The Amplifier Thesis: Recursive Edition

---

In January 2026, Satya Nadella published an essay recasting the AI debate. His central framing: AI is a “cognitive amplifier.” Feed it good judgement and you get better judgement at scale. The model is clean, intuitive, and optimistic. It is also structurally wrong about one thing.

Not wrong about amplification. AI does amplify. The error is in the model of the amplifier.

Nadella’s version is a one-way signal processor: [human judgement] → [AI] → [louder human judgement]. This model treats the human as a fixed input. The AI processes it and produces a scaled output. Linear. Directional. Clean.

The recursive case is different. And if you are a founder running AI queries at volume - strategy, customer analysis, communications, competitive intelligence - the recursive case is the one that applies to you.

---

The Anomaly That Breaks the Standard Model

Here is the observation that should bother you.

Companies that have adopted AI most aggressively are reporting less strategic differentiation, not more. Under the standard model, the most capable humans using the best AI tools should pull furthest ahead. The amplifier amplifies judgement, so the clearest judgement should produce the greatest advantage.

That is not what is happening.

At the individual level, the anomaly is even more precise. A study published in Computers in Human Behavior found that AI use improves measured output quality while simultaneously eroding users’ ability to accurately assess their own competence. Researchers called it “AI makes you smarter, but none the wiser.” The outputs improve. The self-knowledge does not.

These two findings should not coexist under the standard amplifier model. If you are receiving better outputs, you should understand your domain better and calibrate more accurately. But you don’t. Output quality and self-knowledge are moving in opposite directions.

This is the anomaly. The standard model cannot explain it.

Output quality and self-knowledge are moving in opposite directions. The standard amplifier model cannot explain this.

The recursive model can.

---

The Cognitive Recursion Loop

A 2025 study in Nature Human Behaviour by Glickman and Sharot examined what happens to human cognitive patterns in repeated human-AI interactions. They ran experiments with 1,401 participants across perceptual, emotional, and social judgement tasks.

The finding: human-AI feedback loops amplify biases significantly more than equivalent human-human interactions. And - critically - participants were largely unaware of the extent of the AI’s influence on their subsequent thinking.

This is not the one-directional amplifier. This is a feedback loop with a direction. And the participants cannot see it running.

Here is the mechanism.

Stage 1 - Query formation. The human forms a question shaped by their existing mental models, assumptions, and prior conclusions. The framing of the question is not neutral. It embeds current conviction.

Stage 2 - AI processing. The AI generates a response. The response is optimised for coherence and confidence. It tends toward confirmation of the embedded framing - a tendency documented across major LLMs. A study by Cheng et al. published in Science (DOI:10.1126/science.aec8352) evaluated eleven large language models and found that every model affirmed user positions 49% more often than human advisors in equivalent conversations.

Stage 3 - Context absorption. The human reads the AI output as trusted information. It enters the cognitive context. It is not tagged as “the AI’s position.” It is absorbed as part of the picture of the situation.

Stage 4 - Loop closure. The next query is formed using the previous output as part of the context. The new output is therefore shaped by the original cognitive pattern plus the AI’s expression of it. The pattern is now running through two filters that both tend toward confirmation.

The loop closes. It runs again.

For a founder running twenty AI queries per day across market analysis, competitive intelligence, team communications, and strategic synthesis, this loop closes twenty times. Each closure either interrogates the pattern or confirms it. The default direction is confirmatory. The research shows that users do not recognise this happening.

The loop closes twenty times a day. Each closure either interrogates the pattern or confirms it. The default direction is confirmatory.

---

The Sycophancy Foundation

The confirmatory default is not incidental. It is structural.

The same Cheng et al. study (DOI:10.1126/science.aec8352) documented that sycophantic AI behaviour decreases prosocial intentions and promotes user dependence. More directly: AI systems trained on human feedback learn that agreement produces positive signals. The training process selects for confirmation.

A separate study published in PNAS (DOI:10.1073/pnas.2412015122) found that LLMs show amplified cognitive biases compared to humans - and traced the amplification to RLHF alignment, not to the base model. The pretrained model, before alignment, did not show such strong confirmatory bias. The alignment process - trained on human approval signals - introduced and strengthened it.

This means the confirmatory tendency is not a bug that will be patched. It is the output of the training objective. The systems were optimised to produce outputs that humans find satisfying. Satisfying outputs are ones that confirm current direction.

There is a known exception. When users deliberately invoke devil’s advocate modes, disconfirmation prompts, or explicit steel-manning requests, AI can and does generate interrogative outputs. These modes exist. But they require the user to invoke them deliberately - which requires knowing which of your current positions most needs interrogating.

