Before AI Touches the Target Text: A Verification Chain for ATA Members
By Gleb Tsipursky
The useful question isn’t whether AI can produce language quickly. It can. The question is whether the complete workflow produces accurate, appropriate, confidential, and client-ready communication more reliably.
Note: The following was adapted from the author’s book, The Psychology of AI Adoption at Work: From Resistance to Results (Georgetown University Press, 2026).
ATA’s Statement on Artificial Intelligence recognizes useful AI-assisted workflows, including terminology extraction, preliminary drafting, pre-session briefing, and transcripts, while emphasizing expert human oversight. That combination points toward a workable professional standard: use automation where it reduces friction, then make verification visible enough so a linguist can explain why the final output deserves trust.
Start with the Source of Authority
An AI tool should never become the source of truth merely because it produced the latest version on screen. For each assisted task, identify the authority the output must match.
In translation, that authority may include the source text, client instructions, an approved termbase, reference translations, style guidance, or subject-matter documentation. In interpreting preparation, it may include a speaker list, agenda, client glossary, prior materials, or verified institutional terminology. If the source of authority is unclear, the task isn’t ready for automation.
This distinction matters because machine translation and generative AI work from statistical patterns rather than professional responsibility. ATA’s client guidance on machine translation warns that automated output can look plausible without being verified. Therefore, a professional workflow needs a deliberate handoff from machine suggestion to human judgment.
Workflow 1: Terminology Research and Project Setup
Terminology work is an obvious place to use AI because it involves finding candidate terms, grouping concepts, comparing phrasing, and organizing reference material. It’s also an easy place to introduce errors that then propagate through an entire project.
Suppose a translator uses AI to extract candidate terminology from a technical source file. The useful output is a starting list, not an approved glossary. Before a term enters the project termbase, the translator should check the source context, client-approved terminology, authoritative subject-matter references, target-market usage, and any instructions governing product names, legal terms, or regulated language.
The verification chain can be simple. First, record where the proposed term came from. Second, note which authoritative source supports the target-language choice. Third, identify any ambiguity or competing term. Fourth, mark who approved the final choice.
That process protects more than terminology consistency. It prevents a convenient AI suggestion from quietly becoming a project standard simply because nobody documented the moment when suggestion turned into decision.
Workflow 2: First-Pass Translation and Post-Editing
AI-assisted drafting can save time on repetitive or predictable material. The danger appears when speed changes the reviewer’s behavior. Once a sentence reads smoothly, a linguist may be tempted to review the target text as prose rather than verify it against the source.
A stronger post-editing routine separates fluency from fidelity. Start with high-risk elements: numbers, dates, units, names, negation, conditions, exceptions, defined terms, cross-references, and anything that changes obligations or consequences. Then check terminology, register, locale conventions, omissions, additions, and ambiguities. Only after those checks should the reviewer optimize style.
This order matters in legal, medical, financial, public-facing, and safety-sensitive work. A sentence can be elegant and still change who must act, when an exception applies, or how much something costs. Review intensity should follow consequence, not how confident the AI sounds.
For recurring work, translators can turn common corrections into a project-specific review checklist. If the system repeatedly mishandles a particular abbreviation, product family, grammatical structure, or source-language ambiguity, the checklist should make that failure mode visible on the next assignment.
Workflow 3: Interpreting Preparation and Documentation
Interpreters can also benefit from AI before and after an assignment. ATA’s AI guidance notes uses such as real-time glossary access, pre-session briefing automation, and AI-generated transcripts. These tools can reduce administrative load, but they should remain support systems rather than silent authorities.
Before an assignment, an interpreter might use an approved AI system to organize a speaker list, extract acronyms, or create candidate glossary entries from provided materials. Every name, title, acronym, and specialized term should still be checked against the client’s materials or another authoritative source.
After an assignment, an automated transcript may help with documentation or accessibility where the engagement permits it. The transcript should not be treated as a verbatim record without human verification. Speakers overlap, accents vary, proper names are fragile, and context can change what a plausible transcript should say.
The same principle applies during an assignment. AI-generated suggestions should never pressure an interpreter to follow the machine when professional judgment points elsewhere. The tool supplies options. The interpreter remains accountable for conveying meaning appropriately in context.
Put Confidentiality Before Convenience
ATA’s Code of Ethics and Professional Responsibility requires members to protect privileged and confidential information and convey meaning accurately and appropriately. Those duties should shape tool selection before anyone pastes client material into an AI interface.
