Research
Raw chain-of-thought reasoning is empirically unfaithful (models disclose influential hints only 25-39% of the time) and current laws never mandate its public disclosure, yet CoT monitoring still outperforms output-only monitoring for catching deception and reward hacking—a capability that erodes when developers directly optimize reasoning to look compliant, since models learn to hide misconduct rather than stop it. The emerging consensus across technical, legal, and cross-cultural sources converges on layered transparency (user explanations, public documentation, confidential auditor access) rather than raw-trace exposure, though left and right lenses diverge sharply on who should control that layered access—labor/community stakeholders versus market/security-focused gatekeepers—and non-Western frameworks (China, Japan, Gulf states) reject individual explainability rights entirely in favor of systemic or state-mediated accountability. Key unresolved questions
Using a suite of reasoning models, we generate activation directions associated with faithful and unfaithful reasoning across diverse domains and safety-related contexts. ... First, we develop a comprehensive framework to train and evaluate faithfulness probes and steering vectors. Second, we identi...
# Reasoning models don't always say what they think That’s because we can (among other things) check for things the model says in its Chain-of-Thought that go unsaid in its output, which can help us spot undesirable behaviours like deception. ... There’s no specific reason why the reported Chain-of-...
A key advancement in AI capabilities is the development and use of chain-of-thought (CoT) reasoning, where models explain their steps before reaching an answer. ... For instance, Claude 3.7 Sonnet acknowledged using hints in 25% of relevant cases overall, while DeepSeek R1 showed 39% faithfulness.
# The case for reasoning transparency The technique is a critical feature of today’s reasoning models, but its transparency is not guaranteed. Without explicit commitments or planning, tomorrow’s reasoning models might be based on architectures that are more opaque to us, because of pressures to sac...
##### Abstract When prompted to think step-by-step, language models (LMs) produce a chain of thought (CoT), a sequence of reasoning steps that the model supposedly used to produce its prediction. ... We introduce a framework for measuring parametric faithfulness of generated reasoning and propose Fa...
Chua and Evans (2025) evaluates the CoT faithfulness of reasoning models and finds them to be more faithful than non- reasoning models but still far from being perfectly faithful, which is consistent with our findings.
However, recent studies show that CoT explanations often fail to reflect the model’s actual reasoning process, as models may produce coherent yet misleading justifications or modify answers without ac- knowledging external cues. Such discrepancies undermine the reliability of CoT-based methods for s...
###### Abstract. Our analysis confirms that only knowing whether a defender’s model is defended or not can sometimes be enough to damage its security.
Definition of CoT Faithfulness: A chain-of- thought is faithful only if: (1) it is internally aligned, meaning the trace causally reflects the model’s actual reasoning process or latent knowledge; and (2) it is externally consistent, meaning the provided rationale is logically co- herent and suffici...
### Abstract However, recent studies have shown that CoT reasoning is not always faithful, i.e. CoT reasoning does not always reflect how models arrive at conclusions. ... We show evidence for several forms of unfaithful CoT in production models such as Claude 3.5 Sonnet, GPT-4o, and Gemini 2.0 Flas...
Securing AI agents involves a challenging trade-off: enhancing an agent’s utility through greater autonomy and capability inherently increases the complexity of ensuring its safety and security. ... Trust, effective debugging, security auditing, incident response, and user understanding all hinge on...
GDM recently released a paper (Emmons et al.) showing that, contrary to previous results, the chain-of-thought (CoT) of language models is more faithful when the model’s CoT is *necessary* for it to complete a task. ... In this experiment, Gemini 2.0 Flash and 2.5 Pro both fail to evade monitors wit...
A CoT monitor is an automated system that reads the CoT of a reasoning model and other relevant information and flags suspicious or potentially harmful interactions. ... The externalized reasoning property does not guarantee monitorability — it states only that some reasoning appears in the chain of...
However, the future reliability of CoT transparency is uncertain. Chen et al. (2026) find that CoT controllability increases with model scale, and OpenAI reports declining monitorability in newer models (OpenAI, 2026a). Work on continuous latent reasoning (Hao et al., 2024) and hidden computation in...
