Research Literature Synthesis — by an AI council
Synthesise a body of research into what is known and contested. Four rival AIs debate it and hand back: A synthesis of the evidence with consensus, dispute and the gaps.
A synthesis of the evidence with consensus, dispute and the gaps.
Signed off by an independent Final QA audit — with a “verify before you rely on this” list, never a black box.
Why this team
Literature syntheses fail by flattening — treating twelve studies as twelve equal votes — so the Statistician weighs the evidence by its methods, sample sizes and power rather than its conclusions, while the Deep Researcher maps where the field actually stands and what the citation trails reveal. The Data Skeptic hunts the claims that outrun their data and the conflicts nobody declared, the Steelman ensures the minority position gets its best argument rather than a caricature, and the First-Principles Thinker with the Synthesiser turn the mess into a structure a decision can rest on: known, contested, unknown.
Each seat runs on a different frontier model — GPT, Claude, Gemini, Grok — so it’s a genuine cross-model review, not one AI in costumes. How a run works →
What the debate sounds like
An illustrative excerpt — A product team is synthesising the research on four-day work weeks before piloting one, working from a stack of trial reports, company case studies and media summaries.
Sort the stack by evidence class before reading a single conclusion: a handful of coordinated trials with defined measures, a larger pile of self-selected company case studies, and media summaries of both. The trials support cautious claims about wellbeing and retention. The case studies cannot support causal claims at all — companies that volunteer for pilots and publish results are precisely the companies where it was already likely to work. That selection effect is the field’s biggest confound and most write-ups never mention it.
The headline claim — “productivity maintained or improved” — deserves its own audit: in most of the underlying studies, productivity is self-reported or proxied by revenue over short windows, and the pilots run inside a novelty period with public commitment effects. That does not make the claim false; it makes it weaker than its citation count. The synthesis should grade it “suggestive, measurement-limited”, not “established”.
The sceptical minority in the literature deserves its strongest form, not a dismissal: the compressed-hours critique — that some implementations intensify work rather than reduce it, with costs that surface after the study window closes — is raised by serious researchers, not just resisters. And the sector question is real: the evidence base over-represents knowledge work. A synthesis this team can trust must state where their situation falls outside the evidence.
Questions
Do I need to provide the papers, or does the council know the literature?
Provide what you have — abstracts, key findings, links, even a rough list — and state the question the synthesis must serve. The council reasons hardest about the structure and quality of evidence you provide; where it draws on general knowledge, the deliverable marks those claims for verification rather than presenting them as settled.
What does the confidence grading in the output look like?
Claims are sorted into well-supported, contested and unknown, with the reason for each placement — sample sizes, design limits, selection effects, replication status. The most useful line for most decisions is the honest “unknown” list: knowing what the literature cannot tell you prevents borrowing certainty that is not there.
Can it handle a field where the studies genuinely disagree?
Disagreement is where the format earns its keep: the Steelman argues each side’s best case, the Statistician examines whether the disagreement is methodological (different measures, populations, windows) or substantive, and the synthesis says which. Many famous “contradictions” dissolve into definition differences — and some do not, and you need to know which kind you have.
Your material is used only to run your review — never to train public models. Encrypted in transit and at rest. Security & privacy →
Want full control — pick your own minds, set the depth? Open the full council →