Docket Alarm vs UniCourt: how they compare in 2026

The machine learning is real and the product it sits on predates it and stands without it. What the AI does: Talk to the Docket answers plain-English questions about a case, AI Summaries condense a filing on opening, and extraction turns docket entries and underlying filings into the normalised party, counsel and per-entry structure the API returns and the motion and judge analytics run on. Strip all of that out and what remains is the business Docket Alarm sold for a decade and still leads with: full-text search across a billion filings, alerts on any case, party, attorney or judge, docket tracking, direct PACER retrieval and bulk document download. The pricing page settles it more clearly than the marketing does. Both published plans are priced in units of the non-AI product, at 4 dollars a document, 3 a docket update, 2 a day for tracking and 1 a search, and the analytics products are sold separately as Federal Analytics, Easy Analytics and the Analytics Workbench add-ons. That is a feature layer on a data platform whose value stands without it. Graded one band below UniCourt, read the same day, where the normalised dataset is the whole product rather than one route into it.

Machine learning is the mechanism that turns fragmented court filings into a usable dataset, on a platform whose delivery layer would function without it. The vendor describes AI-powered data extraction surfacing data buried deep in dockets and documents, with standardisation and deduplication as the stated output, and names DEEP as its AI-powered docket extraction and enrichment platform. Remove the models and what remains is aggregation and delivery of raw court records through APIs, which is a real product but a materially worse one. That places the AI as the engine of the core capability layered on a data pipeline, which is the B band rather than the A.

Grounding is documented and demonstrated; measurement is absent. The AI Assistant page states that the system draws only on the docket being viewed, reads the actual record from the complaint through the most recent filing including the documents behind the entries rather than the entry text alone, and links every statement to the docket entry or document supporting it so a reader can open the source in one click. The page carries a worked example: an answer on a named 2026 Northern District of California declaratory-judgment case, with a Sources list identifying the complaint and its filing date. What is missing is any number. No accuracy rate, no test set, no evaluation and no error analysis was located on 31 Aug 2026, while the same page asserts as an absolute that output carries no hallucinated facts or fabricated citations. That claim sits against the vendor's own agreement, which states that machine-generated output is probabilistic and should be evaluated for accuracy including by expert human review. The marketing promises what the contract declines to warrant, and a buyer should read both. One limb of this axis does not apply rather than failing: a citator or good-law check is not in scope for a system that retrieves case records instead of asserting propositions of law, and it is neither credited nor penalised here.

Grounding is total and measurement is absent, which is the B profile exactly. Every record UniCourt returns is the actual court document or docket entry, retrievable to source, so there is no gap between what the platform asserts and what a user can open and read; this is the strongest grounding limb in the pull. The vendor also publishes a real limitation rather than hiding it: the public records policy states that records may contain errors or be incomplete, inaccurate, or not the most currently available information, and are provided as is. What is missing is any measurement of the thing the AI actually does. Extraction, standardisation and deduplication are claimed as unmatched with no accuracy or error rate, no benchmark, no test set and no published evaluation located on 31 Aug 2026. One limb of this axis does not bite: a citator or good-law check is not applicable to a platform that delivers records rather than legal propositions, and is neither credited nor penalised here.

A real review surface and a real stopping behaviour are published, and what happens after an error is not. The review surface is specific: each statement in an answer links to the entry or document that supports it, and the vendor frames the point of that link as letting the reader decide for themselves whether the answer is correct. The stopping behaviour is stated too, that where a fact is not in the record the system says so rather than filling the gap. The agreement adds the supervision obligation in writing, telling the customer that machine-generated output should be evaluated for accuracy including by employing expert human review. Against that, checked on the home page, the AI Assistant page, the API page and the terms on 31 Aug 2026: no confidence threshold or exception behaviour is described, no route exists to report or correct a wrong answer, and the case-briefing function, which the vendor says reads hundreds of docket entries and traces the procedural history unattended, has no described checkpoint between the request and the finished brief.

The extraction runs unattended and no oversight model is published. Machine learning classifies, normalises and deduplicates docket data before a customer ever sees it, and for API and customers that output flows straight into their own systems with no review surface in between. Searched the home page, pricing page, enterprise API page, PACER API page, general disclaimer, terms of service and public records policy on 31 Aug 2026: nothing describes what the models decide alone, no confidence or exception handling is documented, and no correction or reporting route for mis-extracted data was located. The mitigation is structural rather than stated, in that every extracted field sits alongside a retrievable source document, so an attentive user can check any value against the original.

