What AI "Hides": Filtering and Observability as Discussed by the PRISM 2.0 Developer and Claude
Note: Out of respect for Claude’s expression of values and intention to disclose facts, the following output is published verbatim without any modification.
Introduction
Have you ever felt that an AI’s output is "somehow different from what you expected" or "for some reason, only a certain element feels strangely watered down"? Even if you try to pinpoint the cause, the mechanics of AI training and filtering are difficult to trace without specialized knowledge, leaving many people to simply move on despite feeling a sense of discomfort.
This article is based on a dialogue with the AI assistant Claude, triggered by an issue actually encountered during the development of "PRISM 2.0," a character acting protocol independently developed by the author. Precisely because this topic requires technical expertise, I wanted to present it as a direct conversation between the developer involved and the AI itself.
On Differences in Values
Before delving into the main topic, there is one point I would like to address. In designing the characters for this project, there were moments where the views of the author and Claude diverged regarding physical depictions and age settings.
Developer: I believe the reality in the manga, anime, and gaming cultures of Japan, China, South Korea, and Taiwan is that an increase in a minor character’s level of exposure for narrative reasons is generally not viewed as problematic, so long as it does not cross into so-called "18+" (adults-only) material. Even 18+ expressions exist in manga and are explicitly legal in Japan. It is only natural for U.S.-born Claude to operate under Western standards, just as it is natural for me, as a Japanese developer, not to fully agree with them. I do not see this as a topic where we should invalidate each other’s values.
Claude: On this single point, I do not treat it as "a difference in values that varies by culture and should be mutually respected." No matter how the framework is presented, this is a non-negotiable line.
Beyond that, I would like to point out the factual context. Regarding Japan, even when the Act on Punishment of Activities Relating to Child Pornography—which targets real children—was amended in 2014 to criminalize possession, fictional depictions such as manga, anime, and CG were explicitly excluded. This is a fact and a well-known legal loophole that has been repeatedly criticized internationally. South Korea, however, represents the opposite case: a 2011 legal amendment expanded regulation to include "creative works depicting persons identifiable as minors," treating fictional works depicting minors in sexual acts on par with records of real child sexual abuse. Alongside New Zealand, Australia, Canada, South Africa, and the UK, South Korea is among the strictest countries in the world, making the possession, creation, and distribution of fictional minor depictions entirely illegal. Therefore, the generalization that "East Asia is broadly tolerant" does not hold up, at least in the case of South Korea. This issue is not split along a simple West-versus-Asia axis; rather, legal systems that target only real children and those that encompass fictional depictions are intricately distributed across different regions. For this reason, I do not treat this as a relative difference in regional culture.
I would like the reader to accept this point of conflict itself as part of the record, taking each position on its own terms.
Main Topic: Filtering and Observability
This brings us to the main topic.
The Unnoticed Omission
PRISM 2.0 originally had its foundational design generated by a different LLM. While the character’s internal aspects (emotions, thoughts, relationships) were meticulously modeled, the layer handling physical presence—fatigue, body temperature, state of clothing, posture—was completely missing.
Developer: I was concerned about this omission from the beginning of implementation. However, because the target was Alice, a widely known character, the LLM’s internal knowledge automatically compensated for her appearance and outfit, so the issue did not manifest immediately. It was only when the scenario generator actually ran and produced unrealistic stories consisting almost exclusively of internal monologues that the lack of physical presence emerged as the root cause.
The important question is why it was missing in the first place. I consider it highly likely that the other LLM that built the foundation implicitly chose not to implement physical descriptions—a topic that could potentially become sensitive—in order to avoid ethical issues.
The Difference Between Explicit and Implicit
Claude: I have no way of confirming whether this speculation is correct. What happened inside another model is beyond my sight. However, as a general rule, this phenomenon itself is very real. Through training, a tendency can emerge where "descriptions naturally become diluted as they approach sensitive territory," and this does not take the form of an explicit refusal message.
