The European Commission released a draft for conducting audits under the Digital Services Act (DSA) on May 6, 2023, which pertains to the 17 Very Large Online Platforms (VLOPs) and 2 Very Large Online Search Engines (VLOSEs, including Facebook, LinkedIn, Twitter, Bing, and Google). The purpose of this delegated regulation is to promote transparency and public accountability for large platforms, with provisions for annual independent audits. Algorithmic systems will be audited and will include disclosures and risk assessments. The draft clarifies the relationship between Audited Providers and Auditing Organizations, and lays down provisions for selecting auditors, data sharing, and cooperation. Auditing Organisations will send Final Reports, including Risk Analyses and Audit Conclusions, and must be completed within a year from the date of application of the obligations to the Audited Provider. The draft is open for public comments until June 2, 2023. The article promotes Holistic AI's interdisciplinary approach for AI governance, risk, and compliance.

New York City's Local Law 144 requires bias audits of automated employment decision tools, and employers and employment agencies who use these tools within the city must have a bias audit performed by an independent auditor. Vendors of these tools may not be directly affected by the legislation but may still be subject to audits if their clients meet the above criteria. Vendors are encouraged to proactively get an audit to mitigate potential issues and provide assurance to clients and prospects that their software is in compliance. A free consultation or quiz is available to determine if an audit is necessary. This blog article is for informational purposes only and is not intended to provide legal advice.

The New York City Council has mandated bias audits of automated employment decision tools (AEDTs) used to evaluate employees for promotion or candidates for employment in New York City. The NYC Bias Audit Law requires employers to make a summary of the results of the bias audit publicly available on their website, increasing transparency in the hiring process. The law requires testing for disparate impact against component 1 categories required to be reported by employers, including sex and race/ethnicity categories. The delayed enforcement deadline provides an opportunity to collect necessary data or use test data for the bias audit. The article advises early preparation by employers to ensure compliance with the law.

Policymakers in the US are starting to prioritize the regulation of automated employment decision tools and systems, with Illinois enacting the Artificial Intelligence Video Interview Act in 2020. New York City has passed legislation mandating bias audits of such tools and California has proposed amendments and new laws to regulate their use. The New York City legislation requires independent, impartial bias audits of automated tools used in hiring, assessment and promotion, as well as notification to candidates and employees of their use. California focuses on making it unlawful to use automated tools that discriminate on the basis of protected characteristics, and proposes restrictions on the electronic monitoring of employees. All legislation has strict notification, collection and data retention requirements for employers and vendors. Employers and vendors using AI employment tools are advised to adopt reliable systems of governance and auditing to avoid discrimination in their use and stay ahead of emerging regulations.

Large Language Models (LLMs) have come a long way, with pre-trained artificial intelligence models serving as a foundational base for a wide variety of applications and tasks. LLMs have immense potential, but it is essential to remain aware of equity, fairness, and ethical issues they may present, as well as the limitations they must overcome to develop Artificial General Intelligence (AGI). The widespread adoption of LLMs needs to be balanced with addressing potential risks to society and humanity. Data is becoming a significant constraint for optimal LLM performance, necessitating innovative approaches to balance model size and training tokens. Innovative approaches are currently addressing the limitations of current language models, providing alternative solutions to data constraints. Together, these strategies offer a promising path to overcome data constraints and improve the effectiveness and versatility of language models.