How to Peer Review a Paper: A Step-by-Step Guide for Your First Review

Have you ever had training on how to peer review a paper? Sooner or later every early careers researcher gets an email asking if they would like to peer review a paper. Before then, there is a good chance your supervisor might forward a manuscript vague “could you have a look at this?” and nothing else. It is true to say that most researchers do their first peer review with no training at all, piecing it together from the reviews they've had on their own papers or through participation in journal clubs.

This guide is the walk-through guide to peer review we wish we'd been given: what the job actually is, what you'll be sent, how to peer review a paper step by step, how to write a report an editor can use, and the newer questions about AI that nobody covered when most of us started.

We wrote it for early career researchers in any area of science, health and medicine. We've had dementia researchers in mind, and there's a box on where to start in our field, but almost everything here applies whatever you work on. It draws on guidance from journals and publishers, the Committee on Publication Ethics (COPE), research on peer review itself, and advice from editors and reviewers.

Why editors need you

If you've ever waited months for reviews on your own paper, you've felt the problem from the other side. Editors are finding it harder and harder to get people to say yes. A study of 20 years of data from the journal Biological Invasions found that acceptance rates for review invitations fell steadily, until by 2023 one reviewer declined for every one who agreed. The same study found that early career researchers were the most likely to accept and senior academics the least. A family medicine journal reported a similar pattern, with acceptance peaking at 56% in 2020 and dropping to 36% in 2024.

So if you've been wondering whether journals really want to hear from someone at your stage, they do. You're the people still saying yes.

What you get out of it

Learning how to peer review a paper makes you a better writer, because once you've had to work out what a confusing methods section is trying to say, you stop writing confusing methods sections. You see how papers are judged from the inside, which takes some of the mystery out of your own submissions. You get to know editors, and editors remember reviewers who are thorough, fair and on time. Plenty of editorial board members started exactly this way. And it counts as service on your CV and in promotion cases, as long as you record it (more on that in getting credit).

The honest bit

It's unpaid, and it takes time. One estimate put the time researchers spent on peer review in 2020 at over 100 million hours, worth close to 400 million US dollars for UK-based reviewers alone, and the same work suggests a typical review takes around six hours. Whether reviewers should be paid is a live argument, and one of our contributors has made the case for paying peer reviewers. For now, it's a choice you make, and saying no is always allowed.

What is a peer reviewer's job? (And what isn't)

This is the bit most guides on how to peer review a paper skip, and it's the source of a lot of anxiety. First-time reviewers often feel they're meant to catch every error, judge whether the paper is publishable and somehow spot fraud too. You're not.

Springer Nature's guidance puts it simply: the main purpose of peer review is to give the editor what they need to make a fair, evidence-based decision, and to show the authors how they could strengthen the paper. So you're writing for two audiences, the editor and the authors, and the editor makes the call.

What is your job

• Judging whether the methods are sound and suitable for the question being asked.
• Checking whether the data support the conclusions. This is the big one. Overclaiming is the most common problem you'll find.
• Checking the reporting is clear and complete, so that another group could, in principle, repeat the work.
• Giving context, if the journal asks for it: does the paper add something, and does it engage with the relevant literature? Some journals, such as PLOS ONE, ask reviewers to judge whether the science is sound rather than how important it is, so check which you're being asked for.
• Flagging concerns about ethics, integrity or anything that doesn't add up, to the editor.
• Saying where your expertise stops. Nature's editors say it's absolutely fine, and often intended, for a reviewer to assess only one technical part of a paper thoroughly. Editors often pick reviewers precisely for one method.

What isn't your job

• Making the decision. You recommend. The editor weighs your report against the others and against the journal's criteria. Reviewers disagree all the time, and that's expected.
• Copyediting. If the writing gets in the way of understanding, say so, with an example or two. But don't correct every typo. Journals have copyeditors, and your time is better spent on the science.
• Turning it into the paper you'd have written. Ask whether this study answers its own question, not whether it answers yours. PLOS advises against asking for work that would really be the next step in the research.
• Investigating misconduct. If something looks wrong, tell the editor. Journals have processes and tools for looking into it. Don't contact the authors or start digging yourself.
• Proving a paper was written by AI. More on this below, but the short version is that noticing problems is your job and diagnosing their cause isn't.
• Checking absolutely everything. Nobody recalculates every number or reads every reference. Do sensible checks, spot-check what looks odd, and be clear in your report about what you did and didn't look at.
• Specialist statistical review, if you're not a statistician. Say the statistics are beyond you and suggest the editor gets a statistical reviewer. That's a useful comment, not an admission of failure.

