Why AI’s “Show Your Work” Problem Just Got a Little Less Impossible
Remember when your math teacher demanded you show your work? Turns out that was excellent training for understanding AI safety. OpenAI just published research on how to catch AI models misbehaving by watching them “think” through their responses—and let’s be honest, it’s about time someone figured out how to peek inside these black boxes before they output something catastrophic.
According to their new paper called “Monitoring Monitorability,” the key is analyzing what researchers call “chain-of-thought” (CoT) reasoning. Think of it as the AI equivalent of showing your work on a calculus exam. The model explains its step-by-step process in natural language before spitting out a final answer.
Here’s what’s interesting: The researchers found that longer, more detailed explanations actually made it easier to predict whether a model would misbehave. More words equal more opportunities to spot the red flags. It’s not a perfect correlation, but it’s a start.
The Black Box Problem Gets a Window
AI systems have earned their reputation as inscrutable “black boxes” for good reason. They process massive amounts of data and detect mathematical patterns that even their creators don’t fully understand. That’s what makes them powerful—and potentially dangerous.
The trick is figuring out when models are taking unwanted shortcuts or, worse, actively deceiving users. It’s like trying to figure out if your teenager is lying to you, except the teenager can process billions of data points per second and has no social awkwardness to give away the deception.
OpenAI’s research doesn’t solve this problem entirely. What it does is create a framework for evaluating the reasoning process itself. They tested their approach on seven different models—four of their own (GPT-5, o4-mini, o3, and o3-mini) plus competitors like Anthropic’s Claude, Moonshot’s Kimi, DeepSeek’s R1, and Alibaba’s Qwen3-8B.
Three Ways to Catch AI in the Act
The researchers developed three complementary approaches to monitoring model behavior:
Intervention focuses on modifying how CoT processes work to make them more transparent. Think of it as installing security cameras in the model’s thought process.
Process establishes whether the model’s outputs are actually truthful. This is basically the AI version of a polygraph test, except it might actually work.
Outcome-property assesses whether catching warning signs in the CoT reasoning actually makes models safer in practice. Because spotting problems is great, but only if you can do something about them.
Here’s the reality: Models with access to complete CoT outputs performed significantly better at predicting misbehavior than those with limited information. More transparency equals better monitoring. Not exactly rocket science, but it’s nice to have data backing up common sense.
The Monitorability Tax
The researchers introduced something they call the “monitorability tax”—a compromise between model capability and safety. You can often switch to a smaller model running at higher reasoning effort to get much better monitorability with only a small hit to performance.
It’s the classic tradeoff. You want your AI to be incredibly powerful, but you also want to sleep at night knowing it’s not plotting world domination. As models get deployed in higher-stakes settings—think healthcare, finance, critical infrastructure—paying this tax becomes less optional and more mandatory.
To be fair, the OpenAI team is upfront about the limitations. This isn’t a silver bullet solution. It’s an early-stage toolkit that needs significant refinement before it becomes truly robust. They’re building the foundation, not the finished product.
What This Means for the Rest of Us
According to OpenAI’s research, maintaining or improving chain-of-thought monitorability requires “a robust and broad set of evaluations.” Translation: We need a lot more work before we can confidently trust these systems in critical applications.
The industry has been scrambling to develop creative safety methods. OpenAI recently unveiled research aimed at getting models to confess their mistakes. Anthropic released an open-source tool called Petri that probes for vulnerabilities. Now we have frameworks for monitoring reasoning processes in real-time.
Despite all this progress, here’s the bottom line: Until developers build truly foolproof models that are fully aligned with human interests—and that’s a massive “if”—we should treat AI systems for what they actually are. They’re pattern-matching machines designed to detect correlations and generate responses, not infallible oracles of truth.
The Real Impact: Defense Against the Dark Context
The correlation between reasoning length and monitorability offers a practical path forward. Longer explanations provide more surface area for detecting problems. More information generally translates to more accurate predictions and, by extension, safer deployments.
But here’s what most people miss: This research highlights how far we still have to go. The fact that we’re celebrating the ability to maybe catch AI models lying some of the time shows just how early we are in this journey.
The researchers tested their framework across multiple models from competing developers. That cross-platform approach matters because AI safety can’t be a proprietary solution. We need industry-wide standards for monitorability, not just individual companies doing their own thing.
Watch for this research to influence how future models are designed. The “monitorability tax” concept suggests that slightly less powerful but significantly more transparent models might become the standard for high-stakes applications. Better to have an AI that shows its work than one that’s brilliant but unpredictable.
In practice, this means organizations deploying AI systems should prioritize transparency over raw capability—at least until the technology matures. The most impressive model isn’t necessarily the safest one, and in domains where mistakes have serious consequences, safety should win that battle every time.