Who Is Doing the Thinking When We Use AI?
Reflecting on Tom Mould’s session about artificial intelligence, professional judgement and safe use
An interpreter receives presentation slides for an unfamiliar technical assignment.
They upload them to an AI tool and ask for a summary, a list of key terms and a simple explanation of the main subject. Within seconds, the system produces something clear, detailed and confident.
The interpreter feels much better prepared.
But what information did they upload? Did they have permission to share it with that platform? Has the AI accurately represented the source material, or added something that only sounds plausible? If the answer contains an error, will the interpreter recognise it?
Artificial intelligence can make professional tasks quicker and easier, but speed and confidence are not the same as accuracy or safety.
During his extended conference session, Artificial Intelligence and Interpreter, Tom Mould offered interpreters a practical introduction to AI, including what current tools can do, why they can appear more intelligent than they are and where professional judgement must remain firmly with the human user.
Tom did not present AI as something interpreters should either embrace without question or avoid completely. His central message was more useful: explore it, understand it and remain careful about what you ask it to do.
AI is not one single thing
The term artificial intelligence is now attached to a wide range of products.
Phones, cars, televisions, washing machines, search engines and office software may all be described as using AI. However, the technology within these products does not necessarily work in the same way or carry the same risks.
Tom distinguished between several concepts, including narrow AI, large language models, generative AI, agentic AI, artificial general intelligence and artificial superintelligence.
Narrow AI usually performs a particular task. A recommendation system may notice patterns in what someone watches and suggest another programme. A smart thermostat may learn when people are usually at home.
This may appear intelligent, but it is not thinking in the human sense.
Large language models, including systems such as ChatGPT, Gemini, Claude and Copilot, work with patterns in language. They produce responses based on what is statistically likely to follow the prompt they have received.
Generative AI creates content such as text, images, audio or video. Agentic AI refers to systems that can be instructed to carry out a series of tasks on someone’s behalf.
Artificial general intelligence would involve human-level intelligence across a broad range of activities. Tom was clear that this is different from the large language models people commonly use today.
Understanding the difference matters because the word “AI” can encourage us to imagine that a system knows, understands and decides far more than it actually does.
A language model produces language
One of Tom’s most important explanations was also one of the simplest.
A large language model is a model of language.
It has been trained on enormous amounts of text and identifies patterns in how words, phrases and ideas commonly appear together. When someone enters a prompt, the model produces a response that resembles the kind of language likely to answer it.
This can create a strong impression of understanding. The system may explain a complex subject, rewrite an email or respond sympathetically to a personal question.
However, it does not understand in the same way as a human reader, interpreter or subject specialist.
Tom compared the process to a highly advanced form of predictive text. The output can be impressive because the model is extremely good at producing natural-sounding language, but it is still generating a likely response rather than independently knowing that the response is true.
This distinction becomes particularly important when interpreters use AI for professional preparation.
A well-written answer can feel trustworthy. Clear formatting, technical vocabulary and a confident tone may make the information appear more reliable than it is.
The interpreter needs to remember that convincing language is one of the system’s main strengths. It is not proof that the content has been checked or understood.
A copilot, not the pilot
Tom suggested that Microsoft’s name, Copilot, offers a useful way to think about the relationship.
An AI tool can sit beside the professional. It may help organise information, suggest questions or provide a starting point. It should not take control of the work.
The interpreter remains responsible for deciding what information can be shared, checking the answer and determining whether it is useful.
This is especially important because AI can gradually become easy to rely upon. Someone begins by asking for help rewriting one email, then uses it for research, preparation, reflection and increasingly complex decisions.
Tom described catching himself treating AI like a physiotherapist while seeking advice about an elbow problem. The system appeared helpful, but a trained professional who could examine him remained a much more appropriate source.
The example showed how quickly the boundary can move. A tool that begins as a convenient assistant can start being treated as an authority because it responds quickly and confidently.
The professional question is not simply, “Can AI give me an answer?”
It is, “Should this be answered by AI, and what will I do to check it?”
Giving the tool enough context
Tom explained that using a large language model is different from carrying out a traditional internet search.
People have become used to reducing questions into a few keywords for a search engine. AI prompts can contain much more context.
A user can explain who they are, what they are trying to achieve, who the audience is, which format they need and what the response should avoid.
The more relevant information the system receives, the more closely its output may match the task.
An interpreter preparing for an assignment might ask for a simple explanation of an unfamiliar process, examples of specialist vocabulary or questions they should consider while researching the subject.
Someone reflecting on CPD could provide their own notes and ask the tool to identify themes or pose questions that encourage further thought.
The conversation can also continue. The user can ask for a clearer explanation, challenge part of the response or request a different structure without beginning again.
