Malt Community SA: G7 Hiroshima AI Process (HAIP) Transparency Report

Malt’s technology relies on an internal proprietary matching algorithm built on internal datasets and is used by over 100,000 companies to find talents from Malt’s community of more than 1,000,000 qualified independent experts. Malt develops a proprietary recommendation system built on a small language model. In addition, Malt integrates third-party generative AI in its software to retrieve a set of relevant criteria from client's hiring needs (prompt – a pdf, a copy/paste of a job description, a long paragraph or just a few keywords), and agentic AI to assist the client throughout their search. All AI extracted hiring criteria or proposed actions require explicit client validation.  

Malt maintains an internal AI risks register which includes risks whether they are algorithmic -derived from the AI model-, operational -derived from the use of the AI model-, or fundamental -impacts individual rights. The risk registers uses a traditional risk-mapping methodology reviewing the gravity scale, and the likelihood scale; and it integrates the lifecycle stages and vets the risks agains the characteristics of trustworthy AI. It is maintained by the AI deployment committee. Risks related to historical data bias crystallization whereby the system replicates legacy (human) biases against certain demographic groups are considered serious/likely and are therefore unreasonable. Risks are analyzed, mitigated operationally and technically, and constantly monitored; they may prevent the model from being released/used. 

The AI Deployment committee has decision authority when it comes to deploying AI models. 

The committee oversees the lifecycle of in-house AI developments and external AI tools being built into Malt's products across two key pillars: (1) new model/feature:  system risk classification, risks associated to the system, data protection and fundamental rights impact assessments, technical documentation check, transparency requirements check, and third-party vendor documentation (if applicable); (2) model/feature monitoring: review risks associated to the system, review incidents (if any), performance reviews and log retention audits (minimum 6-month retention). 

Before submission to the AI Deployment committee, the machine learning/product owner must compile a standardized AI System dossier containing the evaluation framework which includes notably a list of criteria, a description of what is evaluated and the assessment method, the design specification, data protection and fundamental right impact assessments, and a information security report from Malt’s information security team. 

The committee holds binding authority on system deployment under three possible outcomes: (1) approved: sign-off for deployment under specified operational parameters; (2) conditional approval: requires specific remedies (e.g., additional human-in-the-loop controls, refined dataset balancing) prior to go-live; (3) rejected / suspended: immediate block on deployment or mandate to halt an active model due to unmitigated high risk or serious incident reporting.

Safeguards include a mandatory progressive deployment scaling from 5% to 25%, 50%, and 100% user traffic, with reviews at each increment; real-time drift and error monitoring triggers an automated switch to a deterministic, rule-based fallback system. 

Internal incident generates a detailed live incident document that records the incident description, chronological timeline, containment actions, and initial technical resolutions. Upon incident resolution, a post-incident analysis and retrospective are conducted. The incident response team investigates underlying root causes, identifies procedural or technical gaps, and formulates action plans. 

Malt’s AI deployment committee and information security team review internal incident reports related to the AI system, and monitor external incident reports in order to review the risk management of the models in light of the current threats. Security breaches or regulatory non-compliance events trigger external reporting to relevant supervisory authorities within statutory timelines.

The retrospective process produces corrective and preventive measures designed to prevent incident recurrence. Measures are integrated as applicable to the updates of Malt’s AI system at stake, Malt's security, and business continuity policies. 

As of today, Malt’s reports are available to stakeholders based on legal requirements. 

Vulnerability reports can be submitted through Malt’s bug bounty program, following this link. Scientific reports, including on severe risks, can be found on Malt's dedicated research page. 

We further encourage users to file any report regarding AI safety and transparency either through the Malt interface or by sending an email to compliance@malt.com. 

• Reporting to regulatory authorities as relevant
• Publicly
• Information-sharing partnerships (with confidentiality agreements as appropriate)

Technical information is shared with research partners for the purposes of performing research, and will be shared publicly through scientific papers (more information here). Statutory information is shared with regulatory authorities where required. 

