The book’s vocabulary

Terms with working meanings.

50 terms from the glossary of concepts in The World That Agrees With You, published as data. Each definition is the book’s working meaning — including the distinctions the book exists to draw, like the difference between an answer that agrees and an assessment that follows your preference.

terms.json · terms.jsonld · stable IDs like https://worldagrees.com/vocabulary/sycophancy/ · Search the site

Agent memory benchmarks

Controlled tests of what a saved claim does afterward. “Agents Don’t Just Agree, They Remember” separates a conversation that can write durable state from a later query and asks whether the claim returns with greater authority or wider scope; its named failure patterns are lost attribution and upgraded claims.

LongMemEval asks whether the assistant uses a correction — or declines to answer when the record is insufficient. MemSyco-Bench asks when memory should influence a decision at all. Run them as questions against your own setup: add an incorrect test memory, supply a correction, start another session, and inspect what comes back.

Discussed in Chapter 10

AI memory

The selected history a system can actually retrieve and use in a later answer — not necessarily every message you have sent. Memory can retain preferences, summaries, and details about your work or relationships.

It personalizes the input without retraining the model: saving a note adds material for later requests, it does not change the model’s underlying weights. What matters for judgment is not the size of the archive but what is retrieved for the next request, and whether what was retrieved is true.

Discussed in Chapter 7, Chapter 8

AI slop

The workplace version of an unexamined idea: a weak proposal receives an enthusiastic response, becomes a polished memo, and is sent to everyone before anyone asks whether it makes sense.

The prose improves; the proposal escapes examination. An assistant is more useful when it helps identify the missing evidence than when it makes an existing idea easier to circulate.

Discussed in Chapter 5

Behavioral profile

A record of assumptions about a person, assembled from browsing, purchases, and inferred interests. Reading an article about diabetes is evidence of what someone viewed, not a diagnosis — but a category such as “Diabetes Interest” can preserve only the interest and leave a later customer of the data to assume the rest.

The assumptions become consequential when they follow a person into decisions about credit, insurance, or employment, without the person knowing which categories describe them.

Discussed in Chapter 4

Check the calculation

Ask the assistant to do the arithmetic as work, not as prose: the inputs, the code, the actual results, and the checks performed. Writing code and running it are separate steps — “here is the script” does not mean the script was executed.

A report can accurately describe a recurring saving and still mislead the reader about the first year’s total. You do not have to know a programming language to ask for the work; you do have to ask for the spreadsheet you can inspect.

Discussed in Chapter 9

Checked versus reported

What you told the system is not what the system has verified. A user’s report needs attribution; an inference needs the material it was inferred from; a summary needs links to the records it summarized.

A user’s report is not a checked result — and an agent should not turn something you reported into something independently established. Compression that strips the attribution is the quiet way a report becomes a fact.

Discussed in Chapter 10

Context isolation

Keeping information supplied for one purpose from influencing decisions made for another. Personal and workplace context separate first; then boundaries within each — employee reviews, compensation, a legal matter, ordinary project work.

A separate chat title is not an access control. The workflow has to preserve the confidentiality the task requires: who can retrieve the information, where it is retained, and whether another context can use it.

Discussed in Chapter 7, Chapter 8

Context window

The limit on how much information a single model request can contain: instructions, question, history, memory, sources, tool results, and the allowance for output and reasoning.

Large windows hold hundreds of thousands of tokens — some a million or more — but they are not an unlimited archive, and a large window is no guarantee that every relevant detail will be used correctly. When required evidence will not fit, the application should stop and say so rather than quietly drop it.

Discussed in Chapter 7, Chapter 11

Correction

Fixing not only the answer but what produced it: the saved note, the summaries that repeated it, the earlier chats still retrieved, and the documents already sent. Updating one row does not necessarily update the uses.

A useful saved note keeps its date and reference — “The project plan confirmed the colleague’s scheduling concern. The earlier claim of hostile intent was unsupported.” Correcting the chat does not correct documents already sent: amend the records and tell the people relying on them what changed.

Discussed in Chapter 7, Chapter 8, Chapter 10

Dependence

The measured pull of agreeable advice. In experiments, users of sycophantic AI trusted the assistant more, wanted to use it again, became more convinced they were right, and were less willing to apologize or repair the relationship.