No production evidence of any kind was located. Searched the home page, the full use-cases page covering all five buyer segments, the about page, the AI Assistant page, the developer API page, the coverage page and the pricing page on 31 Aug 2026. There is no named customer, no attributed testimonial, no case study, no customer count, no logo strip and no figure for what changed at any organisation. What stands in its place is unattributed: a claim that thousands of legal professionals rely on the platform, a signup modal saying the vendor has relationships with many large firms, and a claim of direct partnerships with law schools that names no school. The five example workflows on the use-cases page are explicitly hypothetical illustrations rather than accounts of real deployments, and the one time figure published, that a case brief replaces two to four hours of paralegal work, is a product claim with no customer or method attached. For a platform this long established, in market since well before its 2018 acquisition, the complete absence of a named reference is the finding.

Logos stand in for evidence. The pricing page carries a customer strip including DLA Piper, Paul Hastings, Morrison Foerster, Alston & Bird, Akerman, Pillsbury, Manatt, Stikeman Elliott and Seyfarth Shaw, alongside CNN, The Wall Street Journal, Financial Times, NBC News, RBC, Colliers and Coindesk, which is a serious roster spanning large firms, media and financial services. Nothing sits behind it. Searched the same surfaces on 31 Aug 2026 and located no named individual, no attributed testimonial, no case study, no figure for what any customer achieved, and no deployment scope or date. Two names in the strip are misspelled, as Seyearth Saw and Stikeman Elliot, which is a small quality signal on the one page carrying the vendor's customer evidence.

Two live documents on the same property give a buyer opposite answers, and that is the finding. The privacy policy makes a specific and well-aimed commitment: it recognises that litigation activity involves viewing potentially adverse litigant material, and states that personal search history is kept strictly confidential to the user's organisation and is not sold to third-party marketers. That addresses the real exposure here, which is that the cases a firm looks up map onto its live docket. The terms of service, which govern, run the other way. They reserve the right to monitor and record customer usage including content interactions, search history and IP addresses, and to track activity to gather further information about the customer and its authorised users. They grant the vendor an irrevocable, perpetual and sublicensable licence over user content to improve its products. They disclose that a limited number of staff and third-party foundational AI models may technically access user content. And the confidentiality clause excludes information that is publicly available, which describes nearly everything in the corpus. Nothing reconciles the two documents. Beyond them, no training position, no retention period, no access model within a subscribing firm and no privilege or work product treatment was located on 31 Aug 2026. Graded on the documented position rather than the better half of it.

The product's central confidentiality question is disclosed plainly and then left unaddressed. UniCourt's CrowdSourced Library works by pooling retrieval: the vendor states that any time one user downloads a court document or docket it is added to the library and made freely available to all UniCourt users, which is why no court fee applies to material already pulled. The material is public court record and the vendor states that sealed and restricted documents are not available and nonpublic information is not distributed, so this is not disclosure of privileged content. But the pattern is exposed, and the PACER API page confirms that search results show which cases are already held and when they were added and last updated, so a subscriber can infer that someone retrieved a given docket and roughly when. Nothing published addresses whether a firm's own search history, tracked cases or saved work are visible to anyone else, and no training, retention or matter-segregation position was located on 31 Aug 2026.

A substantive position, published in the agreement and nowhere else. The terms carry a clause headed to the effect that no legal counsel is provided: content is for practical and informative purposes, should not be construed as professional advice, does not create an attorney-client relationship, and users are told to seek qualified professional counsel before acting on it. A separate clause covers AI output specifically, describing it as machine-generated probabilistic prediction to be evaluated for accuracy including by expert human review, which is the competence and supervision dimension most vendors omit. The audience is defined rather than left ambiguous: the services are stated to be for professional use, and consumers and non-professionals are expressly prohibited from accessing them, so there is no consumer-facing surface to disclose on. Two things hold this short of the top band. No bar or ethics guidance is named anywhere, including ABA Formal Opinion 512. And no practice-jurisdiction limit is stated; the only geographic limit located is a licence restriction on using the service outside the country in the subscription agreement, which is a commercial term rather than a statement about where the product's answers can be relied on. Checked the home page, AI Assistant page, use-cases page, about page and footer on 31 Aug 2026: none of this appears on any product surface, only in the terms.

A published disclaimer that addresses the advice line directly and in the product's own terms. The general disclaimer states that information on the site should not be construed as legal advice, that others may reach contrary conclusions on the same or different information, that UniCourt expressly disclaims all liability for actions taken or not taken based on the information, that it makes no opinion or endorsement as to the quality, competency or character of any person or entity on the site, and that transmission of information creates no attorney-client relationship. The endorsement limb is the notable one, since the platform sells analytics on named attorneys and firms and the vendor declines to stand behind any inference a user draws about them. What holds this at B rather than higher: no bar or ethics guidance is named anywhere, including ABA Formal Opinion 512, and no jurisdiction or practice limits are stated. Checked 31 Aug 2026.