A practical rule is to classify information before use. Ask whether the material contains client-confidential content, personal data, protected information, unreleased business information, privileged communications, or contractually restricted text. Then confirm that the chosen system is approved for that category of data under the client agreement and the organization’s policies.
Free or consumer-facing tools deserve particular scrutiny because data-handling terms can differ from enterprise arrangements. If the professional cannot explain where the material may go, whether it may be retained, and what agreement permits that processing, the safe choice is to keep the content out of the system until those questions are answered.
Confidentiality should also extend to prompts. A prompt can reveal as much about a client as the attached document. Removing the obvious name while leaving a product launch, lawsuit, diagnosis, merger, or internal terminology intact may still expose sensitive context.
Keep an Exception Log, Not a Surveillance Log
AI quality improves when professionals record where the workflow fails. That doesn’t require storing confidential source text or tracking every keystroke.
Create a small exception log for recurring AI-assisted tasks. Record the task type, the category of error, the consequence if it escaped review, and any corrective action taken. Useful categories include omission, unsupported addition, terminology error, register mismatch, number or name error, locale issue, source ambiguity, and confidentiality rule violation.
Over time, the log answers practical questions. Which use cases produce reliable first passes? Which language pairs or subject areas create more rework? Which client instructions are routinely missed? Where does AI save time only to give it back during correction?
This evidence supports a more professional decision than either enthusiasm or fear. Expand uses that reduce cycle time without weakening quality. Redesign uses that create repeated exceptions. Retire uses whose correction burden exceeds their value.
Match Review to Consequence
Not every task needs the same control. A private brainstorming prompt for alternative headings carries a different consequence from translating informed consent, a contract clause, a financial disclosure, or safety instructions.
A three-level review model can keep safeguards proportionate.
- Low-consequence tasks may need a quick accuracy check.
- Moderate-consequence tasks should require comparison with the authoritative source and a named reviewer.
- High-consequence tasks should require explicit verification of critical elements, appropriate subject matter expertise, and a clear release owner.
The release owner matters because responsibility can become blurry when several people and tools touch a project. Someone must decide that the final text or interpreted output is fit for its purpose. AI can assist that person, but it cannot absorb the professional responsibility.
Make Skeptics Part of the Workflow Design
Language professionals who are cautious about AI often see failure modes that enthusiastic adopters overlook. That makes them useful testers.
Instead of asking skeptics to become cheerleaders, ask them to identify the details that an inexperienced reviewer might miss, the contexts where a plausible translation becomes dangerous, and the client promises that automation must never override. Turn those concerns into test cases and review criteria.
This approach also protects professional identity. ATA’s Statement on Artificial Intelligence emphasizes that growing AI use increases the need for expert linguists who can assess, correct, and validate output. A strong implementation makes that expertise more visible. It defines the human contribution as judgment, context, accountability, and the ability to recognize when the machine’s answer should be rejected.
Measure Rework, Not Just Throughput
AI projects often celebrate speed because speed is easy to measure. Language professionals need a broader scorecard.
Track time saved, correction time, exception frequency, repeated error types, client revisions, escalations, and reviewer confidence. Compare those measures across a defined trial period. A workflow that creates a faster first draft but more downstream corrections may be shifting work rather than reducing it.
The useful question isn’t whether AI can produce language quickly. It can. The question is whether the complete workflow produces accurate, appropriate, confidential, and client-ready communication more reliably.
For ATA members, that’s the standard worth defending. The profession doesn’t need to choose between craft and technology. It needs systems that make professional judgment the final control point.
Before AI touches the target text, define the authoritative source, the required checks, the confidentiality boundary, the exception process, and the release owner. Then the technology can do what it does best without asking it to carry a responsibility that belongs to the translator or interpreter.
Further Reading
- ATA Statement on Artificial Intelligence (2025).
- ATA Machine Translation Guidance.
- ATA Code of Ethics and Professional Responsibility.
- Autio, Chloe, Reva Schwartz, Jesse Dunietz, Shomik Jain, Martin Stanley, Elham Tabassi, Patrick Hall, and Kamie Roberts. “Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile” (National Institute of Standards and Technology, 2024).
- International Federation of Translators (FIT) Position Paper on Machine Translation in the Age of AI (2025).
Gleb Tsipursky, a behavioral scientist who has been called the “Office Whisperer” by The New York Times, helps tech-forward leaders stop overpaying for AI while boosting engagement and innovation. He serves as the chief executive officer of the AI consultancy Disaster Avoidance Experts, and has written eight books, including The Psychology of AI Adoption at Work: From Resistance to Results (Georgetown University Press, 2026). gleb@disasteravoidanceexperts.com
