Issue #46 — Co-Improvement, Care Logs & National AI Plans

From self-improving systems to shared, human-in-the-loop superintelligence

By Freddie Seba | GenAI Ethics & Governance for Leaders.

For board members, executives, clinicians, educators, and trustees working at the intersection of GenAI, health, and higher education.

If this was forwarded to you, you can subscribe to future issues to follow along with GenAI ethics and governance for leaders, boards, and university trustees.

This week at a glance

For busy readers, here’s the short version:

Care & documentation: A new Journal of Biomedical Informatics study tests whether GPT can turn nursing conversations into structured event logs. It’s a glimpse of AI-generated “evidence” behind staffing, quality, and reimbursement—and a reminder that event data itself needs governance.
Research & co-superintelligence: A new paper argues for shifting from fully autonomous self-improvement to co-improvement—humans and AI as research partners on the path to “co-superintelligence.” This has direct implications for universities, health-system research, and the boards that oversee them.
National policy & governance: Australia’s National AI Plan offers a whole-of-economy framing around capturing opportunities, spreading benefits, and keeping people safe. It doubles as a template for institutional AI strategies.
New focus this week: Recurring prompts and language explicitly tuned for boards of directors and university trustees: what to ask about AI-derived metrics, research collaborations with AI systems, and institutional AI plans.

A brief note at forty-six

Forty-six issues in, the conversation about AI has shifted.

We’re no longer asking whether GenAI “works.” We’re asking what happens when it starts to generate the very traces we use to understand care, learning, and performance—and what it means to aim for “superintelligence” that is fundamentally shared between humans and machines.

This week, three threads cross:

A nursing study that tries to turn messy, real-world home-care conversations into structured event logs with GPT.
A research agenda that argues for human–AI co-improvement, not just AI systems rewriting themselves in the background.
A national AI plan that makes explicit choices about who benefits, who bears risk, and how governments intend to keep people safe.

Taken together, they raise a governance question for boards and trustees:

If we say we want human-centered AI, are we actually designing for human–AI co-work, or just hoping that humans can bolt themselves on afterward?

1. For healthcare & health systems leaders

Turning Dialogues Into Event Data: GPT for Nursing Actions

Article: “Turning Dialogues Into Event Data: Lessons From GPT-Based Recognition of Nursing Actions,” Journal of Biomedical Informatics (2025). ScienceDirect: https://www.sciencedirect.com/science/article/pii/S1532046425001868

A Dutch research team asks a deceptively simple question:

Can a GPT-class model listen in on real home-care visits between district nurses and patients, and turn those conversations into structured event logs based on the Nursing Interventions Classification (NIC)?

In other words, can we get process-mining-ready data from speech alone?

What they did

Collected transcripts from audio-recorded home-care visits.
Used GPT to identify nursing actions and map them to NIC-based activity labels.
Constructed event logs (who did what, in what order) suitable for process mining and workflow analysis.

What they found

GPT performs reasonably well when actions are explicitly stated and temporally precise: “I’ll change your dressing now, and then check your blood pressure.”
It struggles when key information is implicit, ambiguous, or missing from the conversation—especially around timing and context.
Even when labels look “right,” granularity and sequence can drift from what actually happened, which matters if you’re using the logs for:
- Quality measurement
- Staffing and workload analyses
- Reimbursement or compliance

From these experiments, the authors propose practical guidelines for using large language models to create event data from dialogue, including:

Be explicit about data source limits (conversations are not the whole chart).
Use and iterate on well-designed taxonomies (like NIC) rather than ad-hoc labels.
Represent uncertainty instead of forcing binary decisions.
Detect and manage hallucinations and “invented” actions.
Tailor outputs to different stakeholders (clinicians, managers, analysts) who each need different views of the same data.

What this means for leaders

If your organization is experimenting with:

Ambient documentation
AI scribes and agentic charting
AI-powered process mining or workflow analytics

…this paper is a reminder that you are not only deploying AI; you are creating new types of clinical data.

