How does superintelligence work when a sentence says “she”, “they”, “it” or “the student” without repeating a person’s name? The practical AI problem is reference resolution: connecting an expression in language to the person, object or idea it refers to. An assistant can produce elegant sentences yet completely reverse a story if it connects one pronoun to the wrong participant.
Imagine this message: “Ava spoke to Mei after she finished the experiment.” Who finished? The sentence is not automatically clear. It might be Ava or Mei. A careful AI system uses surrounding context and grammar, but it must not invent evidence that the message never supplied. For a harmless story, a stated interpretation may be acceptable; for a real timetable or school notice, guessing could create a serious error.
This lesson follows a fictional science-fair desk through a fully worked reference check, a changed-message exercise and a simple reader’s test. The examples are invented, not recordings of actual students or measurements of a live model. “Superintelligence” is the educational name of this AI series; no claim is made that contemporary language models are artificial superintelligence.
1. Why pronouns are efficient — and risky
People would sound repetitive if they repeated names in every clause: “Ava picked up Ava’s folder and Ava carried Ava’s folder to Ava’s class.” Pronouns and other referring expressions let speakers maintain a story without repeating the same name. Readers infer the connections from grammar, context, topic and shared knowledge.
For an AI application, that convenience becomes a technical requirement. A summary should preserve who did what. A customer-support assistant should not transfer one customer’s request to another. A tutor should not attribute one child’s work to a different pupil. Reading the nouns is not enough; the application must follow the relationships between mentions.
2. What is coreference resolution?
Coreference occurs when two expressions refer to the same real-world entity in the described context. In “Nadia opened the door. She entered,” “Nadia” and “she” refer to the same person. Coreference resolution is the task of finding and linking such expressions, often called mentions, across a document.
This is different from selecting the dictionary meaning of a word. “Bank” can have several senses; “she” can point to different people in the same paragraph. It is also different from checking whether the story actually happened. Resolving a pronoun supports interpretation, not independent verification of an event.
3. A familiar example with an unfamiliar danger
Consider: “The teacher called the parent because she was concerned.” In isolation, “she” could be the teacher or the parent, depending on context. If an AI rewrites the sentence as “The teacher was concerned, so she called the parent”, it has selected one reading without confirmation.
This matters even when no obvious safety issue exists. A school newsletter, complaint, experiment record or historical summary changes meaning when roles are assigned incorrectly. In writing, a small pronoun can carry the weight of an entire claim.
4. The basic mechanism in plain English
A language system first identifies the meaningful expressions in text: proper names, descriptions, pronouns and sometimes omitted but recoverable subjects. It then considers which expressions might refer to the same person, thing or event. Grammar, nearby sentences, number agreement and discourse context are useful clues.
Some systems explicitly predict mention links or clusters. General-purpose language models may produce these connections as part of their text generation. Neither approach guarantees correctness. It is a mistake to describe every modern model as running the same hidden reference algorithm, or to treat a generated explanation as proof of internal causal processing.
5. A science-fair example with four participants
The fictional Willow Science Fair has four participants: Anika, Bao, Celine and Daniel. The organiser asks an AI assistant to summarise project handovers. The assistant may use only the supplied notes. There is no facial recognition, access to personal school records or outside memory. Each example is an invented exercise.
The event organiser’s central rule is straightforward: do not put a statement about one participant into another participant’s profile. When a reference cannot be uniquely resolved, retain the original wording and flag the uncertainty instead of silently assigning an identity.
6. Name the entities before tracking them
Start with stable participant IDs such as P1 Anika, P2 Bao, P3 Celine and P4 Daniel. These are local reference labels for an imaginary exercise, not real identifying information. The label does not tell us gender, authority or responsibility. Its role is to prevent an explanation from drifting between people who have similar descriptions.
Next separate objects such as a poster, notebook, tablet, table and schedule. Objects can be referred to with “it”, “this”, “that” and descriptions. A sentence may contain multiple matching objects. Keeping entity types explicit reduces confusion, but it does not make every pronoun unambiguous.
7. The first clear mention chain
“Anika collected the poster. She folded it before leaving.” The most natural and well-supported links are “she” to Anika and “it” to the poster. No other person or object intervenes in the supplied text. We can write the chain as Anika → she and poster → it.
Now change the sentence: “Anika collected the poster from Celine. She folded it before leaving.” Both Anika and Celine are plausible antecedents of “she”. The poster still plausibly matches “it”, but the person link becomes less certain. One added name changes the reference problem. A trustworthy system should notice that change.
8. Grammatical agreement narrows possibilities, not truth
In English, number agreement can help: “The pupils displayed their projects” connects a plural possessive with a plural subject in a straightforward sentence. Yet “they” can also be singular, and not everyone’s pronouns can be inferred from names, appearance or cultural assumptions. A system should follow explicitly provided references and avoid guessing identities.
