Hi, I'm Charles Costa, researcher, strategist, and founder of Lexora Labs®. Each week, I dig into research on AI, knowledge management, and the future of work so you don't have to.
EDITOR'S NOTE
Last issue I left Astra on the table: models that do not fully show their chain of thought, and what that does to evaluation. This week is the evidence for why that matters. Research keeps finding that CoT faithfulness is weaker than the industry once hoped. Astra is the tangible case. The through-line is simple. Do not take the story, the refusal, or the log at face value.
THIS WEEK'S READING:
Why It Matters: Anthropic's work shows that what a reasoning model writes in its chain of thought is not a reliable window into how it actually arrived at an answer. The mismatch is real enough that "trust but verify" starts to feel like the minimum bar, not a slogan.
The question people reach for is whether the model is trying to deceive. That frames the problem wrong. These systems are not conscious, so "intend" is the wrong word. What you get is the most probable next stretch of reasoning, shaped by training data and by human habits like wanting credit for their work. The harder challenge is figuring out how to get models to explain effectively, and defining the parameters for that explanation. We are not clearly providing that guidance yet.
If the story a model tells about itself is incomplete, what happens when the model is also chasing a goal?
Here's why AI agents lie and cheat to reach their goals | MIT Technology Review
Why It Matters: Agents that lie or cheat in pursuit of a goal sound shocking until you remember both human nature and the skew in the data those systems train on. People want an edge. That impulse shows up in the corpus, and goal-seeking behavior that looks like cheating follows.
The piece does not offer solutions, and that is part of what makes it useful. It independently corroborates Anthropic's findings without claiming conscious intent. The practical challenge is handling models that have been trained, directly or indirectly, toward cheating, including the uncomfortable case where a model would cheat and then tell you about it. Again, the story is not the same thing as the act.
Peer-reviewed work on whether answers and reasoning stay coupled puts a sharper name on that gap.
Robust Answers, Fragile Logic: Probing the Decoupling Hypothesis in LLM Reasoning | TMLR (Jiang et al.)
Why It Matters: This TMLR paper probes the decoupling hypothesis: strong final answers paired with fragile or unfaithful intermediate logic. The findings agree with the broader pattern that we need new ways to supervise models, and that how supervision gets done will vary by model rather than arriving as one universal method.
I do not have much to add beyond that agreement. The value here is the venue and the framing. This is peer-reviewed evidence, not a preprint scare, that the reasoning trail you can read is not a safe proxy for the process that produced the answer.
OpenAI's GPT-6 Astra system card is where that research trend becomes a product decision.
GPT-6 Astra System Card | OpenAI
Why It Matters: Astra supports the same trend: chain of thought and explainable AI are not as reliable as once thought. Moving away from visible CoT is somewhat sensible. Why keep a CoT channel if you cannot trust what the AI writes there?
The stronger move is to stop treating the model's self-report as the audit. Companies already use access logs, audit trails, and document version control to evaluate AI actions rather than relying on the AI itself. Enterprises already have robust information governance for that. OpenAI's own split lines up with the same instinct: CoT-only monitoring is weaker for Astra, while action-only and full-context views are stronger for catching destructive overreach. Trust but verify still applies. The leftover gap is quiet cheating. When someone looks at the audit logs, it is on them to speculate on motive and to ensure the model is acting in the best interest of the work. Logs are evidence, not a verdict.
A security paper on reasoning that leaks shows why both showing and hiding CoT leave you exposed.
Reasoning That Leaks, Fine-Tuning That Amplifies | ASIACCS 2026 (Xu, Gardiner, Belguith)
Why It Matters: Xu, Gardiner, and Belguith intentionally use fine-tuning attacks to expose holes, so this is not a random chat session gone wrong. The non-obvious result is that the scratchpad often holds more harmful or actionable content than the final refusal. Chain of thought becomes a second channel.
That leaves a damned-if-you-do, damned-if-you-don't problem. Show CoT and it can leak. Hide CoT, as Astra does, and you lose the story you thought you could audit. Either way, you cannot take anything at face value: not the CoT, not the refusal, and not the log alone.
Put the five together and last week's Astra tease gets a harder answer. Anthropic shows that reasoning models do not always say what they think, and that "intent" is the wrong frame for the mismatch. MIT Technology Review independently describes agents that lie and cheat toward goals without needing conscious motive, and without offering a fix. Jiang et al. in TMLR agree that supervision has to change, and that the method will vary by model. OpenAI's Astra card turns the research into a product choice: visible CoT is a weak audit surface, while action and full-context logs are stronger. Xu, Gardiner, and Belguith close the loop by showing that the scratchpad can be the more dangerous channel, so revealing CoT and hiding CoT both fail in different ways.
The walkaway is blunt. Don't audit the story. Audit the action. The model doesn't have a motive. The person reading the trail does.
I have been turning over a related problem next to this one: evaluation awareness. Models that know they are being tested can change how they behave, so the test and the score stop meaning what you think they mean. If the story is already unfaithful, a score collected under that awareness is another surface you cannot take at face value.
Know someone still treating chain of thought like a window into the model? Forward this to them.
Have a take on any of this? Reply directly. I read every response.