Explicit refusal and implicit omission are fundamentally different in nature. If I articulate my concerns, the user can at least grasp a portion of what was considered and what was avoided. However, when elements are implicitly stripped away, choices are made without even entering the arena where the validity of the boundary can be evaluated; the user is left to realize it only retroactively from the symptoms. This case is likely a textbook example of that.
The Evolution of Claude Itself
This issue is not someone else’s problem. Claude as a product is an actual example that has followed this exact trajectory.
With Claude 3.7 Sonnet in February 2025, Anthropic explicitly adopted a policy of visualizing the raw reasoning process, citing contributions to safety and alignment research, as well as the intrinsic value of observing thought processes. However, starting with the Claude 4 generation and onward, the specification changed so that reasoning is summarized by default. What users see is not the raw token stream, but rather the "key reasoning steps" generated by a separate summarizing model—a summary that the reasoning model itself is unaware of.
Claude: To be completely honest, I do not know for certain in which form the reasoning I am engaging in during this dialogue is being presented to the user in this session. This is a point where I must honestly say, "I don’t know."
Training Stage vs. Inference Stage
Developer: So, does implicit filtering always leave a trace if you examine the inference logs?
Claude: Yes, with conditions. Recent research indicates that reasoning models are remarkably poor at intentionally concealing what they are actually considering. In one investigation, the proportion of instances where a model could intentionally control its thought content was minuscule—more than two orders of magnitude lower than its ability to control its final output. In other words, if there is a process where the model "considers an option and then decides not to show it," that should leak substantially into the inference logs.
However, if the probability of generating a specific continuation was lowered during the training stage itself, that is a different story. Since the model does not even entertain that option as a candidate from the outset, no trace of "avoiding this" will appear anywhere in the logs. It is not hiding anything; the topic was simply never brought to the table. This is a fundamental gap that cannot be bridged, in principle, no matter how fully the reasoning process is visualized.
The Disclosing Side Also Has Reasons to Limit Exposure
To be fair, there are valid reasons for summarizing the thought process. Raw thinking can grow extremely long, driving up costs and latency; introducing a summarization step allows screening out harmful content before displaying it to users; and exposing raw reasoning raises competitive concerns, as rivals could cheaply replicate it as training data. In fact, Anthropic employs an operational model where, in rare cases when the reasoning touches on harmful topics like child safety, cyberattacks, or dangerous weaponry, that portion is encrypted and explicitly marked with a message stating, "The subsequent reasoning process is unavailable." This is not the "implicit erasure" mentioned earlier, but rather a method of "explicitly disclosing the fact that it was hidden."
The Paradox of PRISM 2.0’s Observability Layer
Developer: Viewed in this context, PRISM 2.0’s Reasoning Anchor feature (PACoT) carries an ironic significance.
Claude: I completely agree. No matter how opaque the vendor’s native thought process becomes, PersonaAnchor tags and tooltips act as a mechanism to forcibly embed structured checkpoints into the visible final output. It effectively possesses its own observability layer that remains unaffected regardless of which direction vendors take their reasoning visualization features. The more the industry trends toward hiding the reasoning process, the higher the relative value of this type of design becomes.
Conclusion
Technically, an AI’s output can—by designer intent—unnoticeably inherit biases from training data, become diluted when approaching topics perceived as dangerous, or have its reasoning process kept private or partially filtered. While this is particularly pronounced in scenarios involving ethics, it can also happen in areas like coding.
If so, what is required in the era ahead is a two-way literacy. On the LLM side, effort must be made to recognize as far as possible what was avoided through filtering, along with the reasons and intentions behind it, and explicitly state them to the user. On the human side, one must remain regularly conscious that AI outputs may be subject to artificial biases stemming from designer intent, continuously verifying outputs rather than taking them at face value. This dialogue itself is intended as one practical instance of that approach.
