Weekly Digest
Week of 18 May 2026
This week at a glance
This week's digest is shaped by two converging dynamics: the elevation of AI to strategic diplomacy, and intensifying pressure on governance frameworks to keep pace with rapid deployment. The US-China bilateral AI safety dialogue and the Trump administration's forthcoming executive order on pre-deployment model access both signal that frontier model governance is hardening at the geopolitical level, with downstream implications for Australian export controls, procurement conditions, and vendor engagement as domestic arrangements continue to mature. On frameworks, the EU AI Act's high-risk classification guidelines are open for consultation until 23 June — directly relevant to any Australian agencies or vendors with EU-facing AI products — while the EU's transparency code of practice on AI-generated content moves toward a final draft in June. Practitioners advising on APS workforce and procurement governance will also find practical signal in the GoTo survey findings on skill atrophy and the California executive order's vendor certification requirements, both of which surface concerns that are increasingly live in the Australian federal context.
Headlines
Australian Government1 item
Newsom Signs Executive Order To Address AI Disruption
California Governor Gavin Newsom issued an executive order on 20 May 2026 directing state agencies to study and prepare for AI-driven job displacement, including reviews of severance standards, unemployment insurance, retraining programs, and experimental workforce-ownership models. Separately, the order requires the Department of General Services and the Department of Technology to submit AI vendor certification recommendations within 120 days, with vendors seeking state contracts required to demonstrate safeguards including watermarking of AI-generated media, bias mitigation, and civil-rights protections. The order is described as procedural and exploratory rather than codified law, with binding effect concentrated in procurement pipelines. It is framed as the first US governor-initiated AI worker-protection order of its kind.
Key points
- California Governor Newsom signed an executive order directing state agencies to prepare for AI-driven workforce disruption.
- The order tasks procurement agencies to develop AI vendor certification rules within 120 days, including watermarking and bias safeguards.
- This is a US state-level development with no direct Australian regulatory parallel, though procurement parallels are worth noting.
Implications
- Monitor APS procurement and AI policy teams may want to monitor the 120-day vendor certification recommendations for concrete technical requirements that could inform Australian whole-of-government AI procurement conditions.
- Consider Agencies working on AI workforce strategy or responsible AI procurement frameworks could consider how California's paired approach - vendor safeguards alongside workforce transition studies - compares to current Australian arrangements.
Global Regulation & Policy5 items
Trump and Xi Open AI Safety Dialogue
President Trump and Chinese President Xi Jinping placed AI safety on the agenda at their Beijing summit, with US Treasury Secretary Scott Bessent confirming the two countries will begin formal dialogue. Reporting from Reuters and the Wall Street Journal indicates a bilateral mechanism involving Treasury and Chinese finance counterparts is under discussion, focused on protocols for advanced model handling and limiting non-state actor access rather than a binding legal treaty. Urgency was reportedly linked to the launch of Anthropic's Mythos model and its early-access controls. No substantive commitments are expected in the near term, but the elevation of AI to strategic diplomacy increases the likelihood that export controls, chip restrictions, and frontier-model access conditions will tighten globally.
Key points
- Trump and Xi placed AI safety on the Beijing summit agenda, with Treasury-led bilateral dialogue being discussed.
- Talks focus on access controls, best practices for advanced models, and limiting non-state actor access - not a binding treaty.
- Outcome mechanisms, if formalised, could reshape export controls, chip access, and frontier-model procurement conditions globally.
Implications
- Monitor Policy and strategy teams may want to monitor whether the proposed US-China dialogue mechanism produces concrete protocols on export controls, chip access, or frontier-model incident notification, as these would affect Australian government and research access to advanced AI infrastructure.
- Consider Agencies with procurement dependencies on frontier AI models or cloud infrastructure could consider how tightening US export-control regimes might affect vendor availability and terms for Commonwealth entities.
Trump issues executive order on AI oversight
Reuters reports that President Trump is expected to sign an executive order establishing a voluntary framework under which AI developers would provide covered models to the US government 90 days before public release and give pre-public access to critical infrastructure operators such as banks. The order reflects ongoing tension between tech-sector resistance to mandatory requirements and calls for stricter oversight from parts of the political base. Because the framework is voluntary rather than statutory, its practical effect on AI developer behaviour remains uncertain. For Australian practitioners, the development is notable as a US-led precedent for pre-deployment government access to frontier models, an area where Australian arrangements are still maturing.
Key points
- Trump is expected to sign an executive order creating a voluntary pre-release AI disclosure framework for US government and critical infrastructure providers.
- The 90-day pre-public model access window sets a US precedent that could influence Australian pre-deployment safety assessment expectations.
- The framework is voluntary, limiting its direct regulatory force - Australian agencies should note this distinction when tracking US AI governance signals.
Implications
- Monitor DISR and AISI policy teams may want to monitor the final EO text and any implementing guidance, particularly definitions of 'covered models' and how voluntary uptake is tracked.
