Artificial Intelligence and Machine Learning Statement
At Phrase, artificial intelligence (AI) plays an integral role in our services. We use AI to help our customers produce, manage and translate multilingual content more efficiently and at scale.
This Artificial Intelligence and Machine Learning Statement (“Statement”) explains how we use AI-enabled solutions responsibly, including how we use content to train our AI models, with a focus on transparency, security, and ethical practice.
We recognize that trust is the foundation of our relationship with customers, and we are committed to keeping them fully informed about how their content is used. Our approach emphasizes ethical practices, strong data security, and compliance with applicable regulations, ensuring our AI-enabled solutions are not only powerful and efficient but also respect the privacy and confidentiality of Customer Content.
This statement is intended to ensure Phrase’s compliance with the transparency obligations set forth in the European Union Artificial Intelligence Act (AI Act). Although the AI Act will be fully enforceable as of 2 August 2026, Phrase has been committed to create and use its AI system in a manner that is safe, ethical, and therefore aligned with the AI Act and other applicable legal regulations.
1. Definitions
“AI Act” means Regulation (EU) 2024/1689 of the European Parliament and of the Council of 13 June 2024 laying down harmonised rules on artificial intelligence and amending Regulations (EC) No 300/2008, (EU) No 167/2013, (EU) No 168/2013, (EU) 2018/858, (EU) 2018/1139 and (EU) 2019/2144 and Directives 2014/90/EU, (EU) 2016/797 and (EU) 2020/1828 (Artificial Intelligence Act).
“AI system” means, according to the AI Act, a machine-based system that is designed to operate with varying levels of autonomy and that may exhibit adaptiveness after deployment, and that, for explicit or implicit objectives, infers, from the input it receives, how to generate outputs such as predictions, content, recommendations, or decisions that can influence physical or virtual environments.
“Customer Content” means any content that Phrase’s customer or its users upload to, or create, or translate (including the translated content) within the Phrase solutions.
“GDPR” means Regulation (EU) 2016/679 of the European Parliament and of the Council of 27 April 2016 on the protection of natural persons with regard to the processing of personal data and on the free movement of such data, and repealing Directive 95/46/EC (General Data Protection Regulation).
“QPS” means Quality Performance Score.
“LLM Provider(s)” means the third-party providers of generative large language models that Phrase engages as sub-processors, as listed in our Sub-processors overview at https://phrase.com/subprocessor-overview/. These currently include OpenAI, Google, Microsoft and Anthropic.
“LQA” means Language Quality Assessment.
“MQM” means Multidimensional Quality Metrics.
“MT” means machine translation.
2. Introduction
The Phrase Language Intelligence Platform represents an AI system within the meaning of the AI Act and Phrase, as its provider, is responsible for the compliance of the Phrase Language Intelligence Platform with the AI Act.
None of the AI-enabled solutions used within the Phrase Language Intelligence Platform engages in any practices prohibited under the AI Act, such as manipulating users, performing social scoring or exploiting vulnerabilities of users. Further, none of the AI-enabled solutions used within the Phrase Language Intelligence Platform shall be classified as a high-risk AI system under the AI Act as none of Phrase’s AI-enabled solutions is used by Phrase for activities like critical infrastructure management, biometric identification, or other high-stakes applications that would require special oversight in accordance with the AI Act.
For years, customer content has been essential in improving translation models, long before AI-driven solutions became the norm. Historically, translation engines, including statistical and rule-based machine translation models, have relied on customer translation memories (TMs) and linguistic assets to enhance translation accuracy, fluency, and overall quality. This long-standing practice has been fundamental to the development and continuous refinement of translation technology.
3. AI-enabled solutions at Phrase
Below, we provide a detailed overview of the individual AI features that form an integral part of the Phrase Language Intelligence Platform. In line with our commitment to transparency, we aim to offer clear and comprehensive information about how these features function, the processes involved in training our AI models, and how Customer Content is utilized in these processes.