What you'll be sent, and what you won't

Nobody tells you this in advance, so here's what to expect. Every journal and submission system is slightly different, but this is the typical picture.

Before you agree

The invitation is usually an automated email from the journal's submission system, sent on behalf of a named editor. It normally includes the paper's title and abstract, the deadline for your review, and links to accept or decline. In single-anonymous journals you'll often see the author list too. You usually won't get the full manuscript until you've agreed, so the abstract is what you have to decide on.

Once you've accepted

• The manuscript, with author names removed if the journal uses double-anonymous review.
• Figures and tables, sometimes as separate files.
• Supplementary files, if the editor has chosen to share them. Journals differ on whether reviewers are expected to assess these, so check.
• The journal's reviewer guidelines and review form, which tell you what the editor wants to know.
• Sometimes more: a completed reporting checklist, a data availability statement, links to data or code, or a trial registration or protocol.

What you usually won't see

• The cover letter. Authors write this to the editor, and many systems hide it from reviewers by default, partly because it can identify the authors.
• Who the authors suggested or asked to exclude as reviewers.
• The similarity or plagiarism report, which the editorial office normally runs for editors.
• The other reviewers, or their reports, while you're still reviewing. You review independently.
• The editor's own notes and discussions.

After the decision

Many journals send reviewers the decision letter and the other reports once the editor has decided. Not all do. When you get them, read them properly. Seeing what other reviewers spotted that you missed, and what the editor made of it all, is some of the best free training going.

The review form

The form itself is usually simple: a box for the authors, a separate box only the editor sees, and a drop-down for your recommendation. Some journals add rating scales for novelty or clarity, which you'll find either helpful or baffling. Whatever's on it, don't write your review in it. Write it in Word or a notes app and paste it in at the end. Submission systems log you out without warning, and losing two hours of work to an expired session is a rite of passage you can skip.

Who knows who

Journals use different peer review models. In single-anonymous review, the most common kind, you know who the authors are but they don't know who you are. In double-anonymous review neither side knows. In open review, names are shared, and in transparent review the reports are published alongside the paper, sometimes with reviewers' names. Springer Nature has a clear summary of the different models. Whatever the model, write as if your name were on it.

How to get your first invitation

Editors find reviewers in a handful of ways: their own networks, the reference lists of the paper they're handling, database searches for people who've published on the topic, and names suggested by the authors. So the first job is simply to be findable.

• Get your online presence in order. Nature's editors recommend keeping your web page up to date with keywords that describe what you do, and maintaining your ORCID and Google Scholar profiles. If an editor searches for someone who knows your method, you want to come up.
• Tell editors you're willing. Email the editor of a journal in your area with a short note about your expertise, or introduce yourself at a conference. It feels forward. Editors short of reviewers tend to welcome it.
• Co-review with your supervisor, properly and with the editor's knowledge. See the next section but one.
• Join an early career reviewer scheme. Some journals and publishers keep pools of early career reviewers, and eLife has run one for years.
• Review preprints in the open. PREreview's Open Reviewers programme is a mentoring scheme for early career researchers. Trainees publish their first preprint reviews on the platform, giving them a public record of reviewing before any journal has asked.

Before accepting, check your expertise, availability, potential conflicts, the journal’s legitimacy and whether you can review a revised version.

Beware of invitations that are too good to be true

If an invitation comes from a journal you've never heard of, gushes about your “eminent expertise”, or is miles outside your field, look the journal up before you reply. Predatory journals send out review requests too, and reviewing for them adds nothing to your CV and helps a business that exploits researchers. Check whether the journal is indexed in the usual databases, whether its editorial board is real, and whether people you trust have published there.

Before you agree to peer review a paper

Read the abstract and ask yourself five questions.