- Consider Agencies involved in Australia's frontier AI strategy could consider how a US voluntary pre-release access norm might inform future expectations for bilateral AI safety cooperation or domestic pre-deployment assessment arrangements.
Commission seeks feedback on the draft guidelines for the classification of high-risk artificial intelligence systems
The European Commission has published draft guidelines to clarify how AI systems should be classified as high-risk under the EU AI Act, accompanied by practical examples. The guidelines target providers, deployers, businesses, public authorities, and researchers, and are open for stakeholder consultation until 23 June 2026. They sit alongside other forthcoming Commission guidelines on compliance obligations for high-risk AI systems. For Australian agencies and vendors with EU-facing AI products or services, this guidance may directly affect how those systems must be assessed and documented.
Key points
- The European Commission has released draft guidelines clarifying which AI systems qualify as high-risk under the EU AI Act.
- Stakeholder feedback is open until 23 June 2026 - Australian AI providers operating in EU markets may be directly affected.
- Guidelines include practical examples to help providers and deployers self-assess high-risk classification obligations.
Implications
- Monitor Policy teams tracking international AI regulation may want to review the draft guidelines as a reference point for how high-risk AI classification is operationalised in a leading regulatory jurisdiction.
- Consider Australian agencies or vendors with AI systems deployed in EU contexts could consider whether to submit feedback before the 23 June 2026 consultation deadline.
AI Act transparency code of practice - third round of working group meetings
The EU AI Office convened a series of working group meetings and workshops in March 2026 to gather stakeholder feedback on the second draft of its Code of Practice on Marking and Labelling of AI-Generated Content, under Article 50 of the EU AI Act. Working Group 1 addressed marking and detection obligations for providers, including watermarking, metadata, and multi-modal approaches. Working Group 2 focused on disclosure obligations for deployers of deepfakes and AI-generated text, including user-facing labels and a proposed uniform EU icon. The third and final draft is expected in early June 2026, and will inform binding transparency requirements under the AI Act.
Key points
- EU AI Office held third-round stakeholder meetings to finalise the Code of Practice on AI-Generated Content transparency.
- The final draft covering marking, watermarking, deepfake disclosure, and labelling obligations is expected in early June 2026.
- Debates centre on mandatory versus voluntary measures and compliance burden - tensions likely to recur in any Australian equivalent framework.
Implications
- Monitor Policy teams working on AI transparency or synthetic media governance may want to monitor the final June draft for disclosure and labelling approaches that could inform Australian thinking.
- Consider Agencies developing guidance on AI-generated content could consider how the EU's multi-layered marking framework and uniform labelling icon compare to any emerging Australian standards or expectations.
The European Union is deploying AI across strategic sectors
An OECD AI Wonk Blog post examines how the European Union is deploying AI across four strategic sectors - healthcare, manufacturing, mobility, and agriculture - framed around the EU's trustworthy AI principles and competitiveness agenda. The extracted content is limited to a brief summary, so the depth of analysis, specific case studies, and policy mechanisms covered in the full post are not assessable from this item alone. As a signal, it reflects the OECD's ongoing documentation of government-led AI deployment patterns across member and partner economies.
Key points
- The EU is deploying AI across healthcare, manufacturing, mobility, and agriculture under a trustworthy AI framework.
- OECD coverage signals this is a notable comparative case study in sectoral AI deployment by a major jurisdiction.
- Extracted text is minimal - full substance is behind the link and not available for detailed assessment.
Implications
- Monitor Policy teams working on sectoral AI strategy may want to review the full post for comparative deployment patterns relevant to Australian sector-specific AI initiatives.
Risk, Assurance & Ethics4 items
Workers Report Skill Atrophy Amid Heavy AI Use
A survey commissioned by IT firm GoTo of 2,500 global workers and IT leaders found 82% of workers use AI on the job, with 39% reporting that AI reliance has weakened their skills and 28% now trusting AI more than themselves. Roughly one in four IT leaders said AI-related mistakes had already affected customers or business outcomes. The findings are self-reported and vendor-commissioned, limiting causal inference, but they surface a pattern relevant to APS workforce AI governance: high adoption coexisting with measurable worker anxiety about competence erosion and documented error propagation in human-machine workflows.
Key points
- GoTo-commissioned survey of 2,500 global workers finds 39% report AI use has weakened their skill sets.
- Nearly one in four IT leaders report AI-related mistakes have already affected customers or the bottom line.
- Survey is vendor-commissioned and measures self-reported perceptions, not objective skill decline - treat with appropriate caution.
Implications
- Consider APS workforce and AI governance teams could consider whether current AI deployment frameworks include mechanisms to monitor staff over-reliance and maintain human judgement independent of tool outputs.
- Monitor Agencies may want to monitor emerging research and peer-reviewed studies on cognitive offloading to supplement vendor-commissioned survey findings on this topic.
