The Ethics of AI-Assisted Peer Review: How to Critique Other People’s Work Without Outsourcing Your Judgement

The Hidden Burden of the Reviewer

Whether you are an engineering lead looking at a pull request, an academic evaluating a journal submission, or a senior designer giving feedback on a UI prototype, you are likely overworked. Peer review is one of those crucial industry tasks that is vital for quality control, yet almost always pushed to the margins of our working days.

It is no wonder, then, that many have turned to LLMs to lighten the load. It takes thirty seconds to paste a colleague's code into /platforms/openai, or upload a PDF of an academic draft to /platforms/gemini, with the prompt: "Critique this and list the top five weaknesses."

It feels like a win-win. You save two hours of deep cognitive work, and your colleague gets a comprehensive, structured list of feedback. But this convenience hides a minefield of ethical compromises. When you accept the responsibility of reviewing someone else’s creative or intellectual work, you are promising them your human attention. Outsourcing that attention to a machine isn't just lazy; it can actively harm the person on the receiving end, leak confidential intellectual property, and contribute to a bland, homogenised creative landscape.

Here is how to navigate the ethics of AI-assisted review without losing your professional integrity.

The Cardinal Rule: The Privacy Threshold

Before we talk about the quality of the feedback, we have to talk about data.

If someone hands you an unpublished manuscript, a private repository, or a sensitive internal design file, they are trusting you with their intellectual property. Unless you have explicit permission from the author, you should never upload unpublished or proprietary work to a public cloud model.

Many commercial AI tools use user inputs to train future models by default. By pasting your colleague's unreleased research paper or proprietary source code into an LLM, you may be leaking their work into a public training dataset. Even if you use enterprise-grade accounts with data-privacy guarantees, you are still processing confidential data on third-party servers.

If you must use AI for review assistance: * Only use models that have clear, verified opt-outs for data training (refer to privacy documentation via https://googlegemini-support.com or OpenAI support channels). * Anonymise the work. Strip out names, company references, and unique identifying markers before running the text through an LLM. * If the work is highly sensitive, restrict your AI usage to local, open-source models running entirely on your own machine.

The Empathy Void of Algorithmic Feedback

An LLM can spot an unclosed tag, a missing citation, or an inconsistent margin in a design mockup built in /platforms/figma-weave. It is excellent at catching technical errors. What it cannot do is understand the human context behind the work.

AI feedback is notoriously pedantic. Because LLMs operate on statistical likelihood, they tend to flag anything unusual or idiosyncratic as an error. If a writer uses a surprising metaphor, or a programmer uses an unconventional but elegant design pattern, an LLM will often flag it as something to be "corrected" to fit the standard average.

If you simply copy and paste AI-generated critique, you risk crushing the very thing that makes your colleague's work unique. True peer review requires empathy. You need to understand what the creator was trying to achieve, even if their execution wasn't perfect. An LLM only knows what the average of the internet would have done in their place.

Giving feedback is a delicate human relationship. It takes more than a green tick of approval from an LLM to build real trust between peers; it requires knowing when to encourage a wild idea rather than sanding down its edges.

The Danger of the Feedback Monoculture

If every senior developer uses GPT-4o to review pull requests, and every editor uses Claude to review essays, we enter a dangerous loop. We begin training our human creators to write, code, and design specifically to please the LLMs that will review them.

This creates a feedback monoculture. Originality is systematically filtered out because the gatekeepers of our industries have outsourced their taste to the same three algorithms. Over time, our software, our literature, and our designs will begin to look and feel identical—perfectly polished, technically compliant, and utterly soulless.

Your personal judgement is not an administrative chore to be automated away; it is your most valuable asset. Your unique perspective, shaped by your specific career history and personal quirks, is exactly what the creator needs to hear.

A Responsible Protocol for AI-Assisted Review

This doesn't mean you must banish LLMs from your workflow entirely. Instead, use them as an assistant to your thinking, rather than a replacement for it.

1. Do the First Pass Yourself: Always read, run, or view the work first without any AI involvement. Write down your initial, raw impressions. What did you love? What confused you?
2. Use AI as a Specific Sanity Check: Once you have formed your own opinions, you can use an LLM to check for objective, technical blind spots. For example: "Here is a function I am reviewing. Are there any edge cases or security vulnerabilities I might have missed?"
3. Synthesise, Don't Copy-Paste: Never deliver AI-generated text directly to your peer. If the AI pointed out a valid flaw, rewrite that feedback in your own words, contextualising it with your own experience. If you can't explain why the AI's critique is correct, you shouldn't be giving that critique at all.

Peer review is a quiet pact of mutual respect. By keeping your human judgment at the centre of the process, you ensure that our creative and technical industries continue to grow, challenge, and surprise us.