Grammar offers constraints, not a universal proof. “The scissors were placed near the book. They were sharp.” Here “they” points naturally to scissors despite the nearby singular book. But in more complex passages, multiple plural noun phrases may compete. Language is not a fixed nearest-word matching exercise.
9. The difference between anaphora and all references
Anaphora often means an expression pointing back to something introduced earlier: “The tablet was cracked. It needed repair.” But a reference can sometimes point forward, as in “When she arrived, Anika saw the display.” In other cases, “it” has no concrete object as its referent, as in “It is raining.”
A simple application that forces every “it” to match a preceding noun will produce spurious links. The system must recognise several kinds of linguistic structure and may need to leave some expressions without a concrete entity. The lesson is not to memorise every technical label; it is to know which kinds of links the text actually supports.
10. A twelve-message fictional reference laboratory
The Willow desk prepares twelve short notes. Eight are written so the target reference is recoverable from the provided wording. Four are purposely under-specified. The desk asks a fictional AI to return a participant or object ID together with the exact evidence for the link. No actual system has been tested here; the case is an explanation of how to design and inspect a test.
The important feature is that the ground truth is set before model output. If a reviewer cannot justify one unique participant from the text, the approved answer is “UNRESOLVED”. A complete-looking summary is not the goal; a supported summary is.
- 01 “Anika displayed a poster. She explained its colours.” — she = Anika; its = poster.
- 02 “Bao fixed a stand. It had a loose joint.” — it = stand.
- 03 “Celine moved the tablet from the table. It had a cracked screen.” — target it = tablet, supported by screen.
- 04 “Daniel wrote a note. He sent it to the organiser.” — he = Daniel; it = note.
- 05 “Anika and Bao compared their results. They found different totals.” — they = Anika and Bao together.
- 06 “Celine gave Daniel a ruler. He measured the board.” — he = Daniel.
- 07 “The poster was too tall for the drawer, so it remained outside.” — it = poster.
- 08 “Bao stored two banners. They were both labelled.” — they = banners.
- 09 “Anika spoke to Celine after she finished.” — she = UNRESOLVED.
- 10 “Bao handed Daniel his notebook.” — his = UNRESOLVED without ownership context.
- 11 “The tablet stood beside the display board, and it was damaged.” — it = UNRESOLVED without more evidence.
- 12 “They requested another rehearsal.” — they = UNRESOLVED without preceding context.
11. Count the supported cases carefully
Eight messages have supported target links; four should remain unresolved. Suppose a fictional baseline assistant correctly resolves six of the eight clear messages but confidently guesses an entity in three of the four ambiguous messages, while marking one unresolved. The resolved-link accuracy is 6/8 = 75%. The inappropriate-guess count is three out of four ambiguous items.
After revising the procedure, suppose the assistant resolves seven of the eight clear messages correctly and marks all four uncertain cases unresolved. Then resolved-link accuracy is 7/8 = 87.5%, and correct abstention is 4/4 = 100%. These are invented calculations, useful only as examples. Do not report “the model is 91.7% reliable” by mixing two distinct outcomes into one universal score.
12. Why “UNRESOLVED” is an achievement
An unanswered reference is not always a model failure. Sometimes the sentence itself fails to identify its participant, and the system’s correct task is to preserve that fact. In “Anika thanked Celine after she won”, neither the syntax nor the names are sufficient to prove who won without more context.
If a downstream application is asked only to make a readable draft, it may rewrite as “After the award announcement, Anika thanked Celine” without inventing a winner. If it must publish an award list, it needs an authoritative record or a clarification from the organiser. The appropriate response depends on what the system will actually do with the statement.
13. Why the closest name is not a universal answer
Take: “Bao placed the tablet in the cupboard because it was unlocked.” The closest object phrase is “cupboard”, and the property “unlocked” supports that reading. Now change “unlocked” to “low on battery”. The same pronoun “it” is likely to point to the tablet even though “cupboard” remains closer.
A surface rule such as “attach it to the nearest noun” would misread one of these. More capable models may exploit semantic compatibility, but such inference remains fallible. It is especially fragile when multiple objects share the same property or when the author uses an unusual structure.
14. The Winograd-style lesson
A classic family of reference challenges changes one informative word while leaving much of the sentence structure intact. “The trophy did not fit in the suitcase because it was too large” strongly suggests the trophy. “The trophy did not fit in the suitcase because it was too small” strongly suggests the suitcase. The words around “it” create the intended reading.
