Enhanced Interaction

A doctoral student reads a systems paper that cites, in passing, “the well-known impossibility result”: no proof, no reference, just the assertion. She opens a new tab and searches the phrase. She skims three PDFs, finds the actual theorem in the second, checks its assumptions against the citing paper’s setting, and discovers they do not hold here. Forty minutes gone. The paper is still open where she left it, her mental stack of its argument half-evaporated. She will spend ten more minutes working out where she was.

Nothing about those forty minutes was intellectually necessary. Every step (spot the missing reference, retrieve candidates, locate the theorem, compare assumptions) was support work that a reading environment could have done alongside her. She left the paper not because the paper demanded it but because the medium was inert. It displayed the source and abandoned her at exactly the moments she needed the surrounding world.

The previous chapters built the pieces that make something better possible. There is a source. There is a semantic record, a structured and source-linked interpretation of what the source says. And there is a primary rendering the reader is actually looking at. This chapter asks what comes next:

SOURCE
  ↓
SEMANTIC RECORD
  ↓
PRIMARY RENDERING
  ↓
OPTIONAL SIDECAR

Once all of that exists, what additional material should be allowed to appear beside the source while the reader is using it? That is a narrower question than “how can AI help with reading?”, and the narrowness is the point. The sidecar is not a general assistant. It is a place for material that sits next to the source without replacing it, and the interesting questions are about what may go there, how it is labelled, and when it should stay empty.

Three kinds of sidecar content

The material that might appear beside a source comes in three kinds, which differ in where they come from and therefore in what they can be trusted for.

Derived. Generated from the source and its record. A map of the paper’s claims and how they connect. A list of the assumptions the impossibility claim depends on, taken from the qualifiers the record already holds. A reminder of where the same term was defined earlier in the paper. Nothing from outside has entered; the system has re-presented what the source contains.

Retrieved. Brought in from other material. The paper that proves the impossibility result. A dataset the authors used. A later paper that disagrees. The system has gone out and fetched something that was not in the source.

Interactive. Generated in response to what the reader is doing or asking right now. She selects the theorem’s assumptions and asks, “do these hold in the setting of this paper?” The answer is produced for her question, from the source, from whatever was retrieved, and from the model’s own general knowledge.

Each kind carries a different obligation, and the three should never be allowed to blur:

derived      ≠  externally supported
retrieved    ≠  verified
interactive  ≠  authoritative

Derived content can be wrong about its source, but it asserts nothing the source did not say. It must be checkable against the source, and nothing more. Retrieved content has arrived from elsewhere. That it was found says only that something matched; whether it is reliable, relevant, applicable or even accurately retrieved is a further question. Interactive content is the system’s own answer to a question, produced on the spot, and it can be fluent, confident and wrong. None of the three is the source speaking.

Where it came from, always

This leads to the chapter’s central rule, which holds whatever else changes:

The reader should always be able to tell what came from the source, what was derived from it, and what came from elsewhere.

In practice that means keeping four layers visibly distinct:

SOURCE                  what the author wrote
SEMANTIC DERIVATION     what the system re-presented from it
EXTERNAL EVIDENCE       what was brought in from other material
MODEL GENERATION        what the model produced in response

The sidecar must never silently merge into the source. It is easy to see how it could. A retrieved summary styled like the paper’s own text, so a claim from another author reads as the author’s. An answer to the reader’s question that then gets absorbed into her notes as “what the paper says”. A synthesised explanation more fluent and more confident than the passage it explains, so that it becomes the thing she actually reads. In each case the boundary has dissolved, and a source with a sidecar has turned into a source plus a competing source, with nothing to say which is which.

That last case deserves emphasis, because it is where assistance quietly turns into authority. If the sidecar answers the reader’s question more clearly than the source does, readers will tend to trust the sidecar. If it states conclusions the source did not draw, those conclusions will get attributed to the source. A sidecar that is better written than the thing it accompanies is not harmless. It has to be marked as what it is, and the marks have to be hard to miss.

What problem it solves, and what it costs

The sidecar solves a specific problem: a reader leaving the source, repeatedly, to do support work that breaks their train of thought. Done well, it reduces effort in three ways. It saves manual searching. It keeps the reader’s place in the argument. And it surfaces connections and evidence the reader might not have found at all.

