Responsible AI Policy

Last updated: June 10, 2026

This Responsible AI Policy explains how ClearPrecedent, LLC, a Wyoming limited liability company ("ClearPrecedent," "we," "us," or "our") uses artificial intelligence inside ClearPrecedent, where AI helps, where it does not, and the design choices we have made to keep the product a useful tool rather than a substitute for a lawyer.

This Policy works alongside our Terms of Service, Privacy Policy, and Data Handling document. Where those documents and this Policy use the same defined term, the meaning is the same.

1. Why This Policy Exists

ClearPrecedent uses large language models to help users draft legal documents, search case law, and explore how cases relate to one another. AI makes those tasks faster and, when used well, better. It also creates real risks: confident-sounding mistakes, fabricated authorities, oversimplified analysis, and the temptation to skip the human review that legal work requires.

This Policy describes how we manage those risks. It is meant to be specific and verifiable, not aspirational. Where we promise something, we mean it. Where we cannot promise, we say so plainly.

2. What AI Does and Does Not Do Inside ClearPrecedent

What AI does

• Drafting. The Motion Builder uses AI to generate a draft document from facts, prompts, and supporting materials you provide. The draft is structured for a particular court and motion type and includes citations to authorities we have verified against our corpus.
• Search. Case Search uses AI to rank and summarize opinions that may be relevant to a query. The full text of opinions is available so you can read what the cases actually say.
• Synthesis. Judge Intelligence and other features use AI to surface patterns in public materials, such as a judge's published opinions.
• Explanation. Insights articles and Glossary entries use AI as one input among several to explain legal concepts in plain language. Editorial review applies before publication.

What AI does not do

• It does not give legal advice. The Motion Builder produces a draft document, not a legal opinion tailored to your situation.
• It does not select your strategy. Where we surface multiple options, we identify them as options and ask you to choose.
• It does not represent you. No one at ClearPrecedent appears in court for you, signs filings for you, or accepts service for you.
• It does not bind you to a model's reasoning. The model can be wrong. You are responsible for the document you file.

3. UPL Design: How We Built the Product to Be a Tool, Not a Lawyer

This is the most important section of this Policy. It explains the design choices that distinguish ClearPrecedent from a legal-services product, and that we believe keep our work within the scope of permitted software, not the unauthorized practice of law.

The Motion Builder produces a draft, not advice

The Motion Builder outputs a draft document. The draft is clearly identified as a draft. It is structured for your review and editing, not for direct filing. The user interface prompts you to review the draft before downloading or filing.

We do not present the draft as a legal opinion about your specific case. We do not say "you should file this" or "you will win" or "this is the right strategy for your situation." The draft is a working document. The legal judgments embedded in it are yours, not ours.

Citations are checked against our corpus

Citations in a Motion Builder output are automatically checked against our corpus of public opinions, drawn primarily from CourtListener, with the goal of removing unverifiable citations before the draft is delivered to you. We do not guarantee that every fabricated or miscited citation will be caught. We back this up with the hallucination refund described in section 12: if something slips through, the refund is how we stand behind the verification work.

Citation verification does not certify that a case actually supports the proposition the draft cites it for. Verifying the substance of every cited holding is your job, and where stakes are high, an attorney's job. We tell you this in the product and again in our Terms of Service.

You are the author of record

When you download or file a Motion Builder draft, you are the author. We do not sign the document. We do not appear on the caption. We do not enter an appearance. We do not accept service for you. If a court has questions about the filing, those questions go to you, not to us.

This is more than a disclaimer. It reflects how the product is built. The Motion Builder is a drafting tool that produces a document you control. We do not maintain client files, file briefs on behalf of users, or send documents to opposing counsel for them. The Services are software, and the act of being a party or a representative in a legal matter is yours.

We surface options; we do not pick strategy

Where we present multiple possible approaches, we identify them as options and let you choose. Where we score or rank approaches, the score is a relevance signal based on retrieval and pattern matching, not a recommendation to take a particular position in a particular case. We say this in the interface so you can interpret the score for what it is.

