… a document already says what it needs to say, but this actor needs it at a different length or reading level: a help article for someone new to the product, a clinical paper for a patient, or a post in a second language. Rewriting the original for one reader makes it worse for the rest, so the view adapts the phrasing or length for this actor while leaving the source text intact. A text lens helps complete help, where the same documentation serves actors with very different background knowledge, and supplies the coarser text levels that semantic zoom and focus and context step through when operating on prose.

Problem

A single text often serves actors with different background knowledge, reading fluency, or available time, and the same actor may need a quick summary first and full detail next. One fixed version compromises across all of them, while authoring separate versions for every audience and length multiplies what the author has to write and keep up to date.

Five forces shape how a text can be adapted on demand:

• Fixed authoring vs. variable reading depth: systems like StretchText let actors expand or collapse passages by level of detail, but rarely saw adoption because authors had to write and maintain every level by hand. Generating levels on demand removes that authoring bottleneck.
• Fit vs. fidelity: the further a generated version moves from the source phrasing, the more likely it is to drop qualifiers, soften claims, or introduce errors. In clinical research summaries, generated plain-language versions frequently omit the study population or comparator.
• Explicit selection vs. automatic adaptation: the system can infer some needs from role or account settings, but cannot reliably guess an actor’s familiarity with a specific topic. Over-simplifying has a cost: readers who already know a subject tend to overlook caveats and study limitations when given an overly plain version. Automatic adaptation risks serving the wrong level silently; manual controls require the actor to choose before reading.
• Editorial accountability vs. maintenance: authored variants are reviewed by someone accountable for what they omit, but go stale when the source text changes. Generated variants stay in sync with the source, but have no human editor standing behind them.
• How audiences differ: a lens can only reorganise or rephrase information that is already in the source text. Readers differ in three ways:
  • Background knowledge and fluency: the facts stay the same while vocabulary, sentence complexity, or language changes. This is suited to generation from the source.
  • Task or purpose: executing a step, troubleshooting, or evaluating. The underlying facts stay the same, but priority and omissions shift.
  • Applicable rules or context: permissions, pricing tier, platform, or jurisdiction. These require facts outside the passage; generating them invites fabrication, so they belong in authored conditional content instead.

Solution

Let the actor switch the level of detail or reading framing of a document—by passage or as a whole—generating the view strictly from the source text while keeping the original one step away. Use authored content, not generation, for variations that depend on external rules such as permissions, pricing tier, platform, or region.

Where the source is a text of record, such as a contract, policy, or clinical instruction, present the generated view as a reading aid that links back to the binding text rather than replacing it.

In practice

Adjusting detail by passage

Adapting for the audience

Research on this pattern

• Tailoring helps only when it matches the reader: simpler versions served readers new to a topic best, while familiar readers tended to overlook details such as study limitations in overly plain versions.
• Omissions in a shortened version are invisible within the summary itself: dropped details leave no trace, making misrepresentation undetectable without consulting the original text.
• Generated plain-language summaries of medical trials frequently omitted or overgeneralised the populations and comparators studied.
• Minimally lossy text simplification raised comprehension by 4% overall and 15% on medical text across more than 4,500 participants.

Resources & references

• August, Lo, Smith & Reinecke (2024) Know your audience: the benefits and pitfalls of generating plain language summaries beyond the “general” audience — simpler versions served readers new to a topic; familiar readers skipped details such as study limitations in overly plain versions
• Gu, Arawjo, Li, Kummerfeld & Glassman (2024) An AI-resilient text rendering technique for reading and skimming documents — why omissions in summaries go unnoticed; renders the original text with graded emphasis instead of rewriting it
• Joseph et al. (2024) FactPICO: factuality evaluation for plain language summarization of medical evidence
• Fok, Chang, August, Zhang & Weld (2024) Qlarify: recursively expandable abstracts — StretchText’s adjustable detail with generated rather than authored levels
• Huang, August & Sundaram (2026) TermSight: making service contracts approachable — plain-language snippets that link back to the binding contract text
• Brusilovsky (1996) Methods and techniques of adaptive hypermedia — fragment variants, conditional text, and stretchtext as the authored predecessors of these adaptations
• Google Research (2025) Minimally-lossy text simplification with Gemini — the study behind Simplify
• Wikipedia pauses AI summary experiment — editors’ objections to unverified generated summaries published under the site’s name
• Inclusion Europe European easy-to-read logo — human review by readers with intellectual disabilities as a condition of certification