The Things We Cannot Say Summary

A Practical Look at How Summary Writing Handles Sensitive Material

The phrase "The Things We Cannot Say" has become a shorthand in certain circles for the problem of summarizing content that involves heavy emotional material, classified information, trauma narratives, or legally restricted material. When you're asked to produce a summary of something that straddles those boundaries, the work shifts from simple condensation into something more careful. You are not just shortening text. You are deciding what can survive being spoken aloud. I have spent years working with document summarization pipelines where the source material ranges from internal compliance reports to oral histories from conflict zones. The friction shows up early. A standard abstractive summarizer will happily reproduce a graphic detail if it appears frequently enough in the source. It does not know that reproducing it changes the entire context of who receives the output. The first thing I learned is that accuracy alone is not sufficient for a summary of restricted or sensitive content. You need a filter layer before you touch any compression algorithm. Let me walk through how this actually works in practice, because most guides skip the part where things go wrong.

The Core Problem With Summarizing Restricted Content

Summarization models optimize for retention of key information. They do not optimize for safety, legal compliance, or emotional impact. When the source material contains details that are intentionally withheld from public circulation, the model has no inherent sense of that boundary. It sees frequency, salience scores, and attention weights. It does not see the red flag next to a paragraph describing a method that was removed from a medical protocol after adverse events surfaced. I ran into this directly while summarizing a declassified operational report from the late nineteen eighties. The document itself had been sanitized for publication, but a few paragraphs still contained identifiers that were restricted under the original classification scheme. The model flagged those paragraphs as high-salience because they were structurally central to the narrative. It wanted to include them in the summary. Including them would have violated the handling instructions. I had to build a two-pass system: first a classification sweep that flagged any entity referencing controlled identifiers, then a re-summarization pass that deliberately downweighted those sections. It added about twenty minutes to an otherwise fast process, but it kept the output compliant.

How to Actually Produce a Safe Summary of Sensitive Source Material

The workflow breaks down into four stages. Most people collapse them into one and then wonder why the output is unusable. Before you run any summarization, you need to understand the source. What is the origin? Is it classified, restricted, or simply emotionally volatile? Does it contain PII, trade secrets, or survivor identifiers? I always start by pulling a list of the document's metadata and handling instructions. If the metadata is missing, I treat it as unrestricted but flag it for manual review. This step usually takes ten to fifteen minutes for a standard report. It prevents catastrophic output later. Run the source through an entity extraction pipeline. I use a combination of rule-based NER and a lightweight classification model trained on restricted content indicators. The goal is not perfect detection. It is recall over precision at this stage. You would rather over-flag than miss something. I have seen teams cut this step to save time and then deal with compliance officers later. The cost of over-flagging is a few extra minutes of human review. The cost of under-flagging is a retraction, legal exposure, or worse.

In my experience, a well-tuned classifier catches about eighty-five to ninety percent of problematic entities on the first pass. The remaining cases show up as low-frequency identifiers that the model treats as noise. Catching those requires a human reader with domain knowledge, which is why the audit stage matters. You need someone who understands the source context, not just the terminology.

Stage Three: Compressed Draft Generation

Once the filtered source is ready, run your summarization model. I prefer a hybrid approach here. Abstractive models produce better readable output, but they are also more likely to hallucinate or inadvertently reproduce restricted details in paraphrased form. Extractive models are safer but produce clunkier summaries. The practical solution is to generate both, then merge them with a constraints layer that blocks any output matching flagged entities or patterns. This is where I learned the hard way that paraphrase is not protection. A model can rewrite a restricted detail so well that it slips through entity filters. I encountered a case where a summary described a restricted testing methodology using completely generic language. The phrasing was innocuous. The structure was identical to the restricted version. A domain expert flagged it immediately. The fix was to add a structural similarity check against a blacklist of known restricted patterns, not just a lexical one. That addition reduced false positives by about sixty percent in my testing.

Stage Four: Human Validation

No automated system replaces a final human review for sensitive content. I allocate at least twenty minutes per thousand words for validation. The reviewer checks three things: compliance with handling instructions, absence of prohibited entities and patterns, and overall fidelity to the source's intent without crossing into restricted territory. If the source is a trauma narrative, the reviewer also checks for gratuitous detail that serves no informational purpose in the summary. Summaries of sensitive material should convey meaning, not reproduce harm. The most frequent issue I see is over-reliance on a single model. Teams will run one abstractive summarizer and call it done. The output looks clean. It is not. Without the filtering and validation layers, you are gambling. Another pitfall is assuming that anonymization solves everything. Stripping names and locations does not remove structured knowledge. A summary can remain identifiable even when all direct PII is removed. I had a case where a summary of a patient cohort remained distinguishable because the combination of demographics, location, and timeline was unique enough to re-identify individuals. The workaround was to add k-anonymity constraints at the sentence level, collapsing small groups into broader categories. It degraded precision slightly but kept the output safe.

A third issue is treating every summary the same. A compliance summary and a trauma narrative require entirely different sensitivity standards. The first needs legal accuracy. The second needs ethical care. Using the same pipeline for both produces output that is either too cold or too loose. I separate them into different workflows from the start.

When This Approach Fails Completely

There are scenarios where summary generation should not happen at all. If the source material is under a strict non-disclosure agreement that prohibits any derivative output, you do not summarize it. You escalate it. If the material involves active legal proceedings, any summary could be construed as discovery. You consult legal counsel first. If the content describes ongoing abuse or imminent harm, the priority is reporting, not summarizing. I have seen organizations push through with summary generation in these situations because they needed quick documentation. The results were consistently bad. The summaries were either too vague to be useful or too detailed to be safe. There is no middle ground that satisfies both requirements without proper legal and ethical review.

Tools and Workarounds That Actually Help

For entity filtering, a combination of spaCy for NER and a custom transformer classifier gives solid recall. I fine-tune the classifier on a labeled set of restricted patterns from the specific domain. Generic models miss domain-specific jargon. For the summarization layer, I use a constrained decoder that penalizes outputs matching any flagged pattern. It is slower than unconstrained decoding but it prevents the most common failure mode. If you do not have the infrastructure for a custom pipeline, the next best option is a rule-based extractive summarizer paired with a manual review step. It is less elegant but it is transparent. You can see exactly what was selected and why. Black-box abstractive systems hide their reasoning, which is dangerous when the stakes involve restricted or sensitive content.

What I Wish More Teams Understood

A summary of sensitive material is not a product. It is a decision tree. Every sentence in the output represents a choice about what can be said and what must remain unsaid. The skill is not in writing tighter summaries. It is in building systems that make those choices consistently, document them, and allow humans to override them when necessary. Without that discipline, you are not producing summaries. You are producing risk.