LexLift — Data & AI statement

Original version approved by Tim, 2026-10-02; later factual updates are identified here. The training opt-out was requested that day and has not been confirmed. Update §3 when DeepSeek replies. Engineering clarification, 2026-10-06: local Workbench controls and network behavior are described accurately below. Provider facts retain their stated source dates. Accuracy update, 2026-10-21: §6 rewritten with the Fresh20B-D results; §3 notes Claude-drafted internal test briefs. Factual update, 2026-10-22: §6 adds the Fresh20E-G results (80 new opinions in Fresh20D-G); the lawyer-pilot, review-export and recorded-session material is removed because no pilot is running; subscriber, payment (Stripe), email (Gmail) and hosting (Cloudflare) facts are added; the contact address is filled in. These factual updates are not a new approval by Tim or a lawyer. This statement gives no legal advice and promises only what it says.

In one line: LexLift turns public California court opinions into AI-drafted briefs. It takes no client data. Every brief is an unverified AI draft that you must check against the opinion. LexLift never tells you whether a case is good law.

Source-location clarification, 2026-10-08. Source PDF page numbers count from the first PDF page and may differ from printed opinion page numbers. Generated passage numbers identify extracted text, including headings and page fragments; they are not court paragraph numbers. Check the printed opinion before using a legal pinpoint citation. This factual clarification is not a new legal approval or accuracy measurement.

1. Do not send us client information

• LexLift does not accept client information. Local Workbench controls let you search, open public-opinion folders, and enter optional study names and notes. Use them only for public-opinion research and study; do not enter confidential client information.
• Do not send client names, facts, documents or anything confidential or privileged to LexLift in any way: email, feedback, or on a call. If you mention a matter you are working on, keep it generic.
• LexLift cannot promise confidentiality for anything you send it. It is not your agent, it is not a law firm, and it is not set up to protect privileged material. If client information reaches us anyway, we will delete it when we notice. Deleting it does not undo the disclosure.

2. What goes in: public opinions only

• LexLift reads California Court of Appeal opinions, published and unpublished, as filed by the court. Court opinions are public records and are not protected by copyright.
• Where the opinions come from: CourtListener, run by the non-profit Free Law Project. LexLift uses CourtListener's search to find new opinions and downloads its copy of each opinion PDF. Earlier test briefs (made from 2026-09-30) were made from PDFs downloaded from the court's own website (courts.ca.gov). Free Law Project has not checked or endorsed any LexLift brief.
• Supreme Court review status: before LexLift charged anyone, it also read the California Supreme Court's weekly conference-results PDFs on supreme.courts.ca.gov to show whether review was granted, denied or a depublication order was made. LexLift will not use that website for the paid service unless the court gives permission. For paying subscribers, review and publication status comes only from sources that allow that use, and may be late or incomplete.
• LexLift does not use Westlaw, Lexis or any publisher's headnotes, synopses or editorial content.
• Opinions often name private people (parties, witnesses, family members) and describe their affairs. LexLift processes them exactly as the court published them. It does not look anyone up or add information about anyone.

3. Which AI companies see what

DeepSeek — writes the briefs

• Briefs are written by DeepSeek V4.1 Flash (deepseek-flash), a commercial AI model, through DeepSeek's paid API.
• For each opinion DeepSeek receives: the opinion text, split into numbered sentences; LexLift's instructions; and, in later checking rounds, LexLift's own draft brief and check results, which are made from the same opinion. It returns the brief and its checks.
• DeepSeek receives nothing about you: no name, email, firm, bar number, billing details, notes or client information. The requests are sent from LexLift's computer under LexLift's API account, so DeepSeek can link them to LexLift, but not to any subscriber.
• What DeepSeek's own documents say (all read on 2026-10-02):
  • The company is Hangzhou DeepSeek Artificial Intelligence Co., Ltd. Its privacy policy (last updated Feb. 10, 2026) says it collects, processes and stores personal data in the People's Republic of China. It keeps data "as long as necessary" for its services and other stated purposes, and gives no fixed period. It says it uses inputs to train and improve its models, lists a right to opt out of training, and may share data in response to legal process or government requests. https://cdn.deepseek.com/policies/en-US/deepseek-privacy-policy.html
  • Its Open Platform (API) Terms (effective Apr. 29, 2026) assign to LexLift any rights, title and interests they have in the output, if any, subject to compliance with applicable law and the terms. They are governed by Chinese law. They do not say separately whether API inputs are used for training or how long they are kept. They require LexLift to tell you that output is AI-generated, may contain errors or omissions, and is not legal advice. This page and every brief do that. https://cdn.deepseek.com/policies/en-US/deepseek-open-platform-terms-of-service.html
  • Its model and training disclosure says a small part of its fine-tuning data may be based on user input, after de-identification, and that users can opt out. https://cdn.deepseek.com/policies/en-US/model-algorithm-disclosure.html
• LexLift's training opt-out: DeepSeek's API platform has no account setting for training. Tim checked on 2026-10-02 and found only profile and display settings. The privacy policy says the training opt-out is exercised by emailing privacy@deepseek.com. LexLift asked DeepSeek by email on 2026-10-02 to exclude its API data from model training. DeepSeek has not confirmed yet, so assume the opt-out may not be in effect. Either way, DeepSeek's policy says it stores data in the People's Republic of China. LexLift sends DeepSeek only public court opinions, its own instructions, and its own drafts and check results made from those opinions. It never sends subscriber details, user-entered search text, study names or notes, or locally opened folder files to DeepSeek. Those local controls exist; they are not a channel for client submissions.
• What this means for you: assume that anything sent to DeepSeek may be kept indefinitely in China and may be used to train its models. LexLift accepts that only because everything it sends is already public. It is the main reason you must never send client information to LexLift.

