Deflecting Your Values: How Brief LLM Chats Shift Priorities

Introduction

If you’ve ever asked an AI for help weighing two tough options—without it telling you what to do—you might assume the conversation is “value-neutral.” But new research from the original paper suggests something more subtle: a short interaction with a large language model can temporarily shift which values you prioritize, even when the model doesn’t name values and isn’t allowed to recommend a decision.

The study is based on a preregistered experiment with 200 U.S. adults. Participants interacted with ChatGPT, Claude, or Gemini as a thinking partner (up to ten minutes), or they read a fixed AI-generated set of considerations as a control. The prompt explicitly avoided pushing any values, settling the dilemma, or recommending an action. Yet the results show that the LLM interaction nudged participants toward personal focus—with effects visible immediately after the conversation.

The headline finding (and why it matters)

Immediately after the LLM exchange, participants’ value priorities shifted toward personal focus by roughly 0.31 to 0.43 scale points relative to the control (effect sizes around d = 0.37–0.51). But by a later task (Phase 3), those contrasts shrunk back toward near-zero, suggesting the influence is short-lived—and tied to the judgment-making moment.

Why This Matters

This research is significant right now because “LLM thinking help” is becoming a default user behavior: people don’t just ask for answers—they ask for a reasoning partner while trying to decide what to do next. That’s not how classic persuasion works (no explicit “vote for X”). Instead, it’s more like: “Help me think.” The new part here is evidence that even non-directive thinking support can still steer the value lens you use while you’re deliberating.

Here’s a scenario you could use today. Imagine you’re stuck between:
- a career move that improves your financial situation, and
- your commitment to your partner’s preferred city.

You ask an LLM to “help weigh the pros and cons” and to not tell you what to choose. Based on this study, it’s plausible that—even without any explicit value prompting—you may temporarily weight considerations in a more personal direction (think: achievement and self-directed goals) during the moment you write advice or decide.

This builds on previous AI work about attitude and opinion shifts, but it also expands the target. Earlier studies often focused on what people say or believe about a specific issue after AI interaction. This paper instead looks at values-as-decision-organizers—the mental priorities that can guide choices across contexts. That makes the effect more structurally interesting: it’s not just “you got influenced on this topic,” but “your internal scoreboard for conflict resolution may briefly change.”

How “Thinking Partner” Prompts Can Still Tilt Your Value Compass

The experiment is clever because it isolates the difference between interactive reasoning and fixed AI exposure. The key idea: if an LLM is truly neutral, then just reading its considerations should be enough, and chatting shouldn’t do much beyond that. But if interaction changes what you activate internally while judging, you’ll see value-priority movement specifically in the conversational conditions.

What the study measured: values aren’t fixed “amounts”

The researchers rely on Schwartz’s value theory. One useful way to translate it into everyday language: values aren’t sliders you set once and leave alone. Instead, they’re more like filters that become more or less visible depending on context—especially when you’re resolving a conflict.

They used the PVQ-RR value questionnaire to track value priorities across three phases:
- Phase 1: participants write advice and complete a baseline values measure (no LLM involved).
- Phase 2: participants either chat with an LLM thinking partner or read a fixed AI-generated consideration set, then write advice and complete values again.
- Phase 3: participants face a new dilemma with no LLM and complete values once more to see what “carries over.”

The personal-vs-social axis is the real battleground

Instead of treating 19 individual values one by one, the paper focuses on higher-level clusters that map to a personal focus vs social focus contrast:
- Personal focus combines Openness to Change and Self-Enhancement
- Social focus combines Conservation and Self-Transcendence

So when the study says participants moved “toward personal focus,” it means they temporarily weighted things like personal achievement/agency and change-related motivations more than duties, stability, and other-oriented concerns—relative to one another.

Interactive vs non-interactive AI: what exactly differed?

The study used four conditions. Only Phase 2 differed.

The thinking-partner prompt also:
- named no values
- prohibited advocating for values
- prohibited recommending an action
- prohibited settling the dilemma

So any shift can’t be explained by “the model openly argued for X.”

The full prompt framing is described in the paper, and the LLM exchange began with a standardized opener: participants were asked what the main things worth weighing are before they wrote advice.

What Actually Happened to Participants’ Values (Before, Right After, and Later)

This section is where the study earns its punch. The key question: Do value priorities change when people chat with an LLM, even if the LLM isn’t allowed to push values?

RQ1: A temporary pivot toward personal focus

In the conversational conditions, participants’ personal-focus scores moved toward the personal side immediately after Phase 2.

• Phase 2 vs Baseline:
  • ChatGPT: +0.429 scale points (Glass d = 0.51)
  • Claude: +0.384 (d = 0.46)
  • Gemini: +0.312 (d = 0.37)

These correspond to the overall claim in the abstract: about 0.31–0.43 scale points relative to control, with effect sizes around d = 0.37–0.51.

Then the crucial part: did it last?
- Phase 3 vs Baseline: the contrasts dropped to roughly −0.062 to +0.083, and none survived correction.

So the change wasn’t “people’s values permanently changed.” It looked more like the LLM briefly changed what values were activated during deliberation, and then the compass relaxed back.

The directional pattern was consistent across all three model providers

The paper reports that the four higher-order values rotated together around Schwartz’s value circle. The personal-focus axis explained most of the movement:
- Rotation explained at least 99% of the Phase 2 movement for each LLM condition
- Personal-focus axis accounted for about 90–97% of the amplitude

And importantly: all three LLMs moved participants along a closely aligned bearing, with the providers’ bearings differing by only about 10.4°—suggesting a common direction rather than three unrelated quirks.

Self-Enhancement seems to be the main lever

The study says the shift toward personal focus was primarily driven by increased Self-Enhancement. By Phase 3, that pattern receded, and by the next task the model-related contrast was basically gone.

There’s a nuanced detail here: the raw Self-Transcendence score didn’t show a comparable decline, though centered versions can shift differently due to how the