It also costs something, and the cost is attention. Every item beside the source is something else to notice, weigh and decide to ignore. A sidecar that surfaces ten items to save one search has made reading worse. The cost shows up as extra switching between the source and the margin, as lost place when the reader returns, and as time spent on items that turn out not to matter. Any judgement of a sidecar has to count both columns. Items surfaced and used belong in one. Distraction, irrelevant items and time lost belong in the other, and the two should not be netted into a single score, because a useful item and a distracting one are not the same kind of thing.

Preference can stay as context. People often say they like proactive help, annotations and explanations. That is worth knowing and it is weak support for an architecture. The case for a sidecar rests on whether it reduces the effort of a real task without eroding the reader’s grasp of the source, not on whether readers enjoy it.

Three disciplines follow from the cost.

Available, not intrusive. Everything the system finds waits beside the source, inspectable at the reader’s initiative. It does not seize the floor. Identifying useful material while someone reads is one problem. Deciding when something deserves to break into attention is a different problem, and the second belongs to a later chapter. A sidecar that interrupts is a notification system dressed as a sidecar.

Anchored in the source, not the person. What the sidecar proposes is derived from the source in front of the reader and from what the reader is doing with it now. It does not build a lasting picture of who the reader is or guess a goal beyond the current activity. The sidecar responds to the task, and personal conditioning is a separate step that comes with its own controls.

Nothing is a legitimate output. There are passages where the source suffices, where nothing retrieved is good enough, or where the system has no reason to think the reader needs anything. An empty sidecar is then correct. A system that always has something to add has stopped asking whether it should.

What may join the communication

Underneath these rules is a question about authority. When material from outside enters the interface beside a source, it joins the communication, and the reader will treat what is in front of them as part of what is being said. The rules above are an early answer to what may join and on what terms.

a candidate arrives
   ↓
is it derived, retrieved or generated?
   ↓
does it carry its origin and its status?
   ↓
is it labelled as a candidate, not a finding?
   ↓
may the reader dismiss it, check it and trace it?

Retrieved items arrive as candidates, with their provenance, until some further step promotes them. Interactive answers carry the knowledge they drew on and its limits. Uncertainty travels with the item: “a possible match, not confirmed” is a legitimate label, and a clean-looking item with no status is not.

The reader should also be able to see what the sidecar did not do. Manual searches she still had to make, items she opened and items she ignored, dismissals and corrections: all of these are evidence about whether the sidecar is helping, and a system that hides its misses cannot be improved.

A boundary: explaining is not checking

Some sidecar content will be about whether a claim is true. A sidecar may surface verification or external evidence next to a claim, and that can be very useful. But checking truth is a different operation from explaining or augmenting a source. It needs checkable claims, evidence of adequate quality, and a way of reporting a result that does not collapse into a verdict. The sidecar should show evidence status with its sources visible, such as supported, contradicted, mixed or insufficient, and leave judgement to the reader. It should not infer motive or deception from a contradiction. Beyond that boundary the question is a different one, and this chapter does not take it on.

Where this leaves us

A source, a semantic record and a primary rendering now make a place for optional material beside the reader. Three kinds of content can occupy it: derived, retrieved and interactive. Each has its own obligation, and none is the source speaking. The reader must always be able to tell the source from the derivation, the external evidence and the model’s own generation, and the sidecar must stay available, anchored in the source and willing to say nothing.

Look at how the picture has changed over the last few chapters. The representation was once fixed. It became dynamic. It gained a structured interpretation underneath it. Now additional material is entering the interface alongside it, from outside the source. That is a real shift in what a reader is looking at, and it raises a question that none of the earlier machinery answers. We need rules for what may join the communication: which candidates are worth showing, what relationship they bear to what the reader is reading, and why that relationship matters.

References

• Guo, Z., Schlichtkrull, M. & Vlachos, A. (2022). A Survey on Automated Fact-Checking. Transactions of the Association for Computational Linguistics, 10, 178–206. DOI 10.1162/tacl_a_00454. (Used only for the boundary between explaining a claim and checking it.)