Interpreting and Translation in the Police Setting: Human vs AI

From major incidents involving multiple nationalities to everyday interviews with non-native speakers, the quality of these services directly affects evidence integrity, decision-making, and public trust. Whether human or AI, poor quality interpreting or translation outputs in high-stakes scenarios can introduce vulnerabilities that undermine investigations, delay justice, and erode confidence in the system.

Imagine finding yourself in a foreign country, and having to deal – at some point – with the justice system of that country, in a language you neither speak nor understand, and the only way to communicate and be informed is via an interpreter or translated documents.

This can be considered a leap of faith, as there is no other way to have direct access to that particular system. Therefore, the provision of high-quality interpreting and translation services, both human and/or AI, needs to be prioritised to guarantee 
fair justice.

Focusing in on interpreting services in the legal setting, previous research has already established the active role of interpreters in interpreter-assisted interactions, where they can change the course of the legal proceedings.

Nevertheless, especially in the police setting, mainly due to difficulty of access, it is still not clear to which extent interpreters affect conversational dynamics and mingle with interviewing strategies, the quantity and quality of evidence retrievable, and the delivery of justice as a whole.

Furthermore, the spread of AI across the board has brought about new questions to address and new challenges to face. Regarding multilingual interactions, can AI software be a potential way forward and meet the high standards that are vital in policing?

My PhD research at Lancaster University aims to address these issues by means of linguistic analysis, and specifically taking a Pragmatics and Conversation Analysis (CA) approach.

Preliminary findings from the analysis of two interpreter-assisted police interviews in the language pair English-Italian have shown that interpreters do have an impact on a variety of dimensions, including rapport building, interviewing strategies, the perceived credibility and reliability of the interviewee, and the quantity and quality of the evidence obtained, and additionally, that AI software outputs need human post-editing.

In regard to the accuracy analysis of human interpretation, a data-driven, cross-cultural pragmatic approach was taken, where the intended meaning of the input utterance of a primary speaker in the source language is compared against the intended meaning of the output utterance in the target language as rendered by the interpreter. The focus is on the discrepancies between input and output utterances, and recurring modifications in some pragmatic markers have been identified and classified, such as hedges and enhancers, mood, tense, reported speech, presuppositions, and many others.

For instance, if the interviewer’s original question is “Oh, e di Amanda che ha detto?” (So, what did she say about Amanda?), and the interpreter’s rendition is “Did she say anything about Amanda?”, the presupposition behind the original question undergoes a shift, as the original presupposes that the person speaking indeed said something about Amanda – so the question probes about the specific content of this interaction – while the interpreter doesn’t presuppose this, they just ask whether the person speaking said anything about Amanda. This modifies the interviewing strategy adopted by the interviewer, which is originally more leading.

What are the key differences between monolingual and bilingual interactions in the police setting?

The role of interpreters in interpreter-assisted interactions has been studied extensively in different settings, such as legal, medical care, education, asylum seeking, and many others.

Specifically, in regard to interpreter-assisted police interviews, why is this type of interaction different from a monolingual one? What should we pay special attention to?

Firstly, contrary to the neutrality myth surrounding the role of interpreters, they are active participants in the interaction, influencing conversational dynamics in various ways.

For instance, a primary speaker may directly address the interpreter instead of the other primary speaker, or the interpreter might engage in side-conversations with them, ask clarification questions, interrupt, or clarify points. This is because interpreting entails much more than just relaying the content from one language into another.

Interpreting entails much more than just relaying the content from one language into another.

This suggests that interpretation always involves intervention, and we should be focusing on unnecessary types of intervention, but especially being aware of the inevitable shifts that may occur via the interpretation process, without condemning them.

Secondly, the increased vulnerability of primary speakers should be taken into account. 

If we consider the canonical arrangement of interpreter-assisted police interviews, it typically involves interviewers who are native speakers of the dominant language in a given country, an interviewee who is a non-native speaker of that language, and an interpreter, the bilingual participant.