• Do I know enough? You don't need to be the world expert, but you need to be able to judge most of it properly. If you can only judge part of it, accept and tell the editor which parts, or decline.
• Do I have the time? Check the deadline against your diary. If it lands in the week as your annual leave, or during a conference, or (even worse) near a grant deadline say no or ask for more time when you accept. Editors don't mind being asked. What they can't stand is the reviewer who agrees, and then goes quiet, or emails them the due date asking for another month. If you do say yes, consider spreading the work over two or three sessions rather than one long sitting, because you'll spot far more with fresh eyes.
• Do I have a conflict of interest? Nature's editors give the classic examples: you've collaborated with one of the authors, you have a paper competing with this one, or you have financial interests that could look like they'd sway you. Personal friendships and rivalries count too, and so do colleagues in your department. If you're unsure, ask the editor. Even in double-anonymous review you may recognise the work, and if you do, say so.
• Is the journal legitimate? See above.
• Will I be able to review the revised version too? Some journals assume that agreeing to review the first version means you'll look at later versions as well, so factor that in.

How to say no

Quickly, so the editor can find someone else. You don't need to explain at length. If you can, suggest an alternative reviewer. COPE asks that suggestions are based on suitability, not personal favours, but there's nothing wrong with putting forward a capable postdoc or colleague who'd value the experience. Something like this is plenty:

Co-reviewing with your supervisor: do it properly

This is how most researchers learn to review, and it's a good way to learn, as long as everyone does it in the open.

A survey of researchers published in eLife found that around three quarters had co-reviewed a paper, and almost all of them saw it as useful training. The same survey found that about half had ghostwritten a review, even though 81% said ghostwriting was unethical. The difference between the two is simple. Co-reviewing means contributing to a review when you weren't the person invited. Ghostwriting is doing that without the journal being told your name.

Ghostwriting is a problem for three reasons. You do the work and get no credit. The editor doesn't know who actually reviewed the paper. And unless the journal agreed, sharing the manuscript with you may have broken its confidentiality rules. Springer Nature, for example, says anyone who wants to pass an invitation to a colleague must contact the journal first.

How to make it work

• The invited reviewer asks the editor first whether a junior colleague can help. Most editors are happy to say yes.
• Your name goes to the journal, either in a co-reviewer field on the form or in the confidential comments to the editor.
• You do a full review independently, then compare notes with your supervisor. That's where most of the learning happens: seeing what they spotted that you didn't, and why.
• You get credit for it. Web of Science's reviewer recognition service lets reviewers add collaborators to a review record.

If you're the one being asked to “have a look at this”, it's fine to ask: “Shall I write it up so you can let the editor know I helped?” A good supervisor will be glad you raised it. And if you're the supervisor, co-reviewing with a student or postdoc is one of the most practical things you can do as a mentor, which we cover in our guide on how to be a good mentor.

The ground rules

The COPE Ethical Guidelines for Peer Reviewers are the standard most journals sign up to. COPE guidance says that most reviewers too often start without any guidance, so this is what they think reviewers need to be reminded of:

• Keep it confidential, during the review and after it. Don't share the manuscript, discuss it with colleagues, or mention it on social media. Many journals ask you to delete your copies once you're done.
• Don't use what you read. You can't use unpublished findings, methods or ideas for your own work, or to get ahead of the authors.
• Declare conflicts of interest, and ask the journal if you're not sure.
• Judge the work, not the people. Don't let authors' nationality, gender, religion, institution or anything else about them shape your review. That's harder than it sounds. Research has shown that authors' names can have an astonishing influence on reviewers.
• Be objective and constructive. No hostile, sarcastic or personal remarks.
• Do your fair share, and do it on time.

Don't ask for citations to your own work without a good reason

Suggesting relevant papers the authors have missed is part of the job, even when some of them are yours. Demanding citations to your own work as the price of a good review is not. One journal found a reviewer who had asked for an average of 35 extra citations per review, about 90% of them to the reviewer's own papers, and banned them. The test is simple: only suggest a paper, yours or anyone's, if the manuscript would be weaker without it, and say why. Springer Nature warns that it notes any attempt at reviewer-coerced citation against your record.

Using AI as a reviewer: what's allowed

This is the area where the rules are moving fastest, and where first-time reviewers are most likely to slip up without meaning to. The single most important rule is the same almost everywhere.

Taylor & Francis, for example, states that reviewers must not upload unpublished manuscripts, files, images or information into generative AI tools, and must not use AI to generate review reports. Science and The Lancet take a similar line, and the Nature Portfolio's AI policy also bans uploading manuscripts and asks reviewers to declare any AI use in their report.

