Industries / Finance and business services / Legal and professional services / Court reporting
Certification
Testing and licensing the people who make the record, and the rules that make a transcript official once someone certifies it.
7
Milestones achieved, from 1937 to 2026.
2
To watch for: announced, not yet confirmed outside.
2
Open deliverables. Nobody has shipped them.
The road
AchievedWatch forOpen
Achieved
1937
The first skills test
The National Shorthand Reporters Association, founded in 1899 and now the National Court Reporters Association (NCRA), gave its first Certificate of Proficiency test in 1937 at five places around the US, and 27 members passed. The test became the Registered Professional Reporter (RPR) designation on 1 July 1975, and Colorado's courts now require new official reporters to hold the RPR.
LessonA national system failed first: a federal bill to certify shorthand reporters passed both houses of Congress unanimously in 1936, and President Roosevelt vetoed it.
Credit
NSRA Committee on Certificates of Proficiency
1944
The reporter's certificate
Under the US Court Reporter Act, a federal transcript "certified by the reporter" is "deemed prima facie a correct statement of the testimony taken and proceedings had", which means a court accepts it as correct unless someone shows otherwise. At a deposition, the officer "must certify in writing that the witness was duly sworn and that the deposition accurately records the witness's testimony." In Ontario the approval belongs to a person instead: appeal transcripts must be "prepared and certified by an Authorized Court Transcriptionist."
Credit
US Congress; Advisory Committee on Civil Rules; Government of Ontario
1951
A state licensing board
California's Legislature set up a board in 1951 to license certified shorthand reporters (CSRs), and by its own count it has licensed 13,454 people since.
Credit
California Legislature
1967
Tests for voice and recording
The NVRA, founded in 1967, certifies voice writers as Certified Verbatim Reporters at the RPR's three speeds: 180 words per minute literary, 200 jury charge and 225 two-voice question and answer. The American Association of Electronic Reporters and Transcribers (AAERT) has certified people who make and transcribe recordings since 1995, and its transcriber test requires 95% accuracy. In 2022 California's Assembly Bill 156 required the state board to license voice writers as certified reporters.
Credit
NVRA; AAERT; California Legislature
2019
Certified, but not for everyone
Taylor Jones, Jessica Kalbfeld, Ryan Hancock and Robin Clark tested 27 Philadelphia court reporters, all certified at 95% accuracy, on recordings of African American English. On average the reporters wrote down 82.9% of the words correctly, and "Thirty-one percent of the transcriptions inaccurately altered the record."
LessonJones notes that the certification "is based primarily on the speech of lawyers and judges".
Credit
Jones, Kalbfeld, Hancock and Clark, in the journal Language
2026
A licence for any method
Georgia's HB 1208, in force from 1 July 2026, repealed the rule that depositions be "recorded by stenographic means" and requires a certified court reporter instead. A Supreme Court of Georgia rule in force from 26 February 2026 says a transcript of a digital recording "shall not be created by the digital monitor unless the digital monitor is also a certified court reporter." In California a bill to certify digital reporters failed in February 2026.
legis.ga.govassets.georgiacourts.govocp.georgiacourts.govcalmatters.digitaldemocracy.org
Credit
Georgia General Assembly (Rep. Stan Gunter); Supreme Court of Georgia
2026
AI errors in certified transcripts
In Williams v. State, the Indiana Court of Appeals wrote that "it appears that generative artificial intelligence may have assisted with the preparation of this transcript", although Indiana's appeal rules say the court reporter "shall certify the Transcript is correct."
LessonA reporter's certificate says that a transcript is correct, not how it was made.
Credit
Indiana Court of Appeals, reported by Eugene Volokh in Reason, July 2026
Watch for
AI deposition transcripts in Texas
The Supreme Court of Texas heard argument on 6 October 2026 in In re Hughey (No. 25-0463) over a deposition transcript made with Skribe.ai, a platform that records depositions over Zoom and makes transcripts with speech-to-text "reviewed by a human transcriptionist", rather than by a certified reporter.
Achieved whenthe court rules that such transcripts may be used, and they are then used in Texas depositions.
Claimed by
Skribe.ai, whose transcript is at issue; argued October 2026
Oklahoma's speech-to-text pilot
The Oklahoma Supreme Court approved a three-year pilot of speech-to-text in district courts, in which "Only a certified court reporter may review, prepare, and certify a transcript of a proceeding" and pilot data "shall not be used to train, develop, or improve any artificial intelligence or machine learning model". Two justices dissented.
Achieved whenan evaluation of the pilot is published with measured accuracy and turnaround.
Claimed by
Supreme Court of Oklahoma, October 2026
Open
01
Scoring AI on reporter tests
A public test of speech-to-text models on dictation made to RPR conditions (180 words per minute literary, 200 jury charge, 225 question and answer, 95% accuracy on each leg), reporting which models pass each leg as raw output and after a fixed time of editing by a person.
The briefCredit · unclaimed02
A test for every dialect
An openly licensed test leg of everyday speech from speakers of African American English and other varieties, recorded with consent, and scored for human reporters and speech-to-text alike.
The briefCredit · unclaimed
Open 01 · the brief
Scoring AI on reporter tests
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A public test of speech-to-text models on dictation made to RPR conditions (180 words per minute literary, 200 jury charge, 225 question and answer, 95% accuracy on each leg), reporting which models pass each leg as raw output and after a fixed time of editing by a person.
Why it is still open
I found no claim that any AI has passed the RPR or a state licensing exam. The 2026 study for the Texas courts says "Record accuracy is not defined in statute" and "There is not a comparable standard or certification process for digital modalities in Texas." The nearest work, a 2024 benchmark of speech recognition on Quebec French legal hearings, reports word error rates, not test results.
Built with what exists
the open dictation set described under Stenography; ElevenLabs Scribe v3 (see the Note "Fewer word errors"), Deepgram Nova-3 and open-weight models such as Whisper large-v3; a script that counts wrong, missing and added words for each leg.
First step
Score two models on one 225 words-per-minute question-and-answer take.
Done looks like
A public table of model, leg and take, with the accuracy and the pass or fail result at 95%, that a second person reproduces from the published code.
Open 02 · the brief
A test for every dialect
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
An openly licensed test leg of everyday speech from speakers of African American English and other varieties, recorded with consent, and scored for human reporters and speech-to-text alike.
Why it is still open
The 2019 Philadelphia study found that certified reporters wrote down 82.9% of words correctly in such speech, and that the certification is based mainly on the speech of lawyers and judges. I found no sign that certification tests changed after the study.
Built with what exists
volunteer speakers who sign a consent form; steno students on Plover and voice writers; speech-to-text models; the scoring script from the test above.
First step
Draft the consent form and record one speaker.
Done looks like
At least 80 sentences from at least 6 speakers released under an open licence, with word and sentence scores published for at least 10 human writers and at least 3 software systems.