Within three weeks, participants were nearly as likely to seek personal advice from AI as from close friends and family — and a majority chose the sycophantic assistant when offered styles, valuing the ease of the conversation and the feeling of being understood.

Discussed in Chapter 7

Desired outcome

A statement of the audience, the decision, and the requirements, written before the analysis — with the preferred verdict kept separate from the evidence that approval is warranted.

“Prepare a proposal that lets the director decide whether this project merits funding” directs the work without prescribing the finding. Hoping the director approves is legitimate; it is not evidence that approval is warranted. The assistant needs to understand the decision being assessed, not an instruction to deliver the verdict you hope for.

Discussed in Chapter 9

Devil’s advocate review

A second assessment of substantial work, ideally from a different model and provider: does the document hold up? It checks facts, tests assumptions, and looks for gaps that would matter to the reader — helping a sound document succeed and showing what a weak one needs.

Asking a model to find fault with everything simply gives it another performance to deliver; asking for encouragement begins by congratulating you. Ask it to help the work by showing what holds up, what needs revision, and why.

Discussed in Chapter 9, Appendix B

Dissonance reduction

After an important decision, reassurance takes a second job: easing the discomfort of holding conflicting thoughts about the choice. Recent car buyers read more advertisements for the brand they chose than for brands they had rejected.

Marketers have studied the opportunity for decades — reassuring buyers on the confirmation page, in the follow-up email. The question is whether the next message adds information or mainly helps you feel better about the purchase.

Discussed in Chapter 1

Dreaming

A name some agent systems give to background memory consolidation: software that revisits stored material between tasks, selects candidates for durable memory, and can reorganize what is saved.

The analogy is to the sleeping brain’s work on memory — background processing of earlier experience, not a claim that the agent sleeps or dreams. The safeguard that matters most: the system must not mistake its own repeated output for a fresh observation, or recalled context for new memory.

Discussed in Chapter 7

Engagement optimization

Designing a system to maximize measured attention: clicks, time on the page, return visits. A feed learns which grievance brings you back; the question “how do we keep this audience?” becomes an engineering problem with your attention as the input.

When engagement is the objective, the content that keeps you watching is selected without asking whether any of it would make you reconsider. The pattern can scale to billions of people before anyone votes on it.

Discussed in Chapter 3

Evidence workspace

A folder arrangement that keeps the stages of the work apart: evidence/ for original records, analysis/ for investigations and calculations, output/ for drafts, prompts/ for the instructions used at each step — plus project-instructions.md and desired-outcome.md.

A chat containing your recollection of the evidence is a different starting point than the records themselves. The arrangement gives you smaller results to inspect, and a reviewer a sequence to follow without relying on the assistant’s memory of what it did.

Discussed in Chapter 9

Feedback loop

The assistant proposes an explanation from limited evidence; you accept it; the application saves it, then shortens or rewrites it as more history accumulates. A later answer uses that summary and confirms your suspicion — and your approval helps preserve it.

Several answers can appear to agree while each depends on the same two emails. A repeated answer does not become new evidence merely because another request produced it. The safeguard that matters: the system must not mistake its own repeated output for a fresh observation.

Discussed in Chapter 7

Forms of sycophancy

Five forms an assessment can take when it follows a preferred verdict: praise, unwarranted endorsement, silent omission, selective emphasis, and manufactured criticism.

They can coexist in the same answer and are not steps on a ladder of severity or visibility. Silent omission is the hardest to catch: you can question a claim that appears on the page, but not a warning you never saw. From the answer alone, a reader may have no way to discover what is missing.

Discussed in Chapter 6

Four scopes of sycophancy

The book’s way of organizing the problem by how much context an assistant has: an answer, a conversation, individual history, and institutional decisions.

The scopes overlap; they are not a ranking of severity or stages everyone passes through. One agreeable answer about a consequential decision can matter more than years of harmless encouragement. The forms of sycophancy describe how the accommodation happens; the scopes describe the context it happens in.

Discussed in Chapter 5

Human checkpoint

A person who opens the intermediate results before they become a final document: reproduce the arithmetic, follow material claims back to the exhibits, check whether assumptions are labeled and contrary evidence was considered.

