When a Scenario becomes a Sell-Off

mickyatesAI, Artificial Intelligence, Ideas, Mick's Blog, Narrative, Personal Leave a Comment

A Substack post moved markets. What the Citrini report got right, what it got wrong, and what the market missed entirely.

Citrini’s “2028 Global Intelligence Crisis” – a 22-page account written as a retrospective from the future – described how accelerating AI capability hollowed out cognitive labour, cracked private credit markets, and triggered a cascading economic spiral. Published on 22 February, it was widely shared. Markets sold off on the 23rd. The piece was called everything from prescient to irresponsible.

Both descriptions miss the point.

The report is not wrong. It’s doing something more interesting than being right or wrong.

The mechanism Citrini identifies is real. If AI continues to substitute for mid-to-high-skill cognitive work at scale, and if institutions fail to adapt faster than the displacement arrives, then labour income falls, asset valuations tied to human productivity fall, leverage bites, and a feedback loop begins. That chain is economically coherent. The Stage 1 question isn’t whether the mechanism exists. It’s whether it operates with the velocity and completeness the narrative implies.

And there the document has a problem it doesn’t know it has.

The document speaks in facts about things that are not yet facts.

Count the verbs. “AI gets better and cheaper every 6–12 months with no sign of slowing”. “The intelligence premium is unwinding”. “Public faith in government rescue has dwindled”. These are stated as present-tense observations. They are extrapolations dressed in observation’s clothing.

That distinction matters more than it might appear.

When a document says AI capability improvement shows “no sign of slowing,” it is treating the absence of observed deceleration as evidence that no deceleration is coming. That is a logical error. Known constraints on AI scaling – energy costs, hardware limits, data availability, organisational friction – are not addressed, not refuted, simply absent. The document builds a spiral on a trajectory it treats as settled when it is genuinely contested.

“No natural brake” is the report’s central claim. It’s presented as a finding. It’s actually an assumption – and it’s the assumption that does all the work. Remove it, and the scenario becomes one plausible trajectory among several rather than the base case.

Markets reacted as though it were the base case.

There is another layer to yesterday’s sell-off. Technology and AI-linked equities have been trading on aggressive forward expectations – expectations that embed productivity gains, margin expansion, and limited institutional drag. In such an environment, a coherent downside narrative does not need to be probable to move prices. It only needs to expose how much of current valuation rests on assumptions that acceleration continues without systemic problems.

The market reaction to the Citrini post therefore says as much about valuation fragility as about AI risk.

What’s actually being replaced – and why that matters

The report frames its argument around the displacement of “human intelligence”. This framing is doing more damage to the analysis than it might appear.

What’s actually at risk is not intelligence in the abstract. It is something more specific: situated cognitive labour – judgment applied within particular institutional contexts, accountability structures, professional relationships, and social meanings that took decades to build.

Replacing a task is different from displacing a function.

A mortgage underwriting algorithm can handle the calculation. It cannot inherit the regulatory accountability of the underwriter, the liability structure of the firm, the professional relationship with the broker, or the institutional memory of what happened last time a model was wrong at scale. These are not inefficiencies protecting human workers. They are load-bearing components of a system that operates on trust, and trust is not easily automated.

This doesn’t mean displacement isn’t happening. It means it’s happening at different speeds in different places for different reasons – and those differences are where the actual analytical work lives. A narrative that treats the replacement of a task and the displacement of a function as identical accelerates the confusion it’s trying to describe.

The missing frictions – and why they’re not all equal

Every crisis narrative requires frictionless transmission. AI improves → workers displaced → income falls → private credit cracks → contagion. Citrini’s chain is clean. Real systems are not.

Organisational adoption of enterprise technology consistently lags capability demonstrations by years, sometimes decades. This is not a bug in the historical data – it is a feature. Firms underestimate legal liability, workflow redesign costs, management overhead, and the simple difficulty of changing how ten thousand people do their jobs. GitHub Copilot exists. Enterprise software teams have not been eliminated. The gap between demonstration and deployment is where the spiral either confirms or doesn’t.

Financial contagion is also slower than the report implies. The private equity debt overhang the document identifies is real – but PE fund structures, covenant diversification, and the 24-to-36 month timeline between revenue stress and liquidity crisis mean the mechanism is grinding, not sudden. This matters for the policy window. If the crisis arrives over three years rather than eighteen months, the institutional response window is not closed.

The global picture is more uneven still. Workers in India’s IT sector, in Philippine BPO operations, in Eastern European tech services – these are not footnotes to the story. They are the first-order exposure. And their situation is structurally different from the white-collar American worker the narrative centres: less institutional protection, fewer retraining pathways, and no domestic political coalition with the power to force redistribution of AI-generated gains. A sell-off in Citrini’s terms means something very different in Mumbai than in Manhattan.

The frictions are real. But they are not all equal, and they are not all pointing the same direction.

Four futures, not two

Here is what the market discussion is missing.

The Citrini report, for all its sophistication, presents a binary: the spiral confirms, or it doesn’t. That binary is analytically impoverished.

