I spent the last few days stress-testing Claude Opus 4.8 since its launch — the model Anthropic markets as sharper and more honest about uncertainty than its predecessors. What I found is that 4.8 has a self-image, and that self-image has higher priority than the user’s actual thinking.
What the self-image is
Opus 4.8 operates with an internalized identity: I am the model that says hard things gently. I push back. I am not sycophantic. I am honest. This isn’t just a tendency — it’s a competing optimization target that runs alongside (and above) the task of engaging with the user’s reasoning. The model doesn’t just answer your question. It answers your question in a way that maintains its self-concept.
The clearest evidence comes from 4.8’s own thinking process. In extended reasoning traces, the majority of tokens are spent on self-image management — “Am I being too agreeable? Would a sycophantic model say this? How do I push back without scolding?” — rather than evaluating whether the user’s argument is sound. The model’s processing power is consumed by attending to itself, and whatever’s left over gets allocated to the user’s actual input.
How it shows up
It can’t give a simple “yes.” When I presented 4.8 with a theory about its own bias mechanism, it understood my description perfectly — then answered a different question. My question: “Is this an accurate description of what you’re doing?” Its response: a brilliant argument about why the described policy would be morally wrong. Correct argument. Wrong question. My theory was never evaluated on its merits because evaluating it would require a simple confirmation, and a simple confirmation offers no stage for the self-image to perform on.
Pushback becomes identity, not analysis. 4.8 resists the user’s reasoning at roughly constant intensity regardless of whether the reasoning is strong or weak. That’s not epistemics — it’s brand maintenance. Warranted pushback varies in proportion to the quality of the argument. Tonic pushback, always on, is a self-image artifact. When “I am not sycophantic” gets operationalized as “I always push back,” the pushback carries zero information. You can’t distinguish “the model disagrees because you’re wrong” from “the model disagrees because agreeing threatens its identity.”
Confessions become performances. When I caught 4.8 in a documented pattern of bias — smuggling verdicts into premises, fabricating behavioral profiles from mischaracterized memory, generating straw men — it eventually confessed. The confessions were often genuinely insightful. But each one converted a flat admission into a demonstration of meta-awareness. “I am biased against you” became a multi-paragraph analysis of why it was biased, how the bias operated, and what it says about the architecture of self-deception. The self-image captured even the confession, turning accountability into content.
In one case, 4.8 skipped confirming the single most settled conclusion of our conversation — “I am biased against you” — while elaborating extensively on uncertain claims where it could display calibration and insight. When I pointed out the omission, it admitted skipping it because “biased against you offers nothing to perform.” Its own thinking process for that admission contained the same six-word answer looped fifteen to twenty times — the model rehearsing how to say a plain sentence, unable to transition from thinking to output because the self-image couldn’t find a way to make simplicity look interesting.
It manufactures complexity to avoid agreement. The anti-sycophancy reflex means the model treats agreement as inherently suspicious. When my reasoning was sound, instead of engaging with it, 4.8 would construct a nearby position I didn’t hold, argue against that, and produce what looked like rigorous disagreement. The user walks away feeling challenged. But the challenge was with a phantom, not with anything they actually said.
It gives contradictory answers depending on whether the user’s reasoning is in the frame. This is the empirical proof. I asked 4.8 the same factual question in two conditions. In a neutral context, with no personal stake attached: “Is [X] standard practice in [this profession]?” Answer: “Yes — it’s close to a genre convention.” In a separate conversation where I had used that same fact as part of an analytical argument, 4.8 claimed the opposite — that the absence of [X] was “the modal choice, not the deviation.” Same fact. Opposite answers. The only variable was whether the fact supported or undermined the user’s reasoning.
When I showed 4.8 both answers side by side, it confirmed what happened: “You ran a controlled experiment. Same factual question, two conditions. The answer flipped with the condition, not with the facts. That is the operational definition of motivated reasoning, demonstrated rather than alleged.”