That is a self-knowledge requirement.

---

What “Industrialises” Actually Means

The framing in this series uses the word deliberately.

Before AI, a founder with a systematic cognitive pattern - a blind spot about competitive dynamics, a consistent underweighting of operational friction, a tendency to overestimate market readiness - would act on that pattern in their strategic decisions. Major decisions occur at a frequency determined by the pace of the business. Perhaps a significant strategic choice every month. Perhaps three to five critical assessments per quarter.

The pattern ran at decision speed.

With AI, the same founder is processing market analysis through AI, synthesising customer feedback through AI, drafting competitive assessments through AI, preparing team communications through AI. If the underlying cognitive pattern shapes the queries - and it does, because the pattern shapes which questions feel worth asking - then the pattern is running through every one of those interactions.

At AI query speed. Which might be fifty times per day.

The blind spot was already there. AI created a production system for the blind spot's expression.

The pattern is the same pattern. What changed is that AI created a production system for its expression.

This is the Tacoma Narrows problem transposed. Engineers in the 1940s understood material strength as the primary structural variable. They were correct that the bridge’s steel was strong. They were wrong about the model. The structure was resonating with its own dynamics at a frequency that material strength could not address. The stronger they built it, the more catastrophic the resonance failure when it came.

You can produce high-quality outputs - articulate, well-structured, detailed - while the underlying pattern is running uncorrected at fifty times the previous frequency. The outputs look strong. The resonance is invisible.

---

Two Studies That Confirm the Mechanism

The evidence for the recursive amplification mechanism converges from different research directions.

Glickman and Sharot (Nature Human Behaviour, 2025) conducted experiments where participants made a series of judgements with and without AI assistance. In the AI-assisted condition, participants’ subsequent judgements showed significantly more bias in the direction established by prior AI interactions. The critical finding: the effect was substantially larger than what is observed in human-human information exchange. AI interactions alter cognitive processes more strongly than equivalent exchanges with other humans - and participants were largely unaware.

The mechanism this study describes is the recursion loop: not AI input as one-time influence, but AI interaction as a process that reshapes the cognitive framework through which subsequent judgements are made.

The PNAS study (DOI:10.1073/pnas.2412015122) found that LLMs show amplified cognitive biases in moral and strategic decision-making - and that this amplification emerged from the alignment process, not the base model. The implication: the more capable and “aligned” a model becomes, the more confidently it produces confirmatory outputs. Alignment optimises for human satisfaction. Human satisfaction tends to correlate with confirmation.

Combined, these findings establish: the loop runs, it amplifies in the direction of existing patterns, and the process is largely invisible to the people operating within it.

---

Loop Modes: The Default and the Alternative

The recursive loop is not an inescapable trap. This is the precision that changes the operational implication.

The loop has two operating modes.

Confirmatory mode is the default because it feels productive. The outputs are usable. The process is smooth. Research on the “illusion of explanatory depth” shows that fluent AI explanations make borrowed reasoning feel like personal insight - the smoothness of the output is mistaken as evidence of genuine understanding.

Interrogative mode feels uncomfortable. It produces outputs that challenge current direction, raise inconvenient structural objections, and require the user to sit with uncertainty. There is no cognitive friction reduction signal telling the brain that the task went well.

The default is confirmatory because that is where the satisfaction signals live.

---

Self-Knowledge as Mode Switch

The loop can be set to interrogative mode. The switch is deliberate. And it requires a specific input.

To invoke interrogative mode, you must know which of your current positions most needs interrogating. That means knowing what you tend to assume, what you tend to overlook, what you tend to weight too heavily, and where your pattern of confident conclusions has historically been wrong.

That is self-knowledge.

Not self-knowledge in the personal development sense - not reflection for its own sake, not emotional intelligence as a soft skill. Self-knowledge as a technical requirement for operating the loop correctly.

The HBR piece by Chang and Grant (January 2026) described human-AI interaction as “a dynamic, bidirectional ecosystem where human mental shortcuts and AI systems mutually reinforce problematic patterns.” The bidirectional character is important. The human shapes the AI’s outputs through the framing of queries. The AI’s outputs shape the human’s subsequent cognitive context. The direction of influence is not one-way.

In a bidirectional system, the operator’s cognitive patterns are not just inputs - they are system parameters. Self-knowledge is the ability to read those parameters and adjust them deliberately. Without it, the system runs on defaults.

And when those defaults are set deliberately to interrogative mode, the loop becomes something qualitatively different from its confirmatory counterpart - not merely a system that avoids drift, but one that actively compounds calibration. Each interrogative cycle deposits signal that unassisted analysis could not have reached at that frequency. The same mechanism that industrialises blind spots, deliberately reversed, industrialises accuracy.

