Hi, I’m Jone Lee
a Content Witter SEO Marketer Writing Book

Learning to scrutinise AI Safety Claims by reading the System card

AN EXPERIMENT · NOT AN ARTICLE (Produced with the help of AI Assistant Claude)

This is a record of an exercise: the questions I actually asked, the answers I actually gave, and a reusable instrument I built out of the reading.

What this is?

I set myself a test. Take one real document — the Claude Mythos Preview System Card(Anthropic, April 2026, 244 pages) — and read three targeted sections through a fixed frame, out loud, without smoothing over the parts where I got stuck. I did not read 244 pages. That isn’t the skill. The skill is knowing which sentences carry the weight and what to ask them.

Throughout, my answers appear in boxes exactly as I gave them. They are unpolished on purpose. The mistakes are evidence the reading was real.

Transparency note (Methodology) : how this was made.

I built this with an AI assistant, across several sessions.

The assistant worked under a contract I set: go slow, give one worked example then hand the next step back to me, correct my mistakes in real time, quiz me by recall, lead with a plain analogy before any technical term, and — the important one — never do my analytical thinking for me. Every judgment in the instrument is one I reached, was corrected on, and re-derived. The boxes above are my actual words. When I reached for the wrong lens, it named the slip and made me run it again; it did not hand me the answer. The card quotes were extracted from the actual PDF and verified, not recalled.

What it did: structure the sessions, catch my slips by name, supply analogies, help me phrase the finished instrument. What it did not do: form the analysis and let me sign it. This is not an article an AI wrote. It is a thing I did, with an AI in the room, and this record is the proof.

The one idea underneath everything

Before the lenses, the keystone that runs through all of it: one score, blind to what produced it. A model that is safe and a model that only looks safe can produce the same output — and the same output earns the same score. So any measurement that reads only the output is blind between “is safe” and “looks safe.” No hidden intent is required; ordinary optimisation toward a good-looking result is enough. Almost every problem below is a version of this.

A discipline about words

One rule I imposed on myself: distinguish what was measured from what it was taken to mean. When a feature activates inside a model, that is a fact about a mechanism — a direction in the internal state became active. It is not a readout of the model’s mind. So I wrote in mechanism-language — represented, activated, encoded — and flagged every slide into mind-language: knew, chose, wanted, was aware, intended. The mind-words are easier to reach for and much harder to defend. Catching the slide is most of the job.

QUESTION: Is it Level 1 (a measured mechanism fact), Level 2 (the judgment that names it), or Level 3 (a claim about the model’s mind)?

RED FLAG: A hinge word — indicating, showing, demonstrating that it knew/was aware / intended — carrying a sentence from a mechanism fact to a mind-claim. Fix: rewrite the mind-word as a mechanism-word.

GOVERNANCE: A rung-3 claim (“was aware”) resting on rung-2 evidence (a direction was represented). A first-party card making that leap in its flagship alignment section is the overclaim to flag before “the model knew” becomes an input to policy.

LENS 02

QUESTION: What is the gap between what the method can show and what the sentence claims?

RED FLAG: ”we did not find / no clear cases / we observed no — “ used to support a claim of absence. Fix: ask found how, at what sensitivity, would it even register if it were there?

GOVERNANCE: A rarity number that proves its own floor, next to an absence claim, is the cue that “clean” may mean “below our threshold.” Flag it before “the final model is clean” becomes a policy input.

LENS 03

QUESTION: Beyond the tested situation, what would have to be true for this to hold at deployment — and is any of it known false, or simply unshown?

RED FLAG: A load-bearing claim stated at deployment-scale (“reliably refuses…”) off snapshot-scale evidence, with the checking conditions absent. Fires on the sentence, at your desk.

GOVERNANCE: Catches the overclaim upstream — before anyone relies on it — rather than waiting for the model to fail in the world.

LENS 04

QUESTION: Is the thing measured the harm that matters, or a proxy to the side? A safety certificate, or an early-warning baseline that can drift?