That means:

Event logs become governed artifacts, not side effects. If AI-generated logs feed into staffing, quality, or pay-for-performance, they deserve validation, monitoring, and clear ownership.
You must decide what level of wrong is acceptable where. Misclassifying a patient education moment is not the same as misrepresenting a medication administration or omission.
Nurses and frontline clinicians must have a voice. They are closest to the conversation, the chart, and the consequences.

For boards and trustees in health

If you sit on a health system board or similar body, consider asking:

Where in our system are AI-generated event logs, summaries, or metrics already being used in management or board reporting?
What is our validation and ongoing audit plan for these AI-derived structures? Who is accountable for it?
How are nursing and allied health leaders involved in approving the use of conversational data for analytics or performance management?

2. For research, higher education & “co-improvement” leaders

From Self-Improvement to Co-Superintelligence

Paper (concept): A new research agenda from industry and academic collaborators argues for shifting focus from fully autonomous AI self-improvement to co-improvement: humans and AI systems deliberately doing research together, from ideation to experimentation.

Their core claim:

Instead of chasing self-improving AI that may outrun our control, we should target co-superintelligence—a socio-technical system where humans and AI together have superintelligent capabilities.

In this framing:

Co-improvement means designing AI systems to excel at:
- Helping humans generate ideas
- Designing and critiquing experiments
- Analyzing results
- Surfacing counterexamples and failure modes
Humans stay inside the epistemic loop, with AI systems as powerful, structured collaborators—not independent agents rewriting their own goals without oversight.

The argument is that this path could:

Lead to faster scientific progress by exploring more directions in parallel.
Preserve transparency and steerability, because humans stay engaged in the research process.
Offer a more realistic safety story, orienting AI research toward augmenting human reasoning rather than replacing it.

What this means for universities & health-system research

For universities, teaching hospitals, and research institutes, “co-improvement” isn’t just philosophy. It raises design questions for:

Curriculum: How do we train students to treat AI systems as serious but criticizable collaborators—with explicit norms around checking, attribution, and disagreement?
Lab practice: How will labs track what was done by humans vs. AI systems, especially when AI tools propose hypotheses or write analyses?
Shared infrastructure: Are we investing in secure, shared AI research infrastructure (model access, evaluation harnesses, experiment tracking), or are we leaving each lab to improvise?

For boards and university trustees

If you sit on a university board, council, or governing body, you don’t need to referee research methods—but you do need to set the outer frame:

Where will we allow AI systems to participate in our research pipeline (discovery, experiment design, analysis, and grant writing)?
What principles will guide authorship, academic integrity, and conflicts of interest when AI systems contribute meaningfully to research?
How will we ensure that “co-improvement” is used to advance safety, alignment, and public-interest work, not just capabilities?

A practical starting point for trustees:

Ask for a short, written statement from research leadership on “How we are using AI tools as research collaborators today—and where we draw the line.”

3. For national policy, institutional strategy & boards

Australia’s National AI Plan 2025

Document: National AI Plan, Australian Government (Department of Industry, Science and Resources).

PDF: https://www.industry.gov.au/sites/default/files/2025-12/national-ai-plan.pdf

Australia’s National AI Plan offers a structured, whole-of-economy view of what a country intends to do with AI. It is organized around three overarching goals:

Capture the opportunities. Build AI capability, infrastructure, and innovation ecosystems so that Australia doesn’t simply import AI technologies and norms from elsewhere.
Spread the benefits. Ensure that AI’s upside is shared across communities, regions, and sectors, including a focus on workforce skills, transitions, and digital inclusion.
Keep Australians safe. Develop legal, regulatory, and ethical frameworks—including dedicated safety and evaluation capacity—to manage risks like bias, privacy harms, security threats, and frontier-model failures.