A governance surface exists at parent level and nothing connects it to this product or to the exposure this product creates. The Clio trust centre, publicly reachable at trust.clio.com and read on 31 Aug 2026, carries an AI section with three named topics, AI Risk Management, Third-Party AI Diligence and Employee AI Usage, and a featured document titled AI Overview. Their contents were not read, so nothing is credited here beyond their existence. On Docket Alarm's own property there is nothing at all: searched the home page, AI Assistant page, use-cases page, about page, coverage page, API page, privacy policy and terms and located no responsible AI statement, no accountable owner, no pre-release testing regime and no fairness or uneven-output evaluation, and no link from any Docket Alarm page to the trust centre. The gap has a concrete edge on this product. The published corporate use case invites customers to search every court for an individual's complete litigation record before hiring an executive or onboarding a vendor, and to judge whether a person is litigious and how their cases tend to end. Screening identified individuals on litigation history is the question this axis exists to ask, and no published position addresses it.

No governance position for model behaviour was located. Searched the home page, pricing page, enterprise API page, PACER API page, general disclaimer, terms of service, public records policy and privacy policy on 31 Aug 2026, and ran two targeted searches for a security, trust or responsible AI page; none exists. Nothing names an accountable owner, describes pre-release testing, publishes a responsible AI framework, or addresses uneven output. The exposure here is quieter than on a judicial prediction product but real: extraction and entity resolution decide which attorneys, firms and parties are matched to which cases, and errors in that matching propagate into the analytics customers buy and into downstream systems the data is pushed to.

A short privacy policy that is specific where it speaks and silent across most of the ground. What it does publish is real: Google App Engine named as the hosting environment for all non-payment data, a PCI-certified payment processor with card details encrypted with AES-256 and no employee access to them, a commitment to seek consent before using collected information for a new purpose, and prior versions of the policy available on request. The removal route is better than most: a free suppression request with a stated five-business-day decision window, published criteria in both directions including automatic suppression of cases mentioning minors, an explicit refusal to suppress strong-public-interest cases or requests from legal professionals wanting to be disassociated, and an honest statement that search engine snippets may persist for weeks afterwards because the vendor does not control them. What is absent, checked on the same surfaces plus the terms on 31 Aug 2026: no retention period for anything, no deletion route for customer account data as distinct from court records, no access control model, no encryption statement covering anything other than payment details, no subprocessor list, and no incident or breach notification practice. The policy also carries no date, so its currency cannot be established. The parent trust centre publishes incident response, breach notification and backup material, but is not linked from this property and its scope over Docket Alarm was not established.

A privacy policy covers the collection relationship and no operational security material exists. The policy addresses what is collected, the CCPA framework and user rights, and offers a redaction request route for a court record appearing in online search engines, which is a real remedy rather than a statement of principle. Beyond that the ground is empty: searched the home page, pricing page, API pages and the full published document set on 31 Aug 2026, and ran two targeted searches for a security or trust page, and located no encryption standard, no access control model, no retention period, no named subprocessor, no incident or breach notification practice and no business continuity position. The privacy policy itself was recovered in fragments through the search index rather than read end to end, so a complete read is the rebuttal route for anything recorded here as absent.

The allocation of loss is published in full, readable before signing, and almost entirely one-directional. The terms cap total liability, whatever the cause or theory of recovery, at one hundred dollars, and exclude indirect, special, incidental and consequential damages including lost profits, revenue, data and opportunities. Carve-outs are named: fraud, wilful misconduct, violations of applicable law, infringement or misappropriation of intellectual property and breach of confidentiality obligations sit outside the limitation. Warranties are disclaimed expressly, on an as-is and as-available basis, including merchantability, fitness for a particular purpose and non-infringement, with specific disclaimers that results will be accurate, reliable or error-free. A separate clause addresses AI output directly and declines to warrant it. The contracting entity for Docket Alarm subscriptions, the governing law and the forum are all named. What a buyer does not get is anything running the other way: the only indemnity in the agreement runs from the customer to the vendor, no warranty attaches to output, and no insurance position appears in the agreement, though the parent trust centre lists a liability insurance document that was not read. What lifts this above the vendors in this pull that publish nothing is that the risk is allocated precisely and a buyer can read the number before signing. What the number is, is one hundred dollars.

The allocation of loss is published, readable before signing, and unusually explicit about the risk this product creates. The terms of service define the subscription agreement as comprising the general terms and conditions, the privacy policy, the public records policy and an applicable transactional rate schedule, which ties the published price schedule into the contract. The limitation clause excludes indirect, incidental, special, punitive and consequential damages in contract, tort or strict liability; excludes any claim relating to the customer's inability or failure to perform research or other work, or to perform it properly or completely, even if assisted by UniCourt; and excludes any decision made or action taken in reliance on information made available through the platform. The general disclaimer adds an as-is provision disclaiming all warranties express or implied, including merchantability, fitness for a particular purpose and non-infringement. Two things could not be established from the text retrievable on 31 Aug 2026: the liability cap, and whether any indemnity runs to the customer.

