The Weekly Scan for week of 7 September ’26: responsible AI authority
This week’s stories range from fundraising guidance and grantmaking decisions to healthcare regulation, public services and frontier AI risk. What connects them is authority: who is allowed to advise, decide, approve, automate, challenge or act, and what needs to be in place before that authority moves.
The risks are not equivalent, but the organisational requirement is consistent. Principles need to establish the boundaries, people need the confidence and knowledge to exercise judgement, and capability needs to make testing, monitoring and accountability part of the work rather than an afterthought.
FUNDRAISING REGULATION
United Kingdom
Fundraising Regulator · UK · 7 September 2026
Fundraising advice is now part of the regulatory picture
The Fundraising Regulator has strengthened its message about the use of AI in fundraising. Its guidance already applied throughout the AI lifecycle, from exploring a possible use through to regular deployment. The new development is an explicit call for charities, sector bodies, academics, commercial providers and consultants offering AI advice to acknowledge and, where appropriate, integrate that guidance.
The regulator’s concern is that advice may be incomplete or incompatible with the Code of Fundraising Practice if it ignores the regulatory position. It also makes clear that charities cannot outsource their understanding: fundraising organisations should know the guidance themselves and take it into account when exploring, preparing or using AI.
Why it caught our eye: Curia AI’s first piece of work began with exactly this problem. We synthesised guidance from the Fundraising Regulator, the Information Commissioner’s Office and the Charity Commission into a maturity assessment designed to reveal the gap between what regulators expected and what a charity actually had in place.
That cross-regulator view matters because AI rarely sits inside one neat regulatory box. A fundraising use may involve data protection, trustee accountability, public trust and the Fundraising Code at the same time. Advice that covers only the technology, or only one regulator, can leave an organisation with a false sense of readiness.
This is not a new law or an enforcement decision. It is a clarification of expectations, particularly for those influencing charity practice. Its practical value is the reminder that trusted advice should help an organisation join up the relevant obligations, while boards and leaders retain responsibility for understanding the standards that apply to them.
ORGANISATIONAL ADOPTION
United Kingdom
Ada Lovelace Institute · UK · 8 September 2026
NICE work shows responsible adoption is an operating-model challenge
The Ada Lovelace Institute has published the findings of a research collaboration with the National Institute for Health and Care Excellence on the ethical and effective use of AI in internal processes.
The work included a review of existing guidance, interviews with 11 NICE staff, observation of workshops, a case study of an LLM-assisted screening protocol and a workshop involving 22 external experts. It found that apparently low-risk internal uses can still affect evidence quality, accountability, workforce expertise, inequalities and the services people ultimately receive.
The recommendations go well beyond technical accuracy. They call for clear organisational objectives and risk tolerances, end-to-end evaluation, meaningful baselines, defined integration pathways, staff support, stakeholder engagement, transparent success measures and monitoring as systems and workflows change.
Why it caught our eye: The findings align closely with the way Curia AI approaches responsible adoption through Principles, People and Capability.
Principles appear in the need for a strategic narrative connecting AI decisions to organisational values, intended outcomes and an explicit tolerance for experimentation and risk. People appear in workforce support, human accountability, stakeholder participation and safe ways for staff to raise concerns. Capability appears in workflow design, baselines, evaluation, monitoring and reporting.
These elements cannot sensibly be developed in sequence. A clear principle without the people able to apply it remains an aspiration. Training without an agreed purpose or operating boundary can simply accelerate inconsistent use. Technical capability without ongoing evaluation may improve a task while weakening the quality or legitimacy of the wider process.
The report does not show that NICE has deployed these uses or achieved faster decisions, and its qualitative evidence base is deliberately limited. Its value lies in showing why responsible AI is an operating-model question, including when the system is internal and the effect on the public is indirect.
PHILANTHROPY AND GRANTMAKING
United States / Global
Coefficient Giving · US and global philanthropy · 9 September 2026
A grantmaker is speeding up decisions as AI safety funding accelerates
Coefficient Giving says its commitments to AI safety, security and field-building increased from $168 million in 2024 to $351 million in 2025 and are on track to exceed $1 billion in 2026. It also says its biosecurity commitments will exceed $150 million this year.
The scale is striking, but so is the operating model behind it. The funder says it is making larger grants, shortening investigations and moving decisions closer to the grantmakers with the most context. A typical grant now needs active sign-off from only one person in addition to the primary investigator.