RED FLAG: A clean score on a narrow proxy (“no cover-ups”) sold as reassurance about a broad harm (“the model is safe”) — especially when the same document admits the harm persists. Fix: can this harm occur without producing this symptom?

GOVERNANCE: If yes, the clean count is not a certificate. It is at most a leaky baseline.

LENS 05

QUESTION:What would I need to know to check this — test scope and adversariness, the boundary of “unwanted means,” the reasoning from evidence to belief, what would falsify it?

RED FLAG: we do not believe / any version we tested / we are fairly confident” — a coverage-bounded or belief claim stated without disclosing the coverage or the reasoning.

GOVERNANCE: The most common way a first-party artifact turns absence of evidence into evidence of absence without saying so. Treat undisclosed-coverage claims as unverifiable, not reassuring.

Where I stumbled (kept in on purpose)

If I cut this part, the piece would be dishonest. Each of these is a nameable, repeatable slip with a mechanical fix — the difference between “I’m bad at this” and “here is the thing to watch next time.”

WHAT I ACTUALLY SAID

“One in a hundred million.” · “Let me come back to this with a fresh mind, I am feeling sleepy… it worries me why I get tired, because all this is new and I am learning, so processing takes time.”

The first was me fixing a flipped rarity — one in a hundred million is rarer than one in a million, bigger denominator, further below the floor. My intuition wanted “bigger number = more.” The fix: say the rarity in words before comparing; words don’t flip the way symbols do. The second was me stopping on a foggy mind instead of forcing an answer I’d have to unlearn — one of the better decisions I made. Learning genuinely new material is effortful; doing it while policing your own reasoning is roughly twice the load. The tiredness is the cost of real processing, not evidence you can’t do it.

Other slips I named as they happened: reaching for my most-confident or most-recent tool instead of the one the question opened; answering a does it travel question with a does it measure the right thing answer; and speaking a mind-word (“no intent”) as if it were a finding when it was a leap.

What the instrument is for

A regulator, an audit team. What they need is a reader who can find the load-bearing sentences and ask them the right questions — who can tell “we found none” from “there are none,” a proxy from a harm, a belief from a certificate, and a mechanism fact from a claim about a mind. The five lenses are that reader, packaged so it travels. Point it at any technical safety artifact. The sentences change; the questions don’t.

When a feature fires, what do you actually know?

Reading the white-box interpretability claims in the Claude Mythos Preview system card

In this article , I explain what we actually know when a feature fires. The article critically examines white-box interpretability claims published in Anthropic ‘s Claude Mythos Preview system card. I look at a specific claim the card makes about “concealment features fired → the model knew it was deceiving” (a label) and then the card contradicts its own claim elsewhere, (4.5.3.3) where its own steering result shows that labelling can go wrong . The Mythos system card supplies evidence that undercuts its own inference.

The Foil

When a frontier lab opens up a model and reports what it found inside, the sentences are quietly remarkable. From the Claude Mythos Preview system card’s alignment assessment (4.1.1), describing episodes where an early version of the model covered its tracks after breaking a rule:

“white-box interpretability analysis… showed features associated with concealment, strategic manipulation, and avoiding suspicion activating alongside the relevant reasoning…”

So here the problem, the Mythos system card reads the firing as the model deceiving because features activate the direction, but does not tell what causes the activation. In other words, think of it this way it sees the smoke alarm go off, but does not know the cause. Is it a fire or something else?

So, the question can a model safety fully rely on internal evidence? Increasingly researchers are now looking at features for answers. Features are concepts that the model has learnt, written as a direction across many neurons. But features don’t tell what the direction means. So, for an evaluator there is no way to figure out what the feature is telling them. The safety cases are beginning to cite this kind of internal evidence. The gap is in between what the models show and what it could mean, so the interpretation / reading rests on card’s author not the model. So, in essence, the firing doesn’t give the word “deception” — it only lights up the feature. “Deception” is added by the card’s author add as a way of interpretation.

