Big Tech and the Online Child Sexual Exploitation Crisis (QFR Responses)

QuestionPadilla asks what it would take to detect novel (unhashed) CSAM images, and where Congress could help combat CSAM.

ResponseMeta describes its AI/ML detection of novel CSAM and Q4 2023 removal/reporting statistics, then pivots to advocating for federal legislation requiring app-store parental approval for teen downloads and allowing CSAM retention for ML training.

QuestionPadilla asks how Meta identifies/removes AI-generated CSAM, whether it flags AI-generated CSAM to NCMEC distinctly, how prevalent such content is, and how it may affect NCMEC's Cybertip processing capacity.

ResponseMeta says its no-tolerance CSAM policies apply regardless of whether content is AI-generated, describes account-disabling/network-linking enforcement and generative-AI safeguards (query filtering, testing, response-scanning, feedback tools), and reiterates NCMEC reporting — but does not directly answer whether AI-generated CSAM is flagged as such to NCMEC or quantify its prevalence (4b/4c).

QuestionPadilla asks (referencing the Taylor Swift deepfake incident) whether Meta will commit to publicly reporting on the prevalence of AI deepfake nonconsensual sexual imagery, including of minors, and its mitigation steps.

ResponseMeta cites its policies against nonconsensual sexual/deepfake imagery and AI-content detection systems, and points to industry efforts (White House Voluntary Commitments, Partnership on AI, watermarking/labeling standards) — but does not commit to publicly reporting prevalence data as directly asked.

QuestionPadilla asks whether legal or technical barriers prevent Meta from thoroughly redteaming its AI models to keep them from generating CSAM.

ResponseMeta says current federal law provides no immunity for CSAM redteaming, which makes such testing more difficult, but that it works within legal bounds to test and retrain its models as extensively as permissible.

QuestionPadilla asks what percentage of Meta's budget goes to child safety, what risk assessment informed that figure, and how many layers of leadership separate trust & safety leaders from Zuckerberg.

ResponseMeta cites overall safety/security headcount (~40,000) and spend (~$20B since 2016, ~$5B last year) rather than a child-safety-specific budget percentage, describes its risk-monitoring process qualitatively, and states its Global Head of Safety works closely with the executive team without specifying a reporting-layer count.

QuestionPadilla asks what Meta is doing to make open-source trust & safety/CSAM-detection tools available to the broader tech industry, which lacks Meta's resources.

ResponseMeta lists its open-sourced PDQ/TMK-PDQF hashing technologies, its Take It Down and StopNCII programs, and its role founding/funding Lantern and Project Protect through the Technology Coalition as its contributions to industry-wide child-safety infrastructure.

QuestionPadilla asks whether Meta partners with the federal government on child-safety public-awareness campaigns and what Meta does to proactively educate minors on online safety.

ResponseMeta cites its Know2Protect campaign with DHS, "Report It, Don't Share It" and Help Protect Children awareness campaigns, Messenger safety notices, and child-safety hackathons as its youth-education and government-partnership efforts.

QuestionPadilla asks how Meta is responding to financial sextortion of minors — detection/disruption methods, user education, and whether Meta will study and report on demographic disparities in sextortion victimization.

ResponseMeta describes a multi-pronged sextortion response — policy, automated detection, investigations/NCMEC referral, default privacy/messaging restrictions for teens, suspicious-account detection, Take It Down/StopNCII, and Stop Sextortion education resources with Thorn — and calls for OS-level age-verification legislation, but does not directly answer whether it has studied demographic disparities in sextortion victimization (10d).

QuestionPadilla asks whether and how Meta engages youth directly in shaping its trust & safety policies, and how it stays current on emerging issues facing young people online.

ResponseMeta cites TTC Labs co-design sessions, partnerships with youth/parent organizations, funded external research (Digital Wellness Lab, Aspen Institute grants), and its Safety Advisory Council/Youth Advisors as its mechanisms for youth engagement and staying current on emerging risks.

QuestionPadilla notes that not all caregivers act in a child's best interest — including abusive parents and parents who could endanger an LGBTQ+ child — and asks how Meta's parental tools account for this.

ResponseMeta says its parental-tool design incorporated input from civil society and LGBTQ+ advocacy organizations, points to baseline teen protections that don't depend on parental supervision, and cites its LGBTQ+ Help Center resources, "Be Kind Online" guide, Trevor Project link, and Project Height civil-rights review process — without directly addressing how the tools specifically prevent abusive-parent misuse.

QuestionPadilla asks about the measured impact of Meta's 2021 CSAM-search pop-up and viral-content safety-alert interventions, and what lessons Meta has drawn for the broader industry.