Why the rule exists

An unpublished paper is confidential. It may contain unpublished data, new methods, or information about patients or participants. Once you paste it into a public AI tool, you've lost control of where it goes, and you've broken the confidentiality the journal promised the authors when they submitted. There's an accountability issue too. The editor asked for your judgement, and you're responsible for every word of the review.

What may be allowed

Some publishers let reviewers use AI to tidy the language of their own review, as long as no manuscript content goes in. Taylor & Francis, for instance, says AI may be used to help improve review language. Others ban AI entirely, and some ask you to declare any use at all. So the honest answer is that it depends on the journal, and you have to check.

Why outsourcing your judgement is a bad idea anyway

In July 2025, journalists found that a number of preprints contained hidden instructions aimed at AI reviewers, written in white text or tiny fonts, telling any AI reading the paper to give it a positive review. When a Dutch university news service tested one of these papers, the hidden prompts changed the AI-generated reviews. If a reviewer had pasted the paper into a chatbot, they'd have been handing the authors the pen.

Make every major comment actionable: identify the problem, locate it, explain why it matters and suggest what would address it. Be specific and kind.

How to peer review a paper, step by step

There's no single right way to review, but beginners do best with a method. This one follows the approach in Mathew Stiller-Reeve's guide for Nature, which recommends three separate readings, each with a different focus. It also follows a best-practice guide for early career researchers in the International Journal of Epidemiology, which treats a review as a piece of writing with three stages: reading, planning and composing.

Step 1: Set yourself up

Before you read a word of the paper, read the journal's reviewer guidelines and look at the review form. Note what type of article it is (an original research paper is judged differently from a review or a short report) and what the journal's criteria are. Open a blank document for notes. Put two or three blocks of time in your diary over the week rather than one long session, because you'll spot more with fresh eyes.

Step 2: First reading, for the big picture

Read the whole paper through without stopping to pick holes. Then, without looking back, write three or four sentences in your own words: what question it asks, what the authors did, what they found and what they conclude. If you can't do that, that's your first finding, because the paper isn't clear enough yet.

Ask yourself whether there's a fatal flaw: something so fundamental that no amount of revision would fix it, such as a design that can't answer the question. Stiller-Reeve's advice is that if there is, you don't need to review every remaining detail. Explain the flaw clearly and carefully, and submit.

Step 3: Second reading, for the methods and results

This is where most of your time goes. Read the methods and results slowly, with the figures and tables open alongside. Does the design actually answer the question? Is there enough detail to repeat the work? Does every claim in the results trace back to data you can see? Use the section-by-section list in the next part of this guide as a prompt. Write your notes as you go, with page and line numbers.

Step 4: Third reading, for the story and the context

Now go back to the introduction and discussion. Does the introduction set up the question fairly, with the key literature? Does the discussion stay within what the results show, admit the limitations and avoid claiming more than the data can bear? Check the abstract against the paper, and the figures against their captions: you should be able to understand each figure without reading the text. Spot-check a few references, especially any you don't recognise.

Step 5: Sort your notes

Go through your notes and put each point into one of two piles. PLOS puts it well. Major issues are the essential things the authors must fix before the paper can be considered further. Minor issues still matter, but they're smaller and won't change the overall conclusions. Delete anything that's just personal preference about how you'd have done it.

Step 6: Write the report

Use the structure in writing the report. Each major point should say what the problem is, where it is, why it matters and what would fix it.

Step 7: Leave it, then edit it

If you can, sleep on it. Then read your report through, ideally out loud, which Stiller-Reeve recommends because you'll hear anything that sounds harsh or muddled. Check that your recommendation matches your comments. Check you haven't revealed who you are, if the review is anonymous. Then submit it, on time.

What to check when you peer review a paper

Use this as a prompt, not a form to fill in. Not everything applies to every paper.

Title and abstract

Read the abstract again once you've read the whole paper. It's surprising how often the numbers in it don't quite match the results table, or the conclusion is bolder than anything the discussion dares to say. Titles drift too: a study of patients in one clinic can end up with a title that sounds as if it settled the question for everyone.

Introduction

Most papers say what they set out to do in the last paragraph of the introduction. Find that sentence. If you can't, or it's vague, say so, because everything else in the paper gets judged against it. Then ask whether the gap they describe is real. A quick search of the last couple of years often turns up a study that has already answered it, and the authors haven't mentioned it.

