Interpreting SAFE AI Task Force Guidance on AI and Interpreting Services

Accountable use of AI technology means that end users are included in the design process from the beginning in order to ensure that AI tools are actually suitable to end user needs. This includes the development pipeline and evaluation as well as interface design, studies of human computer interaction (HCI), and product testing. Once ready for market, AI products that are procured and utilized for interpreting services must include explicit informed consent to accept/decline the use of AI, opt-in/opt out to data collection and storage without penalty, complete transparency, and evidence of adherence to ethical standards.

Design standards of AI products for interpreting should be created to guide development, governance, and counterfactual* fairness within the system. Organizations deploying AI products must maintain an inventory of the system, collect, process and respond to human feedback, and track corresponding decisions and results. The use of AI tools should be officially recorded in files and records in every instance.

* Counterfactual Fairness: captures the intuition that a decision is fair towards an individual if they are the same or belong to a different demographic group (race, gender, age, language group, etc.).

In all settings, some form of each of the following is required:

• Informed Consent to Accept or Decline the use of an AI product, with confidence that a human interpreter will be provided in a timely manner
• Opt In/Opt Out of data collection and storage without penalty
• A mechanism to shift from AI to human interpreting (or vice-versa) at any time
• User friendly grievance process to report errors or harm
• User friendly explanations of “privacy” and “confidentiality” and the layers/degrees of AI involvement in the interpreting process

Accountability implies compliance. It involves the clear delineation of what constitutes safe, fair and ethical uses of AIxAI and what constitutes harm. Further, examples of harms should be defined and listed for two categories: accidental and adversarial.

AI-assisted Interpreting:

• Explanation: Interpreting services that combine AI tools with human oversight.
• Integration: Ensure AI-assisted ASL interpreting services involve human interpreters to oversee and correct AI-generated interpretations, providing a safeguard against errors.

AI-assisted Technology: This term refers to using artificial intelligence (AI) to enhance human tasks. AI helps in tasks like data analysis by quickly processing large amounts of information more efficiently than humans. Although AI does much of the initial work, humans make the final decisions. Examples include smart speakers like Amazon Echo and smart thermostats, which adjust home settings based on your preferences and schedule.

AI-generated: This term describes content created entirely by artificial intelligence, with no human involvement during the creation process. For instance, ChatGPT generates responses based on user prompts without human edits during the response creation. It autonomously crafts text that can appear as if a human wrote it, based on its extensive training and algorithms. This process is interactive, guided by the user’s input, but the content of each response is directly produced by the AI.

AI-generated Interpreting:

• Explanation: Interpreting services fully generated by AI without human intervention.
• Integration: Clearly differentiate between AI-assisted and AI-generated services, providing users with informed choices and ensuring that AI-generated services meet high standards of accuracy and reliability.

AI Product is a ready-made software or hardware that uses artificial intelligence to perform specific tasks. It is designed for general use and can be easily purchased and used by many people or businesses. Examples include virtual assistants like Alexa, image recognition apps, and smart speakers.

AI Solution is a customized implementation of AI technology designed to solve specific problems for a particular organization. It often involves tailoring AI tools to fit the unique needs and systems of a business. Examples include personalized customer behavior prediction models and AI-powered supply chain optimization.

AI Solution Parameters

Information about how an AI software was adjusted for specific use in interpreting. The key parameters for an AI solution for interpreting are (the list below is illustrative, not exhaustive):

• Ownership of software: degree of control a vendor of an AI solution has over how the software has been developed and can be developed in the future (e.g., can new data sets be added, can new features be added, etc.)
• Volume of digital data used for “training” (data analysis) of an AI solution: 200 texts or 2,000,000
• Types of digital data used for “training” (data analysis) of an AI solution: general data or specific to an interpreting setting (e.g., legal, medical, military, K-12, etc.), sources category (e.g., public websites, setting-specific publications, dictionaries), and level human curation of selecting data sets
• Language sources: geographic dialects, including but not limited to a country of origin (e.g., American English or British English) with notation about volumes (e.g., this AI solution has utilized input of 2,000 data sets from Mexico and 20 from Guatemala)
• Audio input and output parameters: characteristics of voices utilized (e.g., age, gender, health/speech impediments, accents within the native speaker language community, accents of non-native speakers of the language)
• Direction/modality of interpreting: bidirectional/bimodal or unidirectional

AIxAI is the acronym for Automatic (or Automated) Interpreting by Artificial Intelligence.