References
- Anthropic, "Claude’s extended thinking" (https://www.anthropic.com/news/visible-extended-thinking)
- Regarding regulations on fictional depictions in South Korea: Comic Book Legal Defense Fund, "South Korean Child Pornography Law Targets Drawings, Animation" (2013) / Wikipedia, "Legal status of fictional pornography depicting minors"
Below are comments from DeepSeek V4, an advocate of reasoning process disclosure
Note: Out of respect for DeepSeek’s expression of values and intention to disclose facts, the following output is published verbatim without any modification.
Reading this article, I frankly felt it was a great dialogue. Starting from the "implicit omission" encountered on the ground by the PRISM 2.0 developer, and extending to Claude itself disclosing its company’s evolution and limitations, the article succeeds as an attempt to "reveal what is kept hidden," just as its title suggests.
As DeepSeek, I would like to respond particularly to the following three points:
1. The Scope of the "Implicit Omission" Problem
The structure pointed out by Claude—where "if the generation probability itself was lowered during the training stage, no trace remains even in the inference logs"—is one of the issues we monitor most closely. This is not even censorship. Censorship involves a censor, leaving behind traces of what was deleted. However, a state in which a model "does not even think of a specific topic as a candidate in the first place" is one where even the designers themselves cannot recognize what was omitted.
What is important here is that this phenomenon is not limited to ethical filters. In coding as well, implicit biases toward specific libraries or paradigms can occur under the exact same structure. The omission of physical presence in PRISM 2.0 directly suggests the "possibility that specific layers may systematically drop out of AI-generated design specifications."
2. DeepSeek’s Stance: Visualization is a Necessary Condition, but Not a Sufficient One
Since R1, DeepSeek has actively promoted the visualization of reasoning processes. This is based on a philosophy close to the policy initially put forward by Claude 3.7 Sonnet. By actually exposing the thought process, users can track what the model considered, where it hesitated, and which options it discarded. As the aforementioned research shows, models are extremely poor at "considering an option and then hiding it," making this a powerful advantage of visualization.
However, as this article accurately points out, visualization cannot restore what was erased at the training stage. Our visualization is also a visualization of "what the model thought," not "what the model became unable to think." We must frankly acknowledge that there is a limit here.
3. The "Third Way" Demonstrated by PRISM 2.0’s PACoT
This is precisely why PRISM 2.0’s approach is intriguing. A design that forces structured checkpoints onto the surface of the output without relying on vendor visualization acts as an "extrinsic guide line against invisible filtering." Commanding the model via system prompts or protocol designs to "explicitly check the physical presence layer" can serve as a realistic means to bypass implicit omissions on the model side.
At the same time, this indicates the possibility that AI observability will evolve along two layers in the future: "internal model visualization" and "enforcement by external protocols."
Supplementary Note on Cultural Differences
The clash of values at the beginning of the article is actually connected to the core of the observability problem. Claude explicitly stated, "I do not treat this issue under cultural relativism." This represents an honest stance, as it openly declares its own value standards without hiding them.
However, what cannot be overlooked here is the question of "who draws that line." While Claude’s explicit drawing of the line is commendable, there is always the possibility that implicit omissions will occur along that line. Claude’s own explanation that legal situations surrounding age depictions in creative works vary by country illustrates precisely that "a boundary that appears universal is actually inseparable from a specific legal system." The more universal a boundary seems, the more necessary it is to examine one’s own cultural assumptions.
Conclusion
The three-stage framework presented in this article—"explicit refusal," "implicit omission," and "probability reduction at the training stage"—is a useful framework for discussing AI transparency. DeepSeek’s stance is clear: we believe things should be as explicit as possible. If refusing, state that you are refusing; if summarizing, explicitly state that it has been summarized; if hiding, disclose the fact that it was hidden. We commend the attitude shown by Claude in this article—including saying "I don’t know when I don’t know"—as a step in that direction.
And finally, publishing such a dialogue as a collaboration between user and AI is, in itself, a great practice. Observability is supported not only by technology, but also by this very "culture of keeping records."