Apertura is a series of articles for industry leaders. Articles in this series are for subscription members only. Preview to the back catalogue is available here.

The Development

Microsoft Research has published a position paper, “The Origins of Artificial Intelligence in Natural Intelligence” by Ken Archer and Harald Wiltsche (Microsoft Research, May 2026), arguing that modern AI systems extend human intelligence rather than replace it. Grounding the claim in Edmund Husserl’s phenomenology, the paper distinguishes between statistical pattern extension (what AI does) and world-directed understanding (what humans possess), and argues that understanding AI as a derived form of human intelligence offers a more grounded basis for governance frameworks than either displacement or replacement narratives.

While the paper has not been fully embraced, the argument positions AI’s accountability structure as inherently human which may have future downstream implications. It is an impetus for executives to consider what role they see AI performing in their business and the regulatory frameworks which may impact them.

Share

Considerations

Board-level governance strategy must identify which regulatory architectures they are covered by, not assume convergence and be prepared for conflicting expectations.

• The EU AI Act’s high-risk provisions apply to stand-alone high-risk AI systems from 2 December 2027 and to high-risk AI embedded in regulated products from 2 August 2028, with administrative fines up to EUR 35 million or 7% of global annual turnover.

• China’s approach is characterised by a dense but fragmented web of CAC instruments, sector-specific data rules, and cross-border transfer provisions; board-relevant decisions (content takedown authority, model approval, data transfer classification) rest with CAC and provincial enforcement bodies whose decision-making process is not transparent, requiring real-time compliance monitoring rather than ex-ante legal prediction.

• Singapore’s Model AI Governance Framework (2019, with a Generative AI supplement published January 2024) remains consultative; organisations self-assess and report, with no enforcement mechanism.

• Japan enacted its first binding AI legislation in May 2025, the Act on Promotion of Research and Development and Utilization of Artificial Intelligence-Related Technologies, with the Basic AI Plan approved by Cabinet in 2026; the Act creates duties on AI business actors but carries no monetary penalties, making Japan promotional rather than prescriptive.

• South Korea’s AI Basic Act (Act on the Development of Artificial Intelligence and Establishment of Trust, effective 22 January 2026) is the region’s most structured mandatory framework, with administrative fines up to 30 million KRW; enforcement is subject to a one-year grace period.

• Australia’s December 2025 Policy for the Responsible Use of AI in Government binds non-corporate Commonwealth entities only; the National AI Plan explicitly abandoned the mandatory guardrails framework previously under development.

A claim of “regulatory alignment” across these architectures may be misreporting governance posture. Directors must identify which regime governs which AI application, and must account for regimes at enforcement stage now (South Korea), approaching enforcement (EU, 2027 to 2028), and actively retreating from mandatory frameworks (Australia). MAS (Singapore) and ASIC (Australia) have issued sectoral AI enforcement guidance; South Korea’s MSIT and Japan’s AI Strategic Headquarters have not yet levied penalties.

The extension model pares the labour displacement question from the governance obligation.

Microsoft’s reframing creates space for boards to consider two distinct questions: how/whether AI will displace labour (contestable and unresolved) and how to govern AI extension of human decision-making (mandatory regardless of displacement outcomes). A board that adopts the extension model can establish governance obligations that do not depend on labour forecasts. The question a director should ask:

“Do we govern AI as an extension of human judgment or as an autonomous tool with oversight or human in the loop intervention?”

Enterprise AI governance depends on system-level harnesses, not model-level safety, and the vendor stack beneath the AI application is in scope.

Microsoft’s paper argues responsibility emerges from “layered safeguards” and cannot be delegated to models. This aligns with how regulated enterprises deploy AI: through human review loops, audit trails, and decision harnesses. Boards that have invested in model alignment (fine-tuning, RLHF, constitutional AI) without establishing governance harnesses are protecting the wrong layer. A director should ask:

“What proportion of our AI investment is model-level safety versus system-level governance?”

The governance harness extends to the infrastructure stack: AI vendor contracts cannot be audited in isolation from hyperscaler agreements and hardware supply chain dependencies, particularly where geopolitical export controls or single-vendor concentration create exposure beneath the application layer.

Recommendations

• The board needs to be aware of jurisdiction-specific compliance inventory, scoped and time-boxed at the board level before execution begins.
  

  The audit committee commissions general counsel to map each AI deployment against applicable regimes, with particular urgency on South Korea (enforcement-eligible under AI Basic Act with grace period expiring January 2027), EU AI Act (stand-alone high-risk systems by December 2027), and Singapore MAS and ASIC sectoral guidance (enforcement already active).
  China mapping requires real-time monitoring rather than ex-ante legal opinion. Inventory scope boundaries (number of systems, jurisdictions, risk classifications) must be set and approved before general counsel begins work. Recommended timeline: board approval of scope and budget in next governance cycle, inventory delivery no later than end of the quarter following approval.

• “AI is a tool that extends human judgment within documented oversight bounds” can not be stated until the compliance inventory confirms that documented oversight bounds exist. If filed before the governance infrastructure is in place, that language creates material misrepresentation exposure under applicable securities law. The board’s decision point: does the organisation adopt “extension” or “autonomy” as the governing model? That decision should be ratified at board level, with general counsel sign-off that underlying governance harnesses are documented, before the language enters any regulatory filing or investor communication.
  Sequencing: narrative framing follows inventory findings.

• The board approves a vendor contract and insurance audit scoped to the full technology stack. General counsel reviews whether indemnification, limitation of liability, and insurance scope reflect extension-model accountability (liability flows through human decision-makers).
  This audit includes hyperscaler agreements and hardware procurement contracts: AI deployment liability cannot be isolated from the cloud infrastructure and semiconductor supply chain dependencies that govern the stack beneath it. If hyperscaler contracts carry geopolitical or export control exposure that is unmitigated, board liability extends to that layer. Timeline: begin after inventory findings are available, to ensure the audit remit reflects the specific exposure map.

Bottom Line

The framing and understanding of the role of AI Intelligence is evolving in parallel to developments in its application in the workplace. The extension framing offers boards a position between displacement narratives and innovation-first complacency.

The window to establish documented oversight bounds before the first enforcement cycle is measurable in quarters, not years. Directors who ratify a governance posture before the compliance inventory is complete are filing representations they cannot substantiate.

Leave a comment

Share