Methods

• Is the design suitable for the question? An observational study can show an association, not that one thing causes another.
• Who took part, how were they recruited, and who was left out? How was the sample size decided?
• Are the controls or comparison groups appropriate?
• Do the authors report ethics approval and consent?
• Did the authors pre-register the study or follow a published protocol? If so, does the paper stick to it?
• Is there enough detail for someone else to repeat it?
• Have the authors followed the right reporting guideline? The EQUATOR Network lists them, including CONSORT for randomised trials, STROBE for observational studies, PRISMA for systematic reviews and ARRIVE for animal research. Opening the relevant checklist yourself is one of the quickest ways to spot what's missing.

Results

• Does every claim trace back to data in a figure or table?
• Do the numbers add up? Check that participant numbers are consistent between the text, tables and flow diagram, and that percentages match their counts.
• How much data is missing, how many people dropped out, and how was that handled?
• Are effect sizes and confidence intervals reported, not just p-values? Have the authors run lots of tests and only reported the interesting ones?
• Do the figures show what the captions say they show?

Discussion and conclusions

This is where papers most often run ahead of their data. Look out for words like “prevents”, “leads to” or “drives” in a study that could only ever show an association. Then find the limitations. If they get a single sentence near the end, they probably need more, and it's fair to say so. Finally, check who the findings are being applied to. Results from 200 people in one memory clinic don't automatically tell us about everyone over 65.

Data, code and references

“Available from the authors on reasonable request” is a common line in data availability statements, and in practice it often means nobody will ever see the data. If the authors link to a repository or to their code, click the link and see what's actually there. As for references, nobody expects you to check them all. But if one looks unfamiliar, look it up. It should exist, and it should say what the authors say it does.

Separate concerns that affect validity or conclusions from smaller changes that improve clarity or detail. Put the most important issues first.

A note for dementia researchers

A few issues come up again and again in our field. Check how dementia or cognitive impairment was defined and diagnosed, and whether that's consistent across groups. Watch for reverse causation: changes in behaviour, sleep, mood or activity can be early signs of dementia rather than causes of it, often years before diagnosis. And check whether results from animal or cell models are being presented as if they apply directly to people.

Spotting problems: integrity concerns and AI-written papers

Most of the time, when you peer review a paper, it will be honest work with ordinary flaws. Occasionally, something won't feel right. Here's what to look for, and, just as important, what to do and what not to do.

Integrity red flags

• Images that look duplicated, flipped, stretched or spliced, especially western blots and microscopy.
• Data that look too good: perfectly smooth curves, no variability, or results that are suspiciously identical across groups.
• Numbers that can't be right, such as percentages that don't match their totals, or more participants in a subgroup than in the whole study.
• Strange phrasing where a standard technical term should be, sometimes called “tortured phrases”. Software that disguises copied text by swapping in synonyms produces oddities like “bosom peril” for breast cancer.
• A reference list padded with irrelevant citations to the same few authors or journals.

These can be signs of anything from a sloppy mistake to a fake paper bought from a “paper mill”, a business that sells fabricated manuscripts. You can't tell which from where you're sitting, and you're not expected to.

So, is spotting AI your job?

Not quite. It helps to start from what the rules actually say. Many journals allow authors to use AI tools, as long as they disclose it in line with the journal's policy. Sage's guidance for reviewers says AI can be a powerful tool when it's properly disclosed, and that what reviewers should note is irresponsible or undisclosed use. So the question isn't “did they use AI?” It's “is the work accurate, and is it transparent?”

The signs Sage suggests looking out for include:

• References that don't exist, or that have the wrong journal, year or DOI. This is the most concrete sign, because AI tools are known to invent plausible-looking citations.
• Vague, generic statements such as “studies have shown” with nothing cited.
• Smooth prose that says very little, repeats itself or over-explains the obvious.
• Contradictions between one section and another.
• Technical terms used slightly wrongly.
• Leftovers, such as placeholder text or the odd line that reads like a chatbot talking to its user.

What's not a sign: poor grammar or awkward English. Sage is explicit that bad writing doesn't point to AI use, and a large share of the world's researchers write in their second or third language.

