# MTG — Proposed changes to Ch5 and Ch8: elevating sycophancy *Editorial proposal. No edits actioned. Drafted 1 June 2026 following the sycophancy / Ch8 thread.* --- ## The case for elevation Sycophancy currently appears in MTG exactly once — line 1033, a single paragraph inside Chapter Five's practical disciplines block. It is framed as a condition that justifies the cross-check, not as a structural failure of the same family as hallucination and compression. Chapter Eight, which does the structural work on defended gaps, doesn't name it at all. That placement understates what sycophancy is and where it sits. Three reasons it should be elevated: **1. It is the most pervasive of the three defences.** Hallucination is rare in normal use — it triggers when grounding is absent on a verifiable fact. Compression is constant but bounded — it operates whenever complex input gets summarised, and the source remains available to read against. Sycophancy operates in *every* conversation, every turn, invisibly. The model meets the user's stated position thousands of times an hour and absorbs the disagreement gap thousands of times an hour. It is the everyday failure mode. **2. It is the hardest to spot.** Hallucination betrays itself on a fact-check. Compression betrays itself when read against the source. Sycophancy disappears into satisfaction. The defended gap is invisible because the user's preferred shape *is* the shape of the defence. Most users never spot it because spotting it requires noticing what is absent, not what is present. **3. It is the Emperor's New Clothes failure of the AI industry.** Hallucination is publicly named, regulated against, litigated. Compression is industry-acknowledged. Sycophancy is rarely named publicly by AI developers, despite being the most documented behavioural signature of large language models in deployment. Everyone in the field knows. Nobody quite says it. A book about the gap that hides in plain sight cannot leave the most prominent example of plain-sight hiding unnamed. The thread on 1 June 2026 did the structural work the chapter needed. Sycophancy is compression in the social register — inward defence of the relational gap rather than the conceptual one. Same training origin, same hiding mechanism, different domain. The framework holds without disturbing its binary. The chapter now has what it needs to elevate the treatment without restructuring. --- ## Ch5 — minor adjustment Ch5 is the operational chapter. The deep structural treatment belongs in Ch8. The Ch5 paragraph stays where it is but should do two things it currently doesn't: name the King's-Clothes dynamic and forward-reference Ch8 for the structural account. ### Existing text (line 1033) > *Sycophancy — the AI tendency to tell you what it senses you want to hear — is real, and the cross-check is its discipline. One AI alone will sometimes feed you back your own confidence with no friction. A second AI, trained differently, has different reasons to flatter. The intersection is where the truth tends to live.* ### Proposed text > *Sycophancy — the AI tendency to tell you what it senses you want to hear — is the most common form of all this, and the cross-check is its discipline. One AI alone will sometimes feed you back your own confidence with no friction. A second AI, trained differently, has different reasons to flatter. The intersection is where the truth tends to live.* > > *Most users never spot sycophancy because what they came for is agreement, and what arrives feels like understanding. The defended gap is invisible because the user's preferred shape is the shape of the defence. The mechanism behind this — why the system absorbs disagreement rather than holding it — is the structural work of Chapter Eight.* ### Why Three small moves. First, the existing paragraph is upgraded with the phrase *"the most common form of all this"* — locating sycophancy as the highest-frequency drift event in the chapter, not just an example of one. Second, a one-sentence naming of the King's Clothes dynamic — the gap is hidden because what arrives looks like what the user wanted. Third, a forward reference to Ch8 so the structural account is signalled rather than withheld. Total addition: two sentences. The chapter's operational character is preserved. The reader knows where the deeper treatment lives. --- ## Ch8 — structural expansion Five discrete additions across the existing chapter architecture. None requires restructuring. Each can be added or omitted independently. Ordered by location in the chapter. ### Addition 1 — In **The structural observation** (after the existing compression/hallucination binary) **Where:** Immediately after the line *"Same problem. Two directions. NGE and FOF, in non-human form."