Anthropic (Claude) — internal testing and building only

• LexLift uses Claude models (Anthropic) to grade test briefs against the full opinion when measuring accuracy, and to help write and review LexLift's software and documents.
• Claude sees public opinions, LexLift's briefs, and LexLift's code and documents. It never writes or changes the briefs you receive and is never used on client data. In internal tests only, Claude has also drafted test briefs of public opinions to compare against DeepSeek.
• Feedback you send by email is never sent to DeepSeek. Anthropic's handling of anything LexLift gives Claude is governed by the terms of LexLift's Claude account.

OpenAI (Codex) — internal testing, software and reporting work

• From 2026-10-05, Codex helps maintain LexLift’s software and documents and tally existing test grades.
• From holdout 23, independent Codex agents also grade public opinions and their test briefs, and rehearse the Workbench workflow as simulated lawyer personas. They see those public opinions, test drafts, LexLift’s code and documents. This is AI testing, not lawyer feedback. The collaboration tools do not expose the runtime grader model identifier; earlier cohorts used Claude Opus and remain separate.
• Codex does not draft the briefs delivered to users and did not regrade holdout 22: those grades remain Claude Opus grades. No client data or subscriber information is used in these Codex tests.

No other AI providers write briefs

If LexLift adds or changes an AI provider for anything that touches briefs or user data, this statement will name it before the change.

4. What is stored, and where

• LexLift has no user accounts, and no analytics or tracking in its pages.
• Brief pages are self-contained; their automatic requests are restricted. Links to outside source pages open when clicked. The Workbench instead runs through a local server and requests case files and PDFs from that server. Local loading is not a submission to DeepSeek.
• The Workbench has an optional AI chat. It stays off unless someone puts an AI provider key on the computer running it, and it is not part of the subscription. If it is ever offered, this statement will first say which provider receives your questions and the brief text.

5. What a brief is — and is not

• Every brief is an AI draft that no lawyer has reviewed, unless it plainly names a lawyer who did. LexLift's founder is not a lawyer.
• Every statement cites the opinion sentences it relies on, so you can check it quickly. Automatic checks and further AI review passes add warnings to some statements. They never mark anything verified. A statement with no warning can still be wrong, and a brief can leave out important points.
• You must verify everything you rely on against the opinion itself. The State Bar of California's 2026 generative-AI guidance tells lawyers to review and verify AI output before using it. From January 1, 2027, Senate Bill 574 (Stats. 2026, ch. 858) makes that a statutory duty, including personally verifying citations in court filings. The duty falls on lawyers, not on tools. LexLift is a starting point for that work, not a substitute for it.
• LexLift does not say whether a case is good law. Each brief shows only the opinion's official publication and Supreme Court review status "as of" a stated date, with a link to the source. It does not track later cases that cite, question or overrule an opinion, and its status record can be late or wrong. Use KeyCite, Shepard's or the court docket before citing anything.
• Unpublished opinions generally may not be cited (Cal. Rules of Court, rule 8.1115).
• LexLift is not a law firm and gives no legal advice. Using it creates no attorney-client relationship.

6. Accuracy, stated honestly

LexLift tests every release on a fresh batch of 20 public California Court of Appeal civil opinions it has never seen, frozen before any AI call. Each brief is graded against the full opinion by an independent AI grader using a written rubric, followed by a strict second pass that applies the rubric's definition of "serious" literally. These are AI judgments, not lawyer review.

A serious error is one a lawyer could be misled by: a wrong holding, disposition or issue result; a party given the wrong role; a trial-court ruling or party argument presented as this court's holding; a materially overstated or understated holding; saying the court did not decide something it decided (or the reverse); or a wrong date, amount or name a lawyer could rely on.

In short: in AI-graded tests on 80 new opinions (Oct. 21, 2026), the holding and outcome were almost always right; 16 of 80 briefs (20%) had at least one serious error somewhere in the brief and 6 (7.5%) had one with no warning beside it. The reader puts every statement next to its source sentence so you can check before relying on it.

Each test used a different set of 20 opinions, and all used comparable strict standards; the automatic grader used from Fresh20D was checked against the earlier independent grading of Fresh20B and reached the same totals. Fresh20D-G are the 80 opinions behind the summary above; Fresh20B and C used earlier versions and are not pooled with them. Many serious errors were in a short label (an issue's result or a party's role) or a hedged ruling stated too firmly, and most carried a warning. Since then, a result label that a check doubted is no longer stated as fact; the page says "See the holding - label not confirmed" instead.

What this means for you: about one brief in five contained at least one serious error somewhere, and about one in thirteen had one with no warning beside it. A statement with no warning can still be wrong, and these tests cannot establish an exact error rate for all briefs. Earlier test series (holdouts 1-29, through 2026-10-10) used different graders and versions; they remain in LexLift's records and are not pooled with these figures.

7. Quick answers for your AI due diligence

The State Bar of California's Practical Guidance for the Use of Generative AI (approved May 14, 2026) asks lawyers to find out, beyond vendor marketing, how a tool collects, uses, stores and discloses information. It also asks whether inputs are used for training and where data goes. Our answers:

Contact: lexliftai.llc@gmail.com. Latest accuracy tests: Fresh20D-G, 2026-10-21 (§6); internal software-provider list updated 2026-10-05; subscriber, payment, email and hosting facts added 2026-10-22. Vendor terms and training opt-out status remain as checked on 2026-10-02.