AI-Assisted Transcription of Interrogations in Criminal Proceedings – Opportunities, Risks and Defence Strategies

A recent finding from extensive white-collar criminal proceedings shows how routinely artificial intelligence is now being used in investigative work – and how error-prone this is:

In these proceedings, numerous witness interviews were no longer written down in the traditional way by typists, but were generated from the digital recordings using AI-based software of a State Criminal Police Office. At the end of each interview record there is a brief note on the use of AI and a confirmation that a police officer checked the machine transcription and confirmed the accuracy of the transcript.

1. Numerous typical AI errors

On detailed review of these records, however, the defence identified a large number of obvious errors that called into question the accuracy and integrity of the records. These anomalies revealed a broad spectrum of typical AI errors – ranging from attribution problems and distortions of meaning to incompletely transcribed words.

Errors in assigning “speakers”

The records listed “Speakers” (speaker numbers) without all persons present being recorded or correctly assigned.

• In some cases, an investigating officer who was present was mentioned by name in the record but was not assigned a “Speaker” number of their own.
• In individual interviews, a “Speaker” appeared who could not be attributed to any real person in the room.

Questions and answers mixed up

A particularly serious error: questions and answers were attributed to the wrong person:

• For example, a typical opening question from the interviewing officer (“Please explain your professional background to me…”) and the corresponding detailed answer were both attributed in full to the same “Speaker”, namely the officer himself – as if he were questioning himself and answering at the same time.
• In other passages, a statement clearly recognisable as a question from the investigator appears as testimony given by the witness.
• In some cases, the records produce absurd “monologues” in which the investigator appears to provide both question and answer.

Unclear or missing timestamps

Many records included timestamps from the audio recording, while elsewhere this information was missing entirely for no apparent reason. This makes it more difficult to trace the actual course of the interview and to compare audio and text precisely.

Obviously incorrect or distorted words

The records contained a whole series of terms and phrases that obviously cannot correspond to what was actually said, such as unusual or meaningless word combinations. Such wording clearly shows that the AI inserted “supposedly suitable” words on the basis of probabilities which, however, miss the technical context.

Incorrectly recognised or distorted names

Proper names, company names and technical terms were also reproduced incorrectly to a considerable extent.

Completely distorted sentences

In some cases, sentences were reproduced whose meaning can hardly be understood any more. Such passages do not reveal what the witness actually said or whether it is a question from the officer or an answer from the witness.

Obvious gaps

Elsewhere, a passage is marked with “????”, which indicates that parts of the interview could not be transcribed or attributed. These gaps underline that the transcription is not only flawed but in parts incomplete – and at the same time suggest that a check for accuracy and correction was not carried out in practice, or not with the necessary care.

2. AI “hallucinations” and the investigating authorities’ duty to verify

The anomalies in the records clearly point to so-called “hallucinations” typical of AI applications. The AI works with probabilities and language models and supplements or replaces unclear or incomplete input with supposedly suitable content. Particularly in the case of spontaneous, fragmentary statements in interviews, this can result in the actual content of the statement being distorted.

Under the German Code of Criminal Procedure, the following applies: if the record is produced subsequently by verbatim transcription of a recording, the person who produced the transcription or checked a machine transcription must confirm its accuracy and thereby assumes responsibility for its accuracy. In the present case, this duty was evidently not fulfilled, as numerous obvious errors remained unchecked.

In addition, it remains unclear which specific AI processing steps were carried out: the general reference to software using “artificial intelligence” leaves open whether merely a speech recognition program was used or whether more extensive AI functions were employed to reconstruct what was said. The anomalies suggest that the AI independently determined “likely” words on the basis of language models and context and may thereby have replaced what was actually said with an alternative version generated by the AI.

This is, however, highly problematic, above all in the context of complex white-collar criminal proceedings, in which it is of particular importance to whom a given statement – and thus a certain level of knowledge – is to be attributed.

3. Legal consequences of the errors: evidential value, admissibility and duties of investigation

Errors in the transcription of interviews of accused persons or witnesses are, however, subject to judicial review and therefore also offer starting points for the defence.

Right to inspect the original audio recordings

First of all, the defence must have an opportunity to check the accuracy of the records itself. It can therefore assert a right to inspect the original audio recordings of the interviews. This right is based on defence counsel’s comprehensive right to inspect the files, which expressly also extends to audio recordings, which are to be treated as part of the investigation files. Only access to the original recordings makes it possible to prove that the records are inaccurate, a possibility expressly provided for by the Code of Criminal Procedure in the case of recorded interviews (cf. § 168a (2) sentence 3 StPO, where applicable in conjunction with § 168b (2) sentence 1 StPO).

Duty to re-examine and correct the records

It is not, however, the task of the defence to ensure that interview records are transcribed accurately. If such obvious errors are found in records, the defence can demand a renewed, qualified review of the AI transcriptions by the investigating authorities.

Under the statutory provisions, a machine transcription must be checked and its accuracy expressly confirmed (cf. § 168a (5) StPO); this entails responsibility for ensuring that the content of the recording and the record match. Particularly where AI-assisted transcription is used, whose typical sources of error are widely known, a stricter standard for checking the results may be required.

The result of the renewed review and any corrections should be documented and made available to the defence as part of supplementary inspection of the files.

Objection to admissibility and limited evidential value

As a precaution, the defence should lodge an objection to the use of the flawed records as evidence. A record that is obviously inaccurate or distorts the meaning on essential points has, to that extent, no or only limited evidential value.

In such a case, the court must clarify the reliability of the record by way of informal proof (Freibeweis). If doubts cannot be resolved, the record loses its evidential value insofar as the passages in question would operate to the detriment of the accused, because ambiguities caused by breaches of documentation duties must not be held against the accused.

Data protection transparency and information obligations

Finally, data protection aspects must also be taken into account: this applies in particular if witnesses are not sufficiently informed that their statements are being processed and transcribed using artificial intelligence. As a rule, the records do not show that the witnesses consented to such use of AI. This could constitute a breach of the transparency requirements and information obligations under the General Data Protection Regulation, in particular the requirement that data processing be comprehensible.

4. Conclusion: AI in criminal proceedings – evidential advantage or risk for the defence?

These errors in the use of transcription AI illustrate the defence-relevant questions that arise when AI-assisted transcription is used in criminal proceedings:

• The combination of technical AI “hallucinations” and insufficient human oversight can lead to serious errors in interview records – up to and including falsified or entirely fabricated statements.

• If such records are used without being checked, there is a risk of the accused being impermissibly incriminated on the basis of unreliable evidence.

• The defence and the courts are called upon to pay particular attention to the original recordings and the documentation of the AI processing and to critically scrutinise the evidential value of AI-generated records.

In practice, this means that AI can support the criminal justice system, but it does not replace the duty of careful human oversight. It is the task of the law enforcement authorities to ensure that the accuracy of AI-generated work results is checked.

Where this oversight is lacking or merely perfunctory, this opens up substantial points of attack for the defence – up to and including the (partial) loss of the evidential value of such records.

Our expertise at the interface of criminal law, compliance and AI

The case also illustrates how closely modern criminal defence, corporate defence and compliance are now interlinked. Pragal Rechtsanwälte combines many years of experience in complex white-collar and tax criminal proceedings with sound advice on compliance issues – including the legal requirements for the use of artificial intelligence. Particularly in the field of AI compliance, the firm supports companies in deploying technical innovations in a legally compliant manner, identifying risks at an early stage and establishing robust control and documentation processes.

Contact

FAQ: Frequently asked questions on AI-assisted transcription of interviews in criminal proceedings