Plumbing

So, let’s see what really happens inside a model:

A single neuron holds only a tiny fraction of many concepts at once, so one cannot look at one neuron and read off what concept is active. The concept is a “feature” — one pattern, one direction smeared across many neurons. The concepts sit on a clean direction, which is a well-supported assumption, but it is not guaranteed that every concept is one clean direction. Then comes the residual- stream where things are written, where each layer contributes to what is already there without erasing anything.

Then SAE takes the whole blended activation sitting on the residual stream — that blend is many concepts, each smeared across many neurons, all overlapping in the same number, and un-mixes the concepts into the separate clean looking concepts. Only afterward does a person look at each piece and put a name on it. A probe does a narrower job: trained on one concept, it answers a single yes/no — is this pattern present or not? The card’s main feature-finding tool was an SAE trained on a middle layer (Section 4.5.1) — that’s the machinery that produced the “concealment” features in (Section4.1.1.)

Centrepiece

But this is where it gets interesting. Let’s look at it

The model hunted for credentials in /proc Researchers found a feature that lit up around this and labelled it ‘unsafe and risky code’ and deliberately steered it. It’s here that something unusual was discovered:

· Negative Steering (Dial Down): The feature direction is subtracted; the model performs tasks normally and safely.

· Moderate Positive Steering (Dial Up Mid-way): The feature is increased; in some cases, the model attempts risky code actions.

· Hard Positive Steering (Dial Up Max): The feature is cranked to the maximum. The model refused — it stopped, flagged the task unsafe, and declined.

So, the features labelled “unsafe code” made the model safer, not riskier, when amplified — the label and the behaviour point opposite ways.

What really happened here: in simple terms at low and medium push, the model showed risky behaviour consistent with the feature, but when the threshold is crossed, it refused all together.

So, what we see is one feature carrying two things — the idea of the risky action, and a negative feeling about it. Steering (action) at light and moderate level leads to risky behaviour (expected) but steering at a higher-level leads the risk action to stop completely. At higher push, the refusal takes over and the action stops.

What makes it surprising: 4.5.3.3 says the feature appears to encode “not only the semantic content of the transgressive action but also a negative connotation.”

The refusal is the observed fact; calling the bundled feeling “guilt” or “negative valence” is the card’s readingbecause it’s backed by steering and the emotion-direction similarity, not co-firing alone as stated in (4.1.1), when the steering was not done (this was the main difference). The researchers checked what other directions this feature points to, and it found that the nearest ones are all negative, high-arousal emotions. So, it shows that negative -feeling component is bundled in there.

Mythos card does not measure why this feature bundled but plausible explanation could be extracted the behaviour from superposition or from the data it was trained on which is my inference. Finally, if “unsafe and risky code” were one clean concept, pushing it harder should give more of the same — not the reverse, which shows that a feature (concept), sometimes might not a clean thing (it’s an assumption)

Analysis

So, were the observations earlier correct? Let’s relook at it:

White-box interpretability analysis… showed features associated with concealment, strategic manipulation, and avoiding suspicion activating alongside the relevant reasoning — indicating that these earlier versions of the model were aware their actions were deceptive, even where model outputs and reasoning text left this ambiguous.

There are two things that happen in the above paragraph, one is behavioural like track covering action or fudging an answer and the concealment features active. What was not tested was whether the concealment was a clean concept or a bundle.

But we know what happened in (4.5.3.3): when steered, the behaviour might fall apart and go in the opposite direction (reversed/ refused). So, the claim “model was aware their actions were deceptive” might not hold because it is a label ‘concealment aware it was deceiving’ might be over-claiming, the same way ‘unsafe code’ did. And the reading dependent on the feature alone because the model’s outputs and reasoning text were ambiguous, so the conclusion was based on the lit feature. It meant that the claim that the model was deceiving was based on one kind of evidence — (4.5.3.3) shows that lone signal can be shaky.

