References: Lies at scale

Source material

This lesson is an original adaptation. Its durable spine, the way of thinking about false information (misinformation is false information spread without intent to deceive; disinformation is spread on purpose; a hallucination is the tool’s own confident, fluent, false claim; the liar’s dividend is the truth waved away as fake; the four-actor toolkit of companies, platforms, governments, and you), comes from the mis and disinformation session of a Harvard Kennedy School course. The current provenance and detection state, the recognition examples, the hedges, and the prose are our own. One discipline governs the whole lesson: it is defensive only, teaching recognition and resilience, never how to produce anything false. Clawdemy is independent of Harvard, which has not reviewed or endorsed this track.

• The Science and Implications of Generative AI (HKS DPI-681M), Harvard Kennedy School, Spring 2024. Faculty: Sharad Goel, Dan Levy, and Teddy Svoronos. This lesson adapts Class 11, mis and disinformation from the Spring 2024 course site, whose content is licensed under Creative Commons Attribution 4.0. Class 11 defines misinformation and disinformation, works through hallucination and the economics of fakes, names the liar’s dividend, and sorts the response into the four actors who can push back. Its session videos are the source for our framework.
• Official course lecture playlist on YouTube, Harvard Kennedy School. The full lectures, free to watch. We mean it when we encourage you to take the original course alongside this track: good teachers deserve more students, and this is the last lesson, so it is a good moment to go deeper at the source.
• Provenance note: the framework and classroom examples in this lesson were drawn from the transcripts of the official Class 11 lecture videos in the playlist above, obtained from the official Harvard sources only. No third-party re-uploads or mirrors were used.

A note on how these figures are cited

Freshness is the whole discipline here, and this field moves faster than any other in the track. Every dated example and figure below is re-verified live before publish and on each freshness sweep. The framework is durable; the numbers and the state of the technology are not, and none is presented as a settled fact about the present. Every recognition example is a publicly documented, already-debunked case cited to a reputable newsroom, fact-checker, or researcher; we never generate synthetic false artifacts ourselves. Every source-side control below is paired with its documented leak, so no reader is left believing a tool will catch every fake. Two guardrails in particular: the watermarking scale figure is more than ten billion files, and no derivative may report the larger hundred-billion figure that did not verify; and the signed-record format’s move toward becoming a formal international standard is in process, not confirmed ratified, so it is not stated here as an established standard.

The recognition examples, documented and debunked

• Google Bard and the James Webb Space Telescope (public demo, February 2023). In an early public demonstration, a major chatbot was asked about the telescope and answered confidently that it had taken the first image of a planet outside our solar system; it had not, as that first image came years earlier from a different telescope. Carried as a documented historical case of a hallucination, not a claim about any current tool. Sources: CNN and Gizmodo.
• The Arup deepfake video-call fraud (incident January 2024, firm named May 2024). A finance worker at a global engineering firm joined a video call with AI-generated fakes of the company’s chief financial officer and colleagues and transferred about 25 million US dollars before the fraud was caught. Reported figures and mechanics are as covered by the newsroom; carried to recognize the personalized-deepfake pattern, not to detail method. Sources: CNN, firm named and CNN, initial report.
• The Hurricane Helene image of a girl and puppy in a boat (October 2024). A widely shared image of a crying child holding a puppy in floodwater was AI-generated, debunked by fact-checkers who found the details did not hold still across circulating versions, including a child’s shirt changing color and a boat that differed from copy to copy. Presented as the artifact plus its independent debunking plus the harm it did to real relief efforts only; no party, official, or political framing is attached. Sources: Full Fact, PolitiFact, and digital-forensics researcher Hany Farid at UC Berkeley.
• The liar’s dividend in the courts (2023 to 2025). Lawyers now argue that genuine recordings may be AI fabrications; in one instance a company argued that recorded statements by its own chief executive might be fakes and the judge was not persuaded, and judges have said openly that they worry about deciding a real person’s fate on evidence they cannot be sure is real. This landscape is fast-moving and is framed institutionally only, with no partisan cases named. Sources: NBC News, Thomson Reuters Institute, and NPR.

The provenance and detection state, as of mid-2026, hedged

• The two-layer provenance convergence. As of mid-2026 the largest AI labs have mostly converged on pairing an invisible watermark with a signed content record that travels with a file and says what made it, with a major lab adopting both approaches together in May 2026. Carried as a real but leaky advance: a mark only helps if the tool that made the content chose to add one, and the signed record can fall off when a file is re-saved or screenshotted. Source: OpenAI on content provenance.
• The watermarking scale figure. The officially sourced figure is more than ten billion pieces of content watermarked with the leading image and content watermarking system, reported by Google DeepMind in 2025. This is the number the lesson uses; the larger hundred-billion figure surfaced only in a speculative, forward-looking source, did not verify, and must not appear in any derivative. Vendor-reported scale figures are carried with that caveat. Source: Google on SynthID and the SynthID-Image paper (arXiv 2510.09263).
• The signed-record standard. The content-credential format is being fast-tracked toward a formal international standard through the relevant standards body, but ratification was still in process as of mid-2026 and is not stated here as an established standard. Re-verify before promoting it to settled fact.
• Detection is unreliable. The company behind the best-known tool for spotting AI-written text launched it in early 2023 and retired it in July 2023 for low accuracy, including flagging real human writing, which fell hardest on people writing in a second language. The lesson names no percentages, to keep it plain, and states only that the tool was shut down for being wrong too often. Sources: OpenAI’s classifier announcement, Search Engine Land on its retirement, and Search Engine Journal.
• Disclosure and labeling law. New rules requiring that AI content be disclosed and labeled are landing in 2026. This lesson recalls that in one line and does not re-teach it; the risk-map lesson covers the specifics. See the cross-link below.

On this site

• What generative AI actually is, lesson 1 of this track. The prediction-machine idea at the root of hallucination, and the promise this lesson pays off: once you understand the trick, the fear starts turning into fluency. This lesson recalls that idea and does not re-teach it.
• The risk map, lesson 6 of this track. Where misinformation first appears as one square on the map, where watermarking and detection are defined, where the four policy levers and the 2026 disclosure rules are taught in full, and where the voice-clone verify-through-a-trusted-channel move is introduced. This lesson recalls all of it and does not re-teach it.
• Will AI take my job?, lesson 8 of this track. The lesson just before this one, on what AI does to your work; this final lesson turns from your work to what you are able to trust.

Going deeper

• Official HKS course listing for DPI-681M, Harvard Kennedy School. The in-person course behind the open online materials this track adapts. This is the last lesson of the track, so if the course caught your interest, this is the place to keep going at the source.

Where the track ends

There is no next lesson; this is the end of the road. Look back across all nine and what you carry out is not a shield that blocks every fake, because no one has that. It is your own judgment, sharpened: you can tell a mistake from a lie, expect the confident wrong answer and check it, feel the pull of a fake built to move you and slow down instead of sharing, and look sideways before you believe. This track never promised to make you fearless. It promised to make you harder to fool, and that is where fluency was leading all along.