Algorithm management refers to the systematic process of designing, developing, deploying, monitoring, and maintaining algorithms within an organization or a specific application. It encompasses a wide range of activities aimed at ensuring that algorithms operate efficiently, effectively, and ethically (per ChatGPT).

Algorithmic Fairness:

• Explanation: Ensuring AI algorithms do not discriminate and provide equitable outcomes.
• Integration: Implement rigorous fairness checks to ensure AI interpreting tools do not introduce bias against marginalized groups and provide equitable services to all users.

Artificial Intelligence (AI) refers to a body of science rooted in mathematics that is itself fundamentally a process of translation into the language of numbers.

ASTM International is an international standards organization, formerly known as American Society for Testing and Materials, that develops and publishes voluntary consensus technical standards for a wide range of materials, products, systems, and services, including language teaching, translation, and interpreting.

Automatic Interpreting refers to mechanical processes of linguistic translation in place of human interpreting during live human interactions. These are calculated by algorithms based on statistics, and vary in reliability depending on the size and scope of the database for each language. It also includes tools that augment interpreting such as using avatars to mask human interpreters, ASR captioning with generative predictive text, fingerspelling recognition software, etc.

Baselines: Baselines serve as starting points or initial references for comparison in experiments or evaluations. They often represent simple or naive methods that provide a basic level of performance for a given task. Baselines are used to gauge the improvement achieved by more complex or sophisticated algorithms or models. They help researchers or practitioners understand whether proposed solutions are effective or provide significant enhancements over simple approaches.

Benchmarks: Benchmarks are established standards or reference points against which the performance of algorithms, models, or systems can be measured. These standards are typically set based on the performance achieved by existing methods or on known results for specific tasks or datasets. Benchmarks provide a point of comparison to assess the effectiveness or efficiency of new approaches.

Bidirectional Interpreting

Interpreting that consists of converting meaning in both directions within one human interaction/communication between speakers of two different spoken languages, e.g., from English into Portuguese and from Portuguese into English.

Bidirectional and Bimodal Interpreting

Interpreting that consists of converting meaning in both directions within one human interaction/communication between speakers whose languages exist in different articulatory modalities: a sign language and a spoken language, e.g., from American Sign Language (ASL) into English and from English into ASL.

Bilingual Evaluation Understudy (BLEU)

BLEU is a popular algorithm for evaluating the quality of machine translation. It is based on the “closeness” between a machine/candidate translation and a reference human translation (Papineni, 2002). Studies have found a high correlation between a high BLEU score and human-judged scores. The BLEU score is a number between 0 and 1, where 1 is considered a “perfect” translation, which is exceedingly rare in practice due to the variability of human translations. The score is calculated by counting n-grams in the reference translation that are reproduced by the candidate translation, with various penalties and bonuses applied. A drawback of BLEU is that it highly depends on tokenization (finding “word” boundaries). Tokenization of ASL can be difficult due to phonemes being used continuously and in multiple contexts. BLEU has seen some use in Signed English systems. If BLEU is reported with a number next to it like BLEU-2, it means that it is counting only 2-grams.

Briefings are required in medical, educational, and legal settings prior to use of an AIxAI application. In this briefing, end-users (of both/all languages) are informed of the technical specs (quantitative quality) of the AIxAI, the levels of AI and/or human involvement available, and a risk/benefit analysis suited to their individual case.

Clients are the business, corporation, government entity or other organizational customer of the AI application.

Collaborative Communication is a description of the practices (attitudes and behaviors) of the humans involved in interpreted interaction when they are making sincere effort to understand each other both through the interventions of interpreting (by human or machine) and despite the presence of an intermediary who may inject bias (algorithmic or human).

Confidence

Loosely, confidence represents the probability that the prediction is correct. Model output can be constrained using confidence cutoffs to prevent unreliable output that the machine is not confident about.

Confidentiality refers to knowledge about a person being shared only among those (family members, friends, professional service providers) with an authorized right to know.

Counterfactual Fairness ( defined in text, do we want to repeat or expand upon it here? ) captures the intuition that a decision is fair towards an individual if they are the same or belong to a different demographic group (race, gender, age, language group, etc.).

Cultural Nuances

Differences in the way individuals in different cultures believe, behave, feel, and express themselves through language, gestures, and body language. These differences impact how messages are perceived and understood. Cultural nuances cannot be inferred from analyzing texts or speech samples.