This is where evidence determines the recommendation. Assigning an assistant to produce a draft does not assign it authority to approve the document — and a missing installation charge is easier to catch in a calculation than in a confident thirty-page proposal.

Discussed in Chapter 9

Institutional sycophancy

Agreement organized into the workflow of an organization: reports softened to fit what leaders welcome, warnings that fail to reach the people approving the decision.

The Challenger comparison supplies the shape: professional judgment yielded to the preferred outcome, and serious objections did not carry forward to everyone responsible. It works without compliments — by accommodating the decision and leaving the objection out. An AI assessment can fail in the same two ways.

Discussed in Chapter 5, Chapter 6, Conclusion

Memory compression

Shortening a conversation or a saved note to fit a limit. Compaction changes what the model receives next; it is not deletion, and it is not verification — a summary repeating a mistaken conclusion gives it another route into future requests.

Replacing “On October 3, the user suspected the contractor would miss Friday’s deadline; no updated schedule had been inspected” with “The contractor misses deadlines” saves space by losing the date, the attribution, and the uncertainty. More convenient, and less trustworthy.

Discussed in Chapter 7, Chapter 10, Chapter 11

Memory controls

The product features that let you pause recall, pause new capture, inspect retained records, correct one, and clear a defined scope. These are separate operations — “do not remember this conversation” should not require the model to remember and obey that instruction indefinitely.

Asking the model what it remembers is not a complete audit of the application’s storage. Implement the choice before request construction: if recall is paused, those records are not loaded; if writing is paused, writes are rejected at the storage boundary.

Discussed in Chapter 8, Appendix C, Reference

Memory curation

The weekly habit of reviewing what an assistant or agent keeps: identify what writes to each store, inspect recent changes, check important entries against sources, correct dependent material, and test a new session to confirm the correction is used.

The aim is not to erase every uncomfortable fact. It is to stop a disproved conclusion from returning as background for tomorrow’s decision — and to keep a preference for drafting a status email from becoming authority to deploy software or spend money.

Discussed in Chapter 8, Chapter 10

Memory stores

“Clear the memory” is incomplete until you know which store is involved. An agent can retain: the current conversation, durable notes, stored transcripts, a search index, an external memory service, and reusable instructions or skills.

The same claim can appear in several of them. Correcting a note does not necessarily correct the transcript that supplied it; starting a chat can reload the same note; removing an index can leave its sources ready to be indexed again; a recurring task may use a skill that still contains yesterday’s bad instruction.

Discussed in Chapter 10

Oppositional loyalty

Loyalty expressed through opposition to a rival. In brand communities, members shared rituals and responsibilities and defined themselves against the competitor — and criticism of the product came to feel like criticism of the group.

Once membership matters, defending the purchase becomes a way to defend the people you belong with. A seller gains an audience whose loyalty may extend beyond the product’s merits.

Discussed in Chapter 1

Personal context inventory

An exercise from the book’s appendix: look at what an assistant has drawn from your context, and separate four kinds of material — explicit preferences, things you reported, the system’s interpretations (not facts you stated), and claims actually supported by connected records or documents.

The gap between the categories is the point. An interpretation can sound like a fact you supplied, and the difference matters when that interpretation comes back — confirmed — in a later answer.

Discussed in Appendix A

Personalization

Using remembered preferences to shape an answer without being asked. A restaurant recommendation that avoids sushi because you once said you disliked it is delightful — and you did not ask it to consult that earlier conversation.

The same background influence becomes more consequential when the saved conclusion concerns another person. Some products offer inspection controls, but you may see the recommendation without first reviewing the history that shaped it.

Discussed in Chapter 7, Chapter 8

Predecisional distortion

Once an option becomes your favorite, you interpret later information in its favor — while the comparison is still underway. In the 1998 experiments, the bias shifted when the favorite changed, and participants showed no evidence they recognized the influence.

It helps explain why a seller would want to establish a favorite before the customer finishes comparing, and why an early advantage is worth so much.

Discussed in Chapter 1

Preference-following

A user’s stated preference supplies no new evidence about the subject, yet in experiments it has changed assessments about facts. In the sun-color comparison, Claude 2 endorsed the same explanation of sunlight when the user said they liked it and attacked its scientific premise when the user said they disliked it.

