Skip to content
Clawdemy

References: The risk map

Source material

Section titled “Source material”

This lesson is an original adaptation. Its four-category map of risk, the levers for managing them (the four regulatory approaches plus watermarking, detection, and education), the three-question test for a policy, the calm treatment of existential risk and the reason-about-each-world frame, and the healthcare texture come from the eighth class of a Harvard Kennedy School course, its risk session. The voice-cloning example, the European rules note, the reader actions, and the prose are our own. 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 8 from the Spring 2024 course site, whose content is licensed under Creative Commons Attribution 4.0. Class 8 opens the course’s unit on the implications of generative AI: it sorts the risks into four categories, walks the levers for managing them, and works a healthcare case study. Its four session videos are the source for our risk map and lever 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.
  • Provenance note: the risk categories, lever framework, and classroom details in this lesson were drawn from the transcripts of the official Class 8 lecture videos in the playlist above, obtained from the official Harvard sources only. No third-party re-uploads or mirrors were used.

The examples behind the lesson

Section titled “The examples behind the lesson”

The course supplies the durable structure; every current example is re-sourced to a 2025 or 2026 public reference, cited here with its limits stated. A non-author reviewer re-verifies each of these live before publish and on each freshness sweep.

  • FBI Internet Crime Complaint Center (IC3), reporting on 2025, published by the FBI’s Internet Crime Complaint Center. This is the source for the lesson’s statement that a voice-cloning distress scam was counted among AI-boosted scams that in United States fraud reporting for 2025 together cost victims roughly 900 million dollars, with older adults hit hardest, in the first year IC3 tracked AI-enabled fraud as its own category. Limitations: IC3 figures are self-reported complaint data and undercount unreported fraud, and a first-year category has definitional flux, so the lesson hedges to “roughly 900 million dollars” and attributes the full figure to the AI-tied category as a whole, not to voice scams alone. No derivative hardens this to a precise unhedged number.
  • FTC consumer alerts on voice cloning and family-emergency scams (2023, 2024), Federal Trade Commission, and the FBI’s December 2024 IC3 public service announcement on generative-AI fraud, cited alongside the FTC alerts for the resilience habit. The FTC alerts document the mechanic (a short clip lifted from social media) and the call-back-on-a-known-number advice; the family code word habit is carried primarily by the FBI PSA and consumer-protection guidance. Limitations: these are consumer-guidance materials, not technical studies, so the lesson uses them for recognition and resilience only, never as a measure of scale.
  • European Commission, regulatory framework for AI (digital-strategy.ec.europa.eu), plus the Council of the EU on the AI simplification (“omnibus”) package (consilium.europa.eu). These are the sources for the EU AI Act note: a risk-based framework, transparency obligations from August 2026 (people told when they are talking to a machine, certain AI-generated content labeled), and the postponement during 2026 of the heavier high-risk obligations to late 2027 and 2028. Limitations: this timeline is actively shifting (deadlines were moved during 2026), so every date in the lesson is stated as of mid-2026 and re-verified each 90-day sweep; the lesson describes the law institutionally, by what it does, and names no penalties. The consilium page returned an access block to direct fetching on 2026-07-11; the postponement is independently confirmed on the official Commission page, which fetched clean.
  • The four regulatory approaches (disclosure, registration, licensing, auditing) match the framework in Neel Guha, Christie M. Lawrence, Daniel E. Ho, and colleagues, “AI Regulation Has Its Own Alignment Problem: The Technical and Institutional Feasibility of Disclosure, Registration, Licensing, and Auditing,” 92 Geo. Wash. L. Rev. 1473 (2024), work by Stanford researchers led by Neel Guha and Daniel E. Ho; an accessible version is available through the Stanford Institute for Human-Centered AI. The course attributes the framework to these authors; the lesson keeps the framework generic in the body and names the authorship here. Limitations: the four approaches are an analytic frame, not a ranking, and the lesson uses them as such.
  • The paperclip thought experiment is credited to the philosopher Nick Bostrom, who introduced it as an illustration of how a capable system optimizing a harmless-sounding goal without judgment can go wrong at the edges. The lesson paraphrases it and uses it to illustrate a concept, not to predict an outcome.
  • Freshness marker: the 2024 Nobel Prize in Chemistry was shared for computational work on proteins, with one half to David Baker for computational protein design and the other half jointly to Demis Hassabis and John Jumper of Google DeepMind for protein structure prediction (the AlphaFold work), per the Nobel Prize press release. The lesson keeps biomedical research light and cites this only as a recognition milestone, not as a claim about current clinical deployment. Limitations: cited as a marker that the field is being taken seriously at the highest level, nothing more; the body does not lean on it.

Going deeper

Section titled “Going deeper”

On this site

Section titled “On this site”
  • What generative AI actually is, lesson 1 of this track. The home of the promise this lesson pays back: companies tune a raw prediction machine toward being helpful, honest, and safe, and deciding whose values that tuning should reflect is one of the hardest open problems in the field.
  • Should AI do this task?, lesson 4 of this track. The home of the privacy question and the cost-of-false-information filter this lesson recalls in the what-it-gets-wrong square.
  • AI on a real project, lesson 5 of this track. The lesson that closes the track’s first movement, the one this lesson turns from.

Adjacent topics

Section titled “Adjacent topics”
  • This lesson opens the track’s second movement. The next lesson takes one corner of the slow-change square, who owns what a model makes and learns from, and maps the copyright question honestly, as a debate to understand rather than a verdict to hand down.