Insights and Glossary are educational

The Insights articles on clearprecedent.com explain legal topics in plain language. They are educational. They are not advice on your specific case. The Glossary defines legal terms. It is a definitional reference, not a recommendation about whether a term applies to your situation.

Anyone can use ClearPrecedent

We do not require bar admission to sign up. Pro se litigants, paralegals, law students, and curious non-lawyers are welcome. We believe a tool that helps people understand the legal system and prepare drafts is a public good, and that the unauthorized-practice-of-law rules govern what users can do with a draft, not what we can build. As described in our Terms of Service, you remain responsible for complying with the rules that apply to you, including the rules of any court where you file and the unauthorized-practice rules of your jurisdiction.

4. Models We Use

ClearPrecedent uses commercial large language models from established providers. We choose the model for a given task based on quality, cost, and availability. The specific model and the provider used for any given Motion Builder run or other AI task are not exposed to users, because the right model can change as the field changes and because the user-facing contract is about the quality and behavior of the output, not the identity of the engine producing it.

We may add, remove, or substitute models and providers at any time without notice. We will not, however, change our position on no training on customer data, no retention of customer inputs by providers, citation verification, or the other commitments in this Policy without notifying users.

5. Verification Mechanisms

We use several verification mechanisms to keep AI output honest.

Citation verification

The Motion Builder runs automated verification against our corpus before delivering a draft, with the goal of removing citations that cannot be matched to a real authority. Verification catches the great majority of fabricated authorities but is not a guarantee against every error. The corpus draws primarily on CourtListener, which is the largest free repository of U.S. case law. We supplement with public statutes and procedural rules where relevant.

A citation that matches our corpus is real. It is a real case, with a real citation, in a real reporter or court. What the case stands for is a separate question, which you must answer through your own reading.

Full-opinion grounding for Search 2

Our Search 2 ranker reads the full HTML-with-citations text of candidate opinions when ranking results, not just truncated snippets. This is more expensive in compute, but it produces rankings that reflect what the opinion actually says, not just what the first paragraph said. We accept the latency cost to keep results accurate.

Hallucination refund

If a citation in a Motion Builder output is fabricated or materially miscited, you can request a refund of that motion's charge. Section 12 describes the process. Our Terms of Service contain the formal commitment.

Internal evaluation

We maintain an internal test suite for the Motion Builder and for Case Search. The suite includes regression cases drawn from prior issues and from team-generated examples. We run this suite as part of changes to prompts, retrieval, or model routing. We also conduct attorney spot-checks on sample outputs.

6. Human Oversight Is Required

ClearPrecedent works best when a human stays in the loop. Our written commitments and your expectations both depend on this.

You are required to:

• Read the Motion Builder draft before you file it or share it.
• Confirm that the facts in the draft match the facts of your case.
• Check that the law and the citations apply to your jurisdiction at the present time.
• Edit the draft so it reflects what you actually want to say.

We strongly recommend that you have an attorney licensed in the relevant jurisdiction review the draft when the stakes are high, for example when liberty, custody, or significant money is at risk. The Services can prepare a working draft; they cannot replace the judgment of someone who knows your case, your judge, and the local practice.

7. Where AI Reliably Falls Short

We want to be plain about what AI does not do well today.

• Just-decided cases. Opinions issued in the last few days may not be in our corpus yet. Embargoed or sealed opinions are not in any public corpus.
• Novel or unusual jurisdictions. State trial-court practice, magistrate procedures, specialized administrative tribunals, and tribal courts are uneven in our corpus. Output for these forums is more likely to need attorney review.
• Ambiguous facts. AI cannot resolve fact disputes or judge credibility. When the facts are contested, the model is guessing.
• Witness credibility. AI cannot assess whether a witness will be believed.
• Live oral argument. The Services are designed for written work. Preparing for oral argument requires a human, often with knowledge of the bench.
• Tactical judgment. Decisions about whether to file, when to file, what to concede, and what to fight require professional judgment we cannot supply.
• Cases overruled or distinguished after the corpus snapshot. Even with verification, our corpus reflects a point in time. The current treatment of an opinion can change.