This makes AI useful as a sounding board.
However, providing more context can also increase the privacy risk when the information concerns a client, employer, speaker or assignment. A more detailed prompt may produce a better answer while revealing far more than the interpreter was entitled to share.
Good prompting and safe prompting are not always the same thing.
Where might AI help interpreters?
Tom discussed several practical uses for AI within interpreting work.
It may help someone understand the broad context of an unfamiliar subject before beginning more detailed research. It can explain terminology in plain English, suggest related concepts or create a basic glossary for the interpreter to verify.
It can also support administrative and reflective tasks. An interpreter might use it to improve the structure of an email, organise notes, create practice activities or generate questions for a CPD reflection.
For translators, AI may help compare possible wording, reorganise information or identify areas within a source text that require further research.
These uses can reduce the time spent facing a blank page. They give the professional something to question, edit and develop.
That last part is essential.
AI should not remove the interpreter’s own preparation process. The act of researching, comparing sources, identifying ambiguity and making connections is part of becoming ready for an assignment.
A glossary produced in seconds may contain useful terms, but it does not automatically create the deeper subject knowledge needed to recognise them within fast or complex discourse.
Preparation is not only about obtaining information. It is about building enough understanding to work with it.
When the confident answer is completely wrong
Tom gave a direct warning about hallucinations.
A hallucination occurs when an AI system produces information that is false or invented. The answer may be presented with the same confidence and polished language as an accurate response.
Tom shared an example in which an AI tool recommended a hospital that did not exist. The name sounded believable and fitted the kind of answer the model was expected to provide, but it had been fabricated.
This is not limited to casual questions. Tom referred to professionals who had used AI-generated legal references without checking them and later discovered that the cases did not exist.
The risk is especially serious when the user is asking about a subject they do not already understand. Someone with specialist knowledge may notice that a term or explanation looks wrong. A person using AI because the topic is unfamiliar may have no immediate reason to doubt it.
Summaries carry the same problem. An AI-generated summary may omit an important qualification, combine separate points or insert an idea that was not present in the original material.
The user should therefore return to the source.
A citation provided by AI is not enough. The source needs to be opened, checked and read in context.
Tom’s message was clear: the human remains responsible for verification.
Confidentiality does not disappear because a tool is convenient
Interpreters and translators regularly receive information that is private, sensitive, commercially valuable or protected by confidentiality agreements.
Presentation slides may contain internal company plans. An email may include personal details. A medical document may identify a patient. A script may be protected intellectual property.
Uploading this material into an AI platform means sharing it with another system.
Tom warned that users should not assume information entered into an AI tool will remain private. Platforms have different settings, policies and ways of retaining or using data, and these may change.
He gave the example of sending a personal statement to a family member for proofreading. She uploaded it to an AI tool and returned an improved version, but his private information had been shared without him making that decision himself.
The same issue can arise when an interpreter helps a Deaf client with written English. It may feel efficient to paste the message into AI and ask for a clearer version.
However, the interpreter must first consider whether they have permission to share the content and whether identifying details can be removed. They must also check that the system has understood the original message rather than silently changing its meaning.
Removing a person’s name may not be enough when the remaining details still make them identifiable.
Before using AI with professional material, practitioners need to ask:
Does this contain personal, confidential or identifying information?
Do I have permission to upload it?
Can the task be completed without sharing the original content?
Can I create a fictional or anonymised example instead?
What do the platform’s settings and privacy policy say?
Has my organisation approved this tool?
Would I be comfortable explaining this use to the person whose information it is?
Convenience does not remove professional responsibility.
Rewriting can also change the message
AI is often used to make writing sound more professional, formal or grammatically standard.
This may appear to be a low-risk task, but language carries identity, intention and power.
A Deaf person may ask an interpreter or translator for support with an English message. An AI tool could correct the grammar and produce something that reads smoothly.
It could also remove the person’s tone, strengthen a complaint beyond what they intended or add polite wording that weakens an important point.
Because the rewritten version sounds polished, changes in meaning may be difficult to notice.
The professional should not assume that standard English is automatically a more accurate representation of the person.
AI-generated wording remains a draft. The author needs to understand it and agree that it expresses what they wanted to say.
For translators, the same principle applies when AI suggests alternative wording or reorganises a source text. The finished decision remains a translation decision, not a software decision.
Sign language technology is not only a technical problem
Tom widened the discussion from interpreters using AI to the development of AI-generated and AI-recognised sign language.
He drew attention to the report BSL Is Not for Sale, particularly its message that Deaf people must be involved throughout sign language technology.
Sign language AI should not be treated as a technical product developed first and shown to Deaf communities later.