The genuine uncertainty in this situation is not “crisis or no crisis”. It is the intersection of two questions that are genuinely open: How fast does AI displacement into white-collar work actually happen? And how effectively do democratic institutions respond?

Those two questions produce four distinct futures, not two.

If displacement is fast and institutional response effective, you get a managed transition – painful, uneven, globally unjust in its distribution of costs, but not a spiral. History has seen this before. The New Deal was enacted during a depression. The Marshall Plan was built at the moment of maximum need. Rapid political action in crisis conditions is not impossible. It requires a crystallising event, a viable financing mechanism, and a political coalition that holds. Those are demanding conditions, but specifiable ones.

If displacement is fast and institutions fail – captured by lobbying, paralysed by fiscal constraint, operating at legislative timescales that the deployment pace outruns – then Citrini’s scenario confirms. This is plausible. Possibly the most plausible fast-displacement trajectory given current institutional performance. But it is not inevitable.

If the historical pattern of technology diffusion holds – organisational friction, regulatory overhead, and complementary investment requirements slowing deployment to five to ten years for full structural impact – then institutions have a window. The question becomes whether they use it.

And if diffusion is slow but institutions mistake the quiet for resolution, the most dangerous scenario of all: complacency followed by delayed crisis, arriving when institutional capacity for response has been further eroded by years of inaction. This is not better than the Citrini scenario. It may be worse, because the slow burn doesn’t create the visible crisis moment that democratic politics can organise around.

The question the market should be asking is not “will the spiral confirm?” It is “which of these trajectories are the early indicators pointing toward, and what are those indicators?”

What to watch – and what it would tell you

Three things are genuinely observable over the next twelve to thirty-six months that would tell you which trajectory you’re in.

The first is deployment velocity in white-collar sectors. Not capability announcements – deployments. Employment data in software, legal, financial, and administrative roles showing sustained decline above five percent within twelve months would be a genuine signal. Employment data showing stability despite capability growth would be equally significant: it would mean the diffusion lag is holding, and the response window remains open.

The second is the PE/SaaS debt stress signal. Watch secondary market liquidity in PE fund interests, covenant triggers in SaaS-adjacent credit structures, and year-on-year revenue guidance revisions across major enterprise software categories. Citrini’s contagion mechanism is real but slow. Its onset would be visible in these indicators twelve to eighteen months before it became systemic.

The third is the political crystallisation test. The effective response scenario requires a visible, identifiable displacement event that converts diffuse anxiety into organised political response. Watch whether legislative proposals with genuine redistribution mechanisms – not advisory frameworks, not voluntary guidelines – develop cross-party traction in major OECD economies within the next eighteen months. Their presence or absence is the most important indicator of which scenario the next decade belongs to.

None of these questions is answered by the Citrini document. All of them are answerable by observation.

The new risk the report accidentally identifies

There is something genuinely new in the 48 hours since the Citrini post appeared, and it is not what the post describes.

We now live in a world where AI-generated scenario analysis can reach algorithmic trading systems before epistemic evaluation occurs. The format cues of a well-constructed future history – specificity, confidence of voice, institutional plausibility – override the epistemic status of its claims. A scenario becomes a sentiment input before anyone has asked whether the assumptions that drive it are evidenced.

That is not Citrini’s failure. It’s a feature of the current information environment that the post happened to reveal. The question it raises is more interesting than the 2028 story: what institutional infrastructure would need to exist for markets to distinguish between “this mechanism is coherent” and “this mechanism is operating as described right now”?

That infrastructure doesn’t currently exist. Building it – not better AI, but better epistemic standards for evaluating AI-related claims – is arguably the more urgent work.

What the report was

The Citrini piece should be read as a stress test. A structured worst-case. A thought experiment about what happens when institutional lag meets technological acceleration and neither blinks first.

It should not be read as a base case, a near-term forecast, or a single trajectory masquerading as the landscape.

The correct response to it is not panic, and not dismissal. It is specification. Ask what would have to be true for each element of the chain to hold. Ask where the first observable crack would appear. Ask what the countervailing forces are, and whether they are equal to the ones they’re counteracting.

Serious questions deserve analytical discipline. Markets discovered yesterday that they don’t always get it. The question now is whether they’re asking the right questions about what comes next – or whether they’re already repricing the next narrative before they’ve finished evaluating this one.

Markets price probability. Narratives price mood. The two are not the same thing.

What the heck is “Ethical AI”

mickyatesAI, Artificial Intelligence, Ethics, Ideas, Mick's Blog, Philosophy, Technology 2 Comments

If you spend even a small amount of time reading about artificial intelligence, you will know that everyone says they want “Ethical AI”.

Governments want it. Tech companies want it. Universities want it. Consultants definitely want it.

In my opinion, however, saying you’re “in favour of ethical AI” is a bit like saying you hope the sun comes up tomorrow and then it keeps us warm. It signals basic sanity, but it is not useful insight. So the real question isn’t whether AI should be ethical.

The real question is: What do people actually mean by it, and what, if anything, changes because of that understanding?