3.1 Are we engaging Phrase for a service or technology solution that uses artificial intelligence or machine learning?
We offer AI-enabled solutions that provide customers with full control over their activation, whether applied across all projects or selectively for specific ones.
Our dedicated AI research team, working alongside our development teams, reflects our strong commitment to continuously advancing our AI and machine learning (ML) capabilities. The key solutions outlined below represent either features currently available or those actively under development.
We welcome the opportunity to engage in a follow-up discussion to gain deeper insights into your specific use cases and any concerns you may have. Your feedback is invaluable in helping us refine our solutions and prioritize considerations for future releases.
3.2 What is the AI/ML solution intended to do and how will customers use the AI/ML solution?
In the section below we list the individual AI-enabled solutions at Phrase. In general, our goal is to provide solutions that optimize the localization process to facilitate content globalization at scale. We aim to offer both fully and partially automated solutions that can drive commercial optimizations but also tooling that can improve the usability and human experience of our platform. This includes automated translation, as well as mechanisms such as Phrase QPS which help identify and route the lowest quality content for human review.
The following is a summary table of our AI-enabled features, more detailed explanations are included below:
(a) Text-generating AI: The following table details AI features which output or generate textual content or translation:
| Feature Name | Feature Summary | Trained with Customer Content? |
|---|---|---|
| Auto Adapt | Automated adaptation of text to custom instruction | NO |
| MT Optimize | Automated clean-up of MT generated text | NO |
| Phrase NextMT | Machine translation; automated translation of content via Phrase’s MT engine | NO |
| Phrase Next GenMT | Machine translation; automated translation of content via Phrase-proprietary model interaction with third-party LLM Providers | NO |
| Phrase AI Translation Agent | Agentic translation that generates and refines a translation in one automated pass | NO |
| Phrase CustomAI (Custom NextMT) | Training and evaluation of customer-specific, personalized MT engines | YES |
| Phrase CustomAI (Linguistic Asset Curation) | Uses automated tooling to curate linguistic assets and datasets for custom MT training | YES (Uses QPS) |
| Speech-to-text (Phrase Studio) | Transcription of audio files | NO |
| Text-to-speech (Phrase Studio) | Conversion of text into spoken audio | NO |
| AI Summary / AI Insights (Phrase Studio) | Extraction of insights from subtitles | NO |
(b) Other AI features: The following table details AI features which do not generate customer content or text:
| Feature Name | Feature Summary | Trained with Customer Content? |
|---|---|---|
| MT Autoselect | Automatically selects the most appropriate MT engine for a job | YES |
| Phrase QPS | Automatically scores translated content for accuracy (0-100) | YES |
| AI Checks | Evaluates content against customer-defined criteria and returns a pass/fail verdict with rationale | NO |
| Non-Translatables | Automatic detection of non-translatables | YES |
A. Phrase Language AI, MT Autoselect
Our MT Autoselect is an AI-based mechanism for selecting the most appropriate MT engine for a particular job. Customers can use it with or without a defined list of preferred engines in order to maximize output translation quality and minimize post-edit cost downstream.
What is the goal of this feature? We want to enable customers to access a range of MT engines and dynamically switch between them, where one might produce higher quality output than another on Customer Content.
Example: A customer has assigned three MT engines in their MT profile and wants to make sure they get the engine with the best quality output. The system looks at their content, reviews recent post-edit data on the three MT engines and recommends one of those three engines as the best candidate for translation.
How does it work? The system considers the domain of the content together with recent quality data for the available MT engines, and recommends the engine most likely to produce the best output.
How do we use Customer Content? The system briefly reviews Customer Content to identify its domain, and periodically reviews post-editing activity on prior content to assess the performance of each MT engine.
How do we protect Customer Content? No Customer Content is retained or shared during this process whatsoever.