Deaf people need to be involved in leadership, governance, research, data decisions, testing and determining what meaningful access looks like.
Tom discussed the risk of technology companies working quickly because large amounts of funding are available, while using data that does not adequately represent natural Deaf language.
For example, a system may be trained using interpreted television footage paired with captions. This provides interpreted language rather than spontaneous first-language BSL and may not reflect the cultural and linguistic decisions made by the interpreter.
The written captions may also not align directly with individual signs because interpretation is not a word-for-word process.
A large quantity of data is not automatically good language data.
Questions of consent and ownership also remain. Whose signing has been used? Did they agree to that use? Who owns the system created from it, and who benefits financially?
Without Deaf oversight, sign language can be reduced to a dataset and access can become something done to a community rather than designed with it.
The difference between a narrow task and open communication
Tom showed examples of avatars and photorealistic digital signers.
Some systems can produce understandable output for narrow, repeated tasks such as transport information. A train station uses a limited set of predictable messages involving platforms, times, delays and cancellations.
This is very different from producing accurate sign language for any possible subject and audience.
The more open the task becomes, the more variation the system needs to manage. It must deal with context, regional language, grammar, facial expression, movement, cultural knowledge and the fact that several translations may all be reasonable.
Recognising sign language and producing spoken or written English presents further difficulties. The system would need to understand different signers, appearances, signing styles, camera positions and language backgrounds.
Technology may perform one carefully defined task well without being generally capable of interpreting.
This distinction is easily lost in marketing. A successful transport demonstration can be presented as evidence that a system is close to understanding or producing sign language across every setting.
Tom encouraged viewers to ask what they are actually seeing. Is the signing generated by AI, constructed from recordings of a Deaf translator or produced through another form of video editing?
A visually impressive output does not explain how it was created or where its limits are.
The last-mile problem
Tom used the “last-mile problem” to explain why technological progress can appear both impressive and frustrating.
A system may complete a large part of a task relatively quickly. Making the final part reliable across unusual, complex and unpredictable situations can require far more work.
Self-driving technology may work within a carefully mapped area but struggle when faced with an unfamiliar rural road and a situation that was not anticipated.
Sign language AI faces a similar challenge.
Producing something understandable in a narrow setting is difficult but possible. Producing accurate, natural and appropriate language across different people, subjects and contexts is a much larger task.
The final part is not a small cosmetic improvement. It may contain the very features that make the communication linguistically and culturally meaningful.
This helps explain why AI can look extremely advanced in a demonstration while remaining unsuitable for many real-life interactions.
The technology should be assessed according to the situation in which it will be used, not only according to its strongest example.
Will AI take interpreting and translation work?
Tom did not offer a simple yes or no.
Some tasks are more exposed to automation than others. Routine and repeated content may increasingly be produced with AI support, particularly where organisations currently provide no sign language access.
AI may also become part of translators’ workflows, assisting with captions, drafts, review or technical production.
However, interpreting and translation involve much more than generating grammatically possible language.
Professionals consider purpose, relationships, power, culture, audience, consequences and the meaning created across an interaction. They notice when someone is uncertain, when a source message is ambiguous and when a supposedly simple request carries greater risk than the commissioner realises.
AI may affect which tasks people perform and how the work is organised. It may create new roles involving review, quality assurance, language data, governance and Deaf-led technology development.
The danger is not only that a tool becomes capable of replacing part of the work. It is that organisations are persuaded it can do so before it is ready.
A commissioner may see an AI-generated signer and decide that every BSL translation can now be produced cheaply and instantly. A funding body may reduce support for qualified professionals because it believes technology provides an equivalent service.
The profession therefore needs enough AI literacy to challenge inaccurate claims without dismissing every possible use.
Fear can make the technology appear stronger
Headlines about the jobs most at risk from AI often encourage urgency and fear.
Tom suggested that this can lead people to overestimate what the systems currently do. When someone sees smooth language or a photorealistic signer, they may assume intelligence and accuracy that are not present.
Fear can also prevent practitioners from engaging at all.
An interpreter who refuses to learn anything about AI may be less able to identify unsafe practice, advise a commissioner or participate in decisions about future technology.
Critical engagement offers another route.
We can test tools without uploading confidential information. We can compare answers, check sources and learn the language used to describe new systems.
We can listen to Deaf researchers, translators and community organisations raising concerns about ownership, data and access.
Being informed does not require enthusiasm for every development. It gives us a stronger basis for deciding what is useful, what is unsafe and what questions still need answering.
What does this mean for translators?
Sign language translators are likely to encounter AI from several directions.
