LO
DeploymentCognitive automation2019-03-01
The US Department of Justice reports machine translation deployed department-wide since March 2019 for an attorney's initial review, and states that its output does not constitute an official record
Translator / Interpreteroccupation page →
Event date / reported
2019-03-01 · reported 2026-04-14
Evidence stage
DeploymentAn employer has put it into production. Can move the baseline — weighted by scale and how similar the setting is.
Tasks this bears on
Translation where being wrong is expensive
Contracts, regulatory filings, clinical material, safety documentation.
Being augmented✓ Evidence-backed
Where this applies
The entry is the department's own submission to the statutory inventory OMB publishes under the Advancing American AI Act, so the publisher is the employer describing a deployment it runs. The department draws the line itself: the machine output is a tool for initial review by an attorney and, in its own words, does not constitute an official record. The entry gives its operational date only as a month, 03/2019, so it is recorded as the first of that month; the reported date is the April 2026 data upload, which the repository dates as of 13 April 2026. Read across the whole file, the same product appears inside this one department at four bureaus deployed (department-wide from 03/2019, the antitrust division from 01/2023, the criminal division and the public affairs office from 01/2024), one bureau not yet deployed, and one retired. A second agency, the Federal Trade Commission, records the same product deployed since 2022 and gives its reason as manual translation and transcription being slow and of inconsistent accuracy. What the entry does not give is any headcount, any volume, or any measure of how review time changed. It also says nothing about certified translation: it describes an initial read, not an official one, so it does not bear on who signs a certified document.
What this means
An employer in a legal setting runs machine translation as the first pass and writes down, itself, that the output is not the record. For a translator that turns "the machine drafts and a person finalises" from a way of describing the work into a dated entry with a department, a reason and a boundary attached. Note when it starts: March 2019, well before the current generation of models.
What it does not yet show
It says nothing about headcount and nothing about how much review time changed — the inventory records neither. More to the point for this task, it does not reach certified translation at all: it describes an initial read rather than an official text, so who signs a certified document is untouched by it. Inside the same department one bureau has the product and has not deployed it, and one has retired it.
What you can check
Find the rule that governs certified or official text in the field you work in — a court, a notary, a regulator, a clinical trial sponsor — and read whether it requires a named person to sign. That rule, not the quality of the machine, is what sets where you stand in the process.
Does it change the assessment?
No. The impact index is never moved by a single event. What this record did: the 1 linked task judgement above now rest on evidence instead of inference.
Source
OMB — 2025 Federal Agency AI Use Case Inventory (entry DOJ-0244) · verified 2026-09-21 · Claude (VOLO agent) — the 3,611-row individually-reported CSV parsed locally; entry DOJ-0244's own problem_solved and benefits read verbatim, including the sentence 'This machine translation does not constitute an official record, but is a tool to allow initial review of the audio by the attorney'; the four-deployed / one-pending / one-retired spread computed by grouping every Veritone row in the file by bureau and stage, and the FTC's stated reason read from entry FTC-0009 · interpreted 2026-09-21 · Claude (VOLO agent)
Primary source — published by the party that did this, or the authority of record. No co-signature needed.