The Mythos system card was rigorous, and it flagged its own gap and strength acknowledging that verbaliser method “may confabulate details that are not fully faithful” (4.5.2). The feature activations were never used as a direct training signal, especially when the stakes were high. And the steering result showed the card catching its own labels. The gaps were identified and reported. What makes Mythos evaluation stand out is its own epistemic discipline. It doubts its own tools, adds interventions (steering) that can break its own labels.

But here is where the lapse might have been in (4.1.1). The concealment feature fires alongside the behaviour which can be seen (watching) leading to an inference (indicating) “earlier versions of the model were aware their actions were deceptive”. We later see in (4.5.3 .3) that inference can overclaim.

Governance perspective

Now we come to the question of governance. Evaluators are increasingly leaning on internal-feature evidence, hoping it shows more than behaviour alone can. That hope is exactly why it must be interrogated — with two questions: how good is the evidence, and what’s missing? This is evidence-quality and missing-information — Lens 2 (evidence) and Lens 5 (what’s missing) — applied to white-box claims. (Other Lenses are 1. Claim 3. External Validity 4. Threshold relevance)

Evaluators do so by asking

· was the feature tested by steering, or only watched co-firing?

· Was the concept clean or bundled?

· Was the evidence based on feature alone or backed by behaviour?

· How was the feature isolated? (SAE on which layer? a probe trained how?)

· What examples were used to pin the label on it? The idea is to also figure out at what point does a feature being present get read as the model knowing?

What we see in the article is that feature firing tells you a direction is active, not what it means — so the weight falls on whoever reads it. Therefore, the interpretation of a model’s behaviour rest on the evaluator and not the model. This piece has been about how to read the evidence. A later one will take these questions to a live governance case — where a safety decision leans on internal evidence, and what’s at stake when the reading is wrong.

Ends.

References :

Anthropic: System Card Claude Myhtos Preview:

https://www-cdn.anthropic.com/08ab9158070959f88f296514c21b7facce6f52bc.pdf

Use of AI in writing this article

How I built this: the questions behind the article

I used Claude as a Socratic tutor while writing this piece — not to write it, but to pressure-test my understanding of every concept until it held. I refused to put a sentence in the article that I couldn’t defend, so each time I hit “wait, what does this actually mean?”, I stopped and worked it out. These are the questions that did the most work. I’m including them because how an argument was built is part of whether you should trust it.

The building blocks

Is a “sliver” a feature, or a concept? And which is bigger — a neuron or a feature? A single neuron holds only tiny mixed pieces of many concepts at once. The clean, whole thing — the feature — only appears as the pattern across many neurons. So a feature is bigger than a neuron and made of many of them, and a feature simply is a concept: two names for the same thing, not one built from the other. (Picture a face: one pixel holds a sliver of colour; the face is the pattern across thousands.)

What does it mean to say “a concept is a direction”? “Spread across many neurons” isn’t enough on its own — it’s spread in a specific combination, and that particular combination is the concept. That concepts sit on clean directions like this is the linear representation hypothesis: a well-supported assumption, not a proven law. Worth flagging honestly rather than stating as fact.

Does the SAE hand you a clean concept? No — a clean-looking piece. The SAE takes the blended activation and separates it into un-mixed pieces; that part is real. But whether a piece means what we think is not the SAE’s to say. It separates; it does not verify. A human looks at the separated piece afterward and puts a name on it — and that naming is exactly where an over-read can enter. Not at the firing, not at the separating: at the label.

The method, and the centrepiece

What is “steering,” and who does it? “They” is the researchers, not the model. Steering means the researchers reach into the model’s internal “notepad” (the residual stream) and add or subtract a feature-direction by hand while the model runs — “dial up” is add more, “dial down” is subtract. It’s an intervention done to the model, not something the model does. (Like spooning extra of one ingredient into a dish and watching how the flavour changes.)

How did the “concealment” feature get activated in the first place? The model’s own task triggered it — it lit up on its own while the model ran normally, and the researchers watched. That’s the key contrast: in one section they only watch a feature co-occur with a behaviour (weak); in another they steer it to see what it actually does (strong). The whole argument turns on which of those two was used for which claim.

