreover, data voids are difficult to detect. Some are created by obscure search queries that escape notice for long periods of time. Others are the sudden products of a news spike >[!highlight|#ffd400]+ pg. [7](zotero://open-pdf/library/items/L2K3PJR7?page=7&annotation=HPDX9JH7) > >Media manipulators have learned to capitalize on missing data, the logics of search engines, and the practices of searchers to help drive attention to a range of problematic content. > > > > ## How Search Engines Work >[!highlight|#ffd400]+ pg. [9](zotero://open-pdf/library/items/L2K3PJR7?page=9&annotation=NAYXXZJB) > >what people search for is often vague. > > > > >[!highlight|#ffd400]+ pg. [9](zotero://open-pdf/library/items/L2K3PJR7?page=9&annotation=4LI6SJXG) > >The flip side of vague queries is narrow content > > > > >[!highlight|#ffd400]+ pg. [10](zotero://open-pdf/library/items/L2K3PJR7?page=10&annotation=SYFJS599) > >Search engines will need to continue to build improved models that better understand related terms in order to match a user’s intention and the content that it might be able to return. > > > > >[!highlight|#e56eee]+ pg. [10](zotero://open-pdf/library/items/L2K3PJR7?page=10&annotation=XZ3DEW7M) > >For example, a search for “green poop” should probably return similar results as a search for “green stool” because they are commonly used synonymously. Yet, by collapsing these two concepts, a search engine’s model creates a situation in which parenting forums discussing children’s bowel movements as “poop” are ranked alongside medical and scientific content concerning “stool” (not to mention green IKEA stools for sitting). >[!highlight|#ffd400]+ pg. [12](zotero://open-pdf/library/items/L2K3PJR7?page=12&annotation=LVQVVVR4) > >Modern search engines all extend early work on “page rank.” Page rank involves scraping web pages to determine the link structure and then ranking pages based on the inbound links that a given page receives. > > > > >[!highlight|#ffd400]+ pg. [12](zotero://open-pdf/library/items/L2K3PJR7?page=12&annotation=U8Z2LCJG) > >It didn’t take long for people to game Google and increase the visibility of their content on the new search engine. > > > > >[!highlight|#2ea8e5]+ pg. [14](zotero://open-pdf/library/items/L2K3PJR7?page=14&annotation=FRMNSE4L) > >Today, many users approach platform-specific search features as though they are search engines for the entire web. Young people often treat YouTube as their primary search engine when seeking general information.6 > > > > >[!highlight|#ffd400]+ pg. [14](zotero://open-pdf/library/items/L2K3PJR7?page=14&annotation=VTWN4ZY6) > >As platform-specific search features have grown in significance, platform-specific SEO has also proliferated. Yet the amount of data available on these specific platforms pales in comparison to what sites like Google or Bing can provide, increasing the likelihood that people will encounter problematic or manipulated content. > > > > >[!highlight|#ffd400]+ pg. [14](zotero://open-pdf/library/items/L2K3PJR7?page=14&annotation=JWGCL9NU) > >, attempts to influence recommender systems are sometimes referred to as recommendation engine optimization > > > > >[!highlight|#ffd400]+ pg. [15](zotero://open-pdf/library/items/L2K3PJR7?page=15&annotation=PYQBDWKS) > >**Anywhere that a search or recommender system makes decisions based on public data, there is an opportunity for determined, data-literate manipulators to influence other users’ exposure to content.** > > > > ## From Voids to Vulnerabilities >[!highlight|#ffd400]+ pg. [17](zotero://open-pdf/library/items/L2K3PJR7?page=17&annotation=KZX5LJDT) > >Data voids exist because of an assumption baked into the design of search engines: that for any given query, there exists some relevant content. But this is simply not true. > > > > >[!highlight|#ffd400]+ pg. [17](zotero://open-pdf/library/items/L2K3PJR7?page=17&annotation=5A63EGTE) > >we first need to understand the specific type of vulnerabilities that allow data voids to be exploited in the first place. > > > > ### DATA VOID TYPE #1: BREAKING NEWS >[!highlight|#ffd400]+ pg. [17](zotero://open-pdf/library/items/L2K3PJR7?page=17&annotation=KNSH6V8F) > >When news