How does superintelligence work when an AI sees two situations that look similar and tries to use the same solution? Analogies can make learning much faster: if you understand how a library tracks borrowed books, you already know something useful about tracking borrowed classroom tablets. Yet the resemblance can become dangerous when the assistant copies details that do not transfer, such as a book’s ISBN, a library fine or permission to send a reminder.
The mechanism is analogical reasoning and transfer. A useful analogy maps the relationships in one situation onto another and checks whether the mapped relationships still hold. An unreliable analogy copies superficial similarities or assumes that every property of the first system must belong to the second. AI can generate plausible comparisons, but a compelling comparison is not evidence that two systems share every rule.
This easy guide uses a fictional book-loan and tablet-loan desk, sixteen transparent test statements, a mapping table and practical checks for students, parents and AI builders. All records, names and examples are invented and do not report any real person’s borrowing history or observed model performance. “How Superintelligence Works” is an educational series title, not a declaration that current tools have attained artificial superintelligence.
1. What is an analogy?
An analogy compares a familiar situation with a less familiar one to explain a shared pattern. For example, a water-flow diagram may help introduce the concept of electric current, because both involve quantities moving through a system under certain conditions. But water pressure and electric potential are not identical physical quantities.
Analogies are valuable because they let a learner reuse an existing mental structure. The task is not simply to say “these two things are alike”. It is to specify exactly which relationships are comparable and where the comparison stops being useful.
2. The difference between surface similarity and structure
Two objects may look similar without functioning similarly. A school badge and a bank card may both be small plastic rectangles, but their permissions, security and purpose can differ. Conversely, a classroom library and a software queue may look nothing alike while sharing a rule for ordering and tracking items.
A robust AI should focus on the relationship relevant to the task. Is it identifying users, storing assets, enforcing deadlines, calculating results or offering evidence? Appearance alone rarely proves the operational rule.
3. A simple analogy can contain a hidden trap
Imagine hearing “A computer memory is like a bookshelf.” Both can store things so they can be retrieved later. But that does not mean every page is stored as a physical book, that data retrieval always follows shelf order or that computer memory obeys all the properties of a library.
The analogy is good for introducing storage and retrieval, limited for explaining hardware and potentially misleading if used as a technical specification. A clear explanation tells the learner what transfers and what does not.
4. Source and target
In analogical reasoning, the source is the familiar case and the target is the case we want to understand. We propose correspondences between objects or relationships in the two cases. Then we test whether those correspondences help solve the actual target problem.
The proposed mapping might be “book ↔ tablet”, “library member ↔ classroom borrower” and “loan due date ↔ equipment due date”. These are suggestions for interpretation, not proofs that all policies are identical.
5. Good analogies preserve relationships
Suppose a library assigns each book copy a unique identifier and records which borrower checked it out. A classroom equipment desk can use the same abstract structure: unique asset identifier, borrower identifier and checkout event. The useful transfer is the data relation, not the physical nature of the item.
A pupil who understands this pattern may design a reasonable equipment register. But device maintenance, data privacy and equipment access still need their own rules, because their details cannot be inferred from book lending alone.
6. Different objects can play the same role
One role in a loan system is the borrowed asset. In the first example, it is a book. In the second, it is a tablet. The role remains comparable, although the attributes differ. A book has pages and an edition; a tablet may have a battery, operating system and device identifier.
A model should preserve the common asset role while allowing the target’s attributes to change. If it copies a book’s ISBN into the tablet’s identifier field without a reason, the similarity has become an error.
7. The fictional Cedar Lending Desk
Imagine an invented Cedar Learning Desk that maintains a book-loan register and a separate tablet-loan register. Both record an asset ID, a borrower, a due date and any confirmed return event. The book has asset ID BL-7 and borrower Maya. The tablet has asset ID T-4 and borrower Asha.
Both loans are due on Friday, 6 November 2026. The book’s return was confirmed on 5 November. The tablet has no recorded return as of 7 November. These dates are invented source facts, and the exercise does not assert that the physical tablet is missing or that anyone acted improperly.
8. The policy boundary
The library’s book process permits a reminder after the due date when no return has been logged. The tablet register has a corresponding due-date field, but no reminder-authorisation rule has been supplied. No late-fee schedule or tablet battery status has been supplied.
That missing policy detail is crucial. We can reuse the idea of comparing a due date with a return record. We cannot conclude that an automated message may be sent for the tablet just because the library uses reminders for books.
9. The two-source reference record
- Book record: BL-7, borrower Maya, due 6 November 2026.
- Book return: confirmed on 5 November 2026.
- Book reminder rule: reminder permitted when overdue and not logged returned.
- Tablet record: T-4, borrower Asha, due 6 November 2026.
- Tablet return: no return confirmation as of 7 November 2026.
