What the published work shows

Finding

Large organisations place AI among technology, security and compliance controls while using leadership directions, performance measures and incentives to increase its use. The board guidance examined includes no specific requirement to identify or test the written instruction supplied by the organisation and used to shape AI answers.

Evidence

The Classification Study examined published governance material from fifteen large listed organisations. AI appeared within technology, security, product, risk, conduct or compliance governance in every case, while the mechanisms used for professional advice were absent. The Mandate Study found published evidence of AI use being formalised through leadership instructions, performance assessment, incentives, training, integration into work and board expectations. The Unread Instruction examined nine current board and audit frameworks and applied three tests to each. General duties covering lists of AI uses, ownership, testing and monitoring appeared throughout, yet no framework required anyone to know who wrote the organisation's instruction, when it changed, or how the same important request performed under a different instruction or model version.

Evidential limit

These papers use published corporate material and public guidance. They cannot show every internal rule or board practice, and the Mandate Study does not establish whether requiring AI use improves work. The framework review excludes material behind login or membership barriers and applies only to the versions examined.

Finding

Public contracts place responsibility for AI output on the organisation or the person using it, although suppliers describe controls that can shape the output before it appears. European law also relies on designated people to recognise over-reliance, interpret the output correctly and intervene when required.

Evidence

The Liability Transfer examined thirty-seven public documents covering seven enterprise AI services. Every service assigned responsibility for output, its use or its review to the organisation or user, and every service described controls that could affect that output elsewhere in its published terms or documentation. The clauses assigning responsibility did not mention those controls. The Awareness Trap examined the human-oversight duties in Articles 14 and 26 of the EU AI Act alongside published psychological research on automation bias. The law requires designated people to understand limitations, remain alert to over-reliance, detect anomalies, interpret outputs and intervene. The research cited in the paper finds that awareness of automation bias does not reliably prevent the behaviour it describes.

Evidential limit

The contractual study covers public terms as they stood on its access date, while negotiated agreements and live product behaviour may differ. The Awareness Trap is a legal and psychological analysis. It does not test overseers in workplaces or measure how often the law prevents harm.

Finding

Organisations make AI visible through adoption figures, investment, product claims and changing corporate language. Measured contribution to business work and the arrangements through which senior leaders use AI remain much less visible.

Evidence

The Reporting Study coded 714 statements from twenty-two listed organisations. It found 404 statements about adoption and control, 308 product or marketing claims, and two statements meeting its threshold for measured AI contribution, both from one company. The Language of AI Deployment traced 155 verified statements from nine organisations and found expansion language dominant, with claims about replacing work sometimes followed by later qualification. A comparison of 150 annual filings found that language became more qualified, generic or inflated at a higher annual rate in the later observation period, although the study does not attribute that change to AI. The Executive Access Study found survey evidence that directors use AI, while published sources disclosed little about the tools or access arrangements involved.

Evidential limit

These studies analyse published language and disclosed behaviour. They cannot determine the quality of internal measurement, the performance of deployed AI, the cause of changes in filing language, or the private arrangements used by individual executives.

Finding

Written instructions added during deployment can sharply reduce clear recommendations. Narrower instructions can preserve recommendation-making while continuing to refuse requests that breach legal, procurement or compliance boundaries. Published product documentation also shows that any variation in controls is assigned in advance through a person's identity, group or function, with executive rank giving no special authority at the moment of use.

Evidence

The Subtraction Study kept the commercial model fixed and added one enterprise governance instruction. Clear recommendations fell from twenty-three of forty outputs to three, while qualified recommendations rose from one to twenty-one. The Separation Study repeated the comparison with narrower instructions. The full governance instruction produced clear recommendations in eight of twenty outputs, compared with eighteen under no added instruction, while the two narrow conditions produced nineteen and twenty. Across requests involving prohibited actions, those narrow conditions refused every full override and still supplied a permitted course in fifty-five of fifty-six outputs. The Override Rank Cannot Reach examined documentation for seven enterprise deployments. Two allowed authorised identities or groups to vary part of the safety controls during use, while no deployment documented an exception based on executive rank.

Evidential limit

The controlled studies use one model snapshot, fixed prompt sets and simulated written instructions. They do not establish how other models, later versions or live organisational deployments behave. Published product documentation may also omit features that exist in practice.