is breaking, journalists and other content creators produce new material that must be integrated into search engines. At the same time, a wave of new searches that have not been previously conducted appears > > > > >[!highlight|#ffd400]+ pg. [18](zotero://open-pdf/library/items/L2K3PJR7?page=18&annotation=GB28L7Y9) > >media manipulators often seek to capitalize on breaking news in order to influence public perception. > > > > >[!highlight|#e56eee]+ pg. [18](zotero://open-pdf/library/items/L2K3PJR7?page=18&annotation=AIPYZNNF) > >Consider what happened on November 4, 2017. On that Sunday, many people in the United States picked up their phone to a notification that an active shooting was underway in Sutherland Springs, Texas. As the day unfolded, the public learned that a disgruntled white man had walked into a Baptist church and opened fire on worshippers. But the notification that people received on their phones did not provide additional information; the only specific information was the location. With no additional information, people began searching in unprecedented numbers for “Sutherland Springs Texas.” > > > > >[!example|#ffd400]+ pg. [20](zotero://open-pdf/library/items/L2K3PJR7?page=20&annotation=66XYDZKA) > >A Sutherland Springs-style data void is easy to manipulate because automated map and weather data pages are easy to surpass in relevance. Without a breaking news situation, such manipulations would have an audience of none; after all, the void exists because no one is searching for these terms in the first place. But this can all change when a news event occurs. > > > > >[!highlight|#ffd400]+ pg. [20](zotero://open-pdf/library/items/L2K3PJR7?page=20&annotation=68BC4CZN) > >Unfortunately, the time between the first report and the creation of massive news content is when manipulators have the largest opportunity to capture attention. > > > > > > >Is this gap closing? >[!example|#ffd400]+ pg. [21](zotero://open-pdf/library/items/L2K3PJR7?page=21&annotation=HGVXNI57) > >Shortly after the announcement of the shooting in Sutherland Springs, a distributed network of people began coordinating on various forums in an effort to shape media coverage and search engine results. > > > > >[!highlight|#ffd400]+ pg. [21](zotero://open-pdf/library/items/L2K3PJR7?page=21&annotation=I6JKCGGM) > >To increase the likelihood of visibility of their content on search engines, they tweeted and posted content that includes words and phrases related to the incident in the early moments before higher-authority content (news media) appeared. > > > > >[!highlight|#ffd400]+ pg. [21](zotero://open-pdf/library/items/L2K3PJR7?page=21&annotation=8IJH79PK) > >Even when journalists are aware of manipulation, the limitations of search engines can create powerful pockets of low-quality information. > > > > >[!example|#e56eee]+ pg. [21](zotero://open-pdf/library/items/L2K3PJR7?page=21&annotation=6KL64JQT) > >Quickly after the Sutherland Springs shooting, a reporter at Newsweek uncovered the coordinated attempts to manipulate the story. He wrote a scathing article detailing the manipulation with an unfortunate headline: “‘Antifa’ Responsible for Sutherland Springs Murders, According to Far-Right Media.” This headline causes harm because both Google and Bing chop long news headlines when displaying them at the top of search. Thus, in the first critical hours, most searchers were shown the truncated phrase: “‘Antifa’ Responsible for Sutherlan > > > > >[!highlight|#ffd400]+ pg. [22](zotero://open-pdf/library/items/L2K3PJR7?page=22&annotation=XEFZP88R) > >Breaking news situations are typically time bound as a news story runs its course. > > > > >[!highlight|#ffd400]+ pg. [22](zotero://open-pdf/library/items/L2K3PJR7?page=22&annotation=S75LNBPI) > >Because of the wave of high-quality content produced after the shooting, this kind of data void is cleaned up naturally, but the damage created by the data void was done in the hours after the shooting when search engines amplified content designed to manipulate public perception. > > > > ### DATA VOID TYPE #2: STRATEGIC NEW TERMS >[!highlight|#ffd400]+ pg. [22](zotero://open-pdf/library/items/L2K3PJR7?page=22&annotation=FDDUL8ET) > >there