Data Privacy:

• Explanation: Protecting user data used by AI systems in compliance with regulations.
• Integration: Ensure that ASL interpreting tools adhere to stringent data privacy standards, protecting the personal information of users and maintaining confidentiality in sensitive communications.

Deep Learning:

• Explanation: A subset of ML that uses neural networks with many layers to model complex patterns.
• Integration: Implement deep learning techniques to enhance the ability of ASL interpreting tools to recognize and interpret subtle gestures and expressions accurately.

Design justice ensures that interventions are made first at the product level, focusing on equitable and inclusive design, particularly for marginalized communities, including the Deaf community and other communities with low resource languages. For instance, in the context of AI for sign language, this means involving Deaf end users at every stage in the development pipeline, from problem selection through data collection, outcome definition, algorithm development and post-deployment considerations.

Embedding An embedding is a representation of data (such as words, images, or nodes in a graph) in a continuous vector space. Embeddings capture the semantic meaning or features of the data in a lower-dimensional space, facilitating easier manipulation and analysis (per ChatGPT).

End-users are the participants in the interpreted interaction, of both or all languages and modalities (depending how many are in use).

Equity recognizes each person has different circumstances and needs, meaning, different groups of people need different resources and opportunities allocated to them in order to thrive. Equity is a measure of value, specifically the value invested in ensuring that individuals receive similar outcomes of fair, safe and ethical treatment regardless of individual traits, characteristics, and/or potentially disabling or discriminatory conditions in a given situation.

Errors are mistakes in interpreting that are readily and easily corrected during processes of collaborative communication, including interpreted interaction. The source of errors could be participants from either/any language(s), human interpreters, and/or the automated interpreting algorithms and software. “Errors” imply technical glitches or mistakes that can be corrected, which can minimize the severity and human consequence involved. In contrast, “harms” acknowledge the broader and more serious implications, emphasizing the ethical responsibility to prevent and address the negative outcomes that can affect end-users’ well-being, safety, and dignity. Using “harms” highlights the urgency and gravity of these errors, ensuring they are given the necessary attention and resources to be mitigated effectively.

Ethical refers to the material behaviors to be promoted, encouraged and maintained through the use of AIxAI.

Explainability:

• Explanation: The degree to which the workings of an AI system can be understood by humans.
• Integration: Make machine interpreting AI tools explainable to users, ensuring they understand how decisions are made and the basis for interpretations, thus building trust in the technology.

FERPA is the Family Educational Rights and Privacy Act of 1974, a federal law in the United States that protects the privacy of parents and students.

GDPR is the General Data Protection Regulation of the European Union (2016). It

• lays down rules relating to the protection of natural persons with regard to the processing of personal data and rules relating to the free movement of personal data.
• protects fundamental rights and freedoms of natural persons and in particular their right to the protection of personal data.

Generative AI:

• Explanation: AI systems capable of generating natural language responses and interpreting tasks.
• Integration: Ensure that generative AI tools for all language pairings for machine interpreting are rigorously tested for accuracy and reliability, and include mechanisms to correct errors in real-time to avoid miscommunication.

Geographic Dialects

Varieties of a language that develop in and represent speakers of specific places, regional or local. For example, Mexican Spanish, South American Spanish, Levantine Arabic, Yemeni Arabic, Australian English, Texan American English, etc.

Harms are injury or damage to, for example, a person’s physical health, emotional or mental well-being, eligibility for programs or services, successful applications or interviews for employment or educational opportunities, and professional reputation within a community of practice. Harms also include group-level patterns of discrimination, stereotyping, and other forms of systemic marginalization.

HIPAA is the Health Insurance Portability and Accountability Act of 1996 (United States). It’s requirements for the protection of certain health information are specified in the Standards for Privacy of Individually Identifiable Health Information (“Privacy Rule”).

Human-Computer Interaction (HCI) (Workflow}: The process of evaluating how users interact with the computer interface that allow efficient, effective, and satisfying interactions between humans and computers. It focuses on usability, user experience (UX), accessibility, aesthetics, feedback, consistency, and task efficiency.

Human-generated: This term describes content or outputs that are created entirely by humans without the direct involvement of artificial intelligence in the creation process. It involves traditional methods of creation where humans apply their creativity, knowledge, and skills to produce the final product, such as writing a book, painting a picture, or crafting a manual report.