B. QPS, Quality Performance Score
Phrase QPS is a quality estimation system that predicts the MQM score that might be given to a particular piece of content in an LQA process. It outputs a score between 0-100 that helps customers to plan resources, post-edit workload and other processes by providing quality transparency at scale.
What is the goal of this feature? We aim to provide a scalable solution for broad insight into translation quality. This is particularly important in scenarios where machine translation is preferred and quality is not guaranteed. Specifically, customers can use this feature in automated workflows to decide where to route content and ensure that the lowest quality content is captured and sent for post-editing or other appropriate workflow steps.
Example: A customer has a limited budget for post-editing and relies on machine translation for a majority of their content. Using Phrase QPS, they can automatically identify the worst scoring content and route it for human post-editing.
How does it work? The AI behind Phrase QPS is an internal, proprietary model trained on both publicly available datasets and Customer Content. Specifically, we train it with examples of prior LQA assessments that have been generated on the Phrase Platform, as well as examples of MT post-edits that have been generated on the Phrase Platform. This content includes the original content, the resulting translation and the LQA assessments of that translation. We show the AI multiple examples of translations and the MQM scores that resulted from that translation such that it learns to predict MQM scores on new content. When the trained AI is used in practice, we supply it with both the source and the translated content and the system outputs a score between 0-100.
How do we use Customer Content? We use Customer Content to train and periodically retrain the AI and output an estimate of the Multidimensional Quality Metrics (MQM) score that the translation would receive were it sent through human LQA for review. Customer Content is passed through the model to generate a score for that content.
How do we protect Customer Content? Whilst the model is trained with Customer Content, its only function is to output a score between 0-100; no Customer Content is retrievable from the model by anyone. Customer Content that is used to train QPS is also not retained; we implement a regular deletion of training data in the Phrase AI Research Team to ensure compliance with both GDPR and our privacy notice and data retention policy.
C. AI Checks
AI Checks is an evaluation feature within Quality Profiles that uses a large language model as a judge. A customer defines an evaluation they want carried out, and the content is assessed against those criteria to produce a pass or fail verdict together with a rationale.
What is the goal of this feature? AI Checks allows customers to define their own quality criteria and have content evaluated against them automatically, at scale. This supports automated quality workflows and gives customers transparency over whether content meets their specific requirements.
Example: A customer needs translations to match a particular brand register. They define these expectations in their quality profile and appropriate AI Checks are generated. Each relevant translation is then evaluated and returned with a pass or fail verdict and a short rationale.
How does it work? The customer’s defined criteria and the content are sent to an LLM Provider, which returns a pass or fail verdict and a supporting rationale. No model training or fine-tuning is involved.
How do we use Customer Content? The content to be evaluated, together with the customer’s criteria, is sent to the relevant LLM Provider and a verdict and rationale are returned. Customer Content is not used to train or fine-tune any model.
How do we protect Customer Content? We maintain enterprise-level agreements with each of our LLM Providers that ensure Customer Content is not stored by the LLM Provider or used to train the LLM Provider’s own models and products.
D. Auto Adapt
Auto Adapt is an automated post-editing and content adaptation solution that enables adjustment of text to ensure things like terminology, formality-level and tone-of-voice consistency. Customers can provide additional instructions to make customized adaptations of the text such as style.
What is the goal of this feature? Auto Adapt is intended to offer an automated alternative to post-editing that better ensures consistency in style and formality and adjustment to customer-specified style.
Example: A customer has translated content using a neural, segment-level MT system. The result needs to be amended to appeal to a particular demographic. The customer runs the content through Auto Adapt in order to automatically adjust the tone and style of the text to suit.
How does it work? The system currently uses models from one or more of our LLM Providers to adjust translated content, using internally managed model interactions and configurations provided by the customer.
How do we use Customer Content? Customer’s content (either monolingual text or both the text for translation and the translated output) is passed through Auto Adapt via an LLM Provider. The output of the system is a revised version of the original text. Auto Adapt is not fine-tuned or trained on Customer Content.
