I recently had the pleasure of moderating a GALA Interpreting SIG panel on “Choosing the Right AI Interpreting Solution: Challenges and Approaches”.
The conversation brought together four perspectives: technology development, research, language services, and the practical deployment of interpreting solutions within a large international organization. Rather than focusing on whether AI interpreting is simply “good” or “bad,” the discussion explored a much more useful question: When is it the right solution, under which conditions, and for whom?
Several points stayed with me in particular.
1. The right choice depends on the importance and purpose of the meeting
Adam Youngfield, Group Product Manager for AI Advancement in the Publishing Services Department of The Church of Jesus Christ of Latter-day Saints, described a clear distinction between different types of meetings within his organization.
For the major general meetings, where decisions are made and the direction of the congregation for the following six months is established, they rely exclusively on human interpreters. These are high-stakes situations in which accuracy, trust, responsibility, and the ability to manage complex communication are essential.
In many other meetings, however, AI interpreting can help people participate who might otherwise remain excluded.
This distinction is important. The question is not whether AI can replace human interpreters across the board. The more meaningful question is whether a particular form of interpreting is appropriate for a particular communicative situation.
The answer may depend on the stakes, the audience, the consequences of errors, the available resources, and the alternative. In some cases, the alternative to AI interpreting is not human interpreting. It is no interpreting at all.
2. Quality is not only the quality of the translation
Andreea Danielescu, Director of Research and Development at Wordly, stressed the importance of usability.
When evaluating an AI interpreting application, it is not enough to examine the linguistic output in isolation. We must also consider how people actually interact with the system.
Can participants access it easily? Is the interface understandable? Can they select the right language without assistance? Does the system integrate naturally into the meeting? Can people hear the translation clearly? What happens when the network becomes unstable or the audio quality deteriorates?
This leads directly to another essential concept: robustness in real-world conditions.
A system may perform very well in a controlled demonstration and still fail in practice because of poor microphones, background noise, weak connectivity, unfamiliar accents, overlapping speech, or users who do not know how to operate the application.
The success of AI interpreting therefore depends on much more than the underlying translation model. It depends on the entire experience surrounding it.
3. Clients need guidance, not just technology
Peter Argondizzo, CEO of Argo Translation, emphasized the need to educate clients about when AI interpreting is a suitable option and when it is not.
This is becoming a central responsibility for language service providers and technology consultants. Clients may be attracted by cost, scalability, or immediate availability, but they often lack the information needed to assess risk and suitability.
The role of the provider is therefore not simply to sell a human or artificial solution. It is to understand the communicative setting and help the client make an informed decision.
Peter shared a particularly powerful example. After one AI-supported meeting, participants were crying because, for the first time, they had been able to understand and take part in the conversation.
At that point, he said, you know that AI was a good fit.
This example captures something that is often missing from abstract discussions about quality. A solution should also be judged against the situation that existed before it. Even an imperfect system may create enormous value when it opens access to people who previously had none.
4. Vendor quality claims are only the beginning
Tomasz Korybski, Assistant Professor at the Institute of Applied Linguistics at the University of Warsaw, warned against relying solely on vendors’ declarations about accuracy or quality.
A headline number tells us very little unless we know how it was obtained.
In which languages? With what type of speakers? Under what acoustic conditions? On prepared speeches or spontaneous conversation? With which accents, domains, microphones, and evaluation criteria? Were names and specialized terminology included? Were the results measured in a laboratory or during real meetings?
An accuracy figure without context can easily create false confidence.
Anyone selecting an AI interpreting solution should therefore ask not only, “How well does the system perform?” but also, “In which particular cases were those results achieved, and how closely do those conditions resemble our own?”
Independent testing in the intended use case remains essential.
From technological comparison to contextual evaluation
My main takeaway from the panel is that choosing an AI interpreting solution cannot be reduced to comparing models or accuracy scores.
It requires a broader assessment of:
- the importance and risk level of the meeting;
- the needs and expectations of participants;
- usability and accessibility;
- robustness under real-world conditions;
- the availability of human interpreters;
- and the consequences of providing no language access at all.
The future of multilingual communication will not be defined by a simple choice between humans and AI. It will involve a growing range of human, artificial, and hybrid solutions.
The real challenge is learning how to select and combine them responsibly.
That requires realistic evaluation, honest communication with clients, and—above all—a clear understanding of the people and situations the technology is meant to serve.