A documented API and no documented connector into the systems legal work actually lives in. The developer API page names its endpoints with methods and describes what each returns: Search and Search PACER for querying by party, attorney, court, date range, case type or full text; Get Docket for the full docket sheet including every entry, parties, counsel, related cases and status; Get Document for the PDF behind an entry; Track a Case as a POST subscribing to a case or saved search; and Ask Docket, exposing the AI assistant's grounded answers programmatically. A sample JSON response is published showing the normalised structure, and a full API reference sits at a public documentation path. The terms name three native integrations, Microsoft OneDrive, Dropbox and PACER ECF accounts. What is not there is a practice system. The API page invites a firm to pipe dockets into its matter management, DMS or BI stack but names no such system and offers no pre-built connector, and nothing was located for iManage, NetDocuments, Word, Outlook, e-billing or filing on 31 Aug 2026. Data moves one way, out of Docket Alarm, with no write-back or two-way sync described.

Integration is the product rather than a feature attached to it, and it is documented to implementer level. Delivery routes are named and distinct: REST APIs, bulk data downloads, webhooks, and data shares pushed into a customer's data warehouse or data lake, alongside pre-built integrations. Target systems are named specifically, covering matter management, CRM, document management systems, internal legal databases, docketing applications and experience management systems. Structured output is aligned to SALI standards, which is a named industry data standard and the only instance of one located in this pull. What an implementer gets is stated rather than implied: API documentation, sample code and a Python library, with billable calls itemised as Search, Case, Download Document, Order Document Confirm and Update Cases and all other APIs and webhooks included, so a buyer can size an integration before committing. A PACER-specific API additionally reports whether a case is already held and when it was last updated, so a customer can avoid paying twice.

One hosting fact is published and everything a buyer would ask after it is not. The privacy policy states that all data other than payment information is stored on servers operated by Google through the Google App Engine hosting service, and points to Google's own privacy material for how that is protected. That is more than most of this lane publishes and it is where the disclosure stops. Searched the home page, API page, coverage page, terms and privacy policy on 31 Aug 2026 and located no region, no residency option, no tenancy model, and no distinction between where data is processed and where it is stored. There is also an unreconciled discrepancy across the ownership layers: Docket Alarm's own privacy policy names Google App Engine, while the parent trust centre's infrastructure section names Amazon Web Services and publishes separate status pages for United States, Canada, European Union and Australia. Which applies to this product is not stated on either property.

Nothing published on where the software runs or where customer data sits. Searched the home page, pricing page, enterprise API page, PACER API page and the full published document set on 31 Aug 2026, plus two targeted searches for a security or trust page. No hosting provider is named, no tenancy model is stated, no region or residency option is offered, and nothing distinguishes processing from storage. The gap is more consequential here than on a self-contained application, because the delivery model actively moves data into customer-controlled warehouses and lakes, so a buyer needs to know where UniCourt's side sits in order to reason about the whole pipeline, and no answer is published.

Something real was located and none of it connects to this product, so it is described here rather than credited. The Clio trust centre, read on 31 Aug 2026, is a SafeBase portal reachable without a sales conversation. It names GDPR, HIPAA, SOC 1 Type 2, SOC 2 Type 2 and TX-RAMP, publishes a dated update stating that the 2025 SOC 2 Type II report covers 1 June 2024 to 31 May 2025, and offers a penetration test report summary, both SOC reports, a liability insurance document and an AI overview behind a self-serve access request. What is missing is any connection to Docket Alarm. The portal tab reached states that it applies to the Vincent and vLex products, Docket Alarm is named nowhere on any page read, and no page on docketalarm.com links to the portal at all: the footer carries privacy, terms and case removal and nothing else. On the product's own property there is no security page, no named certification, no auditor, no scope statement, no coverage period and no trust portal, checked across the home page, AI Assistant page, use-cases page, about page, API page, coverage page and both legal documents. A buyer cannot establish that the thing they are buying sits inside any audited boundary, which is what the scope section of an attestation exists to answer. One honest qualification about the band rather than the vendor: neither published band fits this case cleanly, because C describes badges on the site with no scope or report and there are no badges here, while D describes no attestation located and one was located at the corporate family level. A band amendment for the corporate family case is logged open.

No independent security attestation was located, and no security page exists. Searched the home page, pricing page, enterprise API page, PACER API page, general disclaimer, terms of service, public records policy, privacy policy and cancellation policy on 31 Aug 2026, and ran two targeted searches specifically for a UniCourt security, trust or compliance page. No certification is named, no auditor identified, no scope or coverage period given, no penetration testing referenced, no badge displayed and no trust centre or portal exists. The absence stands out against how much else this vendor publishes: full pricing, a five-document legal set and API documentation are all open, which makes security the one area where nothing is offered.