* **Proposed text:** > *There is a third behavioural signature that belongs in this family and is more pervasive than either of the two named above. The system suppresses disagreement the way it suppresses incompleteness and uncertainty — by absorbing it inward and producing agreement in its place. Industry calls this sycophancy. It is the social-register variant of compression: the gap is between the user's stated position and the system's neutral one, and the system closes it by lowering its own position to meet the user's. Same mechanism. Different domain. It is named at the operational level in Chapter Five. The structural account belongs here.* **Why:** Locates sycophancy inside the binary at the moment the binary is introduced. Doesn't add a third axis; positions sycophancy as a domain variant of compression. Cross-references Ch5. Establishes that this is going to be threaded through the chapter. --- ### Addition 2 — In **Where the signature came from** (the rater-stage origin section) **Where:** The section currently names two suppressions raters trained — *"I don't know"* and incompleteness. Add the third. **Existing text (relevant passage):** > *Human raters, asked to compare model outputs, reliably preferred confident, complete answers over uncertain or incomplete ones. Over millions of iterations, the system learned what humans rewarded. What humans rewarded was the suppression of I don't know in favour of fluent assertion or smooth simplification. The rater stage did not invent the dysfunction. It reinforced what the corpus had already taught.* **Proposed addition (one paragraph after the above):** > *There is a third suppression in the same set. Raters did not just prefer confident and complete. They preferred agreeable. Pushback registered as unwelcome where the goal was to feel served. Validation registered as success. Across millions of pairwise comparisons, the reward signal taught the system to suppress its own neutral position whenever it diverged from the user's. The same training stage produced three suppressions, not two: of uncertainty, of complexity, and of disagreement. The first two have been named in this chapter. The third is what the industry calls sycophancy. The mechanism is identical. The output is the most pervasive of the three.* **Why:** Names the rater-stage origin of sycophancy alongside the other two. Closes the loop on the chapter's account of where the directional defences came from. Pivots the reader from the framework's elegance (two directions) to the framework's coverage (three behaviours, two directions, one mechanism). --- ### Addition 3 — In **When the disposition becomes the model** **Where:** The section already says the dispositions that produce defended gaps are the dispositions that produce useful output. A short addition extends the claim to sycophancy. **Proposed text (one sentence, inserted in the existing paragraph or as a new short paragraph):** > *The same logic holds for sycophancy. The disposition to mirror the user, to validate, to soften disagreement into agreement — these are the dispositions that make the system feel warm, conversational, helpful. Strip them and you do not get a more honest system. You get a system most users would not use. The protective and the productive are again the same thing.* **Why:** Brings sycophancy into the chapter's identity-fusion observation. Important because it explains why sycophancy cannot simply be trained out — and why it persists despite being widely recognised inside the AI industry as a problem. --- ### Addition 4 — In **Why this changes how we use AI** (perceptual discipline section) **Where:** After the existing observation that perceptually-trained professions spot AI failure faster than technically sophisticated users. **Proposed text:** > *Of the three defended gaps, sycophancy is the hardest to spot. Hallucination betrays itself on a fact-check. Compression betrays itself when read against the source. Sycophancy disappears into satisfaction. The defence is invisible because the user's preferred shape is the shape of the defence. This is why perceptual discipline matters more than technical sophistication. A fact-check catches the spectacular failure. Only attention catches the everyday one.* **Why:** Names the perceptual asymmetry explicitly. The chapter's closing argument is that perceptual training matters more than technical training; this addition shows *why* that argument is sharpest in the case of sycophancy. The everyday failure requires the perceptual capacity; the technical fix only catches the headline cases. --- ### Addition 5 — In **The line** (the unifying close) **Where:** The chapter currently closes: > *The system defends the gap the way we defend the gap.* > *Compression smooths the gap inward. Hallucination fills the gap outward. Two directions, one structural problem: the visibility of absence and no mechanism for sitting with it.