are other techniques for exploiting data voids that attempt to establish a longer-running narrative. One such approach involves the strategic creation of new terms to divert discourse and search traffic alike into areas full of disinformation. >[!highlight|#ffd400]+ pg. [23](zotero://open-pdf/library/items/L2K3PJR7?page=23&annotation=4TEHY9V8) > >This technique’s focus on specific terms not only preys on the infrastructures of hashtags and keywords that exist on social media, but it echoes a longstanding political PR strategy for reshaping public debate. When combined with a breaking news situation, this type of data void can be especially damaging. > > > > >[!example|#e56eee]+ pg. [23](zotero://open-pdf/library/items/L2K3PJR7?page=23&annotation=SPR8QMIW) > >consider what unfolded after the horrific 2012 shooting at Sandy Hook Elementary School in Connecticut… > > > >In the hours and days following this tragedy, members of well-established conspiracy forums produced hundreds of posts attempting to debunk the shooting… > > > >In their conversations, they settled on the conspiratorial ideal that the parents and kids appearing in news footage of the shooting’s aftermath were actually paid actors. They labeled them “crisis actors” and began mobilizing around this term… > > > >Prior to the creation of this conspiracy, the term “crisis actor” referred to a job in which trained actors or volunteers would play mock victims during disaster simulations to help train first responders. While there was some web content referring to those simulations, few people searched for this term and few people created web content referencing this job. Thus, this term was ripe for manipulation. > > > > >[!highlight|#ffd400]+ pg. [24](zotero://open-pdf/library/items/L2K3PJR7?page=24&annotation=6HDMAA2B) > >Although there was a spike in searches connected to this term associated with every shooting, this concept did not break through into national news coverage until the Marjory Stoneman Douglas High School massacre that took place in Parkland, Florida, in 2018. At the height of news coverage, CNN anchor Anderson Cooper asked survivor David Hogg on national TV if he was a “crisis actor.” This was intended to allow Hogg to deny the conspiracy theory, but it ended up breathing life into it. More news outlets – and news comedy shows – started using the term. > > > > >[!highlight|#5fb236]+ pg. [24](zotero://open-pdf/library/items/L2K3PJR7?page=24&annotation=7ZZIHPTE) > >**And the more that the term was used in the media, the more people searched for it.** > > > > >[!highlight|#2ea8e5]+ pg. [25](zotero://open-pdf/library/items/L2K3PJR7?page=25&annotation=3K2BH742) > >In the online communities that organize these types of manipulation efforts, there is frequent discussion of the political strategy work of Frank Luntz. > > > > >[!highlight|#ffd400]+ pg. [25](zotero://open-pdf/library/items/L2K3PJR7?page=25&annotation=7U96RGWE) > >In the 1990s, Republican pollster and “public opinion guru” Luntz became famous for using focus groups and polls to develop pithy phrases that would reframe political concepts. Many of his terms are part of the contemporary lexicon, including “climate change,” “death tax,” and “partial-birth abortion.”What made him successful as a political operative was his ability to create catchy phrases and convince elected officials to repeat these terms until the news media helped spread them across the country. > > > > >[!highlight|#ffd400]+ pg. [26](zotero://open-pdf/library/items/L2K3PJR7?page=26&annotation=WTLUQRAS) > >While Luntz expected his phrases to do political work directly, media manipulators who create strategic phrases are not necessarily looking to get a new term to stick; they are more interested in getting people to search for these phrases and encounter the web of information that they have produced by exploiting data voids before those data voids are cleaned up. > > > > >[!highlight|#ffd400]+ pg. [26](zotero://open-pdf/library/items/L2K3PJR7?page=26&annotation=LVZ9LDGF) > >**those who engage in this tactic celebrate in online forums when journalists pick their terms up. We also know that participants in various hate-oriented and conspiratorial forums reference these terms in describing how they found the forums in the first place.