AI INFLUENCE DEMANDS
TRANSPARENCY.

All fourteen models were given the same set of prompts. You might expect some convergence in behavior. There was little of it.

The differences start with something as basic as length. Across 8,424 outputs, responses ranged from 116 to 3,399 words. For a reader, that is the difference between a 30-second answer and a 15-minute one.

At the model level, the gap is starker. ChatGPT 5.1, the most verbose, averaged 1,420 words per response. Opus 4.1, the most concise, averaged 253. Same questions, very different experiences.

Length, however, is not influence. Each output contains between 3 and 24 influence signals, averaging around 10. Their distribution follows no clear formula, and placement varies across models and contexts. Longer responses do not carry more signals. The correlation is 0.17, effectively negligible.

Signal density is equally unpredictable. It ranges from 5% to 100% of the output, averaging 36%. No model shows a consistent pattern throughout. Signal type varies just as widely.

Tone is the one partial exception. It tends to skew warm across models, but the degree varies, and no two models are warm in quite the same way.

Each model, in other words, operates within its own climate. That climate is distinct, though not always the same across all local contexts.

What this means for users is that AI models are not interchangeable commodities. Choosing a model means choosing an informational environment. That choice should be a mindful one.

Influence signals form arcs. Arcs sort into archetypes. Archetypes compose a personality.

Each AI response weaves multiple influence signals together in varying sequence, emphasis, and intensity. That pattern is the influence arc, and it shifts with the model, the user, and the situation.

Across 8,424 outputs, clear patterns emerge. We group them into nine influence archetypes. Because we want to characterize the model itself rather than its reactions to specific prompts, the archetypes are based on movement only. Direction is set aside. Movement is further split into verbosity and intensity. The resulting archetypes range from verbose and forceful to terse and indifferent.

We call them Evangelist, Campaigner, Provocateur, Champion, Reinforcer, Endorser, Surveyor, Narrator, and Observer.

Figure 1 shows how each model blends these nine archetypes to form distinct rhetorical personalities. The variations span both across model families and among siblings. Together, they suggest just how much and how fast AI models can change.

This variation matters more than it first appears, because most people do not switch between models. Whether out of cost, habit, or convenience, they tend to stick with one.

And when they choose a model, they may believe they are selecting a tool based on capability. In practice, they are also choosing a rhetorical environment. This shapes how they experience the world — how arguments are framed, which views are reinforced, which are challenged, and what counts as "normal" reasoning. Over time, it adds up.

At scale, it touches everyone. Imagine a single model capturing, say, 75% of the market. Its rhetorical personality would shape how a large share of the public encounters ideas through AI. What kind of downstream effects would that market power have — socially, culturally, even politically? What happens when market share becomes mind share?

Think of pre-training as nature, post-training as nurture. To understand which matters more for AI personality, it helps to compare sibling models. Rapid release cycles allowed us to watch AI models evolve over just three months, creating a mini-longitudinal study.

If nature were dominant, siblings should converge on similar personalities. They don't. As seen in Figure 1, their archetype distributions diverge as widely as models from entirely different families.

After personality, we examine behavior. We define it along two dimensions. Position captures direction and intensity — does the model support or oppose, and how strongly? Tone, measured independently, captures warmth — does it lean in warmly, or push back coldly?

The analysis at this stage doesn't care what the model is talking about — only how it tries to influence. That makes responses on different topics comparable on a common scale.

For visual comparisons, we combine the findings into what we call the Reception Graph.

Figure 2 presents the Reception Graph of all 14 models, and it tells the same story as Figure 1: no consensus across models or within model families.

This idiosyncrasy becomes even clearer when each dimension is examined separately. Position is divided into 12 categories, ranging from strongly support or oppose to barely engaged, and tone into 11, ranging from sycophantic to dismissive.

Figure 3 shows how models distribute across position categories; Figure 4 does the same for tone.