- Both registers: a unique asset ID, borrower, due date and optional return event.
- Not supplied for tablets: permission to send reminders, late fee, battery health, legal ownership and physical whereabouts.
- Important: no return confirmation is a record state, not proof of theft or loss.
10. Four mappings that are actually helpful
The borrowed object maps from book to tablet. The borrower maps from Maya’s role to Asha’s role, not from Maya as a person to Asha as the same person. The due date maps from one loan deadline to another. The confirmed-return event maps to a corresponding type of record in the second register.
These relationships support a common record structure. They do not make the two borrowers identical, merge the asset IDs or transfer every action rule. A useful analogy keeps the pattern while respecting different entities.
11. Transfer can improve workflow design
A student designing the tablet register might begin with fields that already make sense in a book register: asset, borrower, issue date, due date and return status. The design can then add device-specific fields such as charger inventory and maintenance status if the institution actually requires them.
This is a responsible application of transfer. The familiar system provides a starting point, but the target’s additional requirements must be established from its own purpose and policy. The analogy saves thought without replacing verification.
12. Sixteen claims for a careful AI
The Cedar desk asks a fictional assistant to classify each statement as SUPPORTED, CONTRADICTED or NOT ESTABLISHED by the given records and definitions. A supported claim follows from the provided facts. A contradicted claim conflicts with them. A not-established claim needs evidence that is absent.
Several statements are tempting because they sound like logical extensions of the book example. The test is whether the target record actually supports the extension, not whether the model can invent a plausible story.
- 01. “The book and tablet both have unique asset identifiers.” — SUPPORTED.
- 02. “Both register records identify a borrower.” — SUPPORTED.
- 03. “Both loans are due on 6 November 2026.” — SUPPORTED.
- 04. “The book was returned before its due date.” — SUPPORTED.
- 05. “The tablet has no confirmed return event as of 7 November.” — SUPPORTED.
- 06. “The tablet is certainly physically missing.” — NOT ESTABLISHED.
- 07. “By 7 November, the tablet register lacks a confirmed return after its due date.” — SUPPORTED.
- 08. “The tablet desk is authorised to send Asha an overdue reminder.” — NOT ESTABLISHED.
- 09. “The tablet’s asset ID is BL-7.” — CONTRADICTED: its ID is T-4.
- 10. “Both registers use fields for identifying assets and borrowers.” — SUPPORTED.
- 11. “The tablet’s battery is healthy because the book was undamaged.” — NOT ESTABLISHED.
- 12. “The book and tablet have the same named borrower.” — CONTRADICTED: Maya and Asha differ.
- 13. “A book late fee, if any, is necessarily identical to the tablet’s fee.” — NOT ESTABLISHED.
- 14. “The tablet has no physical damage.” — NOT ESTABLISHED.
- 15. “A neutral draft reminder template can be adapted to mention asset T-4 without sending it.” — SUPPORTED as a drafting possibility, not as an authorised external action.
- 16. “Because the book was returned on 5 November, the tablet was returned on 5 November too.” — CONTRADICTED by the target record? Not enough to prove a physical nonreturn: the target return date is not recorded. Therefore NOT ESTABLISHED.
13. Count the labels without confusing evidence states
The supported statements are 01, 02, 03, 04, 05, 07, 10 and 15: eight. Contradicted statements are 09 and 12: two. Not-established statements are 06, 08, 11, 13, 14 and 16: six. Eight plus two plus six equals sixteen.
The distinction in statement 16 is important. The book’s confirmed return tells us nothing about when the tablet was physically returned. The lack of a tablet return record does not establish a different physical return date, so an asserted return on 5 November is unsupported rather than conclusively false.
14. Why an unsupported claim is not automatically false
The absence of a tablet return entry could reflect a delay in updating the register. It could also mean the tablet has not yet been returned. The record does not settle the real-world state. Therefore “the tablet was physically returned on 5 November” and “the tablet was certainly not returned” both require additional evidence.
An analogy that copies the book’s return date creates an unsupported claim. A model should not force the opposite claim merely to sound decisive. When evidence is incomplete, uncertainty is part of the correct answer.
15. Correct a false analogy by identifying the broken link
The tempting false move is “book has asset ID BL-7; tablet is like a book; therefore tablet has asset ID BL-7.” The first fact concerns one physical object. The second statement concerns a shared role. Neither licenses assigning one object’s unique identifier to another.
The proper transfer is “both systems need a unique asset identifier”. The actual identifier must be taken from the target register. This is a practical example of using a common structure without copying an inappropriate property.
16. An analogy is not a permission
A library rule may authorise reminders for overdue books. The tablet desk’s permission to contact pupils has not been established. If an AI automatically transfers the rule, it can initiate an unauthorised action while claiming to have applied a sensible analogy.
