White Paper: LegalGuard Social
LegalGuard Social is a free, anonymous web application that applies large language models to pre-publication legal risk screening of social media posts. This paper describes its architecture, screening pipeline, risk taxonomy, and data-handling model.
1. Abstract
A single ill-considered post can expose a publisher to defamation claims, copyright infringement, privacy breaches, and regulatory action. Traditional legal review is slow and costly relative to the pace of social publishing. LegalGuard Social provides an on-demand first-pass assessment: the user submits a draft, selects a platform and jurisdiction, and receives a structured risk verdict with flagged issues and a suggested safer rewrite — in seconds, without an account.
2. System architecture
The application is a React single-page app hosted on the Base44 platform, backed by serverless functions. There is no user authentication layer and no per-user database of drafts. Three concerns are separated cleanly:
Presentation layer
The client renders the composer and the assessment exhibits. All risk visualization is driven directly by the structured response object; no risk logic runs client-side.
Screening service
A backend function receives the draft, validates it, constructs the analysis prompt, and invokes the language model with a response schema that forces a fixed, parseable JSON shape. This keeps output deterministic in structure even though the prose is generated.
Supporting services
An anonymous visit counter records aggregate session totals (a running count plus a rolling 24-hour window of timestamps) and an optional donation flow is handled through Stripe Checkout, with direct Bitcoin and Ethereum wallet addresses offered as an alternative. Neither stores user-identifying data.
3. The screening pipeline
The composer collects the draft text, target platform, and jurisdiction. Inputs are validated in the browser before any request is issued; the draft is held in component state only.
The draft is sent to a backend screening function over HTTPS. No user account, session, or identifier is attached — the request is anonymous by design.
The function assembles a legal-analysis prompt combining the verbatim draft, the platform's content norms, and the selected jurisdiction's law, with a fixed analytical rubric spanning defamation, privacy, intellectual property, harassment, confidentiality, and regulatory categories.
The model returns a strict JSON object: an overall risk level, a publish verdict, a plain-language summary, a list of flagged issues (each with category, severity, explanation, and recommendation), and a safer rewrite of the draft.
The frontend maps the response onto the verdict card and the flagged-issue exhibits. The rewrite is copy-to-clipboard enabled so the user can adopt it directly.
4. Risk taxonomy
Every screening returns one of four overall risk levels, each issue is graded on the same scale, and the verdict states plainly whether the draft is cleared to post:
- Low — no significant legal exposure; notes are stylistic.
- Moderate — plausible defamation, privacy, trademark, or disclosure concerns; revise first.
- High — likely infringement, unverified factual accusations, or confidentiality breaches; do not post as drafted.
- Critical — severe exposure such as explicit defamation, contempt, incitement, or leaked material; do not publish.
Issues are categorized across defamation, privacy, intellectual property, harassment, confidentiality, and regulatory exposure, and each carries its own severity, explanation, and concrete recommendation.
5. Data handling and privacy
The service is anonymous: no accounts, no logins, and no persistent store of drafts. Submitted text exists only for the lifetime of a single screening request. The visit counter retains nothing beyond a total and a 24-hour window of anonymous timestamps. Because no draft is retained, nothing about a user's posting activity can be disclosed, subpoenaed, or breached.
6. Limitations
LegalGuard Social is a screening aid, not a lawyer. Language models can miss context, nuance, and recent legal developments, and their assessments are probabilistic. The tool's verdicts are general legal information, do not create an attorney-client relationship, and must not substitute for advice from a qualified lawyer on matters with serious exposure.