1. The 2024 New Hampshire primary robocall incident involved:
Correct. Voice-cloned audio of Biden was robocalled to ~25,000 New Hampshire voters before the primary, telling them not to vote — the first confirmed AI voice-cloning voter suppression incident in a U.S. election.
Incorrect. The New Hampshire case used voice-cloning — synthetic audio in Biden's voice — deployed via robocalls, not video or text.
2. The C2PA (Coalition for Content Provenance and Authenticity) addresses synthetic media through:
Correct. C2PA works at the creation stage — embedding a cryptographic chain of custody so anyone can verify a file's origin and edit history, rather than trying to detect deception after the fact.
C2PA embeds provenance at creation via cryptographic metadata — it's a chain-of-custody approach, not a detection approach.
3. How many AI-generated "pink slime" news sites did NewsGuard's 2023 AI Misinformation Monitor identify?
Correct. NewsGuard documented 49 pink slime sites in 2023, all identifiable within 2 minutes via WHOIS lookup and Wayback Machine checks — sites with professional designs but no authentic publishing history.
The lesson cites 49 pink slime sites identified by NewsGuard in 2023 — each identifiable within 2 minutes using WHOIS and Wayback Machine lateral reading steps.
4. Statistical watermarking in LLM outputs works by:
Correct. Statistical watermarks work by biasing the probability distribution over token choices during generation in ways that a detector with the key can identify, without noticeably affecting output quality. The limitation is that paraphrasing, translation, or fine-tuning can destroy the signal.
Statistical watermarking biases the token selection process during generation — certain token choices are systematically favored in ways that produce a detectable pattern for authorized detectors. The watermark is statistical, not textual, and doesn't affect perceived text quality. Paraphrasing can remove it.
5. Which reverse image search engine is noted for superior performance on images originating from Eastern Europe or Central Asia?
Correct. Yandex's Russian-language index and facial recognition capabilities make it particularly effective for images originating in Cyrillic-language regions.
Yandex Images — its Cyrillic-language index and superior facial recognition for Eastern European subjects makes it a preferred choice for images from that region.
6. A PolitiFact "Half True" rating most accurately means:
Correct. "Half True" indicates selective accuracy — the stated facts may check out but the overall impression conveyed is misleading because of what's left out.
"Half True" means the claim has accurate elements but omits critical context — the facts check out partially but the overall impression is misleading.
7. What does Peter Gollwitzer's research on implementation intentions show about behavior change?
Correct. Gollwitzer's implementation intention research shows that "If X, then Y" formulations significantly outperform simple goal-setting ("I will verify claims") in producing actual behavioral change.
Gollwitzer's research shows that if-then planning structures ("If I am about to share a surprising claim, then I will apply SIFT first") significantly outperform simple goals in producing behavioral execution.
8. Taiwan's 2024 election media literacy outcomes, documented by Reuters Institute, demonstrated what principle?
Correct. Taiwan's experience — with a school-based literacy curriculum reaching 70% of the school-age population — produced measurably better verification behavior, leading policy researchers to describe it as a form of democratic infrastructure.
Taiwan's case, analyzed by the Reuters Institute and Harvard Kennedy School's Shorenstein Center, demonstrates that population-level media literacy functions as democratic infrastructure — comparable in importance to cybersecurity measures on election systems.
9. In the context of astroturfing, "social proof manipulation" refers to:
Correct. Social proof — the sense that "everyone believes this" — is a powerful driver of belief change. Astroturfing manufactures artificial social proof to pressure real people toward desired positions.
Incorrect. Social proof is specifically about perceived popularity — astroturfing exploits the tendency to update beliefs toward what appears widely held, regardless of the actual distribution of opinion.
10. The IFCN Code of Principles requires signatory fact-checking organizations to commit to:
Correct. The IFCN Code covers five areas: nonpartisanship, transparent sourcing, transparent funding disclosure, transparent methodology, and an open corrections policy.
The IFCN Code requires nonpartisanship, transparent sourcing, transparent funding, transparent methodology, and an open corrections policy — ethical standards, not output quotas.