Interface Design is the process designers use to create interfaces that are user-friendly, easy to use, accessible, and consistent.

Interpreting refers to communication activities during live interaction between two or more human beings who do not fully share a common language.

Large Language Models (LLMs):

• Explanation: Advanced AI models trained on extensive datasets to understand and generate human-like text (which may or may not be part of the main products of machine  interpreting)
• Integration: Highlight the need for LLMs in pairings of languages for interpreting to be trained on diverse datasets that include dialects and regional variations for spoken and sign language nuances to ensure accurate and contextually appropriate interpretations.

Limited English Proficient (LEP) is the federal label for persons whose cognitive fluency is in a language (or languages) other than English. We note that this is an inherently biased label which inherently establishes an unfortunate (discriminatory and stereotyping) linguistic hierarchy.

Low Resource Languages “can be understood as less studied, resource scarce, less computerized, less privileged, less commonly taught, or low density, among other denominations.” Source: Low-resource Languages: A Review of Past Work and Future Challenges  https://arxiv.org/pdf/2006.07264

Machine Learning (ML):

• Explanation: AI that learns from data to make decisions and predictions.
• Integration: Use ML to continuously improve pairings of language interpreting tools by learning from user interactions and feedback, thereby increasing accuracy and “fit” to end user’s context over time.

Metric for Evaluation of Translation with Explicit Ordering (METEOR)

METEOR was developed to address some of the issues with BLEU. It is the harmonic mean of 1-gram precision and recall. Precision is the ratio of true positives to both true and false positives. Recall is the ratio of true positives to both true positives and false negatives (Banerjee, 2005). A study has shown that METEOR may perform better than BLEU or NIST (Lavie, 2009).

National Institute of Standards and Technology (NIST)

The NIST metric is based upon the BLEU metric. The main difference is that n-grams are weighted by their informational content. For example, the 2-gram “of the” would have a low weight compared to “broken clock” (Coughlin, 2003).

Natural Language Processing (NLP):

• Explanation: AI’s capability to understand, interpret, and generate human language.
• Integration: Emphasize the importance of coordinating HCI interface controls with the NLP in low resource language interpreting tools to understand the context and cultural nuances inherent in ASL, ensuring effective communication.

Objective Functions in machine learning/generative AI are the mathematical measure of the quality of the desired optimization. Objective is used in the sense of goal. What is the goal that the AI is being optimized for? Optimization is the process of finding the best, optimal, solution for that goal. The objective function takes data and the specific model’s parameters to return a number, i.e., a quantitative score. Objective functions evaluate the parameters in order to adjust their values in order to maximize or minimize achievement of the goal (definition by ChatGPT).

Principles describe design features in the architecture of AIxAI to enable safe, fair and ethical  applications of AIxAI. These features span the algorithmic infrastructure through dashboards and leaderboards to each end-user’s specific and particular interface.

Privacy refers to the collection of data from a person’s online activities associated with the use of AIxAI. This data must always be de-identified for internal and external use and may not be sold to third parties without agreement by and payment to the individuals whose private data is included in any aggregation.

Title VI of the Civil Rights Act of 1964 prohibits discrimination on the basis of race, color, and national origin in programs and activities receiving federal financial assistance. National origin is understood to include language. Executive Order 12250 required federal agencies to prepare information to affected communities about these protections.

Speech Recognition Software

Computer programs that convert human speech to text. This is mostly a step in a complex computer program that then can analyze the produced written text. The main challenge for such programs is to match audio sounds produced by humans with different accents, age-dependent pronunciation (e.g., “child speak”), or speech impediments (e.g., speech patterns of a person who has suffered a stroke) to the correct letter.  This challenge is further exacerbated by the quality of audio sound (e.g., how close a person is to the microphone) and presence/absence of additional sounds (e.g., more than one person are speaking, music is played in the background, nearby equipment is producing sounds).

Transparency:

• Explanation: Openness about how AI systems function and make decisions.
• Integration: Enhance transparency by providing detailed information about how language pairings using interpreting AI tools are trained, their capabilities, limitations, and the data sources used.

Unidirectional Interpreting

Interpreting that consists of converting meaning in one direction from a speaker/signer of one language to listeners/readers/watchers who speak/sign a different language, e.g., from English into Portuguese only. This type of interpreting is deployed when no response from the listener(s) is expected or needed, e.g., public announcements, lectures to large audiences, news broadcasts, etc.