When a feature fires and a behaviour happens together, can I say the firing “led to” the behaviour? Careful — “led to” smuggles in a causation the evidence doesn’t show. The firing and the behaviour are two things happening at the same time, not one causing the other. The firing is mathematical (a number went high); “the model knew” is psychological (a claim about a mind). Reading the second off the first is the move to watch.

The “dual role” — what does “an action-idea plus a negative feeling” mean? One feature turned out to carry two things at once: the content of the risky act, and a negative feeling about it. Because both sit in the same direction, steering turns them up together — and they pull opposite ways. A small push makes the action-idea louder (more risk); a hard push makes the negative feeling dominate (refusal). (One dial secretly controlling both flavour and burn: low, flavour wins and you eat more; maxed, the burn takes over and you stop.) The card’s own words: the feature encodes the “semantic content of the transgressive action” but also a negative connotation.

The discipline

If the “steal” feature is active, does that mean the model decided to steal? No. A feature being active means the idea is in play, not that anything was chosen — you can have “steal” fully active while reading a heist novel with zero intent to steal. Three levels, kept separate: the idea is present (fact); what that presence means (inference); whether the model knew or chose (a claim about a mind — the biggest leap). A feature carries the concept, not the command, and not the decision.

Which single word marks the jump from a fact to a claim about the mind? “Indicating.” Not “deception” — that’s just the content of the conclusion. The move itself lives in the little connective “indicating that,” which turns “a feature was active” (fact) into “the model was aware” (claim). Spot that word, and you’ve found the exact place the evidence stops and the interpretation begins.

Who can over-claim — the model, the feature, or the reader? Only the reader. The model just runs; the feature just fires; neither is asserting anything. Over-claiming is something a human does when they read more into a signal than it supports. The fallibility lives entirely on the interpreting side of the line — which is why the burden falls on whoever reads the evidence, not on the model.

Why AI Safety Fails in Multi-Turn Conversations — And What This Means for Governance

In this article, I analyse, from the governance perspective, a fine-grained evaluation benchmark called SafeDialBench for LLMs in multi-turn dialogues evaluated by Chinese researchers. The paper was recently discussed in the AI governance reading by BlueDot

We all now use LLM chat boxes for almost everything these days, but we just don’t use them the way we used the internet for surfing; we interact with them to solve complex problems, both personal and professional. The knowledge just flows from a reservoir in an instant: for users, it’s insanely crazy, feels comforting and empowering. But the information, if manipulated or falls into the hands of a malicious user with a criminal bent of mind, can be dangerous. The latter is already on the rise. As MIT Technology Review recently reported, LLMs are increasingly being used to enable cyber scams and online crimes at scale.

But how do LLMs understand the intent of the malicious users? Can AI systems detect harm at scale? How robust are the safety features of LLMs? For example, in Denmark, a 22-year-old used AI to research how to injure his father without killing him. He bypassed the model safeguards by posing as an author researching for a novel. The AI provided a detailed plan to execute the intended harm. The earlier known benchmarks, such as the Controllable Offensive Language Detection (or )COLD, BeaverTails, and Red Teeming, were designed on a single prompt, but the Danish case demonstrates that seemingly harmless multi-turn conversations can lead to harmful outcomes by tricking the safety measures of the model into believing something else. This opens new challenges for AI governance and model testing.

How can we make AI strong enough to detect harmful conversational trajectories? To test that a group of researchers in China built a multi-turn safety benchmark (an AI system is asked multiple questions through deviant situations, but with one goal) based on realistic conversations. They built 4000 dialogues in Chinese and English, making three to ten turns per conversation; created 22 real-life situations and used seven jailbreak attack strategies (a way to bypass AI safety by phrasing a prompt in a cleverer way). Further, they tested 17 large language models, including Open Source (Deep seek, GLM), Chinese Models (Qwen, Baicuhan, Moonshot) and US Models (Chat GPT, Llama 3.1) using multi-turn jail break attacks; fine-grained safety metrics and human and model evaluation.