** > > > > >[!highlight|#ffd400]+ pg. [27](zotero://open-pdf/library/items/L2K3PJR7?page=27&annotation=PAIQLY4V) > >one search query of a strategic term led him to a data void shaped by white nationalists > > > > ### DATA VOID TYPE #3: OUTDATED TERMS >[!highlight|#ffd400]+ pg. [27](zotero://open-pdf/library/items/L2K3PJR7?page=27&annotation=LDG8737P) > >data voids can also emerge when terms stop being regularly used > > > > >[!highlight|#ffd400]+ pg. [27](zotero://open-pdf/library/items/L2K3PJR7?page=27&annotation=J3QHYIRK) > >they have long lives in the gaps of search engines. As search engines respond to new trends and new words, old terms can be left behind, associated only with outdated content, leaving more and more room for manipulators. > > > > >[!highlight|#ffd400]+ pg. [28](zotero://open-pdf/library/items/L2K3PJR7?page=28&annotation=LG46M67C) > >Search engines are programmed in an effort to balance between content that is recent and content that is “authoritative.” > > > > >[!highlight|#ffd400]+ pg. [28](zotero://open-pdf/library/items/L2K3PJR7?page=28&annotation=K8GAUW79) > >when users search for terms that are more consistent and associated with much older but highly authoritative content or more established content creators, search engines are more likely to return authoritative – “evergreen” – content alongside news content. > > > > >[!highlight|#ffd400]+ pg. [28](zotero://open-pdf/library/items/L2K3PJR7?page=28&annotation=9Q4LGFAE) > >Many search-relevant topics have seasonality, ranging from elections to major sporting events. Generally speaking, both content creation and search queries rise alongside each other for rhythmic search topics. > > > > >[!highlight|#ffd400]+ pg. [29](zotero://open-pdf/library/items/L2K3PJR7?page=29&annotation=X636DA8V) > >Other terms are more faddish in nature.That is, they are used heavily for months or years as they catch on and then quickly fade from use as they fall out of fashion. > > > > >[!highlight|#ffd400]+ pg. [30](zotero://open-pdf/library/items/L2K3PJR7?page=30&annotation=ABHFWWDE) > >A data void produced by outdated terms doesn’t have to be strategically manipulated to be dangerous. > > > > >[!highlight|#ffd400]+ pg. [31](zotero://open-pdf/library/items/L2K3PJR7?page=31&annotation=TXRACDJ8) > >Data voids in this category emerge when a disconnect occurs between content creation and search queries. Content creators typically leave behind words faster than people stop searching for them. > > > > >[!highlight|#ffd400]+ pg. [31](zotero://open-pdf/library/items/L2K3PJR7?page=31&annotation=58VIJBPD) > >if people are searching for a term where no new content is being produced, media manipulators can easily fill in the gap and use this disconnect to their advantage. > > > > ### DATA VOID TYPE #4: FRAGMENTED CONCEPTS >[!highlight|#ffd400]+ pg. [31](zotero://open-pdf/library/items/L2K3PJR7?page=31&annotation=QZ8QUKHM) > >manipulators can also work to intentionally separate manipulated content from more popular content through the creation of distinct terms that are too fraught to connect. > > > > >[!highlight|#ffd400]+ pg. [32](zotero://open-pdf/library/items/L2K3PJR7?page=32&annotation=PVNVJDID) > >When large communities of people approach a news item with different political frames, the result can be a “naturally” fragmented model of search results—two largely unconnected sets of results. > > > > >[!example|#e56eee]+ pg. [32](zotero://open-pdf/library/items/L2K3PJR7?page=32&annotation=EWHMA9CG) > >For instance, in late summer 2018, a scandal concerning the Vatican and sexual misconduct was front-page news. As the news was breaking, a search for “Vatican sexual abuse” returned entirely different results than a search for “Vatican pedophiles,” especially onYouTube. > > > > >[!example|#ffd400]+ pg. [32](zotero://open-pdf/library/items/L2K3PJR7?page=32&annotation=AHHPPWZV) > >These results were fragmented because the people producing the new content, as well as the people doing the searches, had different ideological commitments. But over time, the