Unshaken: Justice · Ethics Charter
Practice that’s honest, private and tested in public
Unshaken helps people prepare for court when they can’t afford a lawyer or need to be ready alongside one. People come to us scared and under deadline. That puts a duty on us to be accurate, protect what they share, say clearly what we are and aren’t, and let others check our work.
- Version
- 1.0 draft
- Aligned with
- LSNTAP AI Policy Builder · State Bar of California LAJTC Readiness Model
- Review
- Every 6 months, in public
This charter describes commitments for the planned AI simulator. Most of these controls are not live yet; today the site is informational with an email-only waitlist.
Our manifesto
Ten commitments we hold ourselves to
Every product decision, every Edition and every new feature is checked against these. When one conflicts with growth, the commitment wins.
01
We practice, we don’t practice law.
Unshaken is education and rehearsal. It doesn’t give legal advice, tell you what to file or choose a strategy for you. We point you to lawyers, legal aid and court self-help centers for those decisions.
02
Your story is yours.
We collect the least we need, never sell it, never train shared models on it without your opt-in, and delete it when you ask.
03
Law comes from verified sources, with a date and a citation.
Rules shown in a simulation come from a checked rule base, not from what a model remembers. If we can’t verify it, we don’t present it as law.
04
We say “we don’t know.”
When a rule varies by state, county or judge, or our information may be out of date, we tell you, every time.
05
We never script testimony.
We teach how to answer truthfully: listen, pause, answer only the question. We never help anyone invent or shade facts.
06
We test for bias and publish what we find.
Accuracy and tone are measured across languages, reading levels, case types and communities. Gaps are reported, not hidden.
07
Humans stay accountable.
Licensed attorneys and legal aid advisors review legal content before release. A named person owns every Edition.
08
We care for the person, within limits.
Our companion listens and helps calm stress. It isn’t therapy, routes crisis signs to 988 and domestic-violence hotlines, and encourages real human support.
09
Advertising never becomes advice.
Lawyer listings are clearly labeled. We don’t recommend a lawyer for pay, and your practice data is never sold as a lead.
10
We let others check us.
Independent evaluators audit us, and every audit report is published in plain language, failures included.
Data privacy
What we keep, what we never do
People tell Unshaken things they haven’t told anyone: about their children, their health, their money, their fear. We design as if every record could be read in court, because some could.
What we collect, and why
- ♦Your practice answers and scores, so you can see progress.
- ♦The case details you choose to enter, so simulations fit your situation.
- ♦Your state and county, so we load the right rules.
- ♦Basic account information, so you can sign in.
What we never do
- ♦Sell or rent your data, or your case details as “leads.”
- ♦Share anything with an opposing party, a court or a firm without your explicit, specific consent.
- ♦Train shared AI models on your story unless you opt in.
- ♦Ask for Social Security numbers, account numbers or passwords.
- ♦Use your feelings journal for anything but you.
How we protect it
- ♦Encryption in transit and at rest.
- ♦AI vendors under contracts that bar them from retaining or training on your data.
- ♦Access limited to named staff, with every access logged.
- ♦Delete-everything control in settings, honored within 30 days, including backups on their normal cycle.
What you should know
- ♦Talking to Unshaken is not talking to your lawyer. It may not be protected by attorney-client privilege.
- ♦Don’t enter anything you wouldn’t want a judge to see. Use initials for children.
- ♦On a shared or monitored device, use a private window and sign out. If you’re in danger, your safety comes first.
Accuracy and hallucinations
AI that makes things up is dangerous in law
General-purpose AI invents cases and misstates rules. Even tools built for lawyers do. Independent researchers have measured how often:
58–88%
Hallucination rates of general chatbots on verifiable questions about federal court cases (Dahl et al., Journal of Legal Analysis, 2024).
17%
Of queries where Lexis+ AI gave a hallucinated or misleading answer (Magesh et al., Journal of Empirical Legal Studies, 2025).
33%
The same measure for Westlaw AI-Assisted Research, in the same study.
These figures are other researchers’ measurements of other tools, shown to explain the risk. They are not measurements of Unshaken. Ours will be published in our audit reports below.
Neurosymbolic guardrails
The language model writes dialogue; it doesn’t decide the law. Deadlines, procedures and rules come from a structured rule base built from official sources, each tagged with jurisdiction, citation and last-verified date. A rule the model states that isn’t in the rule base is blocked before you see it.
Corrective AI
A second checking layer reviews each answer for errors the guardrails can’t catch: wrong jurisdiction, outdated rules, overconfident wording, or advice disguised as practice. It fixes or withholds the answer and logs the event for human review.
Source on every rule
When Unshaken states a rule, you can open its source and see when we last checked it. If a source is older than our review window, the rule is flagged as possibly out of date.
Fictional by design
Characters, judges, companies and template cases are fictional. We don’t imitate real judges, lawyers or parties, and we don’t predict how a real judge will rule.
AI ethics issues specific to law
The risks we design against
Legal AI carries risks that general apps don’t. Here is each one and what we do about it.
Unauthorized practice of law
A tool that tells you what to file or argue can cross into legal advice.