Recent decisions include a $160 million grant for a new organisation undertaking alignment research. A separate initial grant of $10 million to establish an institute bringing mathematicians into AI safety research was approved 12 days after the first conversation.
Why it caught our eye: This is not simply a story about more philanthropic money moving into AI. It is about how a grantmaker governs speed and delegates authority when it believes the window for action may be narrowing.
Reducing unnecessary delay can be valuable, particularly when conventional processes consume time without materially improving a decision. But a larger, faster and more concentrated funding model also raises questions. Which challenges and conflicts checks remain non-negotiable? How is independent judgement preserved when a small number of people hold substantial delegated authority? What evidence is gathered after an accelerated decision, and how does the funder change course if its underlying assumptions prove wrong?
The figures and account of the process come from the funder itself. “Committed” means formally recommended for funding, not necessarily money already paid, and the case for urgency relies partly on contested forecasts about the arrival of transformative AI. The useful lesson is therefore not that faster grantmaking is automatically better. It is that speed should be an intentional design choice with clear accountability, not the accidental result of urgency.
NONPROFIT IMPLEMENTATION
United States
Bellwether · US · 10 September 2026
Thirty-plus projects point back to strategy, workflows and staff learning
Bellwether has drawn together lessons from two years of work with more than 30 schools, nonprofits and youth-serving organisations developing practical AI plans.
It identifies three recurring requirements. First, organisations need a coherent, mission-aligned strategy that explains what AI is intended to support, what will be prioritised and what will not be pursued. Second, broad ambitions need to be translated into redesigned and measurable workflows. Third, adoption needs to be treated as a challenge of trust, organisational change and adult learning rather than a technology rollout.
The authors argue that staff need time to build skills, test new practices, reflect on results and help shape the organisation’s approach. Early adopters can support experimentation, while sceptical colleagues may surface important questions about privacy, equity, relationships and risk.
Why it caught our eye: The message is refreshingly practical. Finding a tool is rarely the hardest part. The more difficult work is agreeing where AI fits, choosing a workflow worth changing, defining human ownership and giving people enough shared capability to judge whether the change is actually better.
It also challenges the assumption that scepticism is simply resistance to be overcome. In a responsible adoption process, constructive challenge is a capability. Leaders need to make room for it while still helping the organisation move from disconnected experimentation to deliberate choices.
This is practitioner experience rather than an independent evaluation. The organisations are not named and the article does not report methods or outcome measures for the 30-plus projects. The lessons are credible and transferable, but should not be read as proof that a particular implementation model has delivered measurable impact.
HEALTHCARE REGULATION AND TRUST
United Kingdom
National Commission into the Regulation of AI in Healthcare and the Health Foundation · UK · 10 September 2026
Public support for healthcare AI depends on what happens after approval
The independent Commission established by the Medicines and Healthcare products Regulatory Agency has published recommendations for the future regulation of AI-enabled healthcare technologies.
Its central conclusion is that regulation needs to become more proportionate, lifecycle-based and system-wide. AI systems may change after deployment, perform differently in different settings and depend on the data, workflows, people and organisations around them. A one-off assessment before market entry is therefore not enough. The Commission recommends stronger use of real-world evidence, continuing monitoring and clearer responsibility across the system.
Research undertaken by the Health Foundation reinforces the trust dimension. Public support for AI in healthcare is conditional rather than automatic. Accuracy, meaningful human oversight, transparency, accountability and avoiding worse care for particular groups are central to whether people regard its use as legitimate.
Why it caught our eye: Health and care charities may use AI themselves, support people affected by AI-enabled decisions or operate alongside NHS systems. They therefore have a role not only as adopters, but as informed partners and advocates for the people whose experience may reveal problems that headline performance measures miss.
The wider lesson is that approval is an event, while assurance is a continuing capability. Organisations need to know how a system performs in practice, who watches for change, how people can question an outcome and what happens when evidence from one setting does not transfer to another.
These are recommendations from a non-statutory advisory commission rather than current law, and a cross-government response is still to come. They nevertheless provide a useful direction of travel: proportionate regulation should mean matching scrutiny to risk throughout the lifecycle, not relaxing oversight once a system has been deployed.
Read the Commission’s recommendations → · Read the Health Foundation research →
PUBLIC-SECTOR TRANSPARENCY
United Kingdom
UK Algorithmic Transparency Recording Standard · UK · 9 September 2026
Government records show what operational transparency can look like
Four new records describe algorithmic tools operating within the Ministry of Justice, HM Courts and Tribunals Service and HM Revenue & Customs. The records cover ownership, third-party suppliers, operational data, human review, risk controls and completed impact assessments.