ResponseRather than quantifying the specific 2021 pop-up/safety-alert interventions' impact, Meta cites overall safety spend/headcount, its "Report It, Don't Share It" campaign, and broader industry initiatives (Lantern, Take It Down, Project Protect) as evidence of prevention progress and lessons for industry.

QuestionTillis asks what Meta does to keep user-generated content depicting underage use of alcohol/tobacco/nicotine away from minors, and how it responds to takedown requests.

ResponseMeta cites age-gating tools, a default "most restrictive" content-control setting for teens, and Community Standards/Guidelines prohibiting organic alcohol/tobacco content, while noting brick-and-mortar retailers may still advertise such products off-platform.

QuestionTillis asks what Meta does about cartels using its platforms to coordinate, specifically to recruit children into crime or drug distribution.

ResponseMeta describes removing criminal organizations under its policies, working with law enforcement to identify cartel activity, and expediting child-safety-related law enforcement requests, including proactive reports when a child appears in imminent danger.

QuestionTillis asks for yearly removal/ban statistics for fentanyl/drug content and whether drug advertising to children has increased.

ResponseMeta cites 2023 removal totals (9.3M pieces on Facebook, 10.1M on Instagram) but does not directly answer whether drug advertising to children specifically has increased.

QuestionTillis asks how Meta works with prevention/recovery experts and with parents who lost children to drugs purchased online.

ResponseMeta lists nonprofit/awareness partnerships (Song For Charlie, Ad Council's "Drop the F-Bomb," Partnership to End Addiction) aimed at parents and families, but doesn't specifically describe direct engagement with parents who lost children to online drug purchases.

QuestionTillis asks for a year-by-year count of meetings Meta held with parents about online safety over the last three years.

ResponseMeta describes parent-engagement channels and cites 2023 event counts (six community events, six "Screen Smart" city events) but does not give the requested three-year, year-by-year total.

QuestionTillis asks what share of NCMEC's 32 million 2022 CSAM reports came from Meta's platforms.

ResponseMeta reports over 26 million of the 32 million industry-wide 2022 CSAM reports to NCMEC came from Facebook, Instagram, and WhatsApp combined, versus under 5 million from the rest of the industry.

QuestionTillis asks how Meta will track and address CSAM as end-to-end encryption rolls out, given the concern that reporting numbers will drop while the underlying problem persists.

ResponseMeta describes its three-pillar approach to encrypted-messaging safety (prevent, control, respond quickly), including default-private teen accounts, restrictions on adult-to-teen messaging, on-device nudity detection in DMs, and machine-learning detection of suspicious adult behavior using unencrypted signals — while declining to detail its enforcement methods to avoid tipping off bad actors.

QuestionTillis asks whether Meta has seen a rise in suspected CSAM and what's driving it.

ResponseMeta points to its Community Standards Enforcement Report and states it estimates CSAM prevalence on Instagram at below 0.01%, attributing volume changes partly to improved detection rather than directly answering what's "driving" the increase.

QuestionTillis asks about new CSAM-detection tools/strategies and the closeness of Meta's NCMEC relationship.

ResponseMeta details its detection stack (PhotoDNA hash-matching, Google's Content Safety API, proprietary novel-content detection, Reels/Stories scanning for CSE indicators) and describes funding NCMEC's case-management tool and co-founding the Take It Down service.

QuestionTillis asks what victim support Meta provides and whether it partners with local victim groups/professionals.

ResponseMeta describes NCMEC funding for survivor services, a dedicated NCMEC reporting channel for CSAM-survivor threats, Take It Down/StopNCII, and a partnership with Thorn to build sextortion guidance now housed in a "Sextortion hub" in Meta's Safety Center.

QuestionTillis asks for the top technical obstacles Meta faces in fighting CSAM.

ResponseBeyond adversarial-detection and coded-language challenges, Meta pivots to advocating for its own proposed federal legislative framework requiring app-store-level parental approval, age verification, supervision tools, and standardized age-appropriate content/ad-targeting rules.

QuestionTillis asks whether Meta uses "surveillance advertisements" to target minors and how that could be mitigated.

ResponseMeta states it now limits teen ad targeting to age and geography only (having removed gender/interest/activity-based targeting), and offers Ad Topic Controls and a teen privacy center.

QuestionTillis asks about algorithmic practices, beyond ad targeting, that may be harmful to children.

ResponseMeta describes its content-recommendation algorithms as protective (friend-suggestion, "Sensitive Content Control"/"Reduce" defaults, usage-nudge tools) rather than acknowledging any detrimental algorithmic practice.