Alongside non-incentivized initiatives like dedicated communication channels to encourage feedback reporting by employees when a new model is released, or a new product feature integrating AI; Malt operates an annual pen-test exercice, that includes red teaming, and a bug bounty program to identify and remediate potential security vulnerabilities. More information can be found here.  

Further, Malt has a dedicated in-house research team that collaborates with over four international university research centers on empirical studies and industry research to improve hiring in the labor market. Scientific papers published can be found here. 

In addition, Malt has stand alone incentivized field experiments by making users participate to analyze project briefs by freelancers alongside reviews of anonymized profiles by clients against compensation. Upon release of a new model, or a new product feature that involves AI, users are also provided with early access (no user-centric incentivized initiative ongoing at the moment). 

• ISO/IEC standards (e.g., ISO/IEC 27001, ISO/IEC 42001, ISO/IEC 23894, ISO/IEC 42005)
• NIST AI Risk Management Framework (AI RMF)
• Industry-led best practices or codes of conduct

Malt  contributes to the development of industry tools, evaluation frameworks, and technical best practices for AI risk assessment—particularly focusing on the safe use of AI in recruitment, hiring, and the freelance work domain. 

We collaborate with the broader scientific and engineering community to build standardized tools for assessing AI risks and performance. We share our findings publicly through our research page and engineering blog. For example, Malt researchers recently co-authored and contributed to establishing a standardized, community-driven open-source benchmark for AI in the work domain.

We further participate in the SEISMEC international consortium, where Malt’s pilot focuses on a core issue for digital labour platforms: matching needs to talent efficiently, while ensuring that decisions do not create systematic differences in visibility and access to opportunities between groups of freelancers. The pilot develops and integrates evaluation tools that assess matching systems throughout their lifecycle. 

Last, we participate in technical and regulatory working groups, and we contribute frequently to technical and regulatory panels. 

Malt distinguishes between recommendation systems, generative AI and agentic AI. Generative systems receive additional controls for prompt and output security, factuality, privacy, provider processing and human review. Agentic systems, where applicable, require additional safeguards such as least-privilege access, approved tool lists, action limits, human confirmation for consequential actions and an effective interruption or shutdown mechanism. 

The lifecycle process is reviewed and updated under three circumstances: (1) when a new system version or model iteration is released, including consultation with the AI Deployment Committee for new algorithmic model baselines; (2) at least annually, including extensive independent penetration testing and a broader review of the effectiveness of governance, security, privacy, fairness and monitoring controls; (3) on an ad hoc basis, following an active model defect, vulnerability, operational exploit, material incident, significant performance or fairness degradation, major data or architecture change, new third-party dependency, regulatory development or material research finding.

For the plan and design stage, before developing or deploying an AI system, Malt defines in the design specifications its intended purpose, users, affected stakeholders, expected benefits, foreseeable misuse and potential risks. Systems are assessed in the risk mapping across privacy, security, fairness, transparency, accuracy, human rights, operational resilience and environmental considerations by the AI deployment committee. The review determines whether the system can proceed, requires additional safeguards or should not be deployed. 

In the collect and process data stage, Malt uses professional and marketplace information relevant to the service, such as skills and experience for matching and search. Malt evaluates data quality, completeness, and representativeness. Data is limited, cleaned, and tested, including for potential bias, before use. Users can correct, update, or request deletion of their information. Third-party models and AI services are subject to AI, privacy, security and contractual review. Data shared with external providers is governed by controls such as confidentiality, zero-retention policy, limited access. 

In the build, integrate, and interpret model stage, Malt’s matching systems use layered architectures combining lexical and semantic retrieval, ranking and business rules. Models, prompts, datasets, configurations and third-party dependencies are documented and versioned where relevant. Development and testing are separated from production environments, and access is controlled according to role and need. User-facing explanations focus on the purpose and limitations of AI-assisted search and recommendations rather than presenting the system as an autonomous decision-maker. Proposed candidates, and AI-generated briefs remain subject to human validation. Users can edit, reject or replace AI-generated recommendations and can use keyword search in the directory and type in manually their brief. 