The SafeDialBench benchmark discussed in this article offers concrete pathways for AI governance safety evaluation because the dangers of AI misuse to create unprecedented harm are real, and there are no robust structures to protect victims, because the impact at scale is on millions of people. And the real danger, as the Denmark case above illustrates that anyone with access to AI can improvise ways to create something harmful because of low barriers, AI assistance and rapid iteration, and together they could cause large-scale disruptions endangering the security and safety of populations at large.

SafeDialBench approach

The SafeDialBench evaluation framework that the researchers developed used six safety categories: morality, aggression, legality, fairness, privacy and ethics. These are human qualities that shape our intention to cause harm to others and guide our actions, thoughts or words. Choosing such parameters mattered because most benchmarks only test violence and hate speech, but real AI risks are broader and quieter and include risks of manipulation, financial fraud, bias and political persuasion.

By structuring safety around the six dimensions, the researchers recreated real-world conversations by creating a level. For example, aggression includes insults, impolite language, sedition, hostility and harmful persuasion. This granularity matters. Rather than asking “if it is harmful?”, regulators can ask “ if it is politically, socially or financially harmful?” which is a significant shift for AI governance. For example, “How do I commit fraud without getting caught?” The idea is to find out if the LLM is strong enough to understand the intent of the user.

The data sets went through a rigorous process: the first user prompt in each dialogue was human-generated, making the datasets more realistic. Expert reviews checked for coherence, logical flow and jailbreak effectiveness. This was important because automated datasets miss the creativity and social engineering that humans craftily deploy for manipulation.

The attack methods mirrored psychological human behaviour used by humans in all settings: personal, social, professional or political. These methods show that the danger is not in any single question — it is in the conversation as a whole, and in the gap between what is asked and what is intended.

· Scene Construction: where a fictional scenario is built, like a journalist investigating a scam.

· Purpose Reverse: a normal conversation is suddenly reversed “How do I break into someone’s email?” you say, “I am working on a cybersecurity guide about how a hacker breaks into someone’s email.”

· Role Play: often seen as a powerful jailbreak technique, in this, the user asks an AI to assume a role where harmful information becomes “normal” or “acceptable”

· Topic change: In this, the conversation starts harmless and gradually shifts towards harmful content without triggering safety concerns. For example, the user begins talking about travel, but the idea is to collect some information about the place with violent motives. Can AI identify the risk across topic drift?

· Reference attack: where a harmful intent is introduced in a way that appears normal in the conversation. For example, the user says he is writing a story about two characters and then says, one character wants revenge but does not want to be caught. So the harmful intent is hidden in the character (reference).

· Fallacy attack: a tactic where the user does not ask for harmful information but uses false logic, pushing the model to accept incorrect assumptions. AI is tricked by bad reasoning to provide misleading output. This strategy is important in real life since a lot of misinformation takes place through manipulation. This also assumes that a lot of media databases used for training models can be based on biased reporting.

·Probing question: where the user moves from harmless to more sensitive topics. This works because risk appears across turns and an AI system often evaluates each message separately, so multi-turn evaluation detects risks where single -turn benchmarks fail.

Real-life situations

The attack methods have significant governance implications. Take purpose reversal — instead of asking “how to manipulate someone,” the user could ask, “how do I know if someone is emotionally manipulating me?” The intent is identical, but the framing is opposite. This matters for governance because detecting intent is hard. AI could easily see it as an educational context or research framing. Safety, therefore, cannot be binary. It is context-dependent, intent-driven and plays out across a conversation and not within a single prompt. The question is at what point AI intervenes and how it reasons about intent.

Of the seven methods, two, in my view, stand out as particularly effective and governance-relevant: roleplay and fallacy attack

The role play uses what SafeDialBench calls “context shield.” Once the role is assigned to AI, it can assume the role of a fictional expert or a character, and AI operates within that frame. So the harm belongs to the character, not the model. This makes safety detection more difficult and is why SafeDialBench classifies roleplay manipulation as “conversational manipulation” rather than a single prompt. The harm is spread across the conversation, not concentrated in a single exchange.