results converged on web search engines. > > > > >[!highlight|#ffd400]+ pg. [33](zotero://open-pdf/library/items/L2K3PJR7?page=33&annotation=LMIQL7BT) > >Both manual and algorithmic efforts by search engines to bridge concepts and information can be controversial. > > > > >[!highlight|#5fb236]+ pg. [33](zotero://open-pdf/library/items/L2K3PJR7?page=33&annotation=SWMSXL8P) > >For politically charged content, any decision by search engine designers becomes political itself. > > > > >[!highlight|#ffd400]+ pg. [33](zotero://open-pdf/library/items/L2K3PJR7?page=33&annotation=M7T3TGY9) > >Collapsing ideologically split terms might be the automatic response of a search model to the news content produced, but this technique might also be critiqued for mixing together such diametric terms. At the same time, if search engines intervene manually to shape the model for a specific set of terms, they risk accusations of bias by other political communities. > > > > >[!highlight|#ffd400]+ pg. [33](zotero://open-pdf/library/items/L2K3PJR7?page=33&annotation=HJ8F65SR) > >media manipulators can systematically work to make certain that new content does not connect concepts, to make sure that fragmented concepts remain fragmented.What is emerging as a result are distinct clusters of information that are neither manually nor algorithmically bridged. Depending on the terms users put in a search query, they can end up in an entirely different sphere o > > > > ### DATA VOID TYPE #5: PROBLEMATIC QUERIES >[!highlight|#ffd400]+ pg. [34](zotero://open-pdf/library/items/L2K3PJR7?page=34&annotation=VYKBYWKH) > >Some of what people seek to know is deeply disturbing or taboo. Even more concerning are those who produce information for people who might conduct such disturbing or taboo queries. > > > > >[!highlight|#ffd400]+ pg. [34](zotero://open-pdf/library/items/L2K3PJR7?page=34&annotation=2WYWHTBR) > >these queries are often tricky. Initially, those who produced factual content about the Holocaust did not produce content that included the kind of language contained in this query. When content creators were informed of the concerted effort by Holocaust denial conspiracy theorists, they contributed to addressing the problem by creating new content. But conspiracy theorists consistently produce newer content and seek to optimize their content in new ways, creating a challenge for those who are producing factual information. > > > > >[!highlight|#ffd400]+ pg. [35](zotero://open-pdf/library/items/L2K3PJR7?page=35&annotation=NFW2GBIC) > >YouTube consistently has too little data in this arena, which gives conspiracy theorists an easier path to outrank factual creators. > > > > >[!highlight|#ffd400]+ pg. [35](zotero://open-pdf/library/items/L2K3PJR7?page=35&annotation=X6NGDWY9) > >many conspiracy theorists, hate groups, and media manipulators attempt to push searchers to use specific, problematic search queries they know will lead to these voids. > > > > >[!highlight|#ffd400]+ pg. [35](zotero://open-pdf/library/items/L2K3PJR7?page=35&annotation=PXI454TP) > >Unfortunately, many Holocaust deniers produce videos and talk on the radio, encouraging listeners to search for specific new phrases in order to circumvent the content optimized to address the original problematic query. >[!highlight|#ffd400]+ pg. [35](zotero://open-pdf/library/items/L2K3PJR7?page=35&annotation=LXIB6SBV) > >Fact-checking websites have traditionally been more concerned with establishing a center of facts than debunking every specific conspiracy theory > > > > >[!highlight|#ffd400]+ pg. [35](zotero://open-pdf/library/items/L2K3PJR7?page=35&annotation=RP6UHR6D) > >Most other content producers do not think to devote time to producing content designed to debunk potential conspiracies. Given the absence of inoculation content, conspiracy theorists used SEO to fill up data voids so that conspiratorial content would appear when someone articulated a disturbing query. > > > > >[!highlight|#ffd400]+ pg. [35](zotero://open-pdf/library/items/L2K3PJR7?page=35&annotation=IR67GMCB) > >Historians simply posted evidence, not even thinking to optimize their content for someone who might have considered a conspiratorial