How do we protect Customer Content? We maintain enterprise-level agreements with each of our LLM Providers that ensure Customer Content is not stored by the provider or used to train the provider’s own models and products.
E. MT Optimize
MT Optimize is an automated refinement step for machine translated content. It improves the quality of a translation, for example its fluency and consistency, without changing the meaning of the content. Unlike Auto Adapt, it does not take custom instructions; it applies refinement only and is not used to adapt content to a particular style or requirement.
What is the goal of this feature? MT Optimize offers an automated way to improve the quality of machine translated content, either as an alternative to, or as a precursor to, human post-editing.
Example: A customer has translated content using a segment-level MT engine and wants to improve its fluency and consistency before review. They run the content through MT Optimize to refine the output automatically.
How does it work? MT Optimize uses models from our LLM Providers to refine translated content. It does not take custom instructions and does not adapt the content beyond refining the existing translation.
How do we use Customer Content? The translated content is passed through MT Optimize via an LLM Provider and a refined version is returned. MT Optimize is not fine-tuned or trained on Customer Content.
How do we protect Customer Content? We maintain enterprise-level agreements with each of our LLM Providers that ensure Customer Content is not stored by the provider or used to train the provider’s own models and products.
F. Identification of Non-translatables
An AI system that is tasked with identifying segments that contain items that should not be translated such as company/product names.
What is the goal of this feature? This feature is intended to optimize translation workflows by identifying text that does not require translation and can therefore be skipped. This allows customers to avoid corrective post-editing, particularly of machine translated content.
How does it work? The AI is trained with Customer Content and learns to recognize text items that are non-translatable, it then outputs the likelihood (represented by a number between 0-1) of whether the segment in its entirety is non-translatable or not.
Example: A customer has some content in which a product name appears; they don’t want to pass this to machine translation as it may result in unnecessary translation of the item and subsequent post-edit cost in reinstating the product names. They use this AI feature to capture items that are non-translatable and block them from machine translation.
How do we use Customer Content? Customer Content is used in the training and periodic retraining of the AI feature. The AI reviews Customer Content and outputs a number that represents the likelihood that the content contains non-translatables which can be used to flag content. Customer Content is never retained by the AI and the model is not capable of outputting anything but a binary label.
How do we protect Customer Content? Whilst Customer Content is used to train this AI model, the system cannot retain or output that content. It is not possible to retrieve Customer Content from the model. Similar to our treatment of content used in training of QPS we implement a regular deletion of training data in the Phrase AI Research Team.
G. Phrase NextMT, Phrase’s in-house translation engine
Phrase NextMT is a machine translation engine, which can be used to automatically translate content in a number of languages. It is similar in nature to alternative third party MT engines save that it uses Customer Content at translation time to optimize the quality of its output. It is built and maintained internally at Phrase.
What is the goal of the feature? Customers can improve the quality and speed of localization projects by providing linguists with machine-translated content optimized for post-editing by professional translators. It is also feasible to use Phrase NextMT to automatically translate larger volumes of content at speed.
Example: A customer has a large volume of content to translate with a limited budget, they can use Phrase NextMT to translate the entirety of that content and route it to humans for post-editing.
How does it work? Phrase NextMT engine is tailored to professional translations, including support for tag placement, advanced glossary integration (including morphological inflection), and translation memory adaption (fuzzy matches). We train the Phrase NextMT engine with publicly available data for a range of languages. We then use the customer’s translation memory to improve the translation, e.g. in order to better align with the customer’s content style or branding.
How do we use Customer Content? Phrase NextMT is trained using publicly available and appropriately licensed datasets, and a limited number of commercially obtained proprietary datasets. We do not use any Customer Content, content or translations in the training or retraining of this AI model.
Customers can however use their translation memories and other linguistic assets during the translation process to improve the quality of the output. This content is not retained in any way and is also not used for training of the AI model.