Don't run the paper through an AI detector

It's tempting, but don't. First, it breaks confidentiality, because you're uploading the manuscript to a third-party tool. COPE has looked at exactly this case: a reviewer uploaded a manuscript to an online AI checker and reported that most of it was AI-generated. The COPE council member who presented it said it's nearly impossible to detect AI-generated text properly and definitively, and that the reviewer had breached confidentiality in the process.

Second, the detectors aren't reliable, and they're biased. A Stanford study tested seven popular detectors on essays by non-native English writers and found they wrongly labelled more than half of them as AI-generated, while being almost perfect on essays by native speakers. In an international field, that's a fairness problem, not just a technical one.

What to do if you have concerns

• Check the journal's AI policy for authors, and whether the paper includes a statement on AI use.
• Check the specific things that worried you. Search for a reference you suspect doesn't exist. Looking up a single citation in PubMed or Google Scholar doesn't breach confidentiality.
• Review the science on its merits. If the evidence doesn't support the claims, that's your critique, whoever or whatever wrote the paper.
• Tell the editor, in the confidential comments. Stick to what you observed, not what you suspect. Leave any accusations out of your comments to the authors.

Here's the kind of wording that works:

Protect the manuscript: do not upload it to public AI tools, and obtain the editor’s permission before sharing it with a co-reviewer.

That's it. You've done your job. The journal has research integrity staff, screening tools and processes for what happens next, and COPE has guidance for editors on handling it. If you're not sure whether something is a problem, say that too. Editors would much rather hear a concern that turns out to be nothing than miss one that matters.

How to write a peer review report

Most journals' guidance agrees on the basic shape. PLOS suggests structuring it like an upside-down pyramid, with the most important points at the top. Nature's editors say the best reviews have two things in common: they back up every statement with detailed argument, and they stay constructive even when there's a lot to criticise.

The structure

• Summary. A short paragraph, in your own words, on what the paper asks, what the authors did and what they claim. It shows the editor and authors you've understood it, and sometimes reveals that you haven't.
• Overall assessment. A few sentences on the paper's genuine strengths and your main concerns, so the editor can see your view at a glance.
• Major issues, numbered. For each one, say what the problem is, where it is (page, line or figure), why it matters to the conclusions, and what the authors could do about it.
• Minor issues, numbered. Shorter points that would improve the paper but don't affect its conclusions.
• Confidential comments to the editor. Anything that shouldn't go to the authors: a possible conflict of interest, integrity concerns, the limits of your expertise, or context for your recommendation. Don't use it for a second, harsher review that the authors never get to answer.
• Your recommendation, usually selected on the form. Some journals ask you not to state it in your comments to the authors, because the editor may decide differently.

How long should it be?

Long enough to do the job. A short review of a strong paper is fine, if it's specific. Research suggests the best peer review reports are at least 947 words, but length isn't the point. A page of vague comments is worth less than five sharp ones.

A template you can copy

A worked example: weak review vs strong review

Here's a made-up abstract with some problems planted in it. Before you read on, try spotting them yourself. The quickest way to learn how to peer review a paper is to see it done badly and then done well, so below are two reviews of it.

The weak review

There's nothing rude here, and that's the trap. It's simply no use to anyone. There's no summary, so the editor can't tell whether the reviewer understood the paper. “Consider confounding” doesn't say which confounders, or why they matter. “Could be bigger” isn't a problem the authors can fix for a study that's already been done. The language point gives no examples. The two citations come with no reason, and look suspiciously like the reviewer's own work. And it misses the biggest problem of all, that the conclusion claims cause and effect.

The strong review

Why the second review works

It shows the reviewer understood the study. It says something fair about the paper's strengths. Every major point says what, where, why and what to do. The requests are things the authors can actually do with the data they have. It doesn't ask for a different study. And it's firm without being unkind.

Getting the tone right

Most of us can remember a review that stung. Unkind reviews are common, and they do real damage. In a survey of 1,106 scientists across 46 countries, 58% had received at least one unprofessional review. Everyone got them at similar rates, but researchers from groups under-represented in science were the most likely to say a review had knocked their confidence, their productivity and their career. A careless sentence from you can stay with someone for years.

The good news is that you don't have to become Reviewer 2. In fact, Reviewer 2 may have been framed all along: one study of a political science journal's records found no evidence that Reviewer 2 was any harsher than the others. Reviewer 3 was the more likely culprit.