In a police setting, the conversational power among interviewers and interviewees is usually unbalanced, due to the nature of the interaction, and this imbalance is exacerbated when the interviewee does not speak or understand the dominant language of the interaction. 
Therefore, this augmented vulnerability should definitely be considered when engaging within interpreter-assisted interactions and should result in tailored interviewing strategies.

For instance, it is well-documented that negative polar questions (yes/no questions) pose many issues in multilingual interactions. This is because people from different linguistic and cultural backgrounds may reply with the same answer (affirmative or negative), but intend opposite meanings. For example, with the question “Didn’t they come to the party?”, in English, “Yes” means “They came”, while “No” means “They didn’t come”. In Italian, you might use “Yes” and “No” to say that the person came, maybe specifying “Yes, they came”, or “No, they came”.

It is clear how in an interviewing setting, this may be problematic. Therefore, acknowledging these cross-linguistic discrepancies can help refine interviewing strategies, for instance, by avoiding negative polar questions and instead phrasing them in a more neutral form.

Priority should be given to properly supporting the collaboration between legal and language professionals.

Lastly, unique challenges are faced by legal and language professionals when collaborating, due to their different priorities.

This is why proper training should be ensured to facilitate seamless collaboration, with open communication as one of its key pillars. For instance, briefing and debriefing practices pre and post interview can be considered an example of good practice.

In light of this, priority should be given to properly supporting the collaboration between legal and language professionals so they can deliver high-quality service, and fostering academic research to bridge the gap between academics and practitioners, working towards the shared goal of enhancing current 
policing practices.

Pragmatic Markers
Pragmatic markers refer to every element of language that carries pragmatic force. Pragmatic force encompasses the speaker’s intended meaning and the effect of an utterance on the listener, which goes beyond the literal meaning.

Preliminary findings suggest that human post-editing is still required when it comes to the use of AI translation software.

Is AI a viable alternative in high-stakes scenarios?

Preliminary findings from my research study suggest that human post-editing is still required when it comes to the use of AI translation software.

For instance, to test Google Translate (GT) – which is a Neural Machine Translation (NMT) software, specifically trained on bilingual data – I fed it with the transcripts of the two interpreter-assisted police interviews to draw a comparative analysis between the primary speakers’ inputs, the GT outputs, and the interpreters’ renditions. 

It emerged that in discrepant GT outputs, severe alterations could be identified that could lead to serious consequences for individuals.

In one case, the original input utterance reads “No, like, it could have been me, like the murder could have happened to me.” And GT translated it into Italian as “No, tipo, potrei essere stato io, tipo l'omicidio potrebbe essere successo a me.” (“No, like, I could have done it, like the murder could have happened to me.”). 
It is evident how “it could have been me” and “I could have done it” have different responsibility implications in a police interview setting.

Similar conclusions have been drawn in regard to AI speech-to-text software in legal settings, where its use was recommended to draw initial first drafts only.

Nevertheless, my future research aims to understand how AI translation software can be optimally applied to high-stakes scenarios to speed up the proceedings, but without prejudicing the quality of the outputs.

For policing and justice systems to be truly resilient, they must plan for sustained, high-quality multilingual capability. 

This means investing in skilled human interpreters, integrating AI tools cautiously into linguistic and policing practices, and ensuring robust quality controls for both.

Building resilience also means creating redundancy so when one channel is unavailable or under strain, another can step in without compromising accuracy or fairness. As global mobility, migration, and technology continue to reshape the policing landscape, interpreter-assisted communication should be recognised as a core operational strength that can speed up proceedings, without prejudicing the quality of the outputs.

Download the article

Greta Esposito is a PhD student at Lancaster University in the Department of Linguistics and English Language. Her research interests focus on police interpreting, legal translation, and applications of AI software in legal settings. Her Twitter (X) handle is @GretaEspos1997