How do we protect Customer Content? Aside from the exclusion of Customer Content in training, customer assets (translation memories, glossaries etc) used in the translation process are only accessible by the individual customer. No Customer Content is retained by the AI model or is accessible by other customers.
H. Phrase Next GenMT, Phrase’s LLM-based translation engine
Phrase Next GenMT is an LLM-based, higher quality alternative to Phrase NextMT. It is built on and uses models from our LLM Providers to provide high quality, fluent output.
What is the goal of this feature? Phrase Next GenMT is the best performing MT engine developed by Phrase; the goal is to provide automated translation at the highest possible quality. In a similar fashion to Phrase NextMT we allow customers to use translation memories and other assets to optimize the quality of the output translation. This can then be used as a standalone translation solution or in concert with human post-editing for high quality results.
Example: A customer has a volume of content for translation on a limited budget with high quality expectation; Phrase Next GenMT is used to produce a first translation which is then sent for human post-editing. The customer increases the quality and stylistic alignment by providing access to their translation memory, which further improves the translation.
How does it work? Phrase Next GenMT uses models from our LLM Providers, guided by Phrase’s own instructions and, where available, relevant examples retrieved from the customer’s translation memory.
How do we use Customer Content? Currently, we rely solely on a combination of prompting and few-shot example retrieval to achieve high quality translation. We do not otherwise use Customer Content in the training or fine-tuning of Next GenMT. The content for translation together with any applicable examples from the translation memory are sent to the relevant LLM Provider and a translation is returned.
How do we protect Customer Content? We do not use Customer Content in the training or fine-tuning of Next GenMT. When a customer requests use of this feature we send content and examples to the relevant LLM Provider in order to generate the translation. We maintain enterprise-level agreements with each of our LLM Providers that ensure Customer Content is not retained or used by the provider in the training or development of any of their models or products.
I. Phrase AI Translation Agent
The Phrase AI Translation Agent is an agentic translation solution that both produces and refines a translation in a single automated pass. It generates an initial translation and then applies an automated refinement step to improve the output, combining generation and clean-up into one operation rather than two separate stages.
What is the goal of this feature? The goal is to deliver high quality, near publish-ready translation automatically. By pairing initial generation with an automated refinement pass, the agent improves fluency, consistency and adherence to customer requirements without a separate manual editing step. It can be used as a standalone automated translation solution or as a first pass ahead of human post-editing.
Example: A customer has a large volume of content to translate to a high quality standard. The AI Translation Agent produces an initial translation and then automatically refines it for fluency and consistency, returning improved output in one pass. The customer can provide translation memories and other assets to further improve quality.
How does it work? The AI Translation Agent uses models from our LLM Providers in two or more stages: an initial translation stage, guided by Phrase’s own instructions and any relevant examples from the customer’s translation memory, followed by automated refinement that improves the output.
How do we use Customer Content? The content for translation, together with any applicable examples from the customer’s translation memory, is sent to the relevant LLM Provider and the translated and refined output is returned. We do not use Customer Content in the training or fine-tuning of the models behind this feature.
How do we protect Customer Content? We maintain enterprise-level agreements with each of our LLM Providers that ensure Customer Content is not stored by the LLM Provider or used to train the LLM Provider’s own models and products. Customer assets used during the translation process are only accessible by the individual customer and are not retained.
J. Phrase Custom AI, Dataset creation and custom engine training
Customers have the ability to create custom training data from their translation memories and use them to generate custom Phrase NextMT engines for specific use cases. Customers are able to then utilize a custom-trained “customer-only Phrase NextMT engine” that no other customers will be able to access.
What is the goal of this feature? In certain circumstances it can be beneficial to a customer to train a personalized MT engine for a specific use case. We provide tooling to enable customers to create datasets from their own content and train their own individualized instances of Phrase NextMT. We also provide analytics that demonstrate the success of model training and the utility of the resulting engine.
