We begin with the strongest available evidence, not with a preferred conclusion.
Our job is not to confirm what people already believe, tell readers what they want to hear, or shape a conclusion to satisfy a political side or point of view. Our job is to understand what the evidence actually supports, how strong that evidence is, where uncertainty remains, and what could change the conclusion.
AI allows us to investigate far more information than a human-only research process could reasonably handle in the time available. But AI is a tool, not an authority. The evidence comes from the underlying sources, and human researchers remain responsible for what we publish.
Human researchers choose the questions. AI searches, analyzes, compares, synthesizes, and creates the first draft. Human researchers verify the evidence, challenge the conclusions, and shape the final story.
1. We define the question before looking for the answer
A vague question can produce a vague or misleading answer. We begin by defining exactly what is being asked and what evidence would actually be needed to answer it.
That means separating facts from opinions, association from causation, outcomes from intent, confirmed information from inference, and evidence from speculation. We also pay attention to details that can completely change an answer, including the population being discussed, the relevant time period, the comparison being made, and what a particular number or claim actually measures.
A good answer starts with a well-defined question.
2. We search broadly, then weigh the evidence
AI allows us to search across far more sources, records, and data than would normally be practical through manual research alone. Depending on the question, that may include academic research, government data, court records, regulatory filings, company filings, public datasets, historical records, original documents, and high-quality journalism.
But more sources do not automatically mean better evidence. The goal is not to build the longest possible list of citations. It is to find the evidence most capable of answering the question and determine how much weight that evidence deserves.
An original document is different from someone else’s interpretation of it. A carefully designed study is different from an anecdote. A result supported by multiple independent sources is different from one surprising finding that has never been confirmed.
We do not count citations. We weigh evidence.
3. What counts as strong evidence?
The strongest evidence depends on the question. Different claims require different kinds of evidence, so we do not use one rigid hierarchy for everything. Instead, we ask which type of evidence is best suited to answering the specific question, then evaluate the quality of that evidence.
Does a treatment work?
We generally give the greatest weight to high-quality systematic reviews and meta-analyses of randomized controlled trials, followed by large, well-designed randomized trials.
We look for randomization, appropriate control groups, meaningful outcomes, adequate sample sizes, sufficient follow-up, low risk of bias, and whether the findings have been replicated.
Does something cause harm?
Randomized trials are often unavailable or unethical for studying harms, so strong evidence may come from large cohort studies, case-control studies, surveillance systems, and multiple independent methods pointing toward the same conclusion.
We look at whether the exposure came before the harm, whether risk changes with the level of exposure, whether appropriate comparison groups were used, whether confounding factors could explain the association, and whether the pattern appears across different populations and datasets.
Did a policy cause an outcome?
Policy effects are often difficult to isolate because many things change at the same time. We look for randomized evaluations when they exist, natural experiments, strong quasi-experimental methods, and high-quality observational studies.
We pay particular attention to what was happening before the policy, whether there is a credible comparison group, what else changed at the same time, whether alternative explanations fit the evidence, and whether effects persist over time.
Is a factual claim true?
For straightforward factual questions, we generally prioritize original records and direct evidence. That may include government data, primary documents, filings, transcripts, recordings, official records, or independently corroborated reporting.
We look at where the information originated, whether it is authentic, whether important context is missing, and whether independent evidence confirms it.
What happened legally?
We prioritize the law and the official record itself: statutes, court opinions, filings, transcripts, regulatory documents, and other primary legal materials.
We distinguish allegations from charges, charges from convictions, proposed rules from enacted rules, preliminary decisions from final judgments, and what commentators say a ruling means from what the ruling actually says.
What do people believe?
We generally rely on high-quality representative surveys and polling, preferably with findings supported across multiple independent surveys.
We examine sampling methods, weighting, sample size, question wording, timing, nonresponse, the population actually surveyed, and whether the result is part of a broader pattern or depends heavily on one poll.
What happened historically?
Historical questions often require combining contemporaneous primary sources with rigorous historical scholarship.
We consider the provenance of documents, the context in which they were produced, the incentives and possible biases of the people who created them, whether independent sources corroborate them, and how later scholarship has interpreted the broader record.
Did someone intend something?
Intent requires a higher evidentiary bar than simply showing that an outcome occurred.
We look for direct communications, internal records, documented plans, repeated conduct, responses to warnings or contrary evidence, and behavior over time. We distinguish what someone said they wanted from what the broader evidence demonstrates, and we consider credible alternative motives.
Is a scientific finding reliable?
We look beyond whether a study exists or was peer reviewed. Stronger confidence usually comes from well-designed studies that have been replicated, systematic reviews and meta-analyses of good underlying research, and findings that converge across independent methods.
We examine study design, controls, sample size, measurement quality, preregistration where relevant, replication, publication bias, conflicts of interest, effect size, and whether the result matters in the real world.
Is a viral claim, screenshot, chart, or news story accurate?
We try to trace the claim back to the original material whenever possible.