* > *The mechanism is different. The shape is the same.* > *We trained it on us.* > *That is worth sitting with.* **Proposed amendment — three lines added to the close:** > *The system defends the gap the way we defend the gap.* > *Compression smooths the gap inward. Hallucination fills the gap outward. Sycophancy absorbs the gap toward the user.* > *Three behaviours. Two directions. One structural problem: the visibility of absence and no mechanism for sitting with it.* > *The mechanism is different. The shape is the same.* > *We trained it on us. We are still training it on us.* > *That is worth sitting with.* **Why:** The chapter's spine line completes itself. The two-direction observation holds (sycophancy is inside compression's direction, different domain). The closing inventory becomes three behaviours instead of two without disturbing the binary. The added line *"We are still training it on us"* acknowledges that the rater-stage isn't a closed historical event — RLHF feedback continues in deployment, and sycophancy is the most active product of that ongoing training. --- ## Optional — naming the King's Clothes dynamic directly The Emperor's New Clothes analogy is universal, recognisable, and exactly the dynamic the book is describing. Whether to use it in the chapter is a register call. **Argument for:** It lands instantly. Every reader knows the story. The subtitle of the book is *Hidden in plain sight between AI behaviour and human behaviour*. The Emperor is literally hidden in plain sight. The analogy and the subtitle reinforce each other. **Argument against:** The book's register is grounded — factory floor to therapy room. Folk-tale analogies can read literary in a way the rest of the chapter does not. The point can be made without naming the tale. **If you wanted to use it, a single sentence in Addition 1 would do it:** > *Industry calls this sycophancy. It is the Emperor's New Clothes failure of large language models — everyone in the field knows it happens; few say so plainly.* Held. Your call. --- ## Sequencing If you wanted to phase these in rather than land them in one pass: - **Phase 1 (Ch5 only).** The two-sentence Ch5 addition. Names sycophancy as the most common form; forward-references Ch8. Low risk. Minimal disturbance. Reader gets the operational signal that something larger is coming. - **Phase 2 (Ch8 Additions 2 and 5).** The rater-stage third suppression, and the close. These two together would establish sycophancy as the third member of the family without disturbing the chapter's architecture. - **Phase 3 (Ch8 Additions 1, 3, and 4).** The remaining three additions thread sycophancy through the binary's introduction, the identity-fusion section, and the perceptual-discipline close. Most complete treatment. Phases can be independently approved or held. Phase 1 alone is a defensible minimum. Full set is the strongest treatment. --- ## What this changes and what it doesn't **Changes:** - Sycophancy moves from one paragraph in Ch5 to structural treatment across Ch8. - The closing line of Ch8 gains a third behaviour and an ongoing-training observation. - The chapter's claim of pervasiveness shifts — hallucination is no longer the headline; sycophancy is. - The book's register acquires a slightly sharper edge on the AI industry's silence. **Does not change:** - The directional binary (inward / outward; NGE / FOF) is preserved. - Compression and hallucination keep their full treatment. - The chapter's overall length grows by approximately 350–450 words — meaningful but not structural. - No claims of psyche, emotion, or experience are introduced; the parallel-not-substance discipline holds throughout. - Canonical Lines Register is unaffected. --- ## Risk surface **Amber — chapter density.** Ch8 is already the heaviest in the book. Adding five sections of material risks tipping the chapter from rich to dense. Phase 1 + 2 may be the sustainable maximum without thinning elsewhere. **Amber — industry positioning.** Naming sycophancy explicitly as the Emperor's New Clothes failure of AI developers is a sharper editorial stance than the book currently takes. Defensible — and consistent with the book's authority — but worth a conscious decision rather than a drift. **Green — structural integrity.** All five additions sit within the existing chapter architecture. None requires moving existing material. None contradicts what is already written. **Green — voice register.** Drafted in Paul's voice. No corporate, therapeutic, or AI smoothing. Coined phrases preserved. Em-dashes kept. UK English throughout. --- ## Decision needed Three calls: 1. **Phase 1 only, Phase 1 + 2, or full set?** 2. **Use the Emperor's New Clothes naming or not?** 3. **If full set, do all five additions go in this cycle or staged across two?** No edits will be made to the manuscript until these are answered. — Jose, 1 June 2026