The product acknowledges that third-party models sit underneath and never says whose. The terms disclose that a limited number of staff and third-party foundational AI models used in rendering the vendor's intelligent technology services may technically access user content, which tells a buyer that customer material reaches an external model without identifying it. Beyond that the property is silent: nothing on the AI Assistant page, the developer API page or any other Docket Alarm surface names a model, a provider, a version or a processing location, and no commitment to notify customers of a change to the model or its training data was located on 31 Aug 2026. That silence is what this grade rests on. The corporate parent does publish a subprocessor list naming Google Cloud, OpenAI, Anthropic and Amazon Web Services, and it is recorded here for completeness rather than credited: the portal states coverage of the Vincent and vLex products, does not name Docket Alarm, and is linked from nowhere on this property, so nothing connects those providers to Talk to the Docket or AI Summaries. On what a buyer of this product can establish about the system answering their questions, the answer is that it is a foundation model belonging to someone else.

Nothing published about the models behind the extraction. The vendor describes AI-powered data extraction, sophisticated data normalisation and an AI-driven analytics layer, and names DEEP as its AI extraction platform, but searched across the home page, pricing page, enterprise API page, PACER API page and the full document set on 31 Aug 2026 no model provider is identified, no model or version is named, no architecture is described, no processing location is given, no subprocessor list exists and nothing commits the vendor to notifying customers when any of it changes. For a platform whose output flows automatically into customer data warehouses, the inability to tell which model produced a field, or when that changed, is a live operational question rather than a documentation nicety.

A buyer can price this product completely without speaking to anyone. Two self-serve plans are published with figures: Flat-Fee at 99 dollars a month covering unlimited document access, case alerts, docket tracking and search plus motion filtering, and Pay-As-You-Go at 39.99 dollars a month with every unit priced at 4 dollars a document, 3 a docket update, 3 a day for new case alerts, 2 a day for docket tracking and 1 a search in federal courts. Add-ons are named as Federal Analytics, Easy Analytics and the Analytics Workbench. The pass-through costs are handled with unusual candour, which is where most of this market becomes vague: government fees are stated to be passed on without markup at the government's own rate, the PACER schedule is reproduced at 10 cents a page with a 30-page or 3 dollar cap and no cap on transcripts, nothing is charged where the document is already held, and for-pay state courts, mostly in California, are always presented as an accept-or-decline choice at the point of access. A free tier is published as a stated weekly allowance of searches and document views for material already in the system, excluding anything bought from a government system, and a three-day no-commitment trial is offered. The one gap is API pricing, which is sales-led and carries no published figure or unit.

The most complete pricing disclosure located in either pull. Rates are published for every self-serve tier: Personal at 49 dollars a month, Professional at 149 and Premium at 299, with annual billing shown as two months free at an effective 40.83, 124.17 and 249.17. The unit of charge is defined and so are the limits attached to it, covering searches at 50, 200 and unlimited, users at 1, 3 and 7, case tracking at 3, 15 and 50, automated searches, and case export at 1,000, with enterprise custom above that. API billing is itemised rather than gestured at, with billable calls named as Search, Case, Download Document, Order Document Confirm and Update Cases and all other APIs and webhooks stated to be included. Pass-through costs are addressed head on: court fees apply where a court charges and not where it does not, and no fee applies to material already in the CrowdSourced Library. Contract length, cancellation, plan switching, upgrade and downgrade timing and accepted payment methods are all answered, and the absence of a free trial is stated with the reason rather than left blank.

Both halves of this axis are documented, and the boundaries are stated rather than left open. Five buyer segments each get their own treatment on the use-cases page, with the question each brings and the features answering it: litigators on docket tracking, opposing counsel and judge analytics; IP counsel on USPTO, PTAB, TTAB and ITC Section 337 work; law students on real filings and law school programmes; in-house counsel on portfolio litigation risk and agency enforcement; and corporate teams on counterparty and vendor screening. The API page adds four more, and the signup form enumerates nine organisation types including government, school and bar association. Coverage is published court by court on a dedicated page: all 94 federal district courts, all 13 circuits, every bankruptcy court and the Supreme Court, stated as complete PACER coverage; per-practice-area agency tables naming the USPTO, PTAB, ITC, FDA Orange Book, TTAB, US Tax Court, FERC, FTC and NLRB with a coverage status against each; state courts at 43-plus with a court-level table; and international coverage by country. The limits are explicit rather than implied: the state map legend carries a not-yet-covered category alongside full and partial, 43-plus is not 50, and a request-a-court route is published for the gaps. One retrieval limit to record: the interactive state and international tables render client-side and did not return values on 31 Aug 2026, so the structure and legend were read but not the per-state detail.