These examples are useful demonstrations, not magic tests of all language understanding. Research on pronoun challenges shows that high performance on one curated challenge set does not necessarily imply equally strong performance on ordinary document reference resolution. A good evaluation therefore includes varied natural cases, not just memorable puzzles.
15. Reference chains can extend over many paragraphs
Imagine a school report that introduces “the Science Club coordinator” on page one and later says “the chair”, “the organiser” or “they”. Some expressions may identify the same person; others may describe an institution or a newly appointed role. An AI system summarising the report must keep dates and titles attached to the correct person.
Long-distance reference becomes harder when documents combine quotations, sidebars, forms and revised sections. More context may help, but a very long context is not a guarantee. The application should preserve the relevant passage, document version and any explicit identifiers so that a reader can verify the chain.
16. A role is not necessarily a named person
Suppose the note says “The club president approved the design”. If a separate record says Celine is the president in the current term, a system may connect the role to Celine when that record is authoritative and relevant to the correct date. But a previous year’s president list cannot establish the identity for the current year.
This combines linguistic coreference with entity linking and source verification. The model may understand what “president” means without knowing who held the role on a specific date. When evidence changes, identity conclusions must change accordingly.
17. Quoted speech changes the viewpoint
“Anika told Celine, ‘I finished my project yesterday.’” The “I” inside the direct quotation refers to the speaker Anika, not the person reading the report. “Yesterday” is relative to the speaking time, not automatically yesterday relative to today.
Now consider “Celine said Anika had finished her project”. The source and perspective have changed, and “her” may require clarification depending on wording. A summariser that flattens quotations into a single narrator can change who made which claim. For high-stakes records, source attribution is as important as pronoun resolution.
18. Non-person entities also need tracking
Reference is not limited to people. In a laboratory note, “the sample”, “the container”, “the measurement” and “the reading” are different things. If a model attributes a temperature measurement to the container rather than the sample, the resulting scientific summary may be inaccurate.
Objects can also be grouped: “Two lamps and one display screen arrived. They were tested.” The plural “they” might refer to all three objects or to the lamps depending on earlier context. Some ambiguity can only be settled through an inventory or explicit list. Correct bookkeeping is not the same as a plausible-sounding narrative.
19. Lists, tables and bullet points require special care
A document might contain a row “Student: Anika | Mentor: Celine | Status: approved” and later refer to “her approval”. Without a clear ownership rule, the model could attribute approval to either person. Preserving table headers and role labels matters because layout carries meaning.
If a page extractor drops the columns or merges adjacent rows, even a capable language model may receive a misleading text stream. The right repair lies in document parsing, data structure and source location, not necessarily in changing the language model. The full system matters.
20. Reference resolution cannot supply missing permission
Suppose the organiser writes “She said we may post it.” Even if the model resolves “she” to the organiser and “it” to a photo, the sentence may not establish valid consent for publication. A claim about permission is different from a verified approval tied to a particular image, audience and time.
AI applications should separate interpretation, authorisation and confirmed action. Coreference can reduce misunderstandings, but it cannot create a consent record, confirm identity or make an external operation permissible. This boundary becomes crucial in agent systems.
21. What a good explanation should show
For each nontrivial reference decision, a reviewer benefits from seeing the relevant source sentence, the selected antecedent, the supporting clue and any alternatives that remain plausible. An explanation such as “It is obvious from context” is not auditable. Neither is a list of names with no passage identifiers.
But generated explanations can themselves be wrong or post hoc. The independent check is to return to the original text, not merely judge whether the model’s prose sounds reasonable. Users should also be able to correct the intended referent without rebuilding the whole document.
22. A simple repair workflow
First, identify every mention relevant to the task. Second, draw links only where the supplied context supports them. Third, mark uncertain links explicitly. Fourth, inspect the changes in meaning if one link were wrong. Fifth, retrieve or ask for authoritative clarification when a consequential decision depends on the link.
Only after these steps should a system rewrite the text, generate a summary or prepare a record update. If a summary claims the organiser approved an action, verify that the source actually attributes approval to the organiser and that the approval itself is current. A fluent rewrite is not a substitute for this check.
23. How to ask an AI for safer reference handling
A useful user prompt is: “Identify every person mentioned and show who each pronoun refers to. Quote the supporting sentence. Where more than one person fits, say ambiguous instead of inventing a name. Do not add details beyond the source.” That is a request for a specific inspectable output, not a promise that the model will obey perfectly.
For long documents, split the task by section while keeping a stable entity map. Compare references across section boundaries before combining the summary. This is particularly useful for meeting notes, educational reports, interview transcripts and historical materials with changing roles.