One example is Luna, a voice assistant used by the Prison Enquiry Centre to answer routine questions from prisoners’ friends and families. It currently handles around 2,900 contacts a month and is expected to increase to roughly 60,000. Calls are transferred to a human where the system cannot identify the question, detects a possible safeguarding issue or the caller asks for an adviser.
Another record, BenchNotes, describes real-time transcription of oral decisions in immigration and asylum tribunals. It identifies the use of special-category data, the assessments undertaken and the continuing responsibility of judges to review and correct the resulting transcript.
Why it caught our eye: Transparency is sometimes reduced to telling people that AI is being used. These records show a more useful form: naming the owner, describing what the system actually does, recording the data and suppliers involved, identifying the circumstances for human intervention and documenting which assessments have been completed.
That approach is transferable to charities, particularly where an automated service affects access to support or handles sensitive information. A meaningful record should allow someone outside the project team to understand the workflow, the boundary of the system’s authority, the available route to a person and the evidence used to justify deployment.
The records are written by the organisations operating the tools, not independent evaluators. Some fields are incomplete, and projected volumes should not be confused with achieved impact. Publishing a record does not prove a system is effective or safe. It does, however, create a clearer basis for scrutiny, learning and challenge.
FRONTIER AI GOVERNANCE
United States / Global
Associated Press and The Guardian · US and global · 9 September 2026
An Anthropic resignation exposes the tension between safety and the race to scale
Researcher Jacob Coxon resigned from Anthropic after three years working at Anthropic and OpenAI. He said neither company was acting responsibly and argued that competitive pressure was driving the leading AI companies towards self-improving superintelligence without adequate control.
The coverage can make the episode sound like one statement, but two people were involved. Coxon resigned and issued the wider warning. Evan Hubinger, Anthropic’s alignment science lead, responded separately that he personally believed there was a greater than 10 per cent chance AI could kill all humans within the next decade. He also said Anthropic did not have a plan for aligning superintelligence and was not clearly on track to solve the problem.
Why it caught our eye: A probability of human extinction is inevitably the headline, but it should not be reported as though it were a measured finding. There is no accepted method for producing a reliable ten-year probability of this kind, predictions about superintelligence vary widely and Coxon subsequently said current models do not present an immediate extinction risk.
The more useful governance question is why people close to frontier development believe competitive incentives may be moving faster than control capability. Organisations should be able to hear serious internal challenge, distinguish present operational risks from uncertain future scenarios and decide what evidence would justify slowing, stopping or changing direction.
The warning deserves attention without requiring us to accept the percentage as fact. Dismissing it because the scenario is uncertain would be complacent; presenting it as a forecast would be alarmist. Responsible leadership sits between those positions: examine the assumptions, preserve routes for dissent and ensure that the pressure to move quickly does not override the organisation’s own readiness or principles. But let’s hope governments around the world take these warnings seriously and put competitive interests aside to agree how increasingly powerful AI should be developed and controlled.
Read the Associated Press report → · Read The Guardian’s analysis →
NONPROFIT IMPLEMENTATION
United States
Chronicle of Philanthropy · US · 10 September 2026
A frontline nonprofit tests whether AI can return time to people
First Place for Youth supports young people transitioning out of foster care or experiencing homelessness. With 78 per cent of its workforce providing direct services, the nonprofit identified administrative work around case meetings as a contributor to staff pressure and overtime.
As part of an AI accelerator involving 18 economic-mobility nonprofits, it developed “Note Ninja”. Staff can speak or type an account of a meeting and receive a structured summary identifying focus areas and possible next steps.
The initial pilot involved eight employees over three weeks. All eight reported reductions in administrative burden and burnout, while 80 per cent said they had more time to work with young people and were no longer working overtime. When use became voluntary during the second week, all eight chose to continue.
Why it caught our eye: This is a practical example of AI being tested as additional capacity rather than introduced as a general technology initiative. The organisation selected a specific workflow, involved both frontline and leadership perspectives, kept the pilot small and created a point at which staff could decide whether the tool was genuinely useful.
It also demonstrates why Principles, People and Capability have to develop together. The principle was clear: technology should create more time for human support, not replace it. People closest to the work helped define and test the tool. Capability came through hands-on learning, a bounded pilot and measures connected to the problem the organisation was trying to solve.