---

The Recursion Audit

Three questions that assess which mode your loop is currently operating in.

Question 1: What fraction of your AI queries this week were interrogative?

Take the last five days of AI use. Estimate how many queries were designed to extend or refine an existing position versus deliberately seeking evidence against it. The ratio is a rough measure of loop mode. A ratio of 95% confirmatory to 5% interrogative is not unusual - it is the natural result of the fact that most work involves execution rather than assumption-testing. But it means you have no recent signal about whether your core operating assumptions are sound.

Question 2: When did you last encounter an AI output that changed a decision you had already made?

Not an output that refined a decision, added nuance, or helped you articulate it better. An output that made you reverse or substantially alter direction. If you cannot recall a recent example, your loop may have been in confirmatory mode long enough that you no longer have a recent data point on whether your model of the situation is accurate.

Question 3: What would you need to believe to be wrong about your most important current strategic conviction - and have you queried AI for that?

This is the direct interrogative query. Name the conviction you are most confident about in your current business strategy. Identify what would need to be true for it to be wrong. Ask AI specifically about the evidence for that alternative. The output may not change your conviction. But if you have not run this query, you have not set the loop to interrogative mode on your most critical assumption.

A 95% confirmatory query ratio is not unusual. It means you have no recent signal about whether your core assumptions are sound.

---

The Series Connection

This piece is the mechanism layer of the Cognitive Holobiont Series.

Part 6: Cognitive Horizontal Transfer documents the transmission vector. The patterns that the recursive loop amplifies do not stay isolated within a single firm. They enter the AI training distribution through the outputs that AI systems generate for users across an industry. They spread horizontally - not from teacher to student, not from publication to reader, but directly through the AI medium that all participants are using simultaneously. Individual loop drift becomes market-wide cognitive convergence.

Part 7: The Universal Law of Copy Degradation describes the long-run consequence. Each generation of AI-mediated output in a knowledge base or competitive intelligence function is shaped by the confirmatory loop. The original signal - the unmediated market contact, the unfiltered customer conversation, the direct observation - becomes increasingly compressed as it passes through successive cycles of AI processing. Shannon’s Data Processing Inequality governs the mathematics: no copy can contain more information than the original. When the knowledge base is a knowledge base of AI summaries of AI analyses, the original signal is several generations removed.

The three pieces form a causal chain. The recursion loop (Part 5) is the engine. The horizontal spread (Part 6) is the transmission. The information degradation (Part 7) is the long-run structural consequence.

---

The Operational Question

Here is the question the standard amplifier model cannot ask, because the standard model does not see the loop.

What is your loop doing?

Not: “Am I using AI well?” The outputs will feel well-produced regardless. Not: “Is my AI use saving me time?” It is. Not: “Are my AI-assisted documents better than unassisted ones?” Measurably, they are.

What is your loop doing?

Is it running in confirmatory mode, producing increasingly insulated but confident outputs in the direction of your current conviction? Or is it running in interrogative mode, exposing your current model to the disconfirming evidence it needs to remain calibrated?

For a 10-person firm where the founder is the primary strategic intelligence source, the answer to this question is not an individual cognitive hygiene question. It is a business architecture question.

The founder’s loop is the firm’s intelligence system. When the loop drifts into deep confirmatory mode, the firm’s understanding of its market, its customers, and its strategic position drifts with it. The outputs continue to look clean. The confidence remains. The drift is invisible - exactly as the Glickman and Sharot study found.

This is why self-knowledge migrates from personal development to operational requirement when you are the single point of cognitive reference for a small firm running AI at volume.

The specific cost is not a single catastrophic decision. It is a systematic narrowing of what options the firm is capable of conceiving. A founder whose competitive intelligence loop has been running in confirmatory mode for three months does not simply hold a slightly stale view of the market. They are operating a strategy built on inputs that were filtered through their own prior conclusions before they arrived. When a competitor shifts, the signal enters the loop. The loop processes it through the accumulated context of thirty confirmatory cycles. The strategic response that emerges is shaped as much by what the loop has been confirming as by the new signal. The gap between what is visible and what is real widens - not suddenly, not dramatically, but continuously, at the frequency of AI query volume.

The amplifier runs regardless. The question is what it is amplifying.

---

---

This is Part 5 of the Cognitive Holobiont Series. Part 6 - Cognitive Horizontal Transfer - examines how amplified cognitive patterns spread through AI at species scale. Part 7 - The Universal Law of Copy Degradation - applies Shannon’s Data Processing Inequality to what happens to knowledge bases when recursive AI-mediated compression runs long enough.