Social Media’s Impact on Homeland Security

Question"What restrictions, if any, does Meta place on the type of research employees can perform or the type of data they can access?"

ResponseSays researchers are encouraged to publish externally and aren't restricted on research topics, subject to guideline review.

QuestionPortman asks Boland to explain why the Platform Accountability and Transparency Act is needed and how it would work.

ResponseBoland argues Meta discloses more than Google or TikTok but still far too little, that population-wide averages obscure that roughly 2 million Americans may be receiving a "hate-filled" or radicalizing feed even at a "1%" rate, and that he personally lost trust in what Meta was telling the public and itself about the problem.

QuestionLankford describes Facebook's algorithm rewarding angry, hostile comment exchanges between users and asks whether it's technically possible to let page owners hide public comment threads to prevent commenters from attacking each other.

ResponseBoland confirms hiding public comments is technically feasible, and explains that Facebook's algorithm is purely engagement-driven and "knows no temperature" — it will keep amplifying angry, attacking exchanges because that's what generates engagement, regardless of harm.

QuestionRomney, citing concern for his own kids and grandkids, asks whether other countries have implemented design measures (like TikTok's reported five-second blank-screen gaps in China) to reduce social media's addictive draw, and whether the U.S. is falling behind on protecting kids this way.

ResponseBoland confirms "friction" design techniques exist to reduce compulsive use, but says he's not aware of any U.S. company mandating such friction on its own, and that Europe has moved toward transparency requirements rather than prescriptive design mandates.

QuestionCiting horrifically livestreamed terrorist attacks that inspire copycat violence, Hassan asks whether social media companies can quickly identify and stop terrorist content from spreading in real time.

ResponseBoland says Meta has invested significant AI resources in detecting livestreamed attacks and has improved since Christchurch, but calls it an inherently hard problem.

QuestionPeters asks Boland to confirm he raised internal objections at Facebook about the recommendation algorithm promoting extreme, hateful, and racist content, and what leadership's reaction was.

ResponseBoland confirms he raised internal objections specifically about racist content amplification, was told he was "wrong" without counter-evidence or told it wasn't a current priority, and cites an internal June 2020 polarization-research document (part of the Frances Haugen disclosures) admitting Facebook had "very little understanding of racial, ethnic, or religious polarization."

QuestionCiting the years-long unchecked spread of QAnon and Zuckerberg's own 2018 statement that unchecked engagement optimization favors sensationalist content, Peters asks Cox whether engagement-focused recommendation algorithms inevitably promote extreme content that hasn't yet been formally labeled violative.

ResponseCox says QAnon is now labeled a "violence-inducing network" and banned, and reiterates that Meta works to remove terrorism/extremism content and publishes results, without directly addressing whether engagement-optimized algorithms inevitably promote such content before it's labeled.

QuestionPortman confronts Cox with public reporting that Meta directed content moderators to "err on the side of" assuming a person depicted in sexual exploitation content is an adult when age is unclear, asks him to confirm or deny it, and presses whether he'll commit to reviewing and changing the policy.

ResponseCox does not directly confirm or deny the "err on the side of adult" policy, attributes the approach to guidance from NCMEC to prioritize "known CSAM" content, says he personally wasn't part of that specific conversation, and commits only to having his team "follow up."

QuestionCiting Meta's own reported figures — 1,000+ identified militarized social movements, 270 white supremacist organizations, 2.3 million pieces of organized-hate content removed in Q2 2022 alone — Carper asks what more Meta can do to address the broader threat of these organizing groups.

ResponseCox cites 350 dedicated experts working with law enforcement on terrorist/violence-inducing-conspiracy-network identification, and gives Meta's current hate-speech prevalence figure of 0.02% (down from 50% eighteen months earlier).

QuestionJohnson asks, repeatedly and pressing for a quick answer, whether Meta de-platformed or censored anyone involved in organizing the summer 2020 riots.

ResponseCox is interrupted mid-sentence twice by Johnson and never completes a substantive answer, only invoking "domestic terrorism and extremism" as the applicable policy category before being cut off again.

QuestionSinema asks what Meta is doing to stop cartels from using Facebook and Instagram to recruit teens along the Southwestern Border, and whether Meta will commit to sharing cartel-recruitment information with DHS as quickly as possible.

ResponseCox says Meta prohibits cartels and human trafficking and works with law enforcement to identify and remove cartel content, and conditionally commits to sharing cartel-recruitment information with DHS "provided privacy and legal concerns were addressed."

QuestionPadilla presses past Meta's "95% caught before reported" hate-speech statistic to ask specifically how many times the uncaught 5% was actually recommended to users in their news feeds.

ResponseCox redirects to the same overall prevalence statistic (0.02%) rather than answering the specific recommendation-count question, and commits only to a written follow-up.

QuestionCiting Meta's own 2020 transparency report figure of 5.9 million illegal drug-sale postings found on Facebook and Instagram — 453 times the volume that got the dark-web marketplace Silk Road shut down — Scott asks how many postings are going undetected and why Meta remains operational by comparison.