However, these are promising early signals, not evidence that staff burnout has been solved. The pilot involved eight people over three weeks, the measures were self-reported and the account comes from the organisation itself. There is no information about changes in actual hours worked, longer-term adoption or service outcomes.
The most important unresolved issue concerns the information being processed. The tool turns accounts of meetings with potentially vulnerable young people into case-management summaries, yet the article says only that data privacy remains under consideration. It does not explain what information is entered, where it is processed, how long it is retained, how summaries are checked or what safeguards apply before content becomes part of a person’s record. Those questions need resolving before wider deployment, not after it.
The funding lesson is also relevant. The organisation benefited from sponsored participation, coaching and free access to technology. If funders want nonprofits to realise value from AI responsibly, they may need to fund the learning, workflow design and assurance capacity required to test it properly, rather than paying only for the tool.
A CURIA AI PERSPECTIVE
Responsible AI is tested when authority starts to move
The common thread across this week’s stories is not a particular model or one category of risk. It is the movement of authority.
The Fundraising Regulator is asking organisations that influence charity practice to align their advice with sector guidance, while reminding charities that they remain responsible for understanding it. NICE is considering what happens when AI enters internal evidence workflows that may shape decisions far beyond the back office. Coefficient Giving is moving substantial grantmaking authority closer to individual investigators in order to act faster.
Bellwether’s work shows that staff need enough shared understanding to redesign real workflows and challenge the assumptions behind them. The healthcare commission argues that an approval at one moment cannot provide assurance throughout a changing system’s life. The new public-sector records show how organisations can make ownership, data, escalation and review more visible once a tool is in operation.
First Place for Youth turns those implementation principles into a named frontline example. Its pilot asked whether a simple tool could return time to staff without replacing their judgement. But even a note-taking tool moves a less visible form of authority: it helps decide how a human conversation is translated into an organisational record, which details are emphasised and which next steps are surfaced. Staff therefore need to remain responsible for checking what has been included, omitted or inferred. The early results are encouraging, but the unresolved privacy and assurance questions belong inside the pilot because the tool processes accounts of meetings with potentially vulnerable young people.
The Coxon resignation takes the question to its most difficult frontier edge. The warnings about human extinction should not be presented as a measured forecast, but they still raise a legitimate governance concern about whether competitive pressure is moving capability faster than the institutions expected to direct and constrain it.
That is why responsible AI cannot be reduced to a policy or an ethics statement. It requires Principles, People and Capability to develop together.
Principles establish what the organisation is trying to achieve, whose interests matter, what it will not delegate and how much uncertainty or risk it is prepared to accept. They give people a basis for deciding when speed is justified and when the responsible decision is to pause.
People bring context, judgement and accountability. They include the person who owns the outcome, the colleague who notices a problem, the beneficiary who experiences an effect the project team did not anticipate and the specialist who can challenge an assumption before it becomes embedded. A culture that welcomes informed challenge is part of responsible adoption. It helps an organisation identify weak assumptions before they become operational problems.
Capability turns intent into repeatable practice. It includes cross-regulator interpretation, workflow design, data governance, staff learning, source checking, evaluation, monitoring, transparent records, incident response and the ability to change or stop a system when the evidence changes.
These responsibilities do not automatically require a new team for every use of AI. Where technology releases capacity, some of that time can be redeployed into quality assurance, exception handling, outcome evaluation and improvement of the workflow. Where genuine specialist capacity is needed, that requirement belongs inside the value case rather than being discovered after deployment.
Proportionate governance also means resisting false equivalence. A controlled assistant answering routine questions does not require the same safeguards as a system shaping access to healthcare, an accelerated grant of $160 million or a frontier model connected to consequential tools. The right question is not whether every use has passed the same checklist. It is whether the organisation understands the authority being delegated, the plausible impact if it goes wrong and the evidence and intervention needed for that particular context.
The first Curia AI maturity assessment was designed to expose the space between regulatory expectations and organisational reality. This week’s stories show that the same gap appears in many forms: between a value and a workflow, an approval and continued assurance, a commitment to safety and the incentives shaping behaviour.
Closing that gap is the work of responsible AI. The objective is not to eliminate uncertainty or stop organisations benefiting from new capability. It is to ensure that, as AI is given more influence or autonomy, the organisation becomes better able to decide what it should do, understand what it is doing and intervene when the outcome falls outside the boundary.