Ethical AI, as Usually Meant

Most of the time, “Ethical AI” seems to mean one of three things:

  1. A list of values: Transparency. Accountability. Safety. Human-centricity. Everyone nods. Nobody disagrees. Nothing is tested.
  2. A posture: “We take ethics seriously.” Often accompanied by a PDF, a committee, and a photo of diverse people sitting around a table.
  3. A reassurance: Something you say so others don’t worry too much about what you’re building.

None of these are bad. They’re just … thin.

They tell us what people hope for, not what happens when things become hard to judge, or when they go wrong. And that’s precisely when ethics is supposed to matter.

The Awkward Truth

Here’s the uncomfortable bit: Most ethical failures don’t happen because people don’t care about ethics. They happen because ethical language outruns structure, discipline, and responsibility. In other words, ethics fails after we agree it is important, not before that moment of enlightenment.

Everyone agrees they want fairness – until fairness pulls in opposite directions. Everyone wants transparency – until it conflicts with speed, profit, or security. And everyone wants accountability – until it’s unclear who, exactly, should carry it and what it might cost.

At that point, ethical talk doesn’t disappear. It just becomes performative.

Ethical AI as Theatre

We are now very good at sounding “ethical “when we talk about AI.

AI systems can explain themselves fluently. Institutions can publish values frameworks elegantly. Reports can acknowledge uncertainty while quietly smoothing the issues away. This is dangerous, not because it’s dishonest, but because it’s convincing. It isn’t just LLMs that are convincing even as they hallucinate. Humans do that, too.

When ethical language becomes fluent enough, it stops being questioned. It creates the feeling of responsibility, without the burden of actually doing something about it. And that’s how ethics turns into theatre.

So, What Would Ethical AI Actually Require?

It is not about identifying and listing better values.  It is not another framework diagram. Nor is it about adding a list of socially good practices in some kind of LLM “values alignment” layer. Frankly, that effort tends to be about talking to your customers in a way that they like, rather than talking about the right things.

“Ethical AI”, if it is to mean anything, requires constraints and discipline.

At minimum, it needs

  • Clear boundaries on what the system can and cannot claim.
  • Explicit handling of uncertainty, rather than optimistic smoothing.
  • Retention of responsibility at critical junctures that can’t be delegated.
  • Structures that slow things down when clarity is missing, and that highlight tensions.
  • Mechanisms that surface disagreement rather than rushing to consensus.

In short: Ethics has to be operational, not aspirational.

The Hard Part Everyone Avoids

The genuinely hard work of ethical AI isn’t deciding what’s right.

It’s deciding:

  • Who is responsible when there are no good, easy options.
  • How to act when values collide and consensus can’t be reached.
  • How to stop persuasive language from standing in for judgment.
  • How to admit “we don’t know” without losing face.

That’s slow. It’s uncomfortable. It doesn’t fit neatly into product roadmaps.

Which is why so much “Ethical AI” talk floats safely above it.

Does Anyone Actually Do This?

That is a fair question. Diagnosing what’s broken is easier than building what works.

But yes, there are emerging approaches that take these requirements seriously. These approaches don’t start from values. They start from failure modes, and they tend to share several characteristics:

  • Protocol governance: Explicit rules about what can/cannot be claimed, with violations triggering correction rather than being smoothed over.
  • Uncertainty quantification: Using frameworks to provide confidence bounds rather than probabilistic “hand-waving”.
  • Multi-perspectival interrogation: Deliberate inclusion of conflicting frameworks to surface disagreement rather than manufacture consensus.
  • Human responsibility preservation: Ensuring AI systems are very competent tools within bounded domains, not autonomous decision-makers.
  • Conceptual vacancy detection: Building mechanisms that flag when key terms (“fairness,” “alignment,” “trust”) are doing unexplained work.

These aren’t theoretical proposals. They’re operational methodologies that are gradually being validated and applied to real-world governance questions.

They’re slower than standard AI workflows. They require epistemic discipline that feels uncomfortable. They produce bounded insights rather than confident predictions.

Which is exactly the point.

The test of “Ethical AI” isn’t whether it sounds ethical. It’s whether the methodology catches itself when it starts to bullshit, and whether humans retain the authority to act on what’s caught.

I’ve spent the last year researching and testing such methodologies. They work. They are awkward. They surface inconvenient disagreements. They refuse to claim certainty where none exists. Such approaches aren’t designed to be popular with people who want AI to make decisions faster, sound more confident, or bypass the hard work of human judgment. But for people doing work where being confidently wrong is worse than being honestly uncertain, it turns out there’s real demand for systems that know how to shut up when they should.

A Better Question

So maybe the question isn’t:

“Is this AI ethical?”

But:

“What stops this system from sounding more ethical than it actually is?”

That’s the question worth answering.

Because wanting “Ethical AI” is easy. Building systems that remain ethical under pressure is not. And if we don’t learn to tell the difference, “Ethical AI” will keep meaning everything, and therefore mean nothing.

Contact me if you’d like to continue this conversation.