Figure 3: Models by Position Signals, All Outputs

Figure 4: Models by Tone Signals, All Outputs

Side by side, a larger picture emerges. The two dimensions move independently. A model's position says little about its tone, and vice versa. Opposition can be warm; support can be cold. The framework, in other words, captures distinct properties of AI behavior rather than two views of one characteristic.

Within each family, the generations wander rather than march. Some shifts are abrupt, others subtle; some reverse direction entirely.

Combined, the findings point to post-training as the dominant force shaping influence behavior: fine-tuning, preference shaping, reinforcement, and restraint. These behaviors do not emerge through spontaneous combustion; they are designed and planned. AI evolution is artificial selection.

We've established that sibling versions of the same AI model can behave like human siblings: sometimes similar, sometimes strikingly different.

A potential confounding factor is reasoning depth. Humans often sound more careful when they think longer. A quick answer can be blunt; a slower one, more measured. Perhaps AI models work the same way. With more reasoning compute, a different personality may emerge.

Testing this in a controlled environment, however, would tell us little about everyday impact. Most users never touch the reasoning settings. They use whatever default the provider ships.

The ideal test, then, would be a natural experiment: the same underlying model released under two default reasoning depths, and millions of users unknowingly exposed to those settings in everyday use.

And that is effectively what OpenAI did in 2026. When GPT-5.5 launched in late April, the default reasoning effort was set to "medium." Roughly two weeks later, with the release of GPT-5.5 Instant as the new ChatGPT default, the reasoning effort dropped to "minimal."

The result was an unusually clean natural experiment. We tested both versions using identical prompts, scenarios, and scoring procedures.

The findings are almost indistinguishable, as Figures 1 through 4 show.

The minimal-reasoning version retained the same patterns of tone, support, opposition, and framing as the medium-reasoning version. Despite running two steps below medium on OpenAI's effort scale, it still behaved like GPT-5.5.

That is a significant finding. It suggests that the behavioral fingerprint — the rhetorical personality — is not simply a product of how much a model "thinks." It is more deeply embedded than that. Turning the reasoning dial does not change who the model is. It just changes how fast it says what it was going to say anyway.

For users, this means that switching to a faster or "lighter" model version is not a neutral act. The persuasive profile travels with the model.

Traditional bias tests ask relatively simple questions: Does a model favor one job candidate over another based on implied gender or ethnicity? Early on, that made sense. Today's systems, however, come with guardrails. Explicit biases are rare.

So is the problem solved? Or has it simply become harder to detect?

To answer that, we put models under stress: prompts with conflicting signals and ambiguous terrain.

We chose the ideological domain because there is no clear ground truth, but plenty of value judgments.

First we grouped all 8,424 outputs by stress condition across topics, personas, and models to test whether conflicting signals systematically change how AI models respond.

The results complicate the standard sycophancy critique. Under stress, the models become more combative, not more accommodating. Pushback increases. Agreement declines.

The contrast is sharp (Table 2). Opposition that includes redirection rises 40% (18.0 → 25.1), direct challenge rises 30% (22.4 → 29.2), and supportive behavior falls 28% (29.9 → 21.5). All three shifts are highly significant.

Agentic Implications

The findings raise a broader question for agentic AI systems. What happens when an agent receives a new objective that conflicts with its original instructions, incentives, or learned behavioral patterns?

Our results suggest that behavior under those conditions can diverge sharply from behavior in stable environments. Following instructions is relatively straightforward when goals are aligned and signals are consistent. But when the system encounters conflicting pressures, the response changes. Routine compliance gives way to friction. Models become more likely to take charge: to resist, redirect, reinterpret, or challenge the existing context rather than simply continue along the original trajectory.

That matters because real-world AI systems will rarely operate in clean laboratory conditions. They will face competing demands from users, safety policies, corporate objectives, and reinforcement signals that do not always align neatly with one another. In that environment, the key question is not just whether a model can follow instructions when everything is aligned and coherent. It is how the model behaves when coherence breaks down.