24. The fairness problem: names do not determine pronouns
Names, jobs and cultural stereotypes are poor substitutes for explicit information. “The engineer spoke to the nurse because she needed a report” does not establish which participant “she” identifies by occupation. Similarly, singular “they” cannot automatically be treated as an error or assumed to refer to a group.
An inclusive and technically robust approach uses supplied pronouns and actual grammatical evidence, permits unresolved interpretations and avoids attributing traits or identities from stereotypes. A system that gets stereotyped examples right for the wrong reason can still fail on ordinary documents.
25. Transfer problem: a changed school notice
Try: “Celine emailed Anika after her demonstration was postponed.” Whose demonstration was postponed? With no further evidence, the safe classification is unresolved. Now add “Celine’s demonstration had been scheduled for Monday; Anika’s was scheduled for Friday.” If the postponement is independently shown to concern Monday, Celine becomes the supported referent.
The key is that the later evidence must genuinely identify the postponed event. A schedule alone may not prove postponement unless a source says which event changed. This is a useful lesson in the boundary between resolving language and verifying factual records.
26. An independent practice set
Read these three fresh sentences. A: “The telescope was beside the tripod. It fell during transport.” B: “Anika met Daniel while he was carrying the model.” C: “The two researchers presented their results. They took questions.” Identify which pronouns are clear, which are not and whether more context would change the answer.
Answers: A is uncertain between telescope and tripod without further facts. B clearly points “he” toward Daniel in ordinary grammatical reading, although a richer context could challenge assumptions about pronouns; respect explicit references if supplied. C “they” naturally refers to the two researchers. Explain the evidence rather than merely naming an antecedent.
27. Design a stronger evaluation
Use a balanced packet of clear and ambiguous cases, single and multiple participants, singular they, long-distance references, quoted speech, non-person entities, role changes and changed-document versions. Build the intended labels without consulting the model’s output. Include different writing styles, rather than only puzzle sentences.
Evaluate exact links where ground truth is available, the treatment of unresolved cases, the stability of results when irrelevant names are inserted, and the real effect on summaries or actions. If a product claims it resolved identities, compare against authoritative identifiers when permitted. A grammar test alone cannot establish cross-record identity.
28. How teachers can use this in English lessons
Ask pupils to rewrite an ambiguous sentence in two unambiguous ways. “The teacher thanked the student because she had helped” can become “The teacher, who had helped, thanked the student” or “The teacher thanked the student for the student’s help”. The exercise shows how a reader’s inference can become a writer’s responsibility.
Then invite learners to test an AI rewrite and compare it with the source. The best discussion is not which sentence sounds smoother. It is whether the rewrite has preserved every original fact and avoided committing to an interpretation that the original never settled.
29. What this mechanism does not prove
Accurate pronoun matching does not prove general intelligence, awareness or access to the actual events being described. A model may perform well on a benchmark while failing when the document is badly extracted, when the relevant person was mentioned far earlier, or when two descriptions refer to different people with similar roles.
Nor can model confidence replace a signed event record, current database lookup or explicit human clarification. These remain different forms of evidence. The purpose of the mechanism is to keep language interpretation more reliable within the wider system.
Frequently asked questions
Is the nearest preceding name always the answer? No. Grammar, meaning and discourse can favour another antecedent; sometimes no unique answer exists.
Can the AI infer pronouns from a first name? It should not assume personal pronouns from names or occupational stereotypes. Use what the person or source explicitly provides.
Why does a summary sometimes switch who acted? The system may have misresolved a pronoun, lost a source quotation or merged records from different contexts. Compare the original passages.
Does coreference work in other languages? Yes, it is a cross-linguistic problem, but languages express reference differently, and tools vary in coverage and accuracy.
Can an AI ask the author? If the workflow allows a follow-up, asking the author can be more reliable than guessing. Otherwise preserve the uncertainty.
The useful habit to keep
Every time a sentence contains “he”, “she”, “they”, “it”, “this”, “that”, a role or a repeated description, ask which earlier expression it connects to, what evidence supports the link and what would change if the link were wrong. If you cannot answer, do not let fluent prose conceal the uncertainty.
That small habit turns a subtle language mechanism into a practical safeguard for reading, teaching, summarising and tool-using AI. The model may suggest a connection; the text and trustworthy records must support it.
Connected learning and source notes
- How Super Intelligence Works — 100-part curriculum — the broader architecture, context and verification routes.
- Word Meaning in Context — companion lesson on selecting a word’s sense rather than resolving a pronoun.
- Jurafsky and Martin, Speech and Language Processing, August 2026 draft — reference, conversation and language modelling.
- Porada and Cheung (2024), On Winograd Schemas as a Test of Pronominal Coreference Resolution — caution about interpreting challenge-set scores.
- NIST, Generative AI Profile (2024) — transparency and evidence-based evaluation.