Through the verify and validate stage, Malt verifies AI systems before and after material changes through a combination of automated testing, offline/online evaluation, human review, comparative experiments, security testing and fairness analysis. Malt also evaluates user group performance and potential disparities by including group analysis and counterfactual testing in which relevant characteristics are changed while professional skills and experience are held constant. These evaluations help identify differences in ranking position, visibility and exposure. 

Before release, at the deploy stage, Malt considers the evaluation results. Where appropriate, changes are introduced progressively and monitored against an approved baseline. Production systems maintain logging and traceability sufficient to investigate outputs, model versions, relevant inputs and human interventions, while respecting data-minimization requirements. Malt maintains human oversight for AI-assisted matching and recommendations. Fallback and rollback procedures are designed to allow Malt to isolate a faulty component, revert to a previously validated version or use a deterministic alternative when an AI component or external service is unavailable or behaving unexpectedly.

While in operation, Malt monitors AI systems across data quality, performance, fairness, security, reliability, user feedback and operational impact. Monitoring may include error rates, latency, anomalous behavior, user complaints, and external-provider failures. AI-related incidents may result in additional testing, changes to data or prompts, enhanced human review, rollback, temporary suspension, retraining or retirement of the affected system. Feedback from clients, freelancers, employees, researchers and other stakeholders contributes to continuous improvement. 

Malt’s proprietary matching models are  developed using first-party data generated on the Malt platform and focusing on professional information relevant to the intended purpose, such as skill and expertise. The data is collected in the context of providing and improving Malt’s services. Malt does not rely on public-web scraping for the training of its proprietary matching models.

Malt applies data-quality and data-governance measures throughout collection, transformation and curation. These include documented data source, controlled access, schema and consistency checks, moderation of inappropriate content, validation of professional attributes, monitoring of data freshness. Data pipelines and transformations are subject to quality checks and lineage controls so that material changes can be identified and investigated. In addition, non-AI business rules govern specific operational parameters, including candidate location, availability, and language skills. Where lawful and necessary, protected and aggregated sensitive personal data may be used separately for fairness testing and impact assessment. Data minimization, retention controls and access restrictions are applied to derived datasets, evaluation data and other model-related artefacts. 

To reduce the risk that historical booking patterns reproduce existing biases, Malt evaluates the quality and fairness of training and evaluation data through group analysis, counterfactual testing and audits. This includes testing whether profiles with comparable skills and experience receive different ranking positions when protected characteristics are varied. Data-quality, fairness and representativeness findings are incorporated into model development, deployment reviews and post-deployment monitoring. Material deterioration, evidence of disparate impact, data-quality failures, harmful outputs or changes in the underlying data can trigger additional investigation, corrective action, retraining, rollback or a renewed review by the relevant governance and deployment functions.

The information security team is in charge of controlled environments for red-teaming and adversarial testing, controlled security vulnerability testing, and access controls, including role-based permissions. 

The machine learning team is in charge of isolated and secure test environments, isolation of staging environments, use of “AI-safe” test-data, as appropriate, the use of synthetic, anonymized, or fake data, as appropriate, and human-in-the-loop controls for higher-risk test cases. 

The data platform team is in charge of logging and monitoring. 

• Not applicable

• Not applicable

Malt applies the fundamental data-protection principles under the GDPR throughout the collection, processing, storage and deletion of personal data. 

Personal data used in Malt’s matching and search services is processed for specific purposes, including providing freelancers with relevant project opportunities and helping clients identify freelancers whose professional skills and availability correspond to their project requirements. Malt limits processing to data relevant to these purposes and applies additional safeguards to data used for model development, evaluation and monitoring.

Malt’s privacy governance in the context of the AI system is organized around two main areas. First, privacy by design is integrated into product and AI development, the DPO is consulted before a new model version or material model change is initiated so that the relevant data protection impact assessment can be performed. Second, the upholding of individual's rights throughout the lifecycle stages of the AI system. Freelancers and clients can access, update and correct relevant professional information through their accounts or through dedicated privacy channels. More information is available in Malt’s privacy policy.

Data protection standards align with ISO27001:2022 requirements (see below).

