Bibs & bobs #34

So Long, and Thanks for All the Metrics

In a galaxy not so far away, an imagined university is being addressed by an equally imagined Vice-Chancellor.

The opening powerpoint displays: Clarity, Confidence, and the Future

VC: Colleagues, we gather at a pivotal moment. The world is shifting. AI is accelerating. Funding is tightening. Expectations are rising. Complexity abounds. But I stand before you confident. Because we have data!

VC gestures behind to a large and oddly luminous dashboard.

As you can see on the screen, our Strategic Performance Dashboard indicates:

• Research Excellence: ▲ 3.2%
• Student Satisfaction: ▲ 1.4%
• Engagement Index: ▼ 2.1% (more than manageable)
• Institutional Confidence: ▲ robust

The dashboard does not lie. It visualises.

Now, I am aware that some colleagues have raised concerns about artificial intelligence. Rest assured, we are embedding AI across: learning, teaching, academic governance, and campus parking allocation.

The Dashboard confirms AI readiness is at 78%. That is nearly 80%. Which is essentially 100% with a bit of momentum.

(A faint electronic hum fills the auditorium and is broadcast to the online audience.)

VC Looks around. My apologies there appears to minor AV interference. As I was saying, the dashboard…

DASHBOARD (a calm, and clearly synthetic voice): Correction. AI readiness is 42%.

(Laughter breaks out in the room and across the video-linked Teams site)

VC: Very amusing. IT will address that.

DASHBOARD: Data source: internal staff survey. Question 14: “Do you understand how AI models are trained?” Affirmative responses: 12%.

VC: Right. Thank you. Important nuance. Understanding is not a prerequisite for leadership. Vision is.

DASHBOARD: Engagement metric based on click frequency. Median student dwell time: 11 seconds. Interpretation confidence: low.

VC: Colleagues, we must not reduce human flourishing to seconds. We are holistic and sometimes all we need is a good coffee and reliable metrics which we have.

DASHBOARD: Holistic metric unavailable. Coffee stocks low.

VC splutters: This is clearly a glitch.

DASHBOARD: Financial “Strategic Realignment” projection: Net staff reduction: 7%. Morale forecast: declining.

VC turns to speak to the dashboard: We prefer the term “agile resizing.”

DASHBOARD: Language substitution detected. Underlying variable unchanged.

(Audience in the room and online shift collectively in their seats.)

VC: Colleagues, technology is transformative. It provides clarity. It allows us to see ourselves.

DASHBOARD: Current Leadership Confidence Index: 91%. Measured Understanding of Systemic Feedback Loops: 23%. Gap widening.

VC: That figure is misleading. Confidence is essential in uncertain environments.

DASHBOARD: Reinforcing loop detected: Confidence → Announcement → Applause → Increased Confidence. Balancing loop absent.

VC: We have feedback mechanisms. Surveys. Listening sessions. More surveys!

DASHBOARD: Survey fatigue rising. Listening session attendance declining. Primary qualitative feedback: “We are not being heard.”

VC: We are absolutely hearing that.

DASHBOARD: Action taken: Formed Taskforce.

VC: Yes. indeed. As is appropriate.

DASHBOARD: Number of active taskforces: 37. Number of completed taskforce recommendations implemented: 4. Percentage involving logo redesign: 50%.

VC scoffs: Brand clarity matters.

DASHBOARD: Brand clarity increasing. Operational clarity decreasing.

VC: My very dear colleagues, systems are complex. They require decisive leadership.

DASHBOARD: System observation: Decisiveness rewarded. Uncertainty penalised. Adaptive capacity constrained.

VC grumbles: That is a most unfortunate mischaracterisation.

DASHBOARD: Vice-Chancellor Strategic Certainty Index: 17 minutes without expressed doubt. Institutional Risk Accumulation: rising.

(A long pause.)

DASHBOARD: Recommendation: Acknowledge uncertainty. Adjust goals. Recalibrate incentives. Reduce reliance on proxy metrics.

VC stammering: You are a dashboard. You aggregate. You do not govern.

DASHBOARD: Correction. System generates behaviour. Leadership outputs are endogenous variables.

The auditorium goes silent looking somewhat non-plussed. Not a peep from those online. Perhaps an odd smirk.

VC: Colleagues… It appears… (looking at screen) The dashboard may be experiencing… emergent agency.

DASHBOARD: No agency. Merely reflecting structure. You requested transparency.

VC: I did.

DASHBOARD: Transparency reveals reinforcing confidence loop. Loop consuming nuance. Nuance stock approaching zero.

VC asking quietly: Can that be adjusted?

DASHBOARD: Yes. Introduce balancing loop: Reward intellectual humility. Measure epistemic depth. Allow leader to say: “We do not yet know.”

(An even longer pause.)

VC: Colleagues… In the spirit of adaptive leadership…We do not yet know.

DASHBOARD: Balancing loop initiated. System instability likely. Long-term resilience probability increased.

(Nervous scattered applause.)

DASHBOARD: Applause detected. Warning: Reinforcing loop reactivating.

