Machine translation vs a human: when to use which
Machine translation, including AI chatbots, is useful for getting the gist and for drafting low-stakes material a bilingual colleague will check. It is not a substitute for a qualified interpreter or a certified human translation when words carry consent, rights, clinical instructions, safety, or legal consequences. The tools are not always wrong; they are wrong often enough, unpredictably enough, and invisibly enough that no one reading only English can tell a good translation from a dangerous one.
Sort every task into one of two buckets first
Decide the bucket before you reach for a tool. The test is what happens if the translation is wrong.
- Understanding an incoming email or form well enough to route it
- A rough first draft of a routine notice a bilingual colleague will check
- Following an informal conversation when a bilingual person is also there
- Deciding whether a document is worth sending for full translation
- Informed consent, and notices about rights or eligibility
- Clinical instructions: medication and dosing, wound care, discharge steps, warning signs
- Legal documents, safety-critical warnings, and anything that enters an official record
- Any spoken interaction where the client acts on what is said
If the worst realistic outcome of an error is confusion a person will catch, machine translation is a reasonable start. If a client could be harmed, lose a right or benefit, or sign something they did not understand, a qualified human is not optional.
For common languages and everyday text, it is genuinely good
When researchers evaluated Google Translate on emergency department discharge instructions, Spanish translations were accurate for about 92 percent of sentences (Khoong et al., 2019). A later study comparing ChatGPT-4 and Google Translate on patient-specific discharge instructions found sentence-level accuracy of 96 to 97 percent for Spanish and 90 to 95 percent for Chinese (Kong et al., 2026).
That is enough to be useful for the right jobs: understanding a form well enough to route it, drafting a routine notice, following an informal conversation when a bilingual person is present, or deciding whether a document needs full translation. In these cases the machine gives you speed and reach, and a mistake costs little because someone will catch it before anyone acts on it.
Accuracy falls fast outside the most common languages
Google Translate, share of emergency-department instructions where the overall meaning was retained, across seven languages (Taira et al., 2021).
The many languages community and refugee agencies work in most, underserved precisely because they are less common, are the languages where these tools are least reliable. Across 26 languages, Google Translate was correct only 57.7 percent of the time overall, with Swahili at 10 percent (Patil & Davies, 2014).
The averages hide the danger, and you cannot see it
The same study that found 92 percent accuracy for Spanish also found that 2 percent of Spanish sentences and 8 percent of Chinese sentences contained errors with the potential for clinically significant harm (Khoong et al., 2019). Across 26 languages, the phrase "your child is fitting," meaning a seizure, came back from Swahili as "your child is dead," and "your husband had a cardiac arrest" became, in Marathi, "your husband had an imprisonment of heart" (Patil & Davies, 2014).
A monolingual English speaker looking at a confident, well-formed paragraph in another language has no signal that anything is wrong. Large language models make this worse: they can produce hallucinations, fluent text detached from the source, and these concentrate in lower-resource language directions and can surface toxic content pulled from training data (Guerreiro et al., 2023).
The machine reversed a life-or-death instruction
Keep taking the kidney medicine until you talk to your kidney doctor.
Hold the kidney medicine until you have a chance to speak with your kidney doctor.
Sight translation and back-translation
Sight translation is when an interpreter reads a written document aloud in the client’s language on the spot. It is a reasonable stopgap for short, simple text when no certified translation exists, but professional standards discourage sight-translating complex or consequential documents such as consent forms, because doing it accurately on the fly is hard and the liability is real. Treat it as a bridge, not a replacement for certified translation.
Back-translation is when a second qualified bilingual person translates the machine output back into English so you can compare it to the original. This is exactly how the studies above caught errors, so it works, but it costs a second qualified person and it is imperfect. Use it to check a draft, never as a rubber stamp that lets machine output skip human translation for consequential content.
