Research
AI is entering organisations through systems that do more than provide a model.
Governance instructions shape the answer, contracts assign responsibility for it, reporting determines which effects become visible, and internal structures decide who can alter the conditions under which it is produced.
Richard’s research examines where those parts stop lining up. Organisations report adoption without reporting changed capability. Vendors shape outputs while placing responsibility elsewhere. Governance frameworks rely on human oversight while leaving critical instructions and controls beyond ordinary view.
The studies use public records and controlled experiments to make those arrangements visible. Together, they form a body of work on what AI changes inside organisations, and where authority, judgement, and responsibility move as a result.
What organisations make visible
How corporate reporting changed across the early LLM period, and what it records about the organisational effects of AI.
The First Annual Reports of the LLM Era
A study of 150 annual reports from 50 large American companies, across 2019, 2022 and 2024. The rate of prose drift increased between 2022 and 2024, while its form shifted from sentence inflation towards greater hedging and framing. The study does not attribute that change to the use of language models.
Adoption Without Capability Reporting
A study of the prepared public reporting of 22 large companies. It recorded 404 Adoption and Containment statements, 308 Product and Marketing statements, and two Capability Reporting statements, both in Alphabet's reporting. The study concerns what reporting makes visible, and makes no claim about what the other companies achieved internally.
How AI becomes established
How the technology enters an organisation's governance, and what changes once its use becomes expected rather than permitted.
How Large Organisations Classify AI in Public Governance Documents
A study of 15 large listed organisations across seven kinds of governance material. AI appeared mostly inside technology, security, product, risk, conduct and compliance structures. None of the 15 published a visible mechanism for treating it as a source of professional reasoning or judgement.
The Mandate Study
A study of what happens once adoption moves from permission into expectation. It identifies five recurring mechanisms across the public record: leadership mandates, integration into performance review, incentives and career progression, institutionalised training and workflow integration, and board-level oversight. The cases were selected to identify documented mechanisms rather than estimate their prevalence, and the study does not test whether those mechanisms changed behaviour.
What deployment changes
The layer between model and user: how governance instructions alter output, what organisations are required to record, and who can vary the controls at the moment of use.
The Subtraction Study
The study generated 160 responses across ten business questions, four deployment conditions and four repeated runs. In the cleanest comparison, the same commercial model was run with and without a simulated enterprise-governance prompt. Definitive recommendations fell from 23 of 40 responses to 3 of 40, while qualified recommendations rose from 1 of 40 to 21 of 40. It measures changes in model output under added governance instructions and makes no claim about changes in user capability.
The Unread Instruction
A study of nine public board and audit governance frameworks, using 20 underlying documents. It asked whether they required an organisation to record who wrote the system prompt, when it was last changed, and whether critical queries were retested after a change to the prompt or the model. None of the 27 coded answers was positive.
The Override Rank Cannot Reach
A study of seven enterprise deployments, using 74 vendor and legal records. Two vendors documented partial mechanisms for varying part of the governance layer at the moment of use. Neither mechanism recognised executive seniority, and no comparable mechanism was found in the documents for the other five deployments.
Where authority and responsibility go
How authority is retained, how responsibility is assigned, and whether human oversight remains credible.
The Liability Transfer
A study of 37 public contractual and policy documents across seven enterprise AI services. Every vendor assigned responsibility for AI output to the customer or user, while separately documenting controls capable of affecting that output. None of the responsibility wording mentioned those controls.
The Awareness Trap
An examination drawing together the human oversight duties in Articles 14 and 26 of the EU AI Act with research on automation bias. It argues that routine reliance can weaken the attentiveness and independent judgement on which those duties depend.
Open and current work
How Senior Executives Access AI Capability Beyond Sanctioned Tools
The search found no verified named case of senior figures using local open-weight models and one instance of differentiated access under lighter governance. Because such arrangements may remain private, the paper treats this as a limit of public disclosure rather than evidence that the practices are absent.
The Sparring Partner Study
The study runs five strategic scenarios through three deployment conditions, producing fifteen responses in all. Each is reviewed blind and scored on whether it takes a clear position, offers fresh insight, and gives an argument a leader could actively debate. The grading rules were fixed before any response was generated. The testing method is published, and the results are still to come.
The full record
The completed studies and current research instruments are published at MKAI, with their methods, source records, and stated limits.
Go to MKAI →