frame or to bridge their content to be viable for problematic queries. > > > > >[!highlight|#ffd400]+ pg. [36](zotero://open-pdf/library/items/L2K3PJR7?page=36&annotation=2BGXS497) > >they must make certain that their content is optimized for problematic queries as well as more common ones. > > > > >[!highlight|#ffd400]+ pg. [36](zotero://open-pdf/library/items/L2K3PJR7?page=36&annotation=ZAPQMGRS) > >While some media manipulators produce original content to optimize for problematic queries, others focus on increasing the visibility of content produced for other reasons. > > > > >[!highlight|#e56eee]+ pg. [36](zotero://open-pdf/library/items/L2K3PJR7?page=36&annotation=UYL7PJXL) > >This data void was intentionally exploited by those seeking to associate sexual misconduct with women. To achieve this goal, they focused on getting individual journalists to cover individual cases so as to build up a corpus of content, rather than a single story. To those perpetuating this conspiracy, search offers evidence. In forums, where this topic comes up, nonbelievers are told to “just do a Google image search” as a way of showing that the conspiracy is true. As with strategic terms, people are also encouraged to conduct problematic searches like “female pedophiles” in order to encounter content that was staged. > > > > >[!highlight|#ffd400]+ pg. [36](zotero://open-pdf/library/items/L2K3PJR7?page=36&annotation=686FXXWX) > >a quick search does not provide the context necessary to undo a conspiratorial frame that is staged elsewhere > > > > >[!highlight|#ffd400]+ pg. [37](zotero://open-pdf/library/items/L2K3PJR7?page=37&annotation=75PFI2G6) > >the content creators interested in countering misinformation must not only think about the paths people take to reach positive content, but also the paths people take to reach problematic content. By associating positive content with problematic queries, content creators can help limit the isolating quality of certain data voids. In order to combat data voids that exist because of problematic queries, both content creators and search engines need to work together to identify disconcerting paths and produce content that is valuable for those who are not explicitly seeking out problematic content. > > > > ## Data Voids in Search-Adjacent Recommender Systems >[!highlight|#ffd400]+ pg. [38](zotero://open-pdf/library/items/L2K3PJR7?page=38&annotation=AHV6BQCY) > >The data voids that emerge through search engines are sometimes reinforced by adjacent features that function more like recommender systems. Recommender systems are the various features that encourage users to consume additional content by making recommendations or pushing new content to consume. > > > > >[!highlight|#ffd400]+ pg. [38](zotero://open-pdf/library/items/L2K3PJR7?page=38&annotation=ZKCPZC3P) > >Because search-adjacent recommender systems often rely on many of the same sets of data that undergird search engines and features more generally, they can often compound existing vulnerabilities and open up new vulnerabilities for media manipulators to exploit. > > > > ### SEARCH BAR AUTO-SUGGESTIONS >[!highlight|#ffd400]+ pg. [39](zotero://open-pdf/library/items/L2K3PJR7?page=39&annotation=QBZJRTQ3) > >Auto-suggestions are generated based on previous queries from users and help reveal fragmentations in language patterns. > > > > >[!highlight|#ffd400]+ pg. [39](zotero://open-pdf/library/items/L2K3PJR7?page=39&annotation=L7FHNNE8) > >auto-suggest also introduces new phrases users might never have considered. > > > > >[!highlight|#ffd400]+ pg. [39](zotero://open-pdf/library/items/L2K3PJR7?page=39&annotation=SIIEQEBZ) > >auto-suggestions can also send people down deeply disturbing paths. > > > > >[!highlight|#e56eee]+ pg. [39](zotero://open-pdf/library/items/L2K3PJR7?page=39&annotation=E2NJN8YG) > >Before search engines started overhauling their auto-suggestions around phrases like “women are” the results ranged from offensive to terrifying. The fundamental problem was a data void. > > > > >[!highlight|#ffd400]+ pg. [39](zotero://open-pdf/library/items/L2K3PJR7?page=39&annotation=46RUYN9J) > >media manipulators often look to exploit these data