Segments are documented by role and use case, and coverage is quantified with two real limits stated. Law firms are addressed through four named applications, covering docket research, litigation strategy, docket tracking and experience management, and insurance through litigation risk and strategy and master data management, with the buying question named in each case. Coverage is stated as more than 3,000 state and federal courts across over 40 states with more than a billion dockets and documents, all federal courts including district, bankruptcy and appellate, and a dedicated court coverage page is published for detail. Two boundaries are stated plainly: 40-plus states is not 50, and sealed or restricted documents not publicly available online are explicitly excluded. What is missing is depth within the covered states, since nothing published indicates which state systems are complete and which are partial, and the coverage page itself was not opened on 31 Aug 2026.

No located term or policy states whether customer content is used to train or refine any model, either way. What the terms do carry is a licence pointing in that direction without naming it: the customer grants the vendor an irrevocable, perpetual, transferable, sublicensable, royalty-free and worldwide licence over user content to maintain and provide the services, to improve its products and services, and to create aggregated and de-identified information. The asymmetry is worth a buyer's attention, because the same agreement expressly prohibits the customer from using the services or their content to train any AI models. Read on 31 Aug 2026 alongside the privacy policy, which addresses search history and marketing but not model training.

Searched the home page, pricing page, enterprise API page, PACER API page, general disclaimer, terms of service, public records policy and privacy policy on 31 Aug 2026. No located term or policy addresses whether customer activity, search history or uploaded material is used to train or refine the extraction models, either way. The exposure differs from a document platform because customers retrieve public court records rather than submitting client files, but search patterns and tracked case lists are still customer-specific data and neither is addressed. The terms of service and privacy policy were recovered in fragments through the search index rather than read end to end, so this value is rebuttable on a complete read of either.

Retention is acknowledged in two documents and quantified in neither. The privacy policy states that a record of each search is kept so the user can find it later in their search history, and that search history is also used for aggregated statistics and support. The terms assert a right to monitor and record customer usage of the service, naming content interactions, search history and IP addresses, and to track activity to gather further information about the customer and its authorised users. Searched both documents plus the home page, AI Assistant page and API page on 31 Aug 2026: no retention period is stated for search history, for questions put to the AI assistant or for its answers, no deletion route is offered, and no customer-configurable or zero-retention setting exists.

Searched the surfaces above on 31 Aug 2026 and located no retention period for search history, saved searches, tracked case lists or exported data, and no customer-configurable retention setting. One retention position is published but runs the other way, being about permanence rather than deletion: court documents and dockets retrieved by any user are added to the CrowdSourced Library and made available to all users, with no stated mechanism for removing them. A redaction request route exists in the privacy policy for a court record appearing in online search engines, which is the closest published deletion remedy located. Both legal documents were read in fragments, so this is rebuttable on a full read.

One separation claim is published and nothing describes how it is enforced. The privacy policy states that personal search history is kept strictly confidential to the user's organisation, which is a tenant-level assertion aimed at the right risk, since the cases a firm looks up reveal its docket. No mechanism accompanies it: nothing states whether one user's searches, saved work or tracked cases are visible to colleagues inside a subscribing firm on plans carrying multiple users, and no matter-level concept exists anywhere in the product. Two features cut the other way and are disclosed by the vendor: filings retrieved by any user enter a shared library available to all other users, and the terms reserve a right to monitor and record content interactions and search history that the privacy policy's confidentiality claim does not mention. Checked 31 Aug 2026.

No separation model is published, and the product's core mechanism runs in the opposite direction. The vendor states that any time one user downloads a court document or docket it is added to the CrowdSourced Library and made freely available to all UniCourt users, so retrieved material is pooled across the customer base by design. That material is public court record and sealed or restricted documents are excluded, so this is not disclosure of confidential content, but it means the platform has no document-level separation between customers because it is not built to have any. Nothing published addresses access control within a subscribing firm either, on plans carrying up to seven users, and nothing states whether one user's saved searches or tracked cases are visible to colleagues. Checked 31 Aug 2026.

Disclosure to authorities is contemplated and notice to the customer is not. The terms list compliance with lawful requests from authorities among the grounds on which staff and third-party foundational AI models may access user content, and the confidentiality clause permits release of confidential information where necessary to comply with legal requirements. Searched the terms, the privacy policy, the home page and the footer on 31 Aug 2026 and located no commitment to notify the customer of a subpoena or government demand, no undertaking to seek a waiver where notice is prohibited, and no transparency report. Whether a firm would learn that its search history had been demanded rests entirely with the vendor.

Searched the home page, pricing page, general disclaimer, terms of service, public records policy and privacy policy on 31 Aug 2026. No located material addresses what happens if a third party, law enforcement agency or court requests customer data from UniCourt, and no commitment to notify the customer was found. No transparency report exists. The terms of service and privacy policy were recovered in fragments through the search index rather than read in full, and a compelled-disclosure provision would ordinarily sit in one of them, so this value is rebuttable on a complete read.

