How we grade autism evidence
Every intervention on AutismInsights carries an evidence grade. Every research summary is labelled. Every framework choice is documented here so you can decide for yourself whether to trust what we publish.
The four evidence grades
We use a four-grade scale anchored to two public frameworks: the GRADE working group certainty-of-evidence approach used by Cochrane and the WHO, and the NHMRC evidence hierarchy used in Australian clinical guidelines.
NHMRC Level I–II · GRADE High
Multiple high-quality systematic reviews, meta-analyses, or large randomised controlled trials with consistent findings and low risk of bias.
Examples: ABA / EIBI for early childhood support, speech therapy for language outcomes
NHMRC Level III · GRADE Moderate
Cohort studies, case-control studies, or single RCTs with moderate sample sizes and reasonable methodological rigour. Some inconsistency or risk of bias may be present.
Examples: Occupational therapy for sensory processing, music therapy for engagement
NHMRC Level IV · GRADE Low
Observational studies, case series, or small uncontrolled trials. Findings are suggestive but not yet conclusive. Treat as hypothesis-generating.
Examples: Some dietary interventions, specific anxiety-management programmes
NHMRC Level V · GRADE Very Low
Case reports, expert opinion, or pre-clinical research. Promising signal but insufficient evidence to recommend clinically. Requires further study.
Examples: VR-based therapy for social cognition, novel pharmacological approaches
The grade reflects the strength of the research base for an intervention, not its appropriateness for any individual. A Strong grade does not mean an intervention is right for your child — only that the research evidence is substantial and consistent. A Limited or Emerging grade does not mean an intervention is harmful — only that the evidence is currently insufficient to recommend it broadly.
Editorial process, end to end
Every study on AutismInsights moves through the same five steps. Steps 1–3 are mechanical and reproducible. Steps 4–5 use a large language model under explicit constraints, with full disclosure on every page that displays AI-generated text.
1. Source from PubMed
We ingest peer-reviewed autism research from PubMed daily, covering all studies from 2020 onward (with selective historical coverage of foundational work).
2. Classify by study type
Each study is tagged by design (meta-analysis, RCT, cohort, observational, case report, etc.) using PubMed publication-type metadata.
3. Map to evidence grade
Study design maps to a Strong / Moderate / Limited / Emerging grade using the table above. The mapping is mechanical and identical across all studies — no human discretion at this step.
4. Generate plain-language summaries
Claude (an Anthropic large language model) produces a research summary, parent-facing summary, key findings, limitations, and clinical implications strictly from the published abstract. The model is instructed never to invent findings or speculate beyond the abstract content.
5. Attribute and disclose
Every AI-generated section is labelled "Summary by AutismInsights from published abstract." We never present AI output as if it were the original paper. The DOI / PMID and "View Original Paper" link are always one click away.
Where AI is used, and where it is not
AI is used to generate plain-language summaries of published abstracts: research summary, parent-facing summary, key findings, limitations, and clinical implications on each study page. The model used is Claude (Anthropic), prompted to summarise strictly from the abstract and never invent findings.
AI is not used to assign evidence grades, classify study types, or rank interventions. Those steps are mechanical (study design → grade, via the table above) and identical across all studies.
Every AI-generated section on a study page carries an attribution line reading "Summary by AutismInsights from published abstract. This is not a substitute for reading the original paper." The DOI and "View Original Paper" link are always one click away — we never want you to take our summary on faith.
Editorial commitments
The principles we hold ourselves to. If we ever break them, please tell us at /contact.
We are not a substitute for clinicians
AutismInsights is research intelligence and educational content. It does not provide medical advice, diagnosis, or treatment. Decisions about your or your child's care belong to you and your treating professionals.
Identity-first language
We use "autistic person" rather than "person with autism" by default, following the preference of most autistic adults in published surveys (Bottema-Beutel et al., 2021). We do not use functioning labels.
No paid recommendations
No intervention, provider, or product receives a higher grade because of payment, sponsorship, or affiliate relationship. Provider listings are free; the paid bundles do not influence research grading.
Versioned and dated
This methodology is v1.0, last reviewed 2026-04-27. Changes are logged at the bottom of this page. We re-review the full methodology at least annually.
Update cadence
- Daily: New PubMed studies are ingested overnight (Sydney time).
- Continuous: AI summaries generate within 24 hours of ingestion.
- Quarterly: Intervention evidence grades are re-reviewed against new systematic reviews.
- Annually: The full methodology is re-reviewed; changes are logged below.
Changelog
2026-04-27 · Initial publication
Four-grade evidence scale defined; mapping to GRADE and NHMRC frameworks declared; AI summary disclosure formalised; editorial commitments published; update cadence established.
Questions about how we grade evidence?
Methodology questions, suggestions for improvement, or specific concerns about an intervention grade are welcome. We treat methodology as a living document.