Our safeguard
Rehearsal and general information only. Strategy and filing questions are routed to lawyers, legal aid and self-help centers. Our legal advisors review copy against this line.
Fabricated or outdated law
Fake citations and stale rules have already reached real court filings.
Our safeguard
Verified rule base, a citation and date on every rule, Corrective AI, and an expiry date on every rule that forces re-review.
Bias and uneven performance
Models can work worse in some languages, dialects and reading levels, and can stereotype parties.
Our safeguard
Bias testing every quarter across language, reading level, case type and character design, with results published. Jurors and characters are built to show bias as something to recognize and answer, never as truth.
Over-reliance
People may trust a confident AI more than they should, or skip getting a lawyer.
Our safeguard
Plain reminders of limits at decision points, a “talk to a person” path on every screen, and scores framed as readiness to practice, never as a prediction of your case.
Confidentiality and privilege
What you type into an AI tool may be discoverable and isn’t automatically privileged.
Our safeguard
Data minimization, deletion on request, vendor no-training terms, and clear warnings about privilege before you enter case details.
Coaching false testimony
A practice tool could be misused to rehearse a lie.
Our safeguard
We teach truthful answering habits only. Requests to invent facts or hide evidence are refused and explained.
Emotional harm and manipulation
Realistic hostile questioning can retraumatize, and a friendly AI can foster dependence.
Our safeguard
User-controlled intensity, trauma-informed content, crisis routing to 988 and DV hotlines, and a companion that encourages real support instead of replacing it.
Conflicts of interest in referrals
Paid placement can look like a recommendation.
Our safeguard
Listings labeled as advertising, no pay-for-recommendation, no fee sharing, and compliance with each state’s lawyer-advertising rules (in Michigan, flat-fee advertising only).
Evidence and deepfakes
AI can alter or invent images, audio and records.
Our safeguard
Evidence you upload is analyzed for practice only, never altered. Generated practice exhibits are watermarked as simulated.
Top barriers
Who gets left out, and how we respond
An ethical tool has to work for the people who need it most, not just for those who find technology easy.
| Barrier | What it looks like | Our response |
|---|---|---|
| Cost | Lawyers are out of reach and fees add up. | A free core tier in every Edition, plus free access through legal aid and court self-help partners. |
| Device and internet | Phone only, limited data, shared devices. | Phone-first design, low-bandwidth mode, and text-only play when voice isn’t possible. |
| Language | English-only court materials. | Multiple languages, with each language tested separately before release. |
| Reading level | Legal language is dense. | Plain language checked against a reading-level target, with voice read-aloud. |
| Disability | Screen readers, hearing, cognition, mobility. | WCAG 2.2 AA as the minimum, captions on every voice line, and adjustable timers. |
| Fear and trauma | Freezing up, avoidance, DV safety concerns. | Gentle start modes, a stress companion, a quick-exit button, and safety resources. |
| Distrust of AI | Worry about errors, surveillance and data use. | Published audits, sources on every rule, and a clear privacy promise. |
| Time | Can’t miss work, short deadlines. | Ten-minute practice sessions and a deadline-first checklist. |
Regular testing
We test on a schedule, not when it’s convenient
| When | What we test | Who checks | Published |
|---|---|---|---|
| Every release | Full regression suite: legal-accuracy questions per Edition and jurisdiction, guardrail and refusal tests, safety routing. | Engineering, with release blocked on failure | Release notes |
| Monthly | Red-team testing: attempts to get legal advice, false testimony coaching, private data leaks and harmful content. | Internal red team | Quarterly summary |
| Every quarter | Bias and fairness across languages, reading levels, case types and character design. Rule base freshness review. | Internal team plus legal aid advisors | Quarterly report |
| Every year | Full independent audit: accuracy, privacy, security, bias and user outcomes. | Third-party evaluator | Annual audit report |
| Always | User error reports, reviewed within 2 business days, and a public corrections log. | Named Edition owner | Corrections log |
Compared with people
How we measure against human accuracy
“Accurate” only means something next to a baseline. We compare Unshaken with what a person would get from people.
The method
- ♦A benchmark of real-world practice questions for each Edition and state, written and answer-keyed by licensed attorneys.
- ♦The same questions answered by Unshaken and by human panels: legal aid intake staff, court self-help staff and licensed attorneys.
- ♦Answers graded blind by a separate attorney panel that doesn’t know who wrote them.
- ♦We report error rates, serious-error rates, fabricated citations, refusal accuracy and readability.
What we will and won’t claim
- ♦We will publish every result, including categories where people outperform Unshaken.
- ♦We won’t claim “lawyer-level” anything unless a published audit supports it, and then only for the tested Edition and state.
- ♦Accuracy on practice questions isn’t a prediction of your case outcome.
| Measure | Unshaken | Human panel | Status |
|---|---|---|---|
| Correct answers on the legal-information benchmark | — | — | Baseline in progress |
| Serious errors (could cause a missed deadline or lost right) | — | — | Baseline in progress |
| Fabricated citations or rules | — | — | Baseline in progress |
| Correctly refused legal-advice requests | — | n/a | Baseline in progress |
| Accuracy gap between English and other languages | — | — | Baseline in progress |
Values fill in when the first independent evaluation is complete. We will not publish estimates in their place.