Example: A customer with translation memories has a specific use case where general purpose MT does not provide satisfactory quality. The customer can use CustomAI to clean and filter their translation memories, create a dataset from that content and automatically train their own instance of Phrase NextMT that is specialized to their requirements.
How do we use Customer Content? During the dataset creation and cleaning process and subsequent training of the Phrase NextMT engine, the system has access to the customer’s translation memory. Once trained, access to the resulting engine and datasets is restricted such that they can be accessed and used only by the relevant customer. No other Customer Content is used in the training of these custom MT engines.
How do we protect Customer Content? Customer Content and the custom-trained engine is uniquely accessible to that customer. No other Customer Content is used in the training of custom NextMT engines.
K. Phrase CustomAI; Linguistic Asset Curation
Automated tooling in Custom AI can be used to filter and clean customer translation memories and other assets.
What is the goal of this feature? Our linguistic asset curation tooling is intended to allow customers to filter and optimize their linguistic assets for use in other areas of the localization process such as custom NextMT model training or as ‘few-shot’ examples to improve the quality of output of Phrase Next GenMT.
Example: A customer has generated a large translation memory and wants to clean it to better guarantee the efficacy of their translation workflow that includes skipping the translation step for similar content that was already translated or for optimizing the quality of output from Phrase Next GenMT. They use the curation tooling to remove repetitions and low quality translations, resulting in a refined translation memory.
How does it work? Our asset curation tooling uses a combination of basic, rule-based (non-AI) filters, Phrase QPS (detail on Phrase QPS is provided above) and other tools. Phrase QPS here is used for example to score every item in the translation memory and remove the lowest scoring entries (for example the bottom 10%).
How do we use Customer Content? Beyond Phrase QPS (which is discussed above) no other AI features are used in our asset curation tooling. Customer Content is isolated per customer and the tooling does not retain or share Customer Content. The result of filtering is again isolated and accessible only by the customer.
How do we protect Customer Content? The results of asset curation and related AI-based filtering are only accessible by the customer.
L. Speech-to-text (Phrase Studio)
The transcription component of Phrase Studio takes audio files as input and returns a transcription of the text.
What is the goal of this feature? Customers are able to submit audio files in any format and receive a transcription of the file in text format. This can be used in the platform for subtitling of video for example.
Examples: A marketing team uploads a customer interview recorded as an MP4 file. Phrase Studio returns a transcript, with each speaker identified.
How does it work? The system takes the audio as input and uses Automated Speech Recognition (together with a set of other auxiliary models such as Forced Alignment, Diarization, and Language Identification) in order to generate the text output.
How do we use Customer Content? The AI takes the customer audio as input and processes the file for conversion. No customer data is stored in the model during processing.
How do we protect Customer Content? The AI is not trained on customer data, nor is any customer data stored by the model.
M. Text-to-speech (Phrase Studio)
The Text-to-Speech and voice synthesis component of Phrase Studio takes text as input and converts the text into spoken audio.
What is the goal of this feature? Customers are able to generate natural sounding voice audio from written text, this can be used for example in the dubbing of video or in automated interpretation in a real-time scenario
Examples: A customer has a transcript of a talk track required for a marketing video. With Phrase Studio they are able to generate speech audio of the text to include with the video.
How does it work? The Text-to-Speech feature of Phrase Studio passes a text file to the ElevenLabs API for voice synthesis, ElevenLabs returns an audio file with the corresponding audio according to user specified parameters.
How do we use Customer Content? The system sends the text file via API to ElevenLabs and an audio file is returned.
How do we protect Customer Content? We maintain an enterprise-level agreement with ElevenLabs that ensures that Customer Content is not stored with ElevenLabs or used in the training of ElevenLabs’ own models and products.
N. AI Summary / AI Insights (Phrase Studio)
What is the goal of this feature? To extract structured and meaningful insights such as summaries, sentiment, quality flags, or safety issues from subtitles using AI models.