Commissioners may ask whether AI can reduce production time or cost. Translation platforms may include automated tools, and AI-generated signed content may require professional review.
Translators may also use language models when researching terminology, organising source material or reviewing written scripts.
The same professional principles apply.
AI output should not be treated as the first draft of BSL simply because it has produced an English summary or gloss. The translator still needs to consider audience, register, cultural meaning, visual structure and the purpose of the final translation.
Deaf translators should hold central roles within the design and assessment of sign language technology. Their involvement should include authority and payment, not only testing a nearly completed product.
The profession will also need to decide what quality assurance means when a translation combines human language decisions with automated production.
Who is responsible when the final output is inaccurate? Who approves publication? Is the reviewer given enough time and power to require changes, or are they being used to give the product a professional appearance?
These are professional and ethical questions, not only technical ones.
What should responsible use look like?
Tom’s presentation did not provide a single policy for every interpreter and translator. Organisations, tools and assignments differ, and the technology changes quickly.
It did provide a useful starting point.
Responsible use means understanding that a language model produces plausible language rather than guaranteed truth. It means checking original sources and refusing to upload confidential information simply because the tool makes a task easier.
It also means keeping the human professional in control.
AI can help generate questions, but it should not decide whether someone is competent for an assignment. It can suggest terminology, but it should not be the only source used to verify it.
It can help organise a reflection, but it cannot decide what the professional genuinely learned or how their practice should change.
In sign language technology, responsible development requires Deaf leadership, suitable language data, transparency about limitations and clarity about where a system should not be used.
The goal should not be to appear innovative. It should be to produce something accurate, safe and genuinely useful.
What are we passing on?
The theme for The Together Conference 2026 was Pick It Up, Pass It On.
AI makes passing information on almost effortless.
A summary can be generated in seconds. An email can be rewritten and shared. A claim about new technology can move across social media before anyone checks whether it is true.
This speed increases our responsibility.
We can pass on the idea that AI is either magic or a threat that should be avoided. Both responses prevent more careful discussion.
We can instead pass on a practical form of AI literacy.
Large language models are useful, but they do not understand in the way their language may suggest. They can create false information confidently. Professional and personal data should not be uploaded without careful thought and permission.
Sign language technology must involve Deaf people throughout its development and governance.
Most importantly, interpreters and translators remain accountable for the professional work completed with an AI tool.
The software may assist with the task. It does not inherit our responsibility for the outcome.
New insights need informed scrutiny
The theme for The Together Conference 2027 is New Insights, Shared Purpose.
Artificial intelligence will continue to raise new questions across interpreting, translation, education, access and professional regulation.
Perhaps you are researching how interpreters use large language models, or developing training around privacy, prompting and verification.
You may be a Deaf translator, technologist or community member examining sign language data, ownership or the quality of AI-generated content.
You might have tested AI within a particular workflow and identified both a useful application and a limit that commissioners need to understand.
These experiences could contribute to The Together Conference 2027.
The Call for Papers opens on Monday 5 October 2026. You do not need to decide whether your proposal belongs in the live or extended programme. Submit your idea and the review panel will consider where accepted contributions fit best.
Poster submissions will also be welcomed for emerging research, professional guidance, technology projects, early findings and practice-based reflections.
New technology does not reduce the need for professional judgement.
It gives us more situations in which that judgement needs to be visible, informed and shared.
Using this article for unstructured CPD
You may wish to reflect on one or more of the following questions:
Which AI tools have you used, and what did you believe they were doing?
How would you explain the difference between a large language model and human understanding?
Have you ever accepted an AI response because it sounded clear and confident?
What steps do you take to verify an AI-generated explanation, summary or reference?
Which parts of assignment preparation could AI usefully support?
Which parts of preparation should remain directly with the interpreter or translator?
Have you ever uploaded information without first considering its confidentiality or ownership?
What types of professional material should never be entered into a public AI tool?
Could anonymised information still identify the person or organisation involved?
How might AI rewriting alter a Deaf person’s tone, intention or authority?
What does it mean to treat AI as a copilot rather than the pilot?
How could fear of AI lead practitioners to overestimate its current abilities?
How could refusing to engage with AI leave the profession less prepared to challenge unsafe use?
What role should Deaf people and Deaf translators hold in sign language technology?
What are the risks of training sign language AI mainly on interpreted rather than natural Deaf language?
Where might narrow AI-generated sign language add access, and where would its limitations create unacceptable risk?
How could commissioners mistake an impressive demonstration for a generally reliable service?
What AI guidance or policy would support your own professional practice?
Is there a project, experience or area of research that could become a proposal or poster for The Together Conference 2027?
A short professional reflection on these questions could be recorded as part of your unstructured CPD.
Comments