• Maintain or update security policies aligned with recognized standards (e.g., NIST CSF, ISO/IEC 27001, SOC 2, CIS Critical Security Controls), and review them regularly in response to emerging risks or regulatory guidance
• Maintain vulnerability management processes (scanning, channels for external reporting, patching, remediation)
• Implement incident response, disaster recovery, or business continuity plans, including escalation procedures for high-severity incidents
• Train staff on cybersecurity best practices and AI-specific threats, including how to recognize and respond to misuse or unintended outcomes of advanced AI systems
• Regularly review security measures and update them in response to emerging risks, incidents, or major releases

• Apply multi-factor authentication (MFA) and role-based access controls for advanced AI systems
• Encrypt AI assets (datasets, models, configurations) in storage and in transit as appropriate
• Restrict physical access to servers, GPUs, or devices hosting AI models, systems, and/or applications
• Monitor system activity and log access to detect anomalies or misuse
• Conduct secure AI development and/or deployment practices
• Conduct security testing focused on AI-specific risks (e.g., red-teaming, adversarial testing, stress tests)

• Limit access to proprietary AI assets based on role or need-to-know
• Store model weights, datasets, and algorithms in secure, access-controlled environments
• Track and log access to sensitive AI resources
• Require NDAs or confidentiality agreements for staff and contractors with access
• These responsibilities are delegated contractually to model or infrastructure providers

• Conduct pre-employment screening or role-sensitive vetting
• Monitor for unauthorized internal access or suspicious activity
• Provide staff training on cybersecurity best practices and AI-specific threats
• Implement insider-risk escalation or response procedures

Malt integrates AI cybersecurity into its wider information-security framework rather than treating it as a separate control environment. Malt applies a security framework aligned with ISO/IEC 27001 standards. Malt ensures formalized risk management, continuous security objectives tracking, and regular internal compliance reviews. System access is governed by role-based access control, least privilege principles, and mandatory multi-factor authentication across all corporate systems, developer repositories, and production environments. All platform data is encrypted in transit using strong cryptographic protocols (TLS 1.3) and at rest using industry-standard encryption (AES-256) across databases and file stores. Malt undergoes extensive, independent third-party penetration testing on an annual cycle to proactively identify and remediate code vulnerabilities and configuration flaws. And Malt continuously monitors its information system, including through a bug-bounty program. 

Malt classifies models, databases, prompt templates, and inputs as high-sensitivity assets requiring specialized technical controls such as filters to remove unneeded personal identifiers and confidential information; further, the information system runs on dedicated, private cloud server capacity with restricted network access controls; client project briefs, and user prompts processed via third-party generative LLMs (e.g., Gemini family) are subject to strict zero data retention policies.

• Constraints on the agent's action space (e.g., tool-use permissions, environment sandboxing, allowlists/blocklists of resources, scope limits per task)
• Human oversight or interruptibility mechanisms (approval requirements for sensitive actions, escalation paths, interruption interfaces)
• Monitoring (anomaly detection, multi-agent interaction monitoring)
• Logging of agent actions

• Summary of results from evaluations of safety, security, and societal risks including potential adverse impacts on individuals or groups (e.g. impacts on human rights, privacy or differential outcomes across user groups).
• Technical documentation and instructions for use if relevant.
• Plain-language or non-technical summaries

Malt publishes public-facing information about its AI-assisted search through product explanation available on Malt's dedicated AI page. Further information can be found in Malt's privacy and transparency policies. On an ad hoc basis, Malt will do webinars, or articles that share further information (see examples here). These materials describe, in accessible language: the purpose and main capabilities of the AI systems; how AI assists search, matching, recommendations and, where applicable, the preparation of project briefs; the types of information used, at an appropriate level of detail; the appropriate use of the systems as assistance tools; inappropriate uses, including treating AI outputs as final hiring or contracting decisions, or relying on outputs without human review; the role of human oversight, user control, and available correction or complaint channels; and the way Malt evaluates, monitors and updates its AI systems.

The public information is written for non-technical audiences and is complemented, where appropriate, by more detailed technical and compliance documentation for customers, auditors, regulators and other qualified stakeholders in compliance with the regulation. 