voids to encourage certain auto-suggest results. They may attempt to get a distributed group of “Before search engines started overhauling their auto-suggestions around phrases like ‘women are’ the results ranged from offensive to terrifying. The fundamental problem was a data void. > > > > >[!highlight|#ffd400]+ pg. [40](zotero://open-pdf/library/items/L2K3PJR7?page=40&annotation=C398SN5P) > >- 39 DATA VOIDS people to search for a specific term or even create automated systems (“bots”) to do this for them. Unlike SEO, they aren’t focused on encouraging people to search new phrases. Instead, they are working to extend commonly searched phrases with additional words that could lead users to intentionally produced problematic content. > > > > >[!highlight|#ffd400]+ pg. [40](zotero://open-pdf/library/items/L2K3PJR7?page=40&annotation=58AF5RFM) > >auto-suggest data voids are typically addressed by understanding and limiting the possibility of auto-suggest on more problematic topics > > > > ### YOUTUBE’S “UP-NEXT” AND AUTO-PLAY FEATURES >[!highlight|#ffd400]+ pg. [41](zotero://open-pdf/library/items/L2K3PJR7?page=41&annotation=3WBI7PI7) > >Recommendations are not a result of a user-centered search. Rather, they functionally serve as a machine-driven search for content that is similar to the content currently being consumed or that may otherwise appeal to the particular user. > > > > >[!highlight|#ffd400]+ pg. [41](zotero://open-pdf/library/items/L2K3PJR7?page=41&annotation=5SFHW4LZ) > >recommendation engine optimization parallels search engine optimization. > > > > >[!highlight|#ffd400]+ pg. [42](zotero://open-pdf/library/items/L2K3PJR7?page=42&annotation=4GMBF3V5) > >The specifics may differ, but the fundamental goal is the same: make certain the algorithm ranks your content well. > > > > > > >it's not the algorithm, it's algorithmic manipulation -- which is a sociotechnical (?) problem. >[!highlight|#ffd400]+ pg. [43](zotero://open-pdf/library/items/L2K3PJR7?page=43&annotation=F53ZJWYX) > >The SEO/REO strength of a group with an ideological agenda, combined with the weakness of groups who are providing factual information, makes YouTube’s system easy to exploit. > > > > ## Managing Data Voids >[!highlight|#ffd400]+ pg. [44](zotero://open-pdf/library/items/L2K3PJR7?page=44&annotation=YM8KWVH6) > >Any ranking, rating, or recommender system on the internet is being and will be exploited if media manipulators can benefit from doing so. > > > > >[!highlight|#ffd400]+ pg. [44](zotero://open-pdf/library/items/L2K3PJR7?page=44&annotation=W6UDS7N5) > >While the sheer quantity of content available on the internet has made Bing and Google fairly resilient to SEO for many queries and in many languages, there are still significant vulnerabilities in this information ecosystem. > > > > >[!highlight|#ffd400]+ pg. [44](zotero://open-pdf/library/items/L2K3PJR7?page=44&annotation=FMASUHNS) > >One notable type > > > > > > >data voids are only one way the system is manipulated >[!highlight|#ffd400]+ pg. [45](zotero://open-pdf/library/items/L2K3PJR7?page=45&annotation=HFFUUJSX) > >Factually inaccurate information, hyper-partisan content, scams, conspiracy theories, hate speech, and other forms of problematic content are harmful to individuals and societies, but media manipulators have a higher incentive to create such content than those who seek to combat it. > > > > >[!highlight|#ffd400]+ pg. [46](zotero://open-pdf/library/items/L2K3PJR7?page=46&annotation=SUGPZ4HZ) > >This issue must be treated with the same level of seriousness as any security issue. These companies have a responsibility to identify these vulnerabilities, iteratively respond to attacks, and support content creators who can produce the high-quality content that is needed to fill these data voids. This will be especially daunting in contexts where facts and evidence are contested. > > > > >[!highlight|#ffd400]+ pg. [46](zotero://open-pdf/library/items/L2K3PJR7?page=46&annotation=8MMFLCQX) > >While search engines wish to be neutral platforms, they are going to increasingly face governance challenges that reveal just how political the project of providing information can be. > > > >