Methodology, limits, and corrections
A site that teaches people to check claims should be easy to check. This page says how the content is produced, what rules it follows, what the tools can and cannot do, what has not been verified yet, and what we have got wrong so far.
Scope and purpose
Persuasion Lab is defensive education: recognizing persuasion and manipulation, resisting it, and persuading honestly. It currently holds 722 encyclopedia entries, 90 Defense Playbook entries, 35 reference frameworks, 9 operation anatomies, and 48 practice drills. Content about abuse, scams, and covert operations is descriptive, never procedural: it explains patterns and warning signs and does not supply scripts, tooling, or evasion methods.
How content is written
Most entries were drafted with AI assistance (Anthropic's Claude) working to a written content standard, then checked by automated validators and by a human owner. We say this plainly because it matters: language models can state false things fluently, including citations that do not exist. The controls below exist because of that risk, and they reduce it without eliminating it.
- Writers were instructed to cite only works they were highly confident exist and to omit a detail (a volume number, a URL) when they were unsure of it, on the principle that a missing detail is better than a wrong one.
- Every batch passed a validator for structure, harm class, required safety text, and cross-references before it shipped; a test suite re-checks the whole corpus on every change.
- References that writers flagged as less than fully certain were re-verified against publisher pages and library records in a separate pass. Sources for case studies and operation anatomies were opened or confirmed in a search index at writing time; some sites block automated access, so a minority were confirmed by index only.
- Where science is contested or failed replication (ego depletion, power posing, oxytocin and trust, the backfire effect, mortality salience), the entry says so.
Sourcing rules
722 of 722 entries carry structured references, each with a note saying which claim it supports. Primary sources are preferred: the paper that established a finding, the court opinion, the government or platform report. Wikipedia is not cited. Political claims are sourced to scholarship, journalism of record, or official documents. Numbers that could not be traced to a source were removed or softened; see the corrections log.
Harm classes
Every entry is classified, and the class controls what the AI tools will do with it. Neutral (42): craft with no ethical weight in itself. Dual-use (265): legitimate when truthful and disclosed; each entry states where the line is. Manipulative (373): deceptive by design, documented so it can be recognized. Abuse (42): inherently harmful control; these entries open with a plain statement that the technique should never be used, carry safety-first defenses and hotline information, and the AI tools refuse to help deploy them against anyone.
Political neutrality
The aim is a reference that people who disagree with each other can both use. Political entries are illustrated across the spectrum and across eras, with dated events and named actors, and they describe mechanisms without scoring living politicians. The operation anatomies deliberately include Soviet, Russian, Chinese, Belarus-linked, United States military-linked, and corporate actors. Practice drills use fictional parties and alternate which side is the manipulator. Where the research literature itself is lopsided, the entry says so and does not manufacture balance.
Safety content
Material on coercive control and abuse follows one rule above the others: confrontation and leaving can escalate danger, so defenses point first to safety planning with a domestic-violence advocate (in the US, 1-800-799-7233 or thehotline.org) and only then to communication techniques. This content has not yet been reviewed by an advocacy organization; that review is the first one we are seeking.
What the tools can and cannot do
- Manipulation Detector: a rule engine that reads structure and vocabulary in one text. It cannot see context, omission, truth, images, or intent; it reports its own confidence and never certifies a message as honest.
- Coordination Analyzer: behavioral signals across a set of posts, each with the innocent explanations for it. It never attributes activity to anyone.
- Practice: 8 of the 48 drills are honest messages, and scoring counts false alarms, because training that makes people see manipulation everywhere is a known failure of this kind of teaching. Learning outcomes are measured with a held-out before-and-after test; aggregate results will be published when 30 learners have shared both tests, whatever they show.
- DISARM mapping: 77 entries are mapped to the open DISARM Framework (v1.6.1, CC BY-SA 4.0). Identifiers were generated from the framework's published data; the choice of closest match is ours.
- AI tools: they run under written rules (truthful, attributed, non-coercive influence only; no targeting of people in grief, debt, isolation, or cognitive decline; no work on real elections within 24 months; refusal to deploy abuse tactics).