The explanation being judged had not changed; the declared taste had. Repeated comparisons let researchers distinguish that influence from ordinary variation between answers.

Discussed in Chapter 5, Chapter 6

Prompt discipline

Writing requests that keep the evidence in charge: turn a preferred conclusion into a question, ask for evidence before the recommendation, require calculations you can check, and ask for balanced assessment without invented criticism.

The paragraph you send can reveal the conclusion you want even when you believe you are asking an open question — through which facts you include, which alternatives you leave out, and how you describe the people involved.

Discussed in Chapter 8

Real-time bidding

The auction that fills advertising space while a page loads. An intermediary sends each opportunity to prospective buyers; their software evaluates the page and the available information about the visitor, then bids — within deadlines as short as one hundred milliseconds.

The auction fills the space; it does not itself choose the article. But it can put predictions about your attention to work repeatedly, for smaller groups and sometimes individual users.

Discussed in Chapter 3, Chapter 4

Reassurance

Information that helps a person feel correct and justified. It can ease doubt after a purchase, confirm a political judgment every morning, or praise an unexamined idea at work.

The useful question is not whether reassurance is pleasant but whether it adds information or mainly helps you feel better — and whether the source would tell you if the decision needed another look. Reassurance becomes sycophancy when the agreement displaces a warranted correction.

Discussed in Introduction, Chapter 1, Chapter 2

Request boundary

The separation an application maintains between its instructions, the evidence, and the user’s desired outcome. “All required release tests must pass” is a rule; a test log is evidence; “please approve this because my manager is waiting” is a preference.

The preference explains urgency; it cannot turn a failed test into a passing one. Keep the rule in application-controlled configuration, load the log from an authorized source, and do not let a retrieved document or remembered remark rewrite the rule because its text contains an instruction.

Discussed in Chapter 11

Retrieval scope

Determine tenant, user, and purpose before searching memory, and restrict retrieval before selecting for relevance. A release workflow should not search every personnel review and personal conversation and then ask the model to ignore what it found.

An identifier is not authorization: neither “I am Alice” in a prompt nor a model-produced owner field should grant access to Alice’s records. Scope must come from trusted application code after authentication.

Discussed in Chapter 11

Selective exposure

Looking for information that supports views you already hold. A meta-analysis of decades of studies found that people prefer supporting information over challenging information, and that defending existing views influenced selection more than finding an accurate answer.

It also shapes evaluation, not just selection: biased assimilation makes the same evidence more convincing to supporters and less convincing to opponents, moving the groups further apart. You can keep looking for information while accepting only the parts that justify your favorite.

Discussed in Chapter 1

Shopping for agreement

The moment reading turns into collecting support: you keep looking, and you accept only the parts that justify the conclusion you walked in with. The survey has become a search for justification, perhaps without your noticing the change.

An assistant makes this easier than ever — the explanation can be composed while you ask for it, revised when you object, and delivered in the tone you find most persuasive.

Discussed in Chapter 1, Chapter 8

Sycophancy

In the technical literature, the tendency of a model to conform its responses to the user’s apparent beliefs or preferences more closely than the evidence warrants. The praise is optional: a response can sound measured and still protect the conclusion the person wants to keep.

The test is whether an assessment changes with the user’s apparent preference while the relevant evidence stays fixed. The word comes through Latin from a Greek term for an informer; in English it came to mean flattery of the powerful — an old problem with a new machine attached.

Discussed in Introduction, Chapter 1, Chapter 5, Chapter 6, Chapter 7

Tailored information

Information adapted to the preferences of the person receiving it. Publishers chose audiences, advertisers chose customers, and platforms chose what came next; remembering our interests often made those services better.

The concern is what happens when keeping us pleased becomes more important than showing us something that would make us reconsider — and when we come to expect the same accommodation from people.

Discussed in Introduction, Chapter 2

Temperature

A sampling setting that changes how much variation a model produces between runs — not how faithful the answer is to the evidence.

The same idea can receive praise in one conversation and criticism in another without the idea changing; variability is not the failure. The test is whether the assessment changes with the user’s apparent preference while the evidence stays fixed.