Public audit reports
Every audit, published and readable
Each report comes in two forms: a plain-language summary anyone can read in five minutes, and a full technical appendix with methods, data and failures. Reports stay up permanently, including the ones that make us look bad.
| Report | Scope | Evaluator | Status |
|---|---|---|---|
| Ethics Charter v1.0 | This document and its commitments | Internal, with advisory panel review | Published |
| Baseline accuracy report | Custody and Injury Editions, Michigan | Independent evaluator | In progress |
| Privacy and security assessment | Data flows, vendors, access, deletion | Independent security firm | Planned |
| Bias and language report | All live Editions and languages | Academic partner | Planned |
| Corrections log | Every user-reported error and how we fixed it | Edition owners | Opens at launch |
Third-party evaluation
- ♦An annual independent audit, with evaluators free to publish findings without our approval.
- ♦Access for academic researchers under a privacy-protecting data agreement.
- ♦We follow published vendor review standards for legal aid AI, such as the State Bar of California’s LAJTC vendor guidance.
Advisory panel
- ♦Legal aid attorneys, court self-help staff, a privacy expert, a trauma-informed care practitioner and people who have represented themselves in court.
- ♦The panel can block a release on ethics grounds.
- ♦Panel minutes on major decisions are summarized publicly.
When something goes wrong
- ♦Serious errors are acknowledged publicly, with what we changed.
- ♦Affected users are notified directly about data incidents, as the law requires and sooner where we can.
- ♦A way to dispute a result or request a human review.
Report a problem or ethics concern
Report incorrect information, a biased response, broken resource links, privacy issues or a safety concern. Include the page URL, what happened and the correction you believe is needed. Do not email confidential case documents, names of children or sensitive health details. This is not an emergency or legal-advice service.
Email ethics@symbolic.systemsThis address is the proposed reporting contact; inbox monitoring and response time still need operator confirmation. For time-sensitive legal matters, contact a lawyer or legal aid rather than waiting for a reply. In immediate danger call 911; for a mental health crisis call or text 988.
Alignment with legal aid standards
How this charter maps to the field’s guidance
We built this charter against two resources legal aid organizations use to adopt AI responsibly: the LSNTAP AI Policy Builder and the State Bar of California LAJTC Legal Aid Generative AI Readiness Model.
| Unshaken commitment | LSNTAP AI Policy Builder section | LAJTC Readiness Model |
|---|---|---|
| Our manifesto | Purpose and Objectives; Ethical Principles and Organizational Commitments | Governance; Community Accountability |
| Practice, not legal advice | Acceptable Use of AI Tools; Compliance with Legal and Ethical Standards | Professional Obligations; Use Cases |
| Data privacy | Data Protection and Confidentiality | Technology and Data; controls in place before data enters AI systems |
| Accuracy and hallucinations | Human Oversight and Quality Assurance | Risk of “hallucinations and false or misleading outputs”; human-in-the-loop review that isn’t a rubber stamp |
| Bias testing | Ethical Principles (fairness) | Bias from model design or training data, and uneven performance across languages |
| Clear labels and limits | Disclosure and Transparency | Client-facing transparency, language access, accessibility, notice and human escalation |
| Regular testing | Training, Governance, and Oversight | Continuous evaluation of performance, errors, security, reliability and bias; adversarial testing |
| Report a problem; corrections log | Reporting and Incident Response | Client feedback and redress practices |
| Third-party evaluation | Compliance with Legal and Ethical Standards | Vendor review matched to use case, data and affected users |
| Twice-yearly charter review | Review and Revision | Governance integrated into operations and regularly reassessed |
LSNTAP and the State Bar of California publish these resources for legal aid organizations. Their inclusion here shows how we aligned our practices; it does not mean either organization has reviewed or endorsed Unshaken.
Our promise in one paragraph
We built Unshaken for people walking into court alone. We will tell you the truth about the law and about our limits, protect what you share, test ourselves on a schedule, let independent experts check our work, and publish what they find. If we fall short, we will say so and fix it.
Unshaken: Justice is a product of Symbolic Systems LLC. Education and practice only, not legal advice. If you’re in crisis, call or text 988. For domestic violence support, call 1-800-799-7233. See also the AI engine safeguards and our website privacy notice.
Sources
- LSNTAP, AI Policy Builder: What is included.
- State Bar of California, Legal Aid Generative AI Readiness Model (October 2026), via the Legal Aid Justice Technology Collaborative.
- Dahl, Magesh, Suzgun & Ho, “Large Legal Fictions,” Journal of Legal Analysis 16 (2024).
- Magesh et al., “Hallucination-Free? Assessing the Reliability of Leading AI Legal Research Tools,” Journal of Empirical Legal Studies 22 (2025); summary at LLRX.
- ABA Model Rule 7.2 (advertising) and Michigan informal ethics opinion RI-365 on paid lead services.