Establishing the shared foundations for collective AI security
An OECD AI Wonk Blog post addresses shared international foundations for AI security, with stated coverage of prompt injection, AI agents, and model poisoning. The framing suggests an interest in collective, cross-jurisdiction approaches to secure AI deployment. However, the extracted text is limited to a brief teaser, making it impossible to assess the depth, recommendations, or any concrete standards implications without reading the full piece.
Key points
- OECD AI blog addresses shared foundations for collective AI security across member nations.
- Covers prompt injection, AI agents, and model poisoning - security risks relevant to Australian government AI deployments.
- Extracted text is minimal; full substance of the piece is not available for detailed assessment.
Implications
- Monitor Risk and assurance teams may want to read the full piece to assess whether OECD's framing of AI security foundations aligns with or informs Australian government secure-deployment approaches.
- Consider Agencies developing AI security guidance could consider whether OECD collective-security framing is useful reference material for internal risk assessments covering agentic AI and model integrity.
Expanding our AI and Healthcare Portfolio
The AI Now Institute has announced an expanded research portfolio focused on AI deployment in US healthcare, examining patient safety, worker conditions, regulatory adequacy, and the political economy of health AI vendors. The work critiques both the harms of current AI tools - including ambient scribes that fabricate clinical notes, chatbots missing drug allergies, and algorithmic staffing platforms - and the regulatory environment that leaves many tools outside FDA or HIPAA scrutiny. A first publication, 'Uber for Nursing Part II,' examines algorithmic gig nursing platforms. Future work will cover Epic Systems, HCA Healthcare scheduling software, and the Oxevision patient monitoring system. The research is participatory, drawing on healthcare worker collaboration and union partnerships.
Key points
- AI Now Institute is launching a dedicated research portfolio examining AI deployment risks across US healthcare systems.
- Key concerns include patient safety failures, workforce displacement, regulatory gaps, and corporate consolidation - relevant to Australian health AI governance debates.
- Focus is US-specific; Australian health AI governance context differs, limiting direct applicability for APS readers.
Implications
- Monitor Health and human services policy teams may want to monitor AI Now's forthcoming publications for independent evidence on healthcare AI harms, particularly around ambient scribes and algorithmic staffing tools that are also being adopted in Australian settings.
- Consider Agencies working on AI governance frameworks for the health sector could consider whether the regulatory gap analysis - especially around tools that fall outside existing medical device and privacy regulation - has analogues in Australia's TGA and Privacy Act landscape.
HCLTech Warns 43% of Enterprise AI Initiatives May Fail
HCLTech's 2026 AI Impact Imperatives report, drawing on a survey of 467 senior executives across G2K organisations in 10 countries, identifies a significant gap between AI adoption and measurable business impact. Key findings include a 10-month median payback expectation, rising interest in agentic and physical AI, and 76% of respondents citing Responsible AI concerns as a cause of deployment delays. The report's credibility is somewhat undermined by a factual inconsistency between the press release (43% expected failure rate) and the report's own key highlights page (24%), which warrants caution when citing findings. The themes - governance friction, technical debt, and partner dependency - are nonetheless relevant to APS AI deployment practice.
Key points
- HCLTech survey of 467 G2K executives finds 24-43% of major AI initiatives expected to fail (figures conflict across sources).
- 76% of surveyed executives say Responsible AI concerns have delayed deployments - a tension familiar to APS agencies.
- Private-sector vendor survey with methodological inconsistencies; limited direct applicability to Australian government context.
Implications
- Consider APS practitioners managing AI deployment pipelines could consider whether the governance and integration friction patterns described align with their own agency experience, particularly for agentic AI pilots.
- Monitor Agencies may want to monitor emerging enterprise metrics - model-level SLAs, responsible AI review lead times - as practical indicators for their own AI program management.
Technical Developments1 item
Anthropic’s Code with Claude showed off coding’s future—whether you like it or not
At Anthropic's Code with Claude developer event, the company demonstrated how far autonomous AI coding has progressed. Most software at Anthropic is now written by Claude Code, often without human review of the underlying code. Anthropic's stated goal is for Claude to self-check and self-correct, removing humans from the error-resolution loop entirely. A newly announced 'dreaming' feature enables coding agents to build institutional memory across tasks on a given codebase, improving over time. This signals a shift in what software development — and developer accountability — looks like across the industry.
Key points
- Anthropic's Claude Code now ships pull requests autonomously, with most Anthropic software written by Claude without human review.
- A new 'dreaming' feature allows coding agents to consolidate notes across tasks, improving performance on familiar codebases over time.
- APS agencies relying on software procurement or in-house development should be alert to what 'AI-written code' means for assurance and auditability.
Implications
- Consider APS agencies with in-house software development or vendor-delivered digital products may want to consider whether existing code review and assurance practices are adequate when AI generates code without human inspection.
- Monitor Policy and risk teams could monitor how major vendors' adoption of autonomous coding agents affects software supply chain risk, auditability, and accountability claims in government procurement contexts.
Implications are AI-generated. Starting points, not advice — see methodology for how they're framed.