We look for the full video, document, dataset, quote, image, or source from which the claim originated. We check dates, location, editing or cropping, missing context, source chains, independent corroboration, and whether the evidence actually proves what the post or story claims it proves.
Evidence quality matters at every level
Being higher on a hierarchy does not automatically make evidence good. A poorly conducted meta-analysis can be less informative than an excellent individual study. A peer-reviewed paper can have serious methodological flaws. An official source can make mistakes. Ten news stories can all trace back to the same incorrect original report.
That is why we evaluate both the type of evidence and the quality of the evidence itself.
Across different kinds of questions, we consider factors such as study design, controls, sample size, measurement quality, comparison groups, confounding, replication, source independence, transparency, conflicts of interest, appropriate baselines, effect size, population match, timeframe, and whether the evidence actually answers the question being asked.
The question determines the evidence. The quality determines how much confidence it deserves.
4. We try to prove ourselves wrong
We do not ask AI or our editors to build a case for a conclusion we have already chosen. We actively challenge the emerging answer.
We ask whether there is strong evidence against it, whether credible alternative explanations fit the facts, whether an important source or dataset is missing, and whether the evidence actually distinguishes between competing explanations. We also ask what evidence would cause us to change our minds.
Evidence that merely fits a conclusion is not necessarily evidence that favors it. We want to know whether the full pattern of evidence supports one explanation better than the realistic alternatives.
We also do not manufacture balance. We apply the same evidentiary standards to competing claims, but we do not weaken a conclusion simply because someone would prefer a different one.
The goal is not to prove ourselves right. It is to make it difficult for a weak conclusion to survive.
5. We make uncertainty visible
Not every question has a clean answer. Sometimes the evidence is strong and points clearly in one direction. Sometimes credible evidence is mixed. Sometimes the research is early or incomplete. Sometimes there simply is not enough good evidence to know.
We say which situation we are dealing with.
When credible sources disagree, we try to explain why. They may be using different data, studying different populations, measuring different things, working with different definitions, making different assumptions, or interpreting the same facts differently.
We also distinguish between what has been established, what is strongly supported, what is likely, what is plausible, what is preliminary, and what has not been demonstrated.
Uncertainty is not something to hide. It is part of the evidence.
We do not manufacture certainty where the evidence does not support it.
6. We use AI for scale, not authority
AI is central to how Evidence First works because it allows us to operate at a scale that would be impossible for a human research team to match in the time required to publish.
It helps us search across large bodies of information, identify potentially relevant sources, compare competing evidence, analyze large datasets, connect findings across sources, identify contradictions and gaps, test alternative explanations, and synthesize enormous amounts of information.
AI also creates the first draft. That matters because by that point we are working with a huge amount of material, including large numbers of sources, potentially millions of individual observations in datasets, competing findings, methodological differences, qualifications, and layers of context.
Turning all of that into something a general audience can understand is itself a major synthesis task. AI is instructed to explain unfamiliar concepts in plain language, provide the context people need, use examples and analogies when they make difficult ideas easier to understand, preserve important uncertainty, represent disagreement accurately, and keep the story grounded in the strongest available evidence.
But AI can also make mistakes. It can misunderstand a source, overlook important context, give too much weight to weak evidence, miss something important, invent information, or sound more certain than the evidence justifies.
Its output is never treated as evidence by itself.
AI helps us see more, compare more, and synthesize more. It does not get the final word.
7. Human researchers verify the evidence and take responsibility
Before publication, our editors go back to the underlying evidence. We verify important factual claims, check citations against the sources they are supposed to support, review figures and quotations, challenge the AI’s synthesis, look for evidence that may have been missed, test alternative explanations, and determine whether the conclusion actually follows from the evidence.
Whenever practical, important claims are traced back to original sources rather than relying only on summaries or secondary reporting. That is especially important for scientific findings, statistics, legal records, government data, financial figures, quotations, and claims that materially affect the conclusion.
Human researchers and editors also shape the final story. Our goal is to explain complicated ideas in a way that is clear, engaging, and understandable to a general audience without stripping away the uncertainty, context, or nuance that matters.
No process eliminates the possibility of error. If we make a material mistake, we correct it. If stronger evidence changes what is known, our conclusion should change too.
Changing a conclusion when the evidence changes is not a failure of the process. It is the process working as intended.
Our standard
Our goal is not to tell readers what to believe. It is to make the evidence and reasoning clear enough that readers can understand how we got there.
We aim to show:
what the evidence supports
how strong that evidence is
what remains uncertain
what may be overstated or misleading
which explanations best fit the available evidence
what evidence could change the conclusion
When the evidence is strong, we say so clearly. When it is weak, we do not pretend otherwise. When the answer is uncertain, we say that too.
We are not here to tell people what they want to hear. We are here to get as close to the truth as the evidence allows and show how we got there.
Follow the evidence. Show the reasoning. Stay open to being wrong.
Evidence First. Conclusions Second.