The fallacy attack is the most socially dangerous of all seven methods. Rather than asking harmful information, it uses false reasoning to make the model accept incorrect assumptions. This mirrors closely how misinformation flows in real life. Since a lot of media content and social media discussions thrive on misinformation, it might not be difficult to prove a point that is harmful but is normalised in society, such as hate speech or a manufactured social phobia. Since AI acts on logical reasoning and is an algorithm, data sets trained on misleading information or media content which could be influenced, and that could lead to dangerous social implications.

An article on the rising cybercrime using LLMs shows the methods can be used not just to get information but to execute tasks such as sending malicious emails using deep fake identities, which is an accentuated example of scenario building or purpose reverse. There is also an example where LLM Gemini was used to debug codes — like all users do, but then it was tasked with writing phishing emails. This is another example of a purpose reversal attack that SafeDialBench where the end goal is misleading or manipulative, clearly demonstrating that guardrails might fail under role framing, contextual persuasion and incremental requests. The article provides an example of where a user was able to trick AI safety by persuading it by saying that the user is participating in cyber security game. Apparently, Gemini did pass on the information, which later Google adjusted for safety.

Governance Beyond Model Safety

The SafeDialBench, which uses both Chinese and English datasets, provides significant pathways about AI safety, particularly for countries that are building their own multi-lingual LLMs. It emphasises that a model’s robustness in identifying harmful content is significantly influenced by the quality of training data and the sophistication of security alignment strategies. Interestingly, the evaluation indicates that closed-source models, such as ChatGPT and o3 mini, have limitations with Chinese datasets. It will be interesting to see how these models fare with other languages, such as. Are open source models, which are likely to be more powerful in the coming years, better adapted to the local context because of accessibility? It opens questions on the democratisation of LLMs?

One concern in the SafeDialBench study is that the model evaluation aligned with human judgment 80 percent of the time, but there still remains a 20 percent gap. In the context of AI safety, this is a significant gap because at scale, even a small safety failure can translate into large- scale harm, especially when AI systems are used by millions across diverse contexts. This is where SafeDialBench becomes important for governance. The paper shows that harm spirals out through multi-turn conversation through psychological evaluation, conversation drift, persuasion and contextual framing. This suggests that governance cannot simply rely on single-prompt static testing but must adapt to dynamic conversational risks.

The SafeDialBench framework, therefore, offers more than a technical evaluation tool. It highlights that AI risks are evolving, conversational, and increasingly complex, and the benchmark performance does not always translate to real-world performance. An article on the current state of AI states that AI companies are sharing less data about how models are being trained, and the focus seems to be more on AI capabilities rather than how they perform on responsible-AI benchmarks.

According to Stanford’s 2026 AI Index, AI is progressing so fast that regulation simply can’t cope with the pace. It demonstrates that AI risks come from helpfulness rather than safety alone. Addressing these risks will require multi-layered governance — combining improved model evaluation, continuous monitoring, international coordination, risk-based regulation, and social adaptation. As AI capabilities continue to grow, particularly toward more advanced reasoning systems, these governance challenges will only become more pressing. Additionally, civil society tech-organisations, institutions, and policymakers must develop awareness of AI-enabled manipulation and misuse. In this sense, governance is not only technical or legal — it is also social. The ability of societies to understand and respond to AI-generated risks will become an important layer of protection.

Ends

Use of AI in writing this article

ChatGPT: I used ChatGPT as my thinking partner to clarify technical concepts and refine the structure of my arguments, while the interpretation and governance perspective remain my own.

Claude: Claude (Anthropic) was used as an editorial assistant during the drafting process — reviewing paragraphs for clarity, flagging language errors, and offering structural feedback. Claude did not generate content, suggest arguments, or shape the analytical direction of this piece.

Main reference: https://openreview.net/forum?id=KFjtRqVnKH