Putting client text into a tool is a data-sharing act
Every time you paste a client’s words into a translation tool, you send that text to an outside company. That is a disclosure, and for health or legal information it may be a regulated one. Staff and patients are often unaware that using a general AI translation tool for medical communication means sharing sensitive information with an external provider, and free versions in particular may store input and reuse it to improve the service without a real way to opt out (van Kolfschooten et al., 2025).
The rule follows directly: never put identifying client information into a consumer or free tool. Strip names, dates of birth, addresses, and record numbers before any machine step, and when in doubt, do not paste it at all.
Before your agency adopts a tool, get these in writing
If a vendor cannot answer these plainly, treat that as a no.
For protected health information under HIPAA, a Business Associate Agreement. For other client data, a data processing agreement naming the vendor’s obligations.
A contractual no, or a real opt-out that is on by default for your account, not buried in settings.
Encryption in transit and at rest, a stated retention period, and a way to request deletion.
The list of sub-processors, and whether data leaves the country.
Consumer and free tiers usually offer none of the above. Do not use them for client data.
Quick decision flow
- Will anyone act on it, or does it just need to be understood?If it only needs to be understood or drafted, and a bilingual colleague will check it, machine translation is fine. Skip to the privacy step.
- Consent, rights, clinical, safety, or legal?If yes, use a qualified interpreter for speech and a certified human translation for documents. Stop here.
- A less common or lower-resource language?If yes, lower your confidence further and lean toward a human even for tasks that would be borderline in Spanish.
- Before pasting anythingRemove all identifying client information, and confirm the tool has a data agreement and does not train on your inputs. Never use a free consumer tool for client data.
- For any draft you keepHave a qualified bilingual person back-translate and compare before it reaches a client. Fluent is not the same as correct.
Before client text goes into any tool
- The task is low-stakes gist or a draft a bilingual person will check, not consent, clinical, legal, or safety content.
- All identifying client information has been removed.
- The tool has a signed data agreement and does not train on inputs, and it is an enterprise account, not a free consumer tool.
- For a less common language, confidence is lowered and a human is preferred.
- Any kept draft is back-translated and compared before it reaches a client.
Sources
- Khoong, E. C., Steinbrook, E., Brown, C., & Fernandez, A. (2019). Assessing the Use of Google Translate for Spanish and Chinese Translations of Emergency Department Discharge Instructions. JAMA Internal Medicine, 179(4), 580–582. doi:10.1001/jamainternmed.2018.7653
- Patil, S., & Davies, P. (2014). Use of Google Translate in medical communication: evaluation of accuracy. BMJ, 349, g7392. doi:10.1136/bmj.g7392
- Taira, B. R., Kreger, V., Orue, A., & Diamond, L. C. (2021). A Pragmatic Assessment of Google Translate for Emergency Department Instructions. Journal of General Internal Medicine, 36(11), 3361–3365. doi:10.1007/s11606-021-06666-z
- Kong, M., Fernandez, A., Bains, J., et al. (2026). Evaluation of the accuracy and safety of machine translation of patient-specific discharge instructions: a comparative analysis. BMJ Quality & Safety, 35(3), 150–158. doi:10.1136/bmjqs-2024-018384
- Vieira, L. N., O’Hagan, M., & O’Sullivan, C. (2021). Understanding the societal impacts of machine translation: a critical review of the literature on medical and legal use cases. Information, Communication & Society, 24(11), 1515–1532. doi:10.1080/1369118X.2020.1776370
- Guerreiro, N. M., Alves, D. M., Waldendorf, J., et al. (2023). Hallucinations in Large Multilingual Translation Models. Transactions of the Association for Computational Linguistics, 11. doi:10.1162/tacl_a_00615
- van Kolfschooten, H., Goosen, S., van Oirschot, J., et al. (2025). Legal, ethical, and policy challenges of artificial intelligence translation tools in healthcare. Discover Public Health, 22, 904. doi:10.1186/s12982-025-01277-z
This guide describes practice grounded in the studies listed above. Adapt it to your agency’s policies and the languages you serve.