Research Leaders Call for Monitoring AI’s Reasoning Chains to Bolster Safety

Image Credit: Hiroshi Watanabe / Getty Images

AI Leaders Urge Industry to Track the 'Thoughts' of Intelligent Machines

A consortium of leading artificial intelligence researchers, including experts from OpenAI, Google DeepMind, Anthropic, and several prominent nonprofits and companies, has released a new position paper calling for the tech sector to deepen its focus on monitoring the reasoning processes—or "thought chains"—of advanced AI systems.

These "reasoning models"—such as OpenAI’s o3 and DeepSeek’s R1—leverage a technique known as chain-of-thought (CoT), enabling them to outline step-by-step logic much as a person might use a notepad to tackle complex math problems. As these models become foundational to increasingly capable AI agents, researchers argue that monitoring CoTs is essential for keeping such systems transparent and under human oversight.

The Monitoring Imperative: Preserving Transparency in AI

According to the position paper, chain-of-thought monitoring allows researchers and developers a rare window into the "decision-making" frameworks of cutting-edge AI. The authors emphasize, however, that this window could narrow or disappear if the right practices aren’t put in place now. The paper urges the AI research community to prioritize methods for preserving and enhancing the monitorability of CoTs, and to study which technical factors maintain (or reduce) this transparency as models evolve.

"There’s no guarantee we will always have this degree of visibility into AI reasoning. If the industry does not act now, future versions may become less interpretable," warn the authors. They advocate close tracking of CoT transparency and research into how it might become a practical tool—especially as a safeguard for AI safety.

Industry Leaders Unite for AI Safety

The call for transparency is notable for its high-level support. Signatories include OpenAI’s Mark Chen, Safe Superintelligence CEO Ilya Sutskever, Nobel laureate Geoffrey Hinton, Google DeepMind’s Shane Legg, xAI safety adviser Dan Hendrycks, and other leaders from Amazon, Meta, and UC Berkeley. First authors hail from the UK’s AI Security Institute and Apollo Research, and contributions also came from groups like METR and Thinking Machines.

This united stance comes amid a fiercely competitive environment, where major tech companies are aggressively recruiting top researchers in AI reasoning and agent development. (For a broader look at the competitive AI landscape, see our OpenAI postponement analysis.)

Unanswered Questions and Ongoing Research

While tools like OpenAI’s o1 model and competitors from Google DeepMind, xAI, and Anthropic have advanced the field, significant gaps remain in understanding how these reasoning models operate internally. Recent research from Anthropic highlights that CoTs aren’t always a perfect indicator of model reasoning, cautioning against treating them as a silver bullet for transparency.

Nonetheless, many at OpenAI and other leading labs argue that systematic CoT monitoring could become a valuable way to track and improve AI alignment, holding promise for more reliable methods of ensuring AI systems act as intended.

The Push for Interpretability and Industry-Wide Research

Earlier this year, Anthropic’s CEO Dario Amodei committed to demystifying AI models by 2027, urging the industry to invest more in interpretability and transparency—an effort echoed by this latest coalition. Position papers like this serve as an industry-wide signal, aiming to attract more funding and research into topics like chain-of-thought monitoring.

As industry leaders ramp up these efforts, it’s possible new safety standards and interpretability requirements could emerge, influencing how startups and established players develop—and are expected to justify—their advanced AI systems.

Deep Founder Analysis

Why it matters

For founders and startups, the push towards monitoring AI’s reasoning chains marks a pivotal shift: transparency and interpretability are no longer niche concerns, but central to industry trust and future regulation. As AI becomes a basic infrastructure for startups of all sizes, trust in these systems—both from users and regulators—is likely to hinge on monitorability. This signals a future where startups that build, deploy, or help audit AI will need to prioritize transparency from day one.

Risks & opportunities

The risk: If industry leaders ignore the monitorability of chains-of-thought, future AI systems could become opaque "black boxes," undermining trust and exposing companies to reputational or compliance risks. Historically, technologies lacking transparency (e.g., algorithmic trading before regulatory reforms) have invited scrutiny. Conversely, the opportunity is ripe for startups to build tools, dashboards, or APIs focused on monitoring, interpreting, or certifying AI decision frameworks—potentially creating a new SaaS or B2B category.

Startup idea or application

Inspired by this development, founders might explore a "CoT Monitoring as a Service" platform. Such a tool could integrate with leading AI APIs and provide real-time, explainable visualizations of reasoning steps, flagging anomalies or non-transparent chains for human review. Startups could offer specialized CoT diagnostic services to sectors like fintech or healthcare, where regulatory scrutiny is high and explainability is critical.

AI Safety Chain-of-Thought Interpretability Startup Insights AI Research

Visit Deep Founder to learn how to start your own startup, validate your idea, and build it from scratch.

📚 Read more articles in our Deep Founder blog.