J

JusticeText

The models are the engine of the core capability, on an evidence platform with substantial functions that do not use AI. Speech recognition produces the searchable transcripts, synced to the media, in more than 100 languages that the product leads with, and the MirandaAI assistant, summaries, key moments and PDF analysis are model work. Around them sits conventional evidence tooling that works without models: conversion of proprietary formats such as GTL jail calls and WatchGuard video, syncing of several camera angles by capture time, and clipping, redaction and captioning for courtroom exhibits.

Output is grounded in the source media, and a failure condition is named with a visible behavior, short of any measurement. Transcripts are synced in time to the original audio and video, so any passage can be checked against the recording. The product FAQ names the conditions under which transcription degrades, overlapping voices, static and background noise, and says the transcript clearly identifies areas of lower confidence so reviewers focus there.

No accuracy figure is published, nothing describes how MirandaAI answers are grounded or cited, and the terms disclaim any warranty of transcript accuracy.

A named output carries a categorical limit on use, with a review surface built in. The Terms of Service state that machine generated transcriptions are tools to aid review, not authoritative records of events, and that every transcription must be checked by the user. Users agree not to rely directly on system transcriptions as evidence in a court, arbitration, mediation or other formal proceeding or settlement negotiation.

The product supports that check by flagging passages with low confidence and syncing text to the original media, and exhibits are clipped from the original video rather than the transcript. The limit is specific to transcriptions: MirandaAI's timelines and inconsistency findings carry no equivalent stated limit.

Named deployments with figures, short of any method. The vendor publishes case studies and rollout stories naming the Kentucky Department of Public Advocacy, the Santa Cruz County, Lane County, Travis County, Lehigh County and Siskiyou County defenders, Kansas Holistic Defenders, and private firms including Rossen Law and Nicholson Goetz & Otis. It quotes chief defenders in Harris County and Montana by name. Rossen Law says the product cut video review and trial preparation time by at least two thirds. No figure states how it was measured, over what caseload or period.

Confidentiality is asserted in general terms, while the agreement permits model tuning from customer content. The product page says data remains confidential, the customer controls access and deletion, content is encrypted on U.S. servers, and evidence is never used to train third party machine learning models. The Terms of Service contain no confidentiality clause protecting customer content. The content license is limited to providing the service, but a separate Machine Learning clause authorizes calibrating data models and improving algorithms while processing user content.

The privacy policy permits sharing deidentified aggregate data with research partners and funders. Nothing addresses privilege or defense work product, which is what public defenders' notes, tags and clips are.

A clear position on what the output is and is not, with a defined audience, short of supervision duties. The terms state that transcriptions are review aids and not authoritative records, must be checked, and must not be relied on directly as evidence, and both the terms and privacy policy say JusticeText cannot answer requests for legal advice. The audience is attorneys, and the privacy policy collects each user's law licensing state.

The terms disclaim any representation that the service is appropriate or lawful in a given location, including under court orders. Nothing addresses an attorney's competence or supervision duties when relying on MirandaAI analysis.

No governance position is published on the vendor's own site. The product's Built for responsible review section covers data security only, and no responsible AI policy, testing regime or accountable owner is published. A submission by the vendor hosted by MIT Solve acknowledges that speech recognition engines are less accurate on African American Vernacular English. That is a disclosure of uneven output, but it is not on the vendor's own site and is not dated.

A substantive published position covers most of the ground, short of a subprocessor list and a concrete incident commitment. The privacy policy states that information is stored encrypted on AWS servers in the United States and encrypted in transit, gives users a right to deletion completed in most cases within 30 days, and names categories of providers (hosting, transcoding, storage, databases, crash monitoring, payments) with AWS and Stripe named.

The product page adds multi factor authentication and single sign on and says customers control access and deletion. Breach notification is promised only as required by law, the terms reserve retention and destruction under internal policies that are not published, and no full subprocessor list, including any AI or speech recognition provider, is given.

Liability is addressed only through limitations that disclaim the exposure the product creates. The terms provide the system as is, disclaim any warranty of transcript accuracy, disclaim liability for any legal action or court decision resulting from transcripts or machine learning output, cap liability at the amount paid or $100, and require users to indemnify JusticeText. No vendor indemnity or insurance position is published.

Named connections exist, short of documented depth. The product page shows imports from Google Drive, Dropbox, OneDrive, Box and the Clio practice management system, and the privacy policy describes the cloud storage integrations and Google API use. The platform also takes in the proprietary formats defense discovery arrives in, including GTL jail calls, For The Record courtroom recordings, AV Viewer and WatchGuard. What each connection moves, and whether results flow back to Clio or case management, is not documented.

The hosting model and region are stated, with no options and no tenancy detail. The privacy policy states that information is stored in encrypted form on servers in the United States operated by Amazon Web Services, and the product page repeats that content is stored on encrypted U.S. servers. No tenancy model, deployment option or separate processing location for speech recognition or AI providers is stated.

No independent security attestation located. The vendor describes encryption, multi factor authentication and single sign on but names no audit or standard, which matters for a product holding criminal discovery.

Third party models are acknowledged without being identified. The product page says evidence is never used to train third party machine learning models, and the terms say third party tools execute some functionality and may access user content. No speech recognition engine, language model or provider is named on any page read, and no change notification is described.

The pricing structure is described without figures. The Terms of Service describe monthly and annual subscriptions that renew automatically, usage limits on processing time that can block further processing until the next term, free initial access for a period or amount of usage, and no refunds. No price, rate or plan is published; access starts through an interest form.

The segment and evidence types are described with substance, short of the boundaries. The vendor serves public defender offices and private criminal defense attorneys, and covers body camera footage, interrogations, jail calls, dashcam, courtroom recordings and written discovery, with transcription and translation in more than 100 languages and support for named proprietary formats. It does not state which jurisdictions it serves, which evidence types it handles poorly beyond noisy audio, or whether it suits civil or prosecution use.