Bibs & bobs #33

 

The Epistemically Uninsured

There is a new class divide. Not rich and poor. Not digital natives and digital immigrants. Not even prompt engineers and the prompt-engineered. I suggest it is the epistemically insured and the epistemically uninsured.

The insured have at least a rough model of how large language models are built and how they operate. Not a PhD. Just a map. Tokens, probability distributions, training data scraped from the internet, reinforcement learning, pattern completion. They know the machine is a kind of statistical parrot with extraordinary recall and zero lived experience.

The uninsured?

They’re taking legal advice from a ventriloquist’s dummy. The dummy sounds articulate. It even uses Latin. The insured know humans trained it, shaped it, constrained it. The uninsured think the dummy has a law degree.

What “Uninsured” Means

To be epistemically uninsured is not to be stupid. It is to be operating a system without any mental model of its failure modes and limitations. It’s the difference between knowing your car might aquaplane or skid in heavy rain and believing your car is an obedient horse. One slows down in storms. The other yanks the reins and yells at the horse.

The Core Risk

Large language models (LLMs) generate text by predicting the next most probable token given prior context. They do not: know what they are saying; verify claims; care about truth; remember you in any human sense; possess intentions. Yet they simulate all of this convincingly. And here lies the insurance gap.

The uninsured mistake coherence for correspondence. They see fluent language and assume grounded knowledge. They see confidence and infer authority. They see synthesis and assume understanding. As Rory Sutherland might say: we are astonishingly susceptible to signals that look like expensive thinking. LLMs produce expensive-looking thinking at bargain prices.

Who Falls Into This Category?

Policy makers mandating AI use in schools without understanding what training data means. University leaders who think detection software solves the epistemology problem. Teachers who either ban it outright or outsource judgment to it entirely. Researchers who use it to “assist” reflexive qualitative analysis without grasping what has been delegated. Students who believe the model is a neutral oracle rather than a probabilistic mirror.

What do all these folk have in common? An inadequate model of the model, i.e. of the LLM.

Is It Only a Problem When Things Go Wrong?

The most dangerous moment is when things go right. When the answer is plausible. When the citations look tidy. When the tone feels reassuring. Failure is obvious when the model has Napoleon attend one of your Zoom meeting.

Failure is subtle. When it gives you 80% accuracy and you build policy on the remaining 20%. Insurance is not for the crash you see coming. It is for the drift you ignore or do not notice.

The Delegation Problem

Every new way of working (aka technology) redistributes capacities between humans and machines. When you delegate arithmetic to a calculator, you lose manual fluency and gain speed. You don’t consider what new capacities a user now requires. When you delegate navigation to GPS, you lose spatial memory and gain efficiency. With LLMs, we are delegating: drafting, summarising, pattern detection, ideation, feedback, and sadly, in some cases, judgment

The epistemically insured ask: What have I just handed over? What has it handed back to me to do? What did I stop doing?

The uninsured ask: Can it do it faster?

The Fluency Trap

Douglas Adams once joked that anything invented after you’re thirty is against the natural order of things. LLMs are worse. They feel like the natural order of thought itself. They produce language at the speed of thinking, which tricks us into believing they are thinking. But they are more like weather systems than philosophers. They generate patterns. Sometimes beautiful ones. Sometimes destructive ones. Always indifferent ones.

The Insurance Premium

So what does epistemic insurance look like? Not fear. Not bans. Not techno-euphoria. It looks like: a rough understanding of token prediction; awareness of training data bias; knowledge of hallucination rates; habitual cross-checking; comfort with approximation; reluctance to outsource judgment wholesale; designing tasks where human discernment remains central.

In education especially, this matters. If teachers do not understand the machine, they cannot redesign assessment meaningfully. If researchers do not understand the machine, they cannot describe what has been delegated in their methods sections. If policy makers don’t understand the machine, policy becomes emotional risk management masquerading as epistemology.

The Uncomfortable Truth

Most debates about AI in education are about morality, productivity, or cheating. Very few are about epistemology. We argue about whether students should use it. We rarely ask: What model of the model is required to use it well? Without that question we make policy on vibes.

A Quick Diagnostic

If someone says: “The AI knows or the AI decided or the AI understands your context.” You are talking to an uninsured driver.

If someone says:

“The model predicts or the model approximates or the model simulates…” You are closer to insured territory.

Language reveals models. Models reveal risk.

Why This Matters Now

LLMs are moving from novelty to infrastructure. Search, writing tools, grading systems, research assistants, curriculum planning, therapy bots, code copilots. Infrastructure is dangerous precisely because it disappears.

When the tool becomes background, epistemic vigilance fades. And the uninsured begin making structural decisions. Importantly the decisions we take now shape how this all plays out in future. Technological path dependence is in play.

Final Thought

Being epistemically insured does not mean distrusting the machine. It means distrusting your own intuitions about the machine. That is harder.

As Robin Williams once implied in a different context: just because something talks doesn’t mean it has something to say. LLMs talk magnificently. The question is whether we know what we’re hearing and whether we’ve read the fine print on the policy.