ResponseCox points to Meta's quarterly transparency reporting rather than estimating how much drug-sale content goes undetected, and does not address the Silk Road comparison directly.

QuestionPeters presents a physical poster showing that Facebook auto-generated a home page for the Aryan Brotherhood white-supremacist prison gang that stayed live for 12 years despite years of external warnings, and asks whether this auto-generation feature makes it easier for extremist groups to recruit, and whether taking 12 years to remove it is acceptable.

ResponseCox cites a 97% terrorist-content pre-detection rate but does not directly justify the 12-year delay in this specific case, says Meta "would not have put this page up ourselves" (attributing it to the auto-generation feature rather than a deliberate act), and says he hasn't personally seen this specific example.

QuestionIn his closing question, Peters asks whether Meta consistently measures new features' impact on societal harms including hate speech and terrorism before launch, then presses specifically on Cox's repeated 0.02%-hate-speech-prevalence statistic — asking for the actual total raw view count behind that percentage, not just the rate.

ResponseCox describes a pre-launch safety review process for new products, confirms the 0.02% hate-speech figure, but says he doesn't have the raw total-view-count numbers behind that percentage on hand and commits only to a written follow-up.

QuestionPeters asks how many Meta employees specifically research extreme content and whether that research covers algorithmic promotion of such content.

ResponseMeta says it has 350+ specialists on dangerous-individuals/organizations content and that internal research (including into algorithmic dynamics) informed the 2019 ban on white nationalism/supremacism, without providing a specific headcount answering the exact question asked.

QuestionPeters asks for Meta's hate-speech and overall violative-content takedown rates over the past year, month, and day.

ResponseMeta says it doesn't track an aggregate "takedown rate" as defined, but reports Q2 2022 removals (13.5M pieces from Facebook, 3.8M from Instagram) and a rolling four-quarter hate-speech prevalence range of 0.01–0.03%.

QuestionPeters asks for total QAnon content views/shares before and after Meta's August 2020 policy change, and how many 2017–2020 views/shares were driven by algorithmic recommendation.

ResponseMeta says it did not track QAnon view/share counts before or after the policy change (only removal counts of Pages/Groups/accounts), and does not track how much QAnon exposure came from algorithmic recommendation.

QuestionCiting a Frances Haugen document estimating Meta's hate-speech takedown rate at 3–5%, Peters asks whether that internal estimate was accurate.

ResponseMeta disputes the accuracy of the internal 3–5% estimate cited from the Haugen disclosures, calling it a rough, unverified figure by a single data scientist that conflated prevalence (views) with raw content counts and ignored demotion/prevention measures, and points to its own reported 0.02% prevalence figure instead.

QuestionPeters presses in writing on why Facebook auto-generates pages for banned hate groups, why the Aryan Brotherhood page took twelve years to remove, and why an auto-generated Proud Boys page created on January 6, 2021 remained up.

ResponseMeta acknowledges an "auto generation issue" affecting "a small number of pages," says the flagged pages were removed, but does not directly explain the twelve-year delay or specifically address the Proud Boys page raised in the question.

QuestionCiting reports that Meta directs moderators to use the "Tanner scale" (a puberty-staging tool, not an age-verification tool) and to "err on the side of an adult" for ambiguous-age content, Portman asks Meta to commit, yes or no, to halting both practices.

ResponseMeta does not answer yes or no. It defends continued use of the Tanner scale as "only one indicator" among several, cites the Stored Communications Act as a legal reason it cannot report content without a "good faith belief" the subject is a minor, and says it has developed unspecified "new internal aging guidelines" without committing to end either practice.

QuestionCiting the livestreamed Christchurch mosque attack and bad actors' evolving rhetoric to evade detection, Carper asks how Meta's safety features adapt to keep pace.

ResponseMeta describes its Dangerous Individuals and Organizations policy (expanded in August 2020 to cover militarized movements and QAnon), cites removal statistics (2.3M pieces of organized-hate content in Q2 2022, 97% pre-reported), 350+ dedicated specialists, $5B in annual safety/security spending, and its co-founding of the cross-industry GIFCT database, framing its Christchurch Call to Action response as ongoing rather than describing specific new detection capabilities for evolving rhetoric.

QuestionCarper asks what information Meta proactively shares with law enforcement to help prevent violent attacks before they happen, given how quickly social media can accelerate radicalization.

ResponseMeta describes sharing information with law enforcement reactively (on request or upon identifying a "credible threat"), rather than proactively flagging radicalization patterns before an attack, and does not directly address what "more needs to be done."

QuestionLankford asks how long it takes Meta to detect and remove both previously-known and newly-created CSAM.

ResponseMeta does not give a specific detection/takedown timeframe, instead describing its hash-matching technology for known CSAM, AI detection for new/unknown material, a 24/7 specially trained review team, and use of Google's Content Safety API to prioritize likely child-exploitation content for review.