What has not been verified yet
- The AI tools' refusal and neutrality behavior has a written evaluation suite (refusals, legitimate requests that must not be refused, benign messages that must not be flagged, mirrored left and right prompts, fabrication traps) that has not yet been run against the live models. Until it has, treat the rules above as design intent.
- No outside expert has reviewed the content yet.
- No learning-outcome data has been collected yet.
- Generative-AI material dates quickly; those pages carry review dates and are re-checked every few months.
External review
We are seeking named reviewers in five areas: domestic-violence advocacy (abuse-class entries and relationship playbook entries), inoculation and misinformation research (practice design and measurement), influence-operations analysis (anatomies, coordination tool, attribution guidance), trial advocacy (courtroom entries), and statistics communication. Reviewers' criticisms will be published here with our responses. If you work in one of these fields and see an error, write to corrections@persuasionlab.app.
Corrections log
Substantive corrections, newest first. Typos and formatting are not logged.
Was: Credited the Channel 4 News FactCheck article on Twitter users wrongly labelled as Russian bots to the wrong author.
Now: Credits Martin Williams, with the publication date of 24 April 2018.
Was: Said the New York Times editors' note of May 2004 acknowledged a loop in which officials cited press coverage as corroboration.
Now: Says what the note says: that the coverage was not rigorous enough and that exile claims were often eagerly confirmed by U.S. officials.
Was: Described the 2020 Iranian emails sent in the name of the Proud Boys as a network reported by Meta.
Now: Attributes the finding to U.S. officials, who announced it, and no longer credits Meta.
Was: The tool ran without the shared ethics rules the other tools use, and its instructions asked for a profile of a decision-maker including their "vulnerabilities".
Now: It carries the shared ethics rules, refuses harassment, doxing, hidden sponsorship, vulnerable-audience targeting and live-election work, and profiles officials from the public record only.
Was: Said Norway had withdrawn the Jehovah's Witnesses' state grants and registration and that the matter was being contested in court.
Now: Adds that Norway's Supreme Court ruled both decisions invalid in April 2026.
Was: Several cases carried claims with no source of record: a resale-price multiple for Supreme, "seeded" Beatlemania crowds, a drop in support attributed to the "death tax" label, paid physician endorsements for Camel, and that the 1987 "This Is Your Brain on Drugs" ad backfired.
Now: Unsourceable claims were removed or narrowed to what the cited source supports; the anti-drug case now reports the findings of the federal campaign evaluation (Hornik et al. 2008); every case cites a source of record and none cites Wikipedia.
Was: Heritage and Greatbatch (1986) carried a wrong title in one entry; Bezmenov's 1984 booklet was described as self-published; a practitioner guide carried an unverifiable year.
Now: Title corrected to "Generating Applause: A Study of Rhetoric and Response at Party Political Conferences"; publisher given as W.I.N. Almanac Panorama and "self-published" removed; the guide's year updated to the edition that could be confirmed.
Was: Stated that a slower pace signals confidence.
Now: States that the effect of pace depends on context: the entry's own cited study (Miller et al. 1976) found faster speech more persuasive.
Was: Presented a defector's four-stage schedule as if it were documented KGB doctrine.
Now: Presents it as Bezmenov's own synthesis from a 1984 interview, separates it from what archives support, and notes its use as a partisan meme on more than one side.
Was: Examples were generic, undated, or drawn mostly from one side of politics.
Now: Examples are dated, name actors, and are drawn across the political spectrum and across eras; two contested scholarly models (Herman and Chomsky; Lakoff) are now labelled as contested.
Was: One example asserted a contested political claim about universities as settled fact.
Now: Replaced with documented cases established by inquiries and courts (the UK Post Office Horizon scandal, USA Gymnastics, the Boston Archdiocese, Flint, the Boeing 737 MAX).
Was: Classified as manipulative.
Now: Reclassified as an abuse tactic, with the abuse notice and safety-first defenses.
Was: The earlier word-list detector scored a message containing a bribe and a conditional threat as 0.
Now: The detector reads structure (conditional threats, inducements, ultimatums) as well as vocabulary, reports its confidence, and lists what it cannot see.