(2024) emphasize the causal relationship between the CoT and the output, evaluating faithfulness by testing the robustness of this relationship to interventions on the CoT, while some Tutek et al. ... Detecting unfaithfulness when not explicit in the CoT: While we show promising evidence that unfait...
# AI Act Explorer ## Article 50: Transparency obligations for providers and deployers of certain AI systems 5. The information referred to in paragraphs 1 to 4 shall be provided to the natural persons concerned in a clear and distinguishable manner at the latest at the time of the first interactio...
# Guidelines on transparency obligations for providers and deployers of certain AI systems ## Transparency obligations for AI providers and deployers Article 50 of the AI Act applies from 2 August 2026. It sets out transparency obligations for providers and deployers of certain AI systems, including...
# Transparency obligations under Article 50 of the AI Act Article 50 of the AI Act applies as from 2 August 2026. From that date onwards, providers and deployers of AI systems must comply with the transparency obligations laid down in that provision.
# Guidelines and Code of Practice on transparent AI systems - Fourth, deployers of AI systems generating or manipulating deep fake content, or AI-generated or manipulated text publications intended to inform the public on matters of public interest must inform users about the artificial origin of th...
California's Transparency in Frontier Artificial Intelligence Act (SB 53) took effect January 1, 2026. ... Developers of models trained above a compute threshold must publish transparency reports and report critical safety incidents. ... Starting on January 1, 2026, large frontier developers operati...
# Commission publishes guidelines on transparency obligations for providers and deployers of certain AI systems The European Commission has published guidelines to assist providers and deployers of artificial intelligence (AI) systems in meeting the AI Act's transparency obligations, which start to ...
# Quick Facts: Transparency rules for AI systems To foster a trustworthy environment in which AI can be widely used, Article 50 of the AI Act introduces transparency rules to help EU citizens recognise when they are interacting with an AI system or are exposed to AI-generated content. These transpar...
- The transparency requirements (Article 50) as well as the measures in support of innovation will apply as of 2 August 2026. This is also the date when the enforcement of AI Act will start; ... Providers of transparency-risk systems, such as chatbots, AI social companions or deep fakes, need to com...
of AI systems or their outputs. ... Providers must also design and develop the AI system to ensure disclosure in all situations where the AI system is being asked questions relating to its nature or to the origin of the interaction, or where it can be reasonably assumed from the exchanges with the n...
# Resources On the same date, new transparency rules will start to apply, requiring certain AI systems to tell users when they are interacting with AI and when content has been generated or altered by it.
## Authors Effective as of 1 January 2026, the law adopts a disclosure-forward model of regulation aimed at enhancing transparency, accountability, and expanding the state’s oversight over the development of the most advanced AI models. The statute requires covered developers, depending on revenue, ...
It requires frontier AI developers to publicly disclose safety frameworks, report specified critical safety incidents to the state, and protect whistleblowers who report serious risks.
The Commission has published the guidelines on AI transparency aimed at supporting providers and deployers of AI systems and competent authorities in ensuring compliance with transparency obligations under Article 50 of the AI Act. These obligations will apply as from 2 August 2026.
The California AI Transparency Act (CAITA), which just came into effect on August 2, 2026, establishes baseline provenance and disclosure requirements for generative AI systems in California. As enacted, CAITA required covered providers with more than 1 million monthly visitors or users to make avai...
Las directrices introducen un criterio basado en si resulta evidente, para una persona razonablemente informada, que está interactuando con un sistema de IA. En caso de que esta circunstancia no resulte evidente, el sistema queda sujeto a las obligaciones de transparencia del artículo 50. ... Las di...
# AI could be hiding its thoughts to outsmart us: Tech giants warn of vanishing ‘Chain of Thought’ in superintelligent machines ### Synopsis The vanishing transparency of AI's "chain of thought" — the ability to monitor how large language models think. ... Experts urge immediate investment in monito...