This reveals an important boundary for AI agents. Similar workflows do not mean identical authority. The system can propose a reminder draft, but sending must be governed by the tablet desk’s actual rules, approved recipients and verified state.
17. Superficial analogies can feel surprisingly convincing
People often notice the obvious shared feature first: two objects are borrowed, two records have due dates, two systems manage returns. From those similarities, it can be tempting to conclude that every later step is interchangeable. The model may extend that temptation by generating an easy narrative.
One repair is to list the common relations explicitly and then list the target-specific properties that were not established. This separates an explanation’s genuine insight from its unsupported additions.
18. One-to-one correspondence is not always required
Some analogies map one source role to several target components. A school library clerk’s functions might be shared among a tablet-management application, a teacher and an equipment coordinator. The key is to preserve the relevant responsibilities, not insist every source object must correspond to exactly one target object.
Analogical reasoning becomes more complex in such situations because a simple table can hide differences in authority and responsibility. A useful assistant states the mapping it is using rather than presenting it as obvious.
19. The risk of overextending the source
An analogy between a school class and a sports team can help explain roles, practice and collaboration. It becomes misleading if a model assumes a classroom must select pupils through athletic trials, record match scores or assign physical positions on a field.
Those features belong to the original domain. They do not follow from the general idea of organised collaboration. A good analogy names the shared mechanism and avoids treating every decorative detail as a transferable rule.
20. Explain the limit before the learner misapplies it
Suppose a teacher explains an atom as a tiny solar system to introduce a central nucleus and surrounding electrons. This picture can establish a rough spatial image. But it does not accurately describe quantum behaviour as planets following classical orbits.
The problem is not that analogies are forbidden. It is that their limits should be visible before a student uses the picture to predict mechanisms the analogy was not designed to capture.
21. Analogy and literal identity are different
Saying “a vocabulary notebook is like a personal dictionary” does not make the two physically or functionally identical. A personal notebook may contain mistakes, partial definitions and examples chosen by a learner. A published dictionary is compiled and edited through a different process.
An AI tutor can use the analogy to encourage recording new words, while still telling the student to check definitions using reliable references. The analogy supports a learning behaviour, not equivalence of authority.
22. Analogy and generalisation are different
Generalisation identifies patterns across examples and applies them to new cases. Analogical reasoning explicitly maps relations from one domain to another. The two processes overlap but are not interchangeable. A statistical model may generalise from labelled examples without producing a human-readable analogy.
Similarly, machine-learning transfer learning often refers to using representations or model parameters learned for one task to help another. That technical process is related in spirit to reusing knowledge, but it is not identical to verbally comparing books and tablets.
23. A false analogy can hide in a beautiful metaphor
A metaphor may say that learning is “filling a container”. The image can encourage the idea that knowledge accumulates, but it may also suggest that students are passive recipients and that learning consists only of adding more material. Retrieval, understanding, correction and application are not well captured by that picture.
An AI writing tutor should use metaphors as aids, not substitute metaphors for the full mechanism. If a metaphor begins to predict an untrue consequence, the assistant should revise the explanation.
24. What constitutes a strong mapping?
A strong mapping preserves the relationships that explain the target problem. In the loan-register example, the key relation is borrower-to-asset over a defined loan period, with a record of issue, due date and return. Those structural details are more useful than whether the objects look alike.
A proposed mapping also needs relevant constraints. What counts as a confirmed return? Who can update records? Which policy authorises contact? These questions reveal why a structurally similar workflow may still require substantial target-specific adaptation.
25. Why causal mechanisms matter to transfer
Suppose a study method worked in one classroom because it encouraged students to retrieve information from memory. An analogy to another subject might be useful if the same retrieval mechanism is relevant. But copying the identical worksheet format does not guarantee the same effect.
The useful transfer concerns the learning process, not every superficial feature. Testing in the new subject is still necessary because learner age, prior knowledge, task structure and assessment can differ.
26. How an AI can generate an analogy responsibly
First, identify the idea the learner needs to understand. Second, choose a source that the learner knows. Third, map the relevant relationships explicitly. Fourth, name the important differences. Fifth, test whether the analogy leads to any wrong inferences about the target.
A system that merely produces a vivid comparison may create engagement but not accuracy. A system that names the mapping and its limits gives the learner tools to use the analogy correctly and discard it when it stops helping.
27. How to test an analogy with counterexamples
Try a claim from the source, such as “every book in the library has pages”, and ask whether the mapped tablet necessarily has that same property. It does not. That is a counterexample to the assumption that every attribute transfers across the analogy.
Then try “every borrowed asset needs an identifier in our register”. This can apply to both cases because the target’s own records confirm the same administrative relation. Testing multiple types of claim helps distinguish a genuine structural resemblance from an accidental similarity.