11. The New Hampshire AI robocall voter suppression incident of January 2024 was attributed to:
Correct. Kramer publicly admitted commissioning the calls, claiming he did so to "start a conversation" about AI in elections. The calls reached over 20,000 Democratic primary voters with an AI-generated voice mimicking President Biden, telling them not to vote. The FCC clarified that the TCPA covers AI-generated voice calls.
Political consultant Steve Kramer admitted commissioning the calls. This case illustrates that AI election interference is not exclusively a foreign threat — domestic political operatives with modest resources can deploy voice cloning for voter suppression at scale.
12. The 2019 Gabon presidential video case demonstrated which key aspect of deepfake technology's impact?
Correct. The Gabon case is the canonical example of the liar's dividend — a coup attempt was partially justified by deepfake suspicion of a genuine video. The technology's existence, not a specific deepfake, caused harm.
Incorrect. The key lesson of Gabon is the liar's dividend: the existence of deepfake technology enabled doubt to be cast on an authentic video, with a coup attempt as a real-world consequence.
13. Epistemic paralysis, as a goal of information saturation campaigns, means:
Correct. Epistemic paralysis is the desired outcome of saturation — not belief change, but the collapse of the audience's willingness to try to determine truth at all.
Incorrect. Epistemic paralysis describes the information overload state where distinguishing true from false feels impossible — the goal is cynical disengagement, not belief in any specific falsehood.
14. What does the "S" in SIFT stand for, and why is it the first step?
Correct. "Stop" is first because misinformation exploits emotional arousal to trigger sharing before analysis. A conscious pause interrupts the automatic response cycle.
The "S" stands for "Stop" — a conscious pause to recognize emotional arousal before it can trigger automatic sharing behavior. It is first because emotional reactions are the primary mechanism exploited by misinformation.
15. The RAND Corporation's 2016 "firehose of falsehood" report identified what as the primary distinguishing feature of the Russian state media propaganda model?
Correct. The firehose model is distinguished by volume, multichannel reach, and indifference to internal consistency — aimed at paralysis, not persuasion.
Incorrect. The key feature is the combination of high volume, multichannel delivery, and deliberate inconsistency — designed to overwhelm rather than convince.
16. What did the Stanford HGSE 2022 study find about the time efficiency of lateral reading versus reading within a website?
Correct. The Stanford study found lateral reading was not just more accurate but took only 25% of the time of within-site reading — faster AND more reliable.
Lateral reading produced assessments in 25% of the time of within-site reading AND with greater accuracy — a key finding of the 2022 Stanford HGSE study on fact-checker versus historian verification behavior.
17. The term "burstiness" in AI text detection refers to:
Correct. Burstiness describes how much perplexity varies within a text. Human writing alternates between predictable and surprising sentences; LLM output tends toward consistently low perplexity throughout. AI text detectors use this pattern as a heuristic — which paraphrasing can defeat.
Burstiness is the variation in perplexity across a text passage. It's a heuristic used by AI text detectors: human writing tends to have variable perplexity (some sentences predictable, others surprising), while LLM output tends to be uniformly low-perplexity. This difference is exploitable by paraphrasing.
18. The Internet Archive's Wayback Machine is most useful for fact-checkers when:
Correct. The Wayback Machine stores historical snapshots — when a misinformation publisher deletes content after being called out, the archived version provides a permanent record for documentation.
The Wayback Machine's key value is retrieving deleted content — when publishers remove misinformation, the archive often retains a snapshot.
19. The C2PA standard being developed by Adobe, the New York Times, and the BBC addresses what problem?
Correct. C2PA (Coalition for Content Provenance and Authenticity) embeds verifiable cryptographic metadata into images and media at creation time, allowing downstream verification of origin and editing history.
C2PA is a cryptographic provenance standard — it embeds verifiable metadata into media files at creation so recipients can verify the file's origin and what modifications it has undergone.
20. Microsoft's VALL-E voice cloning paper demonstrated speaker modeling from:
Correct. VALL-E demonstrated convincing speaker modeling from three-second audio clips — a threshold that makes any short publicly available recording of a person sufficient source material for voice cloning.
VALL-E required only three seconds. This is the threshold that makes voice cloning a practical threat for virtually any public-facing individual — the source audio requirement is now shorter than most voicemail greetings.