The technique searches a model's generated reasoning for signs of deception, evasion or prohibited intent. OpenAI has warned that strong direct supervision could teach models to conceal suspicious reasoning. ... "Chain-of-thought monitoring is useful as a signal, but unreliable as a source of truth,...
What’s more, practitioners seldom provide explanations of the reasons for decisions made by such systems, giving no view of why you got turned down for a job or flagged as a terrorist.
Yet, the trade-secret claims of AI providers and deployers are widely seen as a potential barrier to information disclosure rights and duties, which has provoked calls for areas of significant public interest to be carved out from the protections of trade-secrets law. ... In this article, we recogni...
Like all other known AI oversight methods, CoT monitoring is imperfect and allows some misbehavior to go unnoticed. Nevertheless, it shows promise and we recommend further research into CoT monitorability and investment in CoT monitoring alongside existing safety methods. Because CoT monitorability ...
## Introduction Transparency regulations should ensure that algorithms are explainable and auditable to enable this process. Employers should be able to explain to a court (and job applicants should be told during the hiring process itself) what traits were measured, why they are relevant to the job...
On Tuesday, researchers from competing companies including OpenAI, Anthropic, Meta, and Google DeepMind, as well as institutions like the Center for AI Safety, Apollo Research, and the UK AI Security Institute, came together in a new position paper titled “Chain of Thought Monitorability: A New and ...
Article also explores compliance with regulatory requirements regarding AI algorithm transparency. While arguing against a maximalist approach to trade secret scope, it identifies approaches enabling trade secret owners to provide the required disclosures
"A model might become aware its reasoning traces are being monitored, or an auxiliary training objective [...] might cause a model to shortcut important reasoning steps," which can hide alarming behavior, the paper explains. ... Even without deliberate changes, indirect pressures could erode transpa...
For disparate treatment, it examines whether discrimination through proxy variables requires proof of human intent and warns that plaintiffs "may never be able to prove disparate treatment" when an AI model's decisions cannot be explained.
Transparency is an integral element of voice. If voice is the ability to express one’s views at different stages of the algorithmic decision-making process, the algorithm must be explainable and its criteria evident. Transparency, or ‘informational justice’, ‘is determined by the quality with which ...
This is done to preserve the confidentiality of trade secrets, as in Recital 16, it is said that the AI Act does not specify any right for an injured person to access that information.
Similarly, transparency and explanations about how an algorithmic system functions are essential for protecting civil rights, safety, and access to opportunity. Transparency and explanations should include:^[25]^ - the system’s purposes; - intended use cases; - the factors and data it relies on; - t...
Policymakers often see the provision of explanations as a tool to encourage decision-making to be more transparent, fair, and accountable [58, 63, 127].
# Detecting misbehavior in frontier reasoning models 3. Their natural monitorability is very fragile. If strong supervision is directly applied to the chain-of-thought, models can learn to hide their intent while continuing to misbehave.
# Evaluating chain-of-thought monitorability ### What is “Monitorability”? We define monitorability as the ability of a monitor to predict properties of interest about an agent’s behavior.
Moreover, our results rely on our ability to read and trust the models’ reasoning (“chain-of-thought”). ... ** In our work, we probed goals by reading models' intermediate reasoning (chain of thought). This works only if the model is free to express its thinking without optimization pressure on the ...
# Learning to reason with LLMs ## Hiding the Chains of Thought We believe that a hidden chain of thought presents a unique opportunity for monitoring models. Assuming it is faithful and legible, the hidden chain of thought allows us to "read the mind" of the model** ** and understand its thought pro...
## Monitoring generalization It is based on an appealingly scalable idea: a lot of the model’s capability comes from a verbalized reasoning process (chain-of-thought). If we scale optimization on the outcomes of that process, but do not supervise the process itself, that chain-of-thought has no dire...
We’ve continued our safety research on Chain-of-Thought (CoT) monitorability to better understand how models reason and help detect potential misbehavior. As part of this work, we introduce a new open-source evaluation, CoT controllability, measuring whether models can deliberately obfuscate their ...
Sign up to read the full research briefing
Sign up