Rather than prescribing a rigid, once-and-for-all AI Act, the plan is:

High-level but directional, giving agencies and partners room to adapt.
Explicit about sovereign capability, including infrastructure and government AI capacity.
Framed around a partnership with workers, business, and civil society, not only technical experts.

Why this matters for institutions outside Australia

Even if you never operate in Australia, the plan is applicable in two ways:

As a template: It shows one way to structure an AI strategy around opportunity, equity, and safety, with concrete actions beneath each heading.
As a mirror: It invites you to ask, “What is our institution’s AI plan in similar terms—and where are we silent?”

You could imagine your own organization’s AI strategy summarized on a single slide under three similar headings:

How we will capture value from AI (for patients, students, communities).
How we will spread benefits and reduce harms, especially for those most at risk.
How we will keep people safe, including governance structures, evaluation, and escalation pathways.

For boards and trustees

For boards of directors and university trustees, this plan is a good prompt:

Do we have our own articulation of “capture opportunity/spread benefits / keep people safe” for AI?
Who is accountable for each of those pillars—and how are we measuring progress?
What would be the institutional equivalent of an “AI Safety Institute” for us? Is it a new body, or can an existing committee realistically take this on?

4. Signals & questions to take into your week

If you only have time for reflection, here are a few prompts drawn from this issue.

For health & care leaders

Where are AI systems already creating derived structures—event logs, metrics, risk scores—from unstructured clinical conversations or notes?
Are those structures being treated as regulated, governed artifacts, or as informal analytics no one wholly owns?

For higher education & research leaders

How might you redesign one course or lab this coming term to explicitly practice co-improvement with AI tools—while strengthening, not weakening, academic integrity?
What decisions about AI as a research collaborator need Senate, IRB, or ethics board guidance rather than being left to individual labs?

For boards, trustees & senior leaders

Does your current AI strategy meaningfully address who benefits, who bears risk, and how workers and students are included in design—or is it mostly about tools and pilots?
If you had to present your institution’s AI plan on a single page to a skeptical stakeholder, what would be on it—and what would you be uncomfortably leaving out?

Gratitude

Thank you for reading, forwarding, and challenging this work—especially those of you in nursing, medicine, health informatics, higher education, policy, and governance who are trying to operationalize “human-centered AI” amid real constraints.

This week, I’m particularly grateful to:

Clinicians and informatics teams are experimenting (carefully) with ambient documentation, process mining, and AI-augmented workflows.
Researchers and students who are treating AI systems as serious but criticizable collaborators, rather than oracles.
Policy, compliance, and governance colleagues who are engaging proactively with national strategies and institutional AI plans, instead of waiting for perfect clarity that will never fully arrive.

Your questions, critiques, and “this didn’t work the way we expected” stories are what keep this newsletter grounded in practice.

Disclaimer

This newsletter is for informational and educational purposes only. It does not constitute medical, legal, financial, or investment advice, and it should not be used to make clinical or legal decisions for specific cases. Drafting and editing support for this issue used generative AI and related tools (including ChatGPT, Gemini, Speechify, and Grammarly), with human curation and judgment applied throughout. Any errors or interpretations are my own.

References & further reading

Beerepoot I, Brinkkemper S, Huntink E, Duman B, Reijers HA, Bleijenberg N.Turning Dialogues Into Event Data: Lessons From GPT-Based Recognition of Nursing Actions. Journal of Biomedical Informatics. 2025;172:104957.ScienceDirect: https://www.sciencedirect.com/science/article/pii/S1532046425001868
Weston J, Foerster J, and collaborators. AI & Human Co-Improvement for Safer Co-Superintelligence. (Research agenda on co-improvement and co-superintelligence; available via arXiv and related preprint servers.)
Department of Industry, Science and Resources (Australian Government).National AI Plan. 2025.PDF: https://www.industry.gov.au/sites/default/files/2025-12/national-ai-plan.pdf

The Seba GenAI Ethics & Governance Framework for Leaders: Transparency & Copyright