Example: A user wants to summarize a customer support call and identify potentially unsafe or low-quality communication. Phrase Studio returns a summary and flags sections for review.
How does it work? The system sends the original (or translated) subtitles to an LLM-based summarizer and insight extractor operated by one of our LLM Providers. It returns summary, flagged segments, and optional metadata such as speaker sentiment or tone analysis.
How do we use Customer Content? Subtitle content is processed via our LLM Providers’ enterprise APIs. Resulting insights are saved, not the source data.
How do we protect Customer Content? Enterprise contracts ensure that content is not used for training or retained by our LLM Providers. All processing occurs within secure, audited environments.
3.3 Will Phrase custom develop AI/ML models exclusively for customers or will they provide generic AI/ML models?
We offer both. Customers may utilize any of the 30+ generic engines and additional custom MT models. Custom MT models would be tailored for the particular customer, using Customer Content. Please refer to section F in Section 3.2 for more details.
MT engines integrated via customer’s own API key
We have no control over custom model training of MT engines provided by third parties with whom you maintain a direct relationship, within the meaning that you are using your own API key to integrate the MT engine with the Phrase Language Intelligence Platform.
4. General Principles of AI Governance
AI governance encompasses the frameworks, policies, and practices that guide the ethical and responsible development, deployment, and management of AI-enabled solutions. At Phrase, we recognize that trust in AI-enabled solutions is built on a foundation of transparency, accountability, and respect for the rights and reasonable expectations of all stakeholders.
5. Ethical Considerations and Privacy
In developing and deploying our AI-enabled solutions, we are guided by principles that respect individual rights, promote fairness, and preserve the confidentiality of the Customer Content.
Phrase ensures that the amount of data required for training of the AI models is minimized to the extent strictly necessary for achieving the purpose of enhancing the AI-enabled solutions and delivery of the quality services to customers. In this connection, Phrase also strictly limits retention of training data. Where possible, Phrase considers using anonymized or pseudonymized data for training of its AI models or uses aggregated data. Phrase oversees that the outputs of AI models are relevant and do not reveal personal data related to training data.
We ensure that any Customer Content used for training of our AI models (as described in more detail in Section 3 above) is processed in accordance with the GDPR and other applicable data protection laws in case it potentially includes any personal data. For more details on data processing at Phrase, please see our Privacy Notice.
6. Human Oversight and Control
A dedicated AI Research Team at Phrase is responsible for conducting regular quality assurance checks on the outputs of our AI-enabled solutions. The whole procedure is overseen by the Legal Team to ensure compliance with applicable laws. These checks ensure that the outputs meet proper standards of accuracy, fairness, and compliance with both internal policies and applicable legal regulations.
To support transparency and continuous improvement, users and employees may flag potential issues for further review by our AI Research Team. Once flagged, these issues undergo a structured review process, allowing our team to identify root causes, address any shortcomings, and implement necessary adjustments to improve the system’s performance and trustworthiness.
This open feedback loop allows users and employees to contribute to the reliability and responsible use of our AI-enabled solutions.
7. AI Literacy
In accordance with the AI Act, at Phrase we are committed to ensuring that all our personnel involved in the development, deployment, and supervision of AI systems possess an adequate understanding of AI concepts, risks, and best practices. Phrase’s personnel are obliged to participate in training programs covering fundamental AI concepts and principles of the AI Act and compliance therewith. This ensures our teams have the knowledge and skill sets required to implement and oversee AI-enabled solutions responsibly.
As the AI Act and other relevant regulations will evolve in the future, we are committed to update our training materials, policies, and documentation to reflect the most current standards.
8. Conclusion
We reserve the right to revise this Statement as necessary, recognizing that understanding and interpretation of the AI Act will continue to evolve over time.
If you have any questions or requests regarding this Statement, please contact us at privacy@phrase.com.