Publicly available technical information is available on our research page and engineering blog.

Malt has a dedicated AI page whereby Malt shares in non-technical, plain-language summaries of the description of Malt's AI system and how it works, as well as our principles guiding AI developments, and security, privacy, and AI safeguards. 

Detailed information, including evaluation results may be shared confidentially with customers, enterprise partners, auditors, regulators, research collaborators or other qualified stakeholders where this is necessary to demonstrate responsible use or support risk mitigation. Depending on the recipient and purpose, this may include technical documentation, evaluation reports, fairness analyses, security-assessment findings, model limitations, monitoring results and remediation plans.

Malt provides information to competent authorities where required by law or where regulatory engagement is appropriate. Regulatory information may include system descriptions, data-flow information, risk and impact assessments, testing results, incident information, corrective actions and relevant governance records.

Where appropriate, Malt also engages with public institutions, civil-society organizations, industry groups and AI-safety or responsible-AI initiatives. These collaborations support the assessment of matching quality, fairness, bias, explainability and the broader socio-economic effects of AI-assisted access to freelance opportunities. An example can be found here. 

Malt shares research findings through peer-reviewed publications, conference presentations, technical or engineering communications, product and policy documentation. These materials may describe evaluation methods, performance results, fairness and bias findings, known limitations, security and privacy safeguards, human-oversight arrangements and mitigation measures. See our research page here. 

Malt maintains comprehensive, publicly accessible terms and conditions of the platform, and privacy policy that explicitly governs the collection, processing, and protection of personal data, user inputs, and system-generated outputs across the platform. Additional information can be found in the dedicated AI page.

Malt does not train its AI system on external datasets - the source of the data used for the training of the AI system is Malt's platform. The source of data used for the training is detailed in Malt’s publicly accessible privacy policy that explicitly governs the collection, processing, and protection of personal data. The terms and conditions provide for the use of the user inputs, and system-generated outputs across its platform. Additional information is available on Malt's dedicated AI page. 

Further information about specific training datasets can be shared confidentially through protected information sharing, regulatory reporting, and direct engagement with civil society or academic partners. 

Malt analyzes and publishes anonymized, aggregated market data to inform broader economic understanding regarding freelance labor dynamics, emerging skill demand, and digital workforce transformation across sectors and European markets. See the latest Malt tech trends report here. 

The governance structure distinguishes between advisory, operational and final decision-making responsibilities. System documentation and assessments are reviewed when a new model or system version is released; and ad hoc reviews are triggered by a major incident, a material change to the model, prompt, data, architecture, provider or intended use, a significant deterioration in accuracy, fairness, security or user outcomes.

Malt’s Board has an advisory role. It assesses and advises on issues such as algorithmic fairness, and may recommend additional safeguards. Malt also involves a Freelancer Advisory Board of seventeen freelancers selected to reflect diversity in expertise, geography and socio-economic background. The board is consulted annually on progress, ethical considerations and emerging trade-offs between business objectives and fairness, ensuring that those affected by platform decisions have a formal channel into governance.

The AI Deployment Committee has operational decision-making authority for AI deployments and material changes. It reviews the relevant risk classification, impact assessments, data and model documentation, evaluation results, security and privacy controls, human-oversight arrangements, monitoring plans and fallback or rollback procedures. It may approve a system or model version, approve it subject to conditions and remediation, require additional testing, reject deployment or suspend an existing system. 

Product, data and machine-learning engineers (owners) are responsible for implementation. They operate the system, maintain the technical documentation, implement approved controls, monitor performance and fairness, investigate anomalies and ensure that remediation actions are completed. 

New employees complete AI, data protection, and security trainings during onboarding. This includes an introduction to responsible AI, data-protection principles, information security, acceptable use of AI tools, human oversight and the organisation’s corporate AI policy. Employees are required to review and formally acknowledge the relevant policies.

