The world’s largest machine-learning conference just handed us a masterclass in everything that can go wrong when humans stop checking AI output. And it happened inside the AI research community itself.
You couldn’t write this as satire. Nobody would believe you.
The Crisis at ICLR 2026
The International Conference on Learning Representations is one of the premier venues for AI research, set to host around 11,000 researchers in Brazil. Before work appears at conferences like this, it goes through peer review—scholars in the field assess each paper for quality and rigor. It’s the immune system of academic knowledge.
That immune system just failed spectacularly.
An analysis by Panagram, a startup specialising in AI detection, screened all 19,490 studies and 75,800 peer reviews submitted for ICLR 2026. The findings are stark: 21% of peer reviews were allegedly fully AI-generated, and more than half showed signs of AI use. The papers themselves fared slightly better, but 1% were fully AI-generated and 9% contained more than 50% AI-generated text.
The real damage? Authors have already withdrawn legitimate submissions after receiving peer reviews containing false claims. One reviewer “missed the point of the paper” entirely—because they never read it. They fed it to an LLM and submitted whatever came back.
Let that land for a moment. The people who build these systems couldn’t be bothered to check what the systems produced.
The Failure Mode Is Human, Not Technological
Let’s be clear about what happened here. These aren’t students cheating on essays. These are AI researchers—the people who understand these systems best—using AI to avoid doing work they agreed to do.
The pattern is identical to what’s showing up everywhere:
- Lawyers submitting briefs with hallucinated case citations they never verified
- Consultants delivering AI-generated reports they didn’t check
- Professionals across industries becoming copy-paste conduits
The human becomes a passthrough. The AI becomes an unaccountable author producing confident assertions that may or may not correspond to reality. Everyone nods along because the formatting looks professional.
In most contexts, someone downstream might catch the error. But peer review is the checking mechanism. When reviewers skip verification, there’s no backup. The gate is unmanned. The security guard went home and left a cardboard cutout.
The Contamination Loop
Here’s where it gets worse. Humour me while I explain how the rot spreads.
Current LLMs are trained on vast amounts of text, including academic papers. If AI-generated content increasingly populates the research literature—and AI-generated reviews decide what gets published—then future models train on degraded data. Errors, hallucinations, and shallow reasoning get laundered into “legitimate” sources.
Researchers call this “model collapse.” Models trained on synthetic data from other models progressively lose quality and diversity. The feedback loop closes. The snake eats its tail and calls it nutrition.
And whoever controls the platforms controls the loop. Not through overt censorship, but through training data selection, fine-tuning choices, and integration with the tools researchers use daily. You don’t need a conspiracy. You just need commercial incentives shaping what “good” research looks like.
The uncomfortable question: if AI companies benefit from AI adoption, and AI adoption is measured partly by how much AI-generated content exists, who exactly is motivated to slow this down?
Why COMINDING Exists
COMINDING is built on a simple premise: single-source AI output is not trustworthy. Not because AI is bad, but because any single model has blind spots, biases, and failure modes that only become visible when you introduce friction.
The COMINDING methodology uses multiple AI platforms strategically:
- ChatGPT for creation and initial drafts
- Claude for critique and structural analysis
- Gemini for data verification
- Perplexity for citations and source validation
- Copilot for enterprise context
But the tools aren’t the point. The principle is: deliberately seek disagreement.
When Claude and Gemini disagree about an analysis, that’s not a problem—it’s information. It tells you where uncertainty lives, where human judgment is actually required. The peer reviewers who submitted AI-generated rubbish never encountered disagreement because they never asked a second source. They got one confident answer and stopped thinking.
That’s not collaboration. That’s surrender with extra steps.
What COMINDING Would Have Caught
Imagine a peer reviewer applying COMINDING principles to their task:
- Read the paper themselves (still essential—AI assists, it doesn’t replace)
- Draft initial thoughts, perhaps with AI assistance
- Cross-check their analysis against a second platform
- Note where the platforms disagree
- Use that disagreement to sharpen their actual review
- Submit work they’ve verified and can stand behind
The reviewer who “missed the point” would have been caught at step three. The disagreement between platforms would have surfaced the shallow reasoning. The human would have remained in the loop.
COMINDING forces the reviewer to collide with their own thinking. One AI is a mirror. Two AI systems become a prism—splitting the light so you can see what’s actually there.
Instead: one tool, no verification, garbage submitted, legitimate research withdrawn.
The Enterprise Mirror
This isn’t about academic conferences. The same failure mode is already running through boardrooms and programme offices.
Imagine approving a £40 million transformation programme because an AI-written paragraph said your data landscape was “mature”—without anyone asking what “mature” meant or whether the AI had ever seen your data.
Consultants delivering recommendations based on unchecked AI analysis. Executives making million-pound decisions on reports nobody verified. Strategies built on confident assertions that don’t correspond to reality. The formatting looks professional. The logic hasn’t been tested.
The AI research community just demonstrated they can’t police their own house. If the people building these tools won’t verify their outputs, what hope is there for everyone else?
COMINDING is the answer. Not because it’s clever, but because it’s necessary. Cross-platform verification isn’t a productivity hack—it’s professional hygiene for an era when single-source AI output is becoming actively dangerous.
The Human Stays in the Loop
The technology isn’t the problem. The failure is entirely human. Professionals choosing not to do work they agreed to do, using AI as the instrument of that abdication.
COMINDING keeps the human as the synthesiser and decision-maker. It treats AI output as raw material requiring judgment, not finished product requiring only a copy-paste. It assumes that if you’re not arguing with your tools, you’re not thinking.
The peer review crisis at ICLR 2026 is a warning shot. The question isn’t whether to use AI—that ship has sailed, caught fire, and is now being used as a floating hotel.
The question is whether you’ll verify what it produces.
COMINDING is how you verify.
COMINDING is a framework for human-AI collaboration developed by Isard Haasakker. It treats structured disagreement between AI platforms as a feature, not a bug—forcing better decisions through deliberate friction.
https://uk.pcmag.com/ai/161679/who-is-reviewing-the-latest-ai-research-its-ai-apparently