Give each reading a purpose: understand the study, examine the methods and results, then assess whether the claims fit the evidence and context.

How to be honest without being harsh

• Follow the golden rule. Write the review you'd want to receive. PLOS suggests only writing comments you'd be comfortable signing your name to, because the authors will read every word.
• Talk about the paper, not the people. “The analysis doesn't account for…” rather than “the authors have failed to…”.
• Name the strengths, specifically. “Interesting paper” means nothing. “The recruitment of a community sample rather than a memory clinic one is a real strength” tells the authors what to keep.
• Turn criticism into a request. “The paper would be stronger if…” lands better than “this is inadequate”, and it's more useful.
• Cut the sarcasm, rhetorical questions and exclamation marks. They never read the way you meant them to.
• Remember a real person will read it. Quite possibly a PhD student who did most of the work.

Being kind doesn't mean going soft. If the paper's main conclusion doesn't hold, you must say so, clearly. Kind and honest aren't opposites. For the other side of this, read when peer review feels personal and how to fight back against a peer-review bully.

Making your recommendation

Most journals ask you to choose from four options, though the wording varies.

• Accept: ready to publish as it is, or with trivial changes. Rare for a first submission.
• Minor revision: sound work that needs light changes before it's ready, such as clarifications, extra detail or toned-down wording.
• Major revision: potentially publishable, but needs substantial work, such as new analyses, a rethink of the conclusions or a lot more explanation.
• Reject: the problems can't be fixed through revision, or the paper isn't right for this journal.

Make sure your recommendation and your comments tell the same story. If you recommend major revisions but every point is minor, or reject without explaining a problem that can't be fixed, the editor won't know which to believe. If you recommend rejection, still give the authors useful comments. They'll take the paper somewhere else, and your review can make that version better.

Don't be put out if the editor goes a different way. They're weighing several reviews and their journal's priorities, and reviewers disagree all the time. That's the system working, not your review being ignored.

When the paper comes back

If the authors revise, you'll usually be sent the revised manuscript (often with changes tracked), a response letter answering each comment point by point, and your original report. Sometimes you'll see the other reviewers' reports and responses too.

Your job this time is narrower. Check that each of your points has been dealt with, either by a change or by a reasoned explanation of why not. It's fine to accept a good argument. Authors don't have to do everything a reviewer asks, and sometimes they're right. Don't raise new major issues that you could have raised the first time, unless the revision has introduced them or you've spotted something serious you genuinely missed. Moving the goalposts at round two is one of the most frustrating things a reviewer can do. A second-round review is usually much shorter, and that's as it should be.

Getting credit when you peer review a paper

Reviewing only helps your career if you can show you've done it, and anonymous work is easy to lose track of. The main tool is the Web of Science Reviewer Recognition Service, which used to be called Publons. It lets you record and verify your reviews even when they're anonymous or the paper was never published, and it can export your review history to your ORCID profile automatically. Many journals offer a tick box on the review form to add the review to your record. Some also send certificates.

On your CV, add a peer review section listing the journals you've reviewed for and roughly how many reviews. Don't list manuscript titles, which are confidential. In narrative CV formats, such as the Résumé for Researchers, reviewing belongs in your contribution to the research community. Our guide to academic promotion criteria explains how this kind of service fits into a promotion case.

Free training on how to peer review a paper

A word of honesty first. In a randomised trial at the BMJ, short training courses had only a slight effect on the quality of reviews, although trained reviewers did find more of the errors. A course can show you how to peer review a paper in principle. Practice, feedback and reading other people's reviews are what make you good. With that said, these are worth your time:

• Focus on Peer Review from Nature Masterclasses: free, about three and a half hours, taught by journal editors, with a certificate.
• Web of Science Academy, ACS Reviewer Lab and Elsevier's Certified Peer Reviewer Course, all free and all listed in the European Association of Science Editors' peer review training toolkit. The Web of Science Academy includes a course where you write a review and then revise it with a mentor.
• PREreview Open Reviewers, a mentoring programme for early career researchers that ends with published preprint reviews.
• PLOS's peer review checklist and template, short enough to keep open while you work.
• The COPE Ethical Guidelines for Peer Reviewers, the ethical standard most journals follow.
• Ella August and colleagues' staged writing guide to peer review in the International Journal of Epidemiology, written specifically for early career researchers.