Malt also provides continuous and role-based targeted workshops, practical guidance, policy updates and refresher sessions. The content is adapted for technical, product, data, security, legal, privacy and operational teams. Depending on the role, training may cover: AI governance and risk classification; privacy by design and data minimization; fairness, bias and responsible use of professional data; secure development and protection of AI assets; model, prompt and data lifecycle management; incident response, vulnerability reporting and escalation; appropriate and inappropriate uses of generative and agentic AI.

Employees can raise concerns through the “Speak Up” reporting channel. This channel can be used to report suspected misuse, discriminatory outcomes, privacy concerns, security vulnerabilities, unsafe model behavior or breaches of Malt’s AI policy. The reporting channel may provide for anonymity here. Reports are handled according to the applicable escalation and incident-management procedures, with access limited to the compliance and security teams.

Malt manages AI-related incidents through its general incident management policy, which is integrated with Malt’ information security management system (ie. tracking system), the information security policy and the data breach policy. Every incident is recorded in a live incident document which includes a root cause analysis report. Once the incident is resolved, a post-incident analysis is conducted to further understand root causes, control gaps and measures required to prevent recurrence.

Escalation thresholds are based on severity and impact. Escalation may involve the system owner, ML or engineering teams, information security team, the DPO, legal, senior management. A critical impact on users, legal obligations, finances, productivity or reputation may trigger escalation to the executive committee. Incidents are shared with external parties—such as regulatory authorities and impacted users—when legally mandated or to prevent unintended outcomes.

Incident records and the personal-data breach register are used for audits, accountability, trend analysis, risk-register updates, control improvements, training and future deployment decisions. Major incidents, material system changes, new vulnerabilities, significant evaluation findings and regulatory developments trigger an ad hoc review of the relevant AI risk assessment and controls by the AI deployment committee.

Malt participates in research collaborations with academic and external research communities; and Malt engages in public body consultation calls such as the consultation on the draft guidelines for the classification of high-risk artificial intelligence systems, technical or regulatory panels, regulatory working groups. These collaborations support independent evaluation, the development of common benchmarks and methods, and the assessment of broader effects of AI on work, matching and access to professional opportunities. 

Further, Malt research scientists share findings through conferences, workshops, lectures and other scientific or technical forums (example can be found here and here). Malt shares its research and good practices through several channels; it publishes papers at world renown conferences (RecSys) and on its research website, including work on AI evaluation in the work domain, and fairness across groups. 

In addition, Malt’s engineering blog provides practical explanations of how its technology, data and product teams design, implement, evaluate and operate AI-enabled systems. These publications share technical approaches, architecture decisions, performance results, implementation lessons and operational considerations with the wider engineering and research community. 

Malt shares sufficient methodological and contextual information to support meaningful review while protecting personal data, confidential information, security-sensitive details and proprietary assets. This enables external review, discussion and comparison with approaches developed by other organizations.

• ISO/IEC standards (e.g., ISO/IEC 27001, ISO/IEC 42001)
• NIST AI Risk Management Framework (AI RMF)
• Industry-led best practices or codes of conduct
• Other international or regional standards (please specify): EU AI Act/GDPR

Malt uses a combination of technical standards, regulatory frameworks and industry best practices to structure AI governance framework. 

The AI Act and GDPR provide AI-system classification, risk and impact assessments, data governance, privacy and fundamental rights by design, human oversight, transparency, technical documentation, monitoring, and individual rights. 

ISO 27001 provides the foundation for Malt’s information-security management system which covers for example access control, encryption, vulnerability management, secure development, supplier due diligence, incident response, business continuity and continuous improvement. 

The risk management framework is a combination of traditional risk-mapping that integrates the NIST AI Risk Management Framework, particularly the AI lifecycle stages and trustworthy AI characteristics. The technical risks are addressed through industry-led, and technical research works on model and data evaluation, fairness testing. 

Malt informs users when they are interacting with AI through product labelling (see for example the "AI Search" product), user-interface information through the AI page, and dedicated disclaimers when they are using the conversational AI where applicable. This informs users that AI is being used to assist the interpretation of project requirements, the retrieval and recommendation of freelancer profiles and, where applicable, the preparation of project information. 