You should treat these areas as places where extra review is warranted.

8. Bias Acknowledgment

AI systems trained on legal data inherit the shape of that data. Published opinions over-represent certain jurisdictions, certain courts, certain doctrines, and certain kinds of litigants. Federal courts of appeals are heavily represented. State trial-court orders are sparsely represented. Civil litigation around large commercial parties produces more written law than disputes between individuals.

We try to mitigate this. We index broadly. We use retrieval that surfaces less-cited but on-point authority. We use full-opinion grounding to reduce summary-level bias. We invite user reports of bias in outputs.

We cannot eliminate it. If a topic is under-litigated in the public record, the Services may have less to offer. Be especially careful with output on civil-rights claims for marginalized litigants, on emerging technologies, and on areas where the controlling authority is administrative rather than judicial.

9. Reporting Issues

If you receive a Motion Builder output that contains an error, please tell us. Email support@clearprecedent.com with:

• The motion ID or a link to the output.
• A description of the error, including the specific text or citation at issue.
• What you believe the correct content should be, if you know.

We review every error report. We use error reports to fix individual outputs where we can, to improve verification, to update prompts and retrieval, and to add cases to our internal test suite so the same error does not recur. Error reports are not used to train AI models on your inputs; the no-training commitment applies.

For fabricated or materially miscited citations, you can also request the hallucination refund described in our Terms of Service.

10. No Training on Customer Data

We do not use your inputs, your uploaded documents, your prompts, or your Motion Builder drafts to train any AI model, our own or any third-party provider's. This applies to:

• Pretraining.
• Fine-tuning.
• Reinforcement learning from human feedback or any related technique.
• Any other model-improvement process.

For model improvement we use public CourtListener opinions, public statutes and rules, and our own internal evaluation corpus, which is built from synthetic and public materials, not from user content.

We have selected AI providers whose commercial API terms commit them not to train on customer inputs. Those providers may briefly retain inputs and outputs for trust-and-safety and abuse-monitoring purposes (typically up to thirty days) and then delete them; they do not use them for any other purpose. If a provider's standard terms or behavior cannot be kept consistent with this commitment, we will move off that provider.

This is repeated in our Data Handling document and in our Terms of Service. It is meant to be enforceable.

11. How We Evaluate and Improve the System

Model improvement at ClearPrecedent is a continuous process. The pieces include:

• Internal test suite. A library of inputs and expected outputs covering core motions, key search queries, citation verification, and edge cases.
• Regression checks. When we change prompts, retrieval, or model routing, we run the test suite and compare results against the prior version.
• Attorney spot-checks. Sampled outputs are reviewed by attorneys on a periodic basis. Findings feed back into the prompts and the test suite.
• User reports. Issues you report are triaged and, where appropriate, become new test cases.
• Corpus updates. Our corpus is refreshed as CourtListener data updates and as we expand coverage of statutes and rules.

Where we make a material change to how AI is used in the product, we will update this Policy and, when appropriate, notify users in the app.

12. Refund Mechanism for Hallucinations

This section restates the customer-facing promise. The formal terms are in our Terms of Service.

If a Motion Builder output contains a fabricated citation, or a citation that materially miscites the case it claims to support, you can request a refund of that motion's purchase price. Material miscites are mistakes that would mislead a careful reader: a wrong holding, a wrong court, a wrong year, wrong parties. Stylistic variations in citation form do not qualify.

To request a refund, email support@clearprecedent.com within 30 days of the motion's purchase with the motion ID and the specific citation or citations at issue. We respond within 10 business days. Approved refunds are returned to the original payment method.

13. Updates to This Policy

We will update this Policy as the product evolves and as we learn from experience. Material changes will be highlighted to users through email or in-app notice. The "Last updated" date at the top reflects when the current version took effect. Prior versions are available on request.

Questions?

For questions about how AI is used in ClearPrecedent, to report an issue with an output, or to request a hallucination refund, write to us at support@clearprecedent.com.