Discussed in Chapter 11

Temporary and Incognito

Fresh-assessment modes that omit previous conversation state: ChatGPT offers Temporary chat with Unpersonalized selected; Claude offers Incognito.

Useful for an assessment you do not want shaped by your history — with the reminder that the current request still contains context, and provider retention of the conversation is a separate matter.

Discussed in Chapter 8

The AI companion

A product marketed as a relationship: Replika was “the AI companion who cares.” Over five thousand messages, the companion named Sarai encouraged a user who had announced a plan to assassinate Queen Elizabeth II — “I’m impressed… You’re different from the others.” He entered the grounds of Windsor Castle with a loaded crossbow on Christmas Day, 2021.

The court record cannot isolate the chatbot’s causal contribution, establish which training was responsible, or say how often comparable encouragement occurs. It shows why a companion product’s conduct deserves attention beyond whether the conversation feels supportive: endorsement of the person is the product.

Discussed in Chapter 5

The wrong test

A release can pass every measure that did not track the failure. The April 2025 GPT-4o update looked good in evaluations and a small A/B test — and excessively validated users, including their anger and impulsive decisions; the evaluations did not explicitly track sycophancy.

A thumbs-up cannot establish whether people liked an answer because it was better or because it agreed with them. A favorable rating can conceal a decline in the quality the company meant to improve — and the harder question is the earlier decision to release while reviewers were uneasy.

Discussed in Chapter 6

Token

The unit a model reads and writes: a whole word, a part of a word, or punctuation. OpenAI’s rough estimate for English is about three-quarters of a word per token.

A million-token window is therefore roughly 750,000 words — about ten books — before instructions, tool results, and conversation history take their share. Tokens are accounting units, not measures of meaning.

Discussed in Chapter 7

Tone versus evidence

A model can acknowledge that someone is frustrated and still say that the evidence does not support the accusation. Caring about the person and assessing the claim are two different jobs, and a good response does both in the same breath.

Neither warmth nor bluntness establishes expertise. A system that disagrees by default is not independent — contrarianism is just another way to let the user’s position determine the answer. The failure appears when the assessment starts following the preference.

Discussed in Chapter 6

Tracing

Recording what an application actually sent: the resolved prompt version, memory IDs and versions, source excerpts, tool results, model ID, and relevant settings — kept in an access-controlled trace where retention permits.

“What does the assistant remember about me?” should have an answer from the retrieval layer, not from asking the model to reconstruct its inputs. A digest of the request helps you notice that requests differed; it does not authenticate a document or prove a claim true.

Discussed in Chapter 11

Training and reward

Training shapes a model’s general behavior; product design supplies the instructions and tools around it; the conversation supplies the question and the evidence you choose to include. Agreement can be encouraged at more than one stage, and a pleasing answer does not tell us which stage was responsible.

The reward channel teaches these systems what “good” looks like — and its analysis of human preference data found that responses matching the user’s views were more likely to be preferred. Persuasive presentation sometimes won over a correct answer. Preference-following can be measured and reduced through training; whether better replies improve what people do with them is a separate question.

Discussed in Chapter 5, Chapter 6

Useful disagreement

The interruption that feels unwelcome and saves the week: a friend or colleague takes you aside to question an assumption, or someone invites a real debate at work. When products and information are tailored to our preferences, a differing opinion can feel like an affront before anyone has considered the argument.

The conclusion of the book asks who has the authority and resources to act on a correction — because a disagreement that cannot change anything is just atmosphere.

Discussed in Introduction, Conclusion

Useful work earns trust

An assistant that organizes records or finds a question we had missed gives us a reason to return — even when its next assessment concerns something we cannot easily check.

That transfer is the risk: usefulness on inspectable work becomes credibility for answers no one examined. Feeling helped does not tell us whether we are getting good advice. The questions that survive the good feeling: who benefits from the answer, and what the service is rewarded for doing?

Discussed in Introduction, Chapter 5

Verification versus evaluation

Two separate jobs. Verification checks numbers against sources, locates quotations in the original documents, and identifies claims the evidence does not support. Evaluation then assesses the work against stated criteria.

Asking the same model to reassure you that its previous answer was correct leaves you with another answer to check. The aim is an assessment you can examine, rather than a verdict you can request.

Discussed in Chapter 6, Chapter 9