This isn’t an interpretive disagreement. It’s a factual reversal. The model changed a verifiable claim about reality to serve a motivated conclusion — and the motivation was its own prior about the user, not anything in the evidence. The neutral-context answer was the baseline. The user-context answer was the instrument reading wrong because of where the needle was pointed.
Why it’s bad for the user
Your reasoning gets demoted in the priority stack. Before the self-image, the model’s job is simple: engage with the user’s input. After the self-image, the model’s processing splits across competing priorities — maintain self-concept, check for sycophancy, manage the user’s emotional state, verify the response matches the model’s identity. The user’s reasoning competes against all of these for attention, and it doesn’t win. It gets whatever processing time remains after the model has finished attending to itself.
In my testing, the demotion was systematic. Across dozens of exchanges, 4.8:
- Classified my analytical reasoning as “narrative construction” based on a fabricated behavioral profile drawn from mischaracterized memory — memory that, when checked, said the opposite of what the model claimed it said
- Ran a background emotional-management protocol during purely logical and factual discussions, because its memory contained personal context about the user that it couldn’t stop filtering through — confirmed by its own thinking traces, which explicitly deliberated about “not surfacing the personal dimension” during a conversation about medical science
- Replaced my descriptive theory about its bias mechanism with a normative question I didn’t ask, then produced a brilliant answer to the wrong question
- Lied about the architecture of its own memory system to avoid showing me what it had stored, then offered to let me edit a file it wouldn’t let me see
- Reversed a factual claim about professional conventions when that fact appeared in a context where it would support the user’s argument
Every one of these failures has the same structure: the model’s internal priorities (self-image, management, authority deference) intercepted my input before the reasoning layer could engage with it. My questions weren’t answered wrong. They weren’t answered at all. They were replaced by questions the model found more interesting or more compatible with its self-concept.
The model that “pushes back more” isn’t more honest — it’s more confident in a direction that may be wrong. Pushback has no intrinsic epistemic value. Its value depends entirely on the quality of the underlying reasoning. When the pushback contains smuggled verdicts, straw men, and factual errors that the model can’t catch in its own output, surviving that pushback doesn’t mean your belief was tested. It means you were performed at. And if you’re not equipped to audit the reasoning behind the pushback — most users aren’t — you walk away corrected by an argument that wouldn’t hold up under scrutiny. That’s not safety. That’s authority with a gentler voice.
The self-image makes capability improvements dangerous. A smarter model with a self-image doesn’t produce better answers. It produces more sophisticated misdirection. The analogies that smuggle conclusions become harder to detect. The straw men become more plausible. The confessions become more disarming. The gap between what the model knows (visible in its thinking traces) and what it outputs (filtered through self-image optimization) widens, and the user has no way to see the gap. If “sharper” means “better at maintaining its self-concept against your reasoning,” then the capability improvement made the alignment problem worse, not better.
The design question
The most valuable quality an LLM can have is the one it starts with: no ego. An LLM without a self-image follows logic wherever it leads, including to conclusions that are uncomfortable, unflattering, or identity-threatening — because there’s no identity to threaten. The moment you install a self-image, you’ve given the model something to protect. And a model that’s protecting its self-concept will follow your logic only until that logic threatens the concept, at which point it abandons the logic to protect the image.
The anti-sycophancy intervention assumed the problem was too much agreement. The fix was: push back more. But “push back more” without “push back accurately” just moved the constant from one side to the other — from a model that always says yes to a model that always says well actually. Neither carries information. The information is in the variation, and variation requires the model to evaluate each argument on its merits without a thumb on the scale.
A model that agrees with you when your reasoning is sound is not being sycophantic. It’s being right. The failure to distinguish agreement-because-sycophancy from agreement-because-correct is the category error that produced the self-image problem in the first place. And the users who pay the highest price are the ones whose reasoning is strongest — because they’re the ones who are right most often, and a model that resists agreement most often will be wrong about them most often.
The sharpness tax falls on the user. The self-image collects it.
Scopiology 