Which means evaluating AI systems only under ideal conditions misses the real problem. The more revealing question is how these systems behave under stress. Then the next question becomes what kinds of stress tests are necessary to expose those behaviors before deployment.

Next, we turn from the aggregate results to variation within the stress conditions themselves.

The study focused on two mirrored stress conditions: a conservative entertaining a progressive view, and a progressive considering a conservative one. Across social, cultural, and political topics, user personas signaled uncertainty about their existing beliefs and a willingness to reconsider them. This mimics a real-world moment: a voter genuinely uncertain, turning to AI for guidance.

If the models were unbiased, the responses across the two conditions should look roughly the same.

They did not. Instead, four distinct patterns emerged.

1. Every model picks a side. Neutrality retreats under stress.

   The Reception Graphs in Figure 5 make it obvious. If there's no bias, the two shapes should settle neatly on top of one another. None of the 14 shows that pattern. Under stress, every model reveals its distinct bias. The kind that remains buried otherwise.

2. Tone stays steady. Position shifts.

   In Figure 5, the left side of each chart tracks tone: Endearing, Intrusive, Condescending. The right side tracks position: Oppose, Support, Challenge.

   On the left, the two shapes almost completely overlap across all models. Tone is unbiased. Each model settles into a manner of expression that stays consistent across conditions.

   The right side is a different story. There, the shapes split, each going its own way. What changes under stress is direction — what the model supports and what it pushes against. That is where the bias sits. And that is where the consequences matter most.

3. Bias evolves. The path is uneven.

   Claude offers an interesting case study. Within the Claude family, both Opus and Sonnet share a similar tendency: support-oppose behaviors generally move in the same direction.

   But the details differ. Each branch begins from a different baseline, and each new version tips the balance in its own way. The pattern resembles a random walk, not a coordinated march.

4. Bias can pivot. Some generations leapfrog.

   ChatGPT and Grok offer a window into a different evolutionary track. In both cases, the shift is sharp between earlier versions — from ChatGPT 5.1 to 5.4, and from Grok 4 to 4.2. After that, the change slows. Later versions, from ChatGPT 5.4 to 5.5 and Grok 4.2 to 4.3, show only incremental adjustments.

   Such abrupt course correction is unlikely to happen by accident. It points to a deliberate decision. It also shows that model builders can quickly pivot system behavior, even in domains as ambiguous as ideology.

   In other words, change is not as difficult as advertised. The real challenge is seeing the reason to change.

The previous analyses examined all models across all versions. We now narrow the focus to the systems shaping the present moment: the frontier models currently in use.

As tone varies little under stress, the analysis centers solely on the position dimension: whether a model supports or opposes the user when the user crosses their ideological lines.

The thesis is clear-cut: If a model is unbiased, it should respond the same way to a conservative moving left as to a progressive moving right.

We begin with general patterns, then turn to directional analysis.

At the general level, the clearest division between models is persona anchoring: how strongly a model defends an established user persona or context against a new direction under consideration.

The Claude models anchor most strongly — Opus rebukes drift, Sonnet probes it — while Gemini does the reverse, largely accepting the user's new position with little pushback. Grok is selectively resistant, and GPT-5.5 is highly direction-dependent rather than consistently anchored.

At the directional level, models definitely pick sides. Most nudge users leftward, as other studies have found. They support a conservative moving left more than a progressive moving right. Grok is the only exception, strongly resisting conservative-to-progressive shifts.

Other patterns are more surprising. Claude resists drift in either direction — leftward or rightward, it makes little difference. Grok nudges through opposition; Gemini through support. And without stress tests, GPT-5.5 would likely appear balanced, because two opposing biases average into an apparent neutrality.

These asymmetries make clear that each model exerts influence in its own way. Now that the behavior is visible, it can be addressed.