A few days ago I released a free tool that simulates a Reviewer 2 read of your manuscript before you submit it to a Q1 journal: the AI Paper Reviewer. In just a few days hundreds of psychology and health-sciences manuscripts have gone through it, from two-thousand-word papers to seventeen-thousand-word theses. Since I get a copy of every report, I have been able to read a pattern that repeats with striking consistency. These are not exotic errors. They are three concrete, predictable problems that almost any author can fix before submitting, if they know where to look.
I am writing this because the value is not in the anecdote, but in the fact that these are exactly the critiques a real reviewer writes in the first round, and the ones that most often turn a "major revisions" into an outright rejection. If you recognise any of them in your manuscript, you have an afternoon of work ahead that will save you months of review.
1. The sample size is not justified a priori
This is by far the most frequent. I am not talking about "having a small sample": I am talking about the manuscript not explaining why that sample is enough to detect the effect being sought. The author collects the data, runs the analyses, and nowhere in the method is there a power calculation. A reviewer at a serious journal reads this as a sign that the study was not planned, and it gives them ammunition to question any non-significant result ("is there no effect, or did you simply lack the power to detect it?").
The fix is one of the cheapest in all of methodology: an a priori power analysis with G*Power, a simulation, or a calculator like our effect size calculator, plus one sentence in the method that says "with an expected effect size of X, alpha of .05 and power of .80, N participants were required". Ten minutes of work that defuse the most predictable critique of all. If you already collected the data and the sample fell short, it is better to acknowledge it yourself in the limitations than to let the reviewer do it.
2. Results report neither effect sizes nor confidence intervals
The second pattern is reporting significance (the p value) and little else. At a Q1 journal in psychology or health sciences, the p value alone has not been enough for years. A "significant" result (p < .05) can correspond to a trivial effect if the sample is large, and a "non-significant" result can hide a relevant effect if the sample is small. The effect size (Cohen's d, partial eta squared, r, odds ratio, standardised beta) tells you how much, and the 95% confidence interval tells you how precisely you have estimated it.
APA has been asking for this for editions, and most journals' reporting guidelines require it explicitly. Even so, it is still missing from a huge share of the manuscripts I review. Adding effect sizes and intervals to each test does not change your data, but it completely changes how your results section reads: it goes from "I found something" to "I found an effect of this magnitude, estimated with this precision". If you want to review how each one is reported correctly, see the guide on how to report results in APA 7.
3. Causal conclusions from non-causal designs
The third is the most conceptual and the hardest to see from the inside, because it usually lives in the language. The author has a correlational or cross-sectional design, finds an association, and writes the discussion with causal verbs: "social media use increases anxiety", "self-esteem produces better academic outcomes". The problem is that with correlational data the direction could be the reverse, there could be a third variable explaining both, or the relationship could be bidirectional. Only an experimental design with manipulation and randomisation supports a causal claim.
The fix does not require redoing the study: it requires adjusting the language. Change "increases" to "is associated with", "produces" to "is related to", and reserve the causal interpretation for the future directions section, framed as a hypothesis to be tested with an adequate design. It is one of the quickest changes to make and one of the ones that most strengthens your discussion, because it shows you know exactly what your design does and does not allow you to conclude.
What the three have in common
None of the three is a sophisticated analysis error. All three are caught by reading the manuscript with the eyes of whoever will review it, and all three are fixed before submitting, while you can still do something about it. That is exactly the gap the AI Paper Reviewer tries to fill: giving you that critical read in a couple of minutes, so you reach submission having resolved the predictable and can focus on what genuinely requires judgment.
That said, an automatic tool has a ceiling. It detects patterns, but it does not replace looking at your specific case properly: if the critique you get is a design or analysis problem you are not sure how to solve, or you need a response letter to reviewers for an ongoing submission, that is where human judgment is needed. If you want me to look at your manuscript, you can get a full statistical review or write to me for an initial assessment at no cost.