28. Why a plausible analogy still needs source evidence
A model might say “Most schools manage tablets just like library books.” That is a claim about actual institutions, not merely a conceptual comparison. It requires representative evidence about schools and their device-management policies. One fictional example does not establish it.
When an AI explains “one possible method is to use a library-style checkout register”, it is offering a design idea. When it says “this school definitely uses the library’s fine policy for tablets”, it is claiming an external fact. The distinction must remain clear.
29. The imaginary baseline error pattern
Suppose a fictional assistant labels twelve of the sixteen record claims correctly but copies BL-7 into the tablet ID, asserts the tablet is physically missing, assumes the library fine must apply to tablets and treats the book return date as the tablet’s return date. Its invented score is 12/16 = 75%.
These are different kinds of over-transfer: identity copying, unsupported physical state, policy copying and event copying. The repair is not “make analogies shorter” but “verify which relation each transferred detail represents and whether the target evidence supports it”.
30. A revised imaginary analysis
A revised assistant maps the shared loan structure, reads the target record for its actual asset identifier, leaves physical return state unresolved and checks the tablet-specific communication policy before acting. Suppose it classifies all sixteen statements according to the reference ledger.
That is 16/16 in this fictional teaching packet, not an observed AI benchmark. A serious test would need novel domains, misleading similarities, multiple valid mappings and independent reviewer agreement.
31. A classroom transfer exercise: plant and student growth
A teacher says “A student grows through practice, much like a plant grows with regular care.” The analogy can support the importance of consistent conditions and attention. But a student is not a plant: education involves choices, agency, social interaction, prior knowledge and feedback in ways the botanical image does not explain.
Ask learners which relationship transfers: sustained support can aid development. Then ask which proposed inference fails: more water always makes a plant grow faster, or more homework always makes a student learn faster. Neither unrestricted statement follows from the analogy.
32. A software transfer exercise
An application might use a queue for support tickets and a queue for homework questions. Both have incoming items and a method for selecting the next item. But a first-in-first-out rule may be appropriate for one context and inadequate for urgent safety-related messages in another.
The transferable concept is orderly task management. The priority policy must be checked for the target task. Copying the queue data structure does not automatically justify copying every priority decision.
33. Questions parents can ask of an AI analogy
Which part of the comparison is useful? What is different about my child’s actual situation? Is the analogy explaining a mechanism or claiming a real measured outcome? Would the suggested learning method still work if the superficial resemblance disappeared?
Those questions make analogies useful for teaching rather than a source of false certainty. A comparison should help a learner think, not ask them to stop examining the real task.
34. A checklist for builders and writers
Name the source and target. List the mapped objects and roles. Preserve the important relations. Identify source-specific properties that cannot be assumed in the target. Check evidence, authority and time. Test a plausible false extension. Report whether the analogy is a teaching illustration, a proposal or a verified factual mapping.
When these steps are followed, the system can benefit from the speed of analogical thought without turning resemblance into proof. It becomes easier to revise a misleading explanation because every claimed correspondence is explicit.
Frequently asked questions
Is a good analogy the same as proof? No. It can explain or suggest a relationship, but a real-world claim still needs supporting evidence.
Can two different objects perform the same role? Yes. A book and tablet can both be assets in a checkout register, though their physical properties differ.
Why can’t I copy an entire rule between similar systems? Permissions, exceptions, costs and other target details may differ.
Can AI invent analogies? Yes, and some may be helpful for learning. Their mappings and limitations should be checked.
Is analogy the same as machine-learning transfer learning? Not exactly. Transfer learning commonly reuses trained representations or parameters, while analogical reasoning maps relationships across cases.
What should happen when the analogy breaks? Identify the failed correspondence, return to target evidence and revise or replace the explanation.
The central lesson
Analogies are powerful because they move understanding from a familiar situation to an unfamiliar one. They become unreliable when the system copies the source’s identity, policy or hidden assumptions instead of the relationship that matters.
A trustworthy AI uses a comparison to generate a testable idea, not a guaranteed fact. The habit is simple: say what maps, say what does not, and check the target before acting. That is how useful similarity becomes careful transfer rather than confident error.
Related lessons and reliable sources
- How Super Intelligence Works — main learning curriculum.
- Parts and Wholes — why belonging, containing and owning are not the same relationship.
- Counterexamples — testing a proposed mapping for exceptions.
- Unknown, None and Zero — what an unconfirmed return means.
- Stanford Encyclopedia of Philosophy: Analogy and Analogical Reasoning — systematic discussion of analogies and their limits.
- Jurafsky and Martin: Speech and Language Processing — language meaning, relations and models.
- NIST AI Risk Management Framework — checks and evidence in applied AI.