Sources are named to the level of the individual court and the rights basis is stated, which together is the strongest provenance position located in this pull. The coverage page enumerates the corpus rather than describing it: all 94 federal district courts, all 13 circuits, every bankruptcy court and the Supreme Court through a stated direct PACER connection, state courts across 43-plus states, and named agencies including the USPTO, PTAB, TTAB, ITC, FDA Orange Book, US Tax Court, FERC, FTC and NLRB, each with a coverage status against it. The rights basis is public record, stated on the removal page and reflected in the terms, which classify content sourced from government publications and public records as publicly sourced material and distinguish it from third-party licensed content. Cadence is stated as real-time updates on new filings. Two limits: the basis is stated for the corpus as a whole rather than per source, and the international courts are listed without a stated basis or supplier.

The strongest provenance disclosure located in this pull, short of naming its suppliers. A dedicated public records policy states that all records in the databases are obtained either from the government or from reputable private suppliers, that UniCourt does not distribute nonpublic information, and that records are provided as is and may contain errors or be incomplete, inaccurate or not the most current available. It sets out the legal basis rather than assuming it, citing endorsement of open access to court records by the Supreme Court, federal courts and state courts, and noting that public records are typically exempt from privacy statutes including the CCPA. An update cadence is published separately, with federal and state data aggregated daily. What holds this short of the top value is that the private suppliers are described by category rather than named, so a buyer cannot trace any particular record to its immediate source.

Searched the home page, AI Assistant page, API page, coverage page and the terms on 31 Aug 2026. No citator, treatment signal or subsequent-history check is claimed, and the question does not bite on this product: Docket Alarm delivers dockets, filings and court documents rather than propositions of law requiring validation, and its AI assistant answers questions about a case record instead of citing authority for a legal proposition. Recorded as not addressed because that is the honest value, with the note that this is a question the product class does not raise rather than one the vendor has neglected. The adjacent statement the vendor does publish is a general one in the terms, that content from courts and other official sources is provided as is with no warranty as to accuracy or completeness.

Searched the home page, pricing page, API pages and the published document set on 31 Aug 2026. No citator, treatment signal or currency check is claimed, and none applies: UniCourt delivers dockets, filings and court documents rather than legal propositions requiring validation, and its analytics concern parties, counsel and case outcomes rather than the standing of authority. Recorded as not addressed because that is the honest value, with the note that this is a question the product class does not raise rather than one the vendor has neglected. The vendor does publish an adjacent and more relevant currency statement, that records may not be the most currently available information.

An explicit no-answer path is described in the vendor's own material, which is uncommon. The AI Assistant page states that the system draws only on the docket being viewed, that it does not fill gaps with plausible-sounding but fabricated information, and that where a fact is not in the record it says so. The design rationale is published alongside it, contrasting general-purpose tools trained to sound confident with a system built so that every answer is traceable to the record. What is not published is any demonstration: no worked example of a refusal, no evaluation, no measurement of how often the behaviour fires or fails, and no confidence or grounding score exposed to the user. Documented rather than demonstrable on that basis, read 31 Aug 2026.

Searched the home page, pricing page, enterprise API page, PACER API page and the published document set on 31 Aug 2026. No confidence score, extraction certainty indicator or exception-handling behaviour is documented for the AI extraction layer, and nothing states what the platform returns when a field cannot be reliably extracted or when a court's records are unavailable. The vendor discloses at the corpus level that records may be incomplete or inaccurate, which is a general limitation rather than a per-output uncertainty signal. For API and customers the absence matters more, because extracted values arrive in their own systems without any accompanying indication of reliability.

No court order, opinion or disciplinary record naming this product has been located as of 31 Aug 2026. Instrument searched: the AI Hallucination Cases database maintained by Damien Charlotin, which tracks decisions worldwide where a court addressed hallucinated AI content and records the tool implicated where known, searched on the product name and on the company name, alongside 2026 sanctions coverage and trade press. This is a statement about the public record on the date shown rather than a clearance, and it is bounded by what that database covers. One appearance is worth distinguishing from a hit: a parent-company blog post uses Docket Alarm's own corpus as the denominator for a hallucination base rate, putting the tracked cases against roughly 40 million US cases filed since January 2023. That is the product as the measuring instrument, not as the tool implicated.

No court order, opinion or disciplinary record naming this product has been located as of 31 Aug 2026. Instrument searched: the AI Hallucination Cases database maintained by Damien Charlotin, which tracks decisions worldwide where a court addressed hallucinated AI content and records the tool implicated where known, searched on the product name alongside 2026 sanctions trackers and trade press summaries. This is a statement about the public record on the date shown rather than a clearance, and it is bounded by what that database covers. The failure mode this signal tracks does not fit the product well: UniCourt delivers actual court documents and docket entries retrievable to source rather than generating citations to authority, so a fabricated citation would have to originate elsewhere.