Malt’s AI documentation explains the purpose of the AI features, the types of assistance they provide, their main limitations and the continuing role of the user. Users are informed that AI outputs are recommendations or drafts and should be reviewed. They can modify search criteria, edit or reject generated content, use manual search and challenge or report an inappropriate result.

The disclosure mechanisms are reviewed when new AI features or model versions are introduced, when the system’s degree of autonomy changes, or when user feedback, incidents or regulatory developments indicate that additional transparency is required.

Malt invests in research and technical development to improve the safety, security, fairness, transparency, explainability, reliability of AI systems used in the work and freelance domain. Malt’s research team includes in-house researchers with academic expertise (see more information here). Malt also collaborates with external academic and research communities. This includes participation in the EU-funded SEISMEC consortium, where the pilot develops and integrates evaluation tools that assess matching systems throughout their lifecycle. This makes risks visible early and supports evidence-based go or no-go decisions before changes are deployed at scale.

Key areas of research and investment include: (i) data quality and representativeness by studying the quality, coverage and limitations of data used for professional matching and ranking. This includes work on multilingual data, profile and project representation, historical selection effects and the risks of using marketplace outcomes as training or evaluation signals (eg. An Efficient Long-Context Ranking Architecture With Calibrated LLM Distillation: Application to Person-Job Fit.); (ii) fairness, safety and responsible use by investigating algorithmic bias and the potential effects of AI-assisted matching on access to professional opportunities. Research includes the evaluation of implicit preferences, group disparities, counterfactual profiles and the possible reproduction of historical or user-driven biases (eg. Evaluating LLM Behavior in Hiring: Implicit Weights, Fairness Across Groups, and Alignment with Human Preferences.); (iii) explainability and interpretability by reviewing how matching and ranking systems can be understood and evaluated, including the factors influencing relevance scores, model behavior and ranking outcomes. This supports technical diagnostics, user explanations and the design of appropriate human oversight (eg. WorkRB: A Community-Driven Evaluation Framework for AI in the Work Domain.).

Research outputs are shared publicly through papers, technical articles and conferences. More detailed evaluation results, datasets or security information may be shared confidentially with research partners, customers, auditors or regulators where appropriate. Some information remains internal where disclosure could create privacy, security, intellectual-property or competitive risks.

Malt has a strong corporate social responsibility culture since day one (more information can be found here). Malt researches and invests in methods intended to reduce the socio-economic and environmental risks associated with AI-assisted matching. 

With regards to socio-economic risks, Malt’s research examines how algorithmic matching can affect access to professional opportunities, and the distribution of work. This includes research on user preferences, group disparities, and the potential for recommendation systems to reproduce existing inequalities (see above). 

With regards to environmental considerations, Malt follows a task-appropriate model-selection approach. Large generative models are not used where a smaller, specialized model can perform the task adequately (see for example this tech conference talk). Malt continuously strengthens its work by measuring energy consumption and carbon impact across material AI workloads.

Malt promotes AI literacy through a combination of internal training, user education and public-facing information.

For employees, the Malt AI training path provides onboarding-level training on AI governance, responsible use of AI, privacy, information security and the organization’s AI policy. The AI training path also provides for role-based learning for technical, product, data, security, legal and operational teams. 

For freelancers and clients, Malt provides in partnership with continuous learning institutions specialized trainings “AI for all” where users are invited to automatize, analyze and innovate through AI (see this series of webinars for example). Furthermore, Malt hosts dedicated resources such as "AI for freelancers" to provide guidance/tips on AI literacy and "tips", as well as "AI projects" showcasing success stories. 

Malt also shares public research and technical explanations through its research website and engineering blog. These materials provide broader information on AI evaluation, fairness, matching, model behavior and the potential impact of AI on work and professional opportunities.

Malt pursues AI research and product initiatives that aim to generate broader societal benefits, particularly by supporting fairer access to professional opportunities, reducing bias in digital matching and improving transparency in AI-assisted work platforms. Examples include the Malt Trends Seniors report. These initiatives are intended to support more inclusive, transparent and accountable digital labor markets.