Author: Nita Farahany Source: nitafarahany.substack.com Published: March 5, 2026 Series: Persuasion and Manipulation (Part 3 of 3)
"Aristotle's Algorithm" is a March 5, 2026 Substack essay by Nita Farahany documenting Class 7.3 of her Advanced Topics course, the third and final installment of the Week 7 Persuasion and Manipulation series. The class asks whether any existing regulatory framework reaches the full problem of AI-driven persuasion, and proposes the UK Competition and Markets Authority's Online Choice Architecture taxonomy as a behavioral-analysis alternative to US-law categories such as gambling, deception, and unfairness. The title alludes to Aristotle's Rhetoric and its ethos / pathos / logos taxonomy of persuasion, reframed for algorithmic persuasion through the CMA's framework.
Summary of argument
The class follows Class 7.1 (The Persuasion Exchange), which examined what persuasion feels like from the inside, and Class 7.2 (The Statute That Couldn't Stretch), which described how Facebook v. Duguid read the TCPA out of relevance for AI persuasion. Building on both, Class 7.3 poses two questions: whether a single regulatory framework captures the whole problem, and whether any framework currently in hand actually reaches the persuasion techniques becoming available.
Farahany's answer, set up by the structure of the course, is that no current US framework reaches what she calls the stacked version of persuasion. This frames the remaining course question of what such a framework would look like and whether it is constitutionally feasible.
Empirical anchor: Salvi et al.
The class opens with the Salvi et al. experiment conducted at EPFL and the Bruno Kessler Foundation and published in Nature Human Behaviour in May 2025. The preregistered study enrolled 900 participants, each of whom debated either a human or GPT-4. Some AI opponents received six basic demographic variables about their counterpart: age, education level, ethnicity, employment status, and political affiliation. With that demographic data, GPT-4 was more persuasive than a human debater 64% of the time, and the odds of shifting an opponent's position were 81% higher than a human's.
Farahany emphasizes that the inputs were not browsing history, purchase patterns, or psychological profiles, but only six variables, and that the persuasion advantage scales quickly from there.
The CMA Online Choice Architecture taxonomy
The UK CMA's framework, as Farahany presents it, reframes the regulatory question away from "is this gambling?", "is this unfair?", or "does this violate a specific statute?" and toward how the design of the choice environment changes the decisions people make. Farahany characterizes this as a pivot from legal categorization to behavioral analysis. The taxonomy organizes 21 online choice-architecture practices into three categories.
| Category | Practices | Documented effect sizes |
|---|---|---|
| Choice structure | Defaults, ranking, partitioned pricing, virtual currencies, sludge, forced outcomes | Defaults increase selection rates by ~27%; ranking captures ~95% of clicks on top results |
| Choice information | Drip pricing, reference pricing, framing, complex language, information overload | Drip pricing weakens competition by shifting attention to headline prices |
| Choice pressure | Scarcity claims, prompts, messengers, commitment, personalization | (effect sizes vary by tactic) |
Farahany uses loot boxes as the worked case study for how the categories combine. She writes:
"Loot boxes deploy all three categories simultaneously. Virtual currencies (choice structure) obscure real spending. Undisclosed probabilities (choice information) prevent informed comparison. And the reward system itself (choice pressure) uses what psychologists call a variable ratio reinforcement schedule, the same mechanism that makes slot machines addictive."
Her central pedagogical point is that the stacking of techniques is what makes loot boxes, and by extension AI persuasion systems, particularly potent: no single technique would be as effective, and the combination produces an effect greater than the sum of its parts. This is the analytical lens she asks students to bring to AI persuasion.
The framework also introduces three competitive harms that the CMA links to these practices: distorting consumer behavior, weakening competition, and maintaining or leveraging market power. Farahany notes that the third harm is underexplored in US law, and that the shift from "was this transaction fair?" to "was this market functioning?" poses a different question with a different answer, one that current US AI regulation does not address.
Pedagogical structure
The class culminates in a mapping exercise in which students take their work from the Persuasion Exchange (Class 7.1) and map it onto the CMA taxonomy, then map loot boxes the same way, and then ask whether any current US legal framework reaches the stacked version of persuasion. The exercise leads to Farahany's conclusion that no current US framework reaches stacked persuasion, which sets up the course's continuing question about the design and constitutional feasibility of such a framework.
The essay is the most detailed application in this set of pages of the CMA's Online Choice Architecture taxonomy to AI persuasion. Farahany presents the taxonomy as a more comprehensive analytical framework than US-law-derived categories such as gambling, deception, and unfairness, and as portable to comparative-law analysis.
Provenance and confidence
The page draws on a single source, treated as Farahany's pedagogical position. The CMA taxonomy is independently verifiable, but its application to AI persuasion is Farahany's analytical move. The Salvi et al. finding that GPT-4 with demographic data was 81% more likely to shift an opponent's position is a primary-record empirical claim. Confidence is medium, reflecting the single-source basis and the distinction between the verifiable taxonomy and Farahany's application of it.
Relationships
- part-of: Nita Farahany Advanced Topics course (Class 7.3 of ~30)
- previous: Inside My Advanced Topics Class 7.2: The Statute That Couldn't Stretch (Farahany, March 2026)
- depends-on: Inside My Advanced Topics Class 7: The Persuasion Exchange (Farahany, March 2026), Inside My Advanced Topics Class 7.2: The Statute That Couldn't Stretch (Farahany, March 2026) — direct prerequisites
- related: Inside My AI Law & Policy Class 15: Why Governing AI Synthetic Media is So Hard (Farahany, October 2025), Inside My AI Law & Policy Class 15: When AI Learns to Manipulate (Farahany, October 2025) (earlier-course synthetic-media manipulation coverage extended here to choice-architecture-based manipulation), Regulating Under Uncertainty (the CMA framework as an example of behavioral-analysis-based regulatory design), AI Mental Health and Psychological Harm (choice-pressure category overlaps with mental-health-harm patterns), Raine v. OpenAI, Inc., Garcia v. Character Technologies, Inc. (the choice-pressure "messengers" and "commitment" patterns are alleged in both complaints), Agentic AI, \"For All Issues So Triable\" — Dean W. Ball (Hyperdimensional, August 2025) (compatible tort-as-discovery framing for the persuasion problem)
- instance-of: AI Safety Cases and Frameworks