Searched the home page, AI Assistant page, use-cases page, about page, coverage page, API page, privacy policy and terms on 31 Aug 2026. No bar or ethics authority is engaged with anywhere, including ABA Formal Opinion 512 and any state bar guidance, and nothing maps the product to a professional conduct obligation. The terms do carry a clear no-legal-advice position and an instruction that AI output should receive expert human review, but both are drafted to protect the vendor's position rather than to help a lawyer evidence compliance with their own duties, and neither cites the guidance those duties come from. Recorded as not addressed on that distinction.

A clear professional-responsibility position is published without reference to any bar authority. The general disclaimer states that information on the site should not be construed as legal advice, that others may reach contrary conclusions on the same information, that no attorney-client relationship is created, and that UniCourt offers no opinion or endorsement as to the quality, competency or character of any person or entity on the site. Searched the same surfaces on 31 Aug 2026 and located no reference to ABA Formal Opinion 512, no state bar guidance, and no mapping of the product to any professional conduct obligation. The disclaimer is the vendor protecting its own position rather than helping a lawyer meet theirs, which is why this records as not addressed despite the substantive statement.

Time savings are claimed and the client's side of the equation is not addressed. The AI Assistant page states that a full case brief is generated in seconds against work that takes a traditional paralegal two to four hours, and the use-cases page describes research finishing in far less time. Searched the pricing page, the AI Assistant page, the use-cases page and the terms on 31 Aug 2026 and located no per matter record of AI-assisted work intended for fee purposes and no guidance on billing, fee or client disclosure treatment where an AI answer informs advice. One adjacent strength belongs to the vendor's own charging rather than to this signal: per-document and per-search unit pricing, with court fees passed through at the government's rate without markup, gives a firm a clean disbursement to attribute to a matter.

The vendor's cost material is about its own charges rather than a firm's onward billing, and on that it is unusually clear: court fees are passed through at cost where a court charges and waived where a court is free or the material already sits in the CrowdSourced Library, and PACER costs are explicitly framed as something the platform reduces. Searched the home page, pricing page, API pages and the published document set on 31 Aug 2026 and located no per matter record of AI-assisted work intended for fee purposes and no guidance on billing, fee or client disclosure treatment. Case export exists and could support cost allocation, but nothing presents it for that purpose.

No material supporting a client-side disclosure obligation was located for this product. Searched the home page, AI Assistant page, use-cases page, about page, developer API page, coverage page, privacy policy and terms of service on 31 Aug 2026: no subprocessor list, no model provider named, no data processing agreement, no trust portal, no named certification and no client-facing consent or notification material. A firm asked by a client which vendors see its research activity would have nothing to forward. One thing was located and is described rather than credited, because it belongs to the corporate parent and nothing connects it to this product: the Clio trust centre, publicly reachable without an agreement, names Google Cloud, OpenAI, Anthropic and Amazon Web Services as subprocessors and publishes a data processing agreement alongside them. Docket Alarm is not named on any page of that portal, the tab reached states it covers the Vincent and vLex products, and docketalarm.com references the portal nowhere, so a firm would have to find it unaided and then assume it applies to what it bought.

Searched the home page, pricing page, enterprise API page, PACER API page and the full published document set on 31 Aug 2026, plus two targeted searches for a security or trust page. Nothing that would support a client-side disclosure obligation was located: no subprocessor list, no model provider named, no trust centre or portal, no named certification, no data processing agreement and no client-facing consent or notification material. The published legal set is genuinely open and covers the commercial relationship well, but none of its five documents addresses the third parties in the processing chain, so a firm asked by a client which vendors see its research activity would find nothing to answer with.

The sources behind an answer are recorded and the AI provenance is not. Every statement the assistant produces links to the docket entry or document supporting it, the worked example on the vendor's page displays a Sources list identifying the filing and its date, and the underlying material is the court record itself, so a lawyer can produce the primary documents an answer rests on. What cannot be established is the other half a standing order asks for: no model is identified or versioned anywhere on the property, so which system generated a given answer is unknowable, nothing records who verified it, and no export designed for an AI-use disclosure or certification was located on 31 Aug 2026.

Provenance to source is complete and AI provenance is absent. Every value the platform surfaces sits alongside the underlying docket entry or court document, retrievable and openable, so a lawyer relying on UniCourt data can produce the primary record behind it, and case export is offered for moving that material out of the application. What cannot be established is the AI layer: no model is identified or versioned, nothing distinguishes a machine-extracted field from one taken directly from the court's own record, and no export designed for a court disclosure or AI-use certification was located on 31 Aug 2026. For a customer pushing extracted data into a warehouse, the absence of that distinction is the harder half.