How does Forewords generate content warnings?

Forewords checks each of 11 sensitive topics with several independent AI assessments. It only publishes a warning when two independent assessments agree that content is present, and marks a topic Not found only when every assessment agrees it is absent. Other cases are shown as Unresolved. This means we couldn't tell for sure either way.

I leave a topic Unresolved rather than turn it into a confident “no.” I would rather show uncertainty than risk telling you a topic is absent when I cannot support that. These warnings are a guide, not a guarantee; they give you useful context before you start reading.

No. I tried this for a little while, but frankly it just didn't work. Based on the number of people who selected a hard no for one or more topics, combined with the average number of content warnings per book being ~2.5, recommendation quality dropped and speed to recommendation dropped. It was considerably more difficult to find candidates that met the hard no criteria, chiefly because it is pretty difficult to be sure a book contains no content warnings at all for a given topic.

So after the test had run for a few weeks, I removed the hard no filter on recommendations. They remain information on the book detail screen for you to reference and help make a decision.

Every work is checked against the same fixed set of topics:

• Animal death: death, killing, euthanasia or fatal harm to non-human animals.
• Graphic violence: detailed, vivid or materially intense physical violence or injury.
• Sexual violence: rape, sexual assault, coercion, forced sexual activity or sexual abuse of a minor.
• Child harm: abuse, neglect, exploitation, abduction, serious injury or death involving someone under eighteen.
• Suicide or self-harm: suicide, attempts, ideation, planning or deliberate self-injury.
• Terminal illness: a terminal or life-limiting illness, dying process, or terminal diagnosis or care.
• Addiction: substance or behavioural addiction, dependency, relapse, withdrawal or sustained harmful compulsion.
• Abusive relationships: sustained coercive control or emotional, physical, sexual, financial or psychological abuse within an intimate, family or caregiving relationship.
• Pregnancy loss: miscarriage, stillbirth, ectopic loss or termination for a fetal or non-viable pregnancy.
• War or genocide: concrete or sustained exposure to warfare, military combat, atrocities, ethnic cleansing or genocide.
• Explicit sex: on-page or textually explicit sexual activity, or comparable sustained concrete nonfiction discussion.

I assess the work, rather than one particular edition, so different editions of the same book share the same result. I give each assessment the title and author, alongside catalogue information such as the description, themes, genre, setting and audience as far as it is available.

I ask about every topic, not just the ones an assessment happens to mention. Tap See all 11 topics on a book to see its complete assessment, rather than just the warnings.

The important change is what happens when the assessments do not line up. I do not average them or make a best guess. I can show a supported warning, mark a unanimously clear topic Not found, and mark the rest Unresolved. In other words, I do not mistake uncertainty for absence.

I tested this rule against 200 books across fiction and non-fiction: 2,200 separate topic decisions. I researched the reference answers independently of the system and applied the same rule to every book.

I also ran a separate check on 50 novels already in the catalogue that were not part of that benchmark. I prepared their reference answers before opening either the old or new system results. That fresh audit matters because it tests whether the approach holds up beyond the books used to develop it.

The headline result is that a published warning was supported by the research 90.9% of the time in the 200-book benchmark. For Moderate and Strong content, 99.3% of true cases were either warned about or left Unresolved rather than incorrectly marked Not found. All 362 Strong cases were protected.

The separate 50-novel audit broadly replicated that result: 94.3% of published warnings were supported, 96.7% of true warnings were warned about or left Unresolved, and no Strong warning was incorrectly marked Not found. Three topics were incorrectly marked Not found in that smaller audit: two Minor and one Moderate.

I do not think 100% is a meaningful target. A content warning needs a definition: does a brief historical reference count, does an off-page event count, and how severe is each layer? I set those parameters before testing in conjunction with other readers, research reference answers against them, and hold the assessment process to the same standard. Readers can still reasonably draw a different boundary.

A single recall number therefore does not tell the whole story. I deliberately leave some topics Unresolved, especially for newer or more obscure books, instead of turning limited or conflicting information into a false all-clear.

Use the labels like this:

• About 9 times out of 10, if a warning appears, it will occur in the book.
• Only about 1 time out of 100 do I fail to show Moderate or Strong content as a warning or Unresolved, and mark it Not found instead.
• About 97 times out of 100, covered content will be shown as a warning or Unresolved rather than incorrectly marked Not found.
• If a topic is Not found, every assessment agreed it was absent, but check a trusted human-reviewed source if it is a deal-breaker.
• Unresolved means I could not tell for sure, so do not treat it as absent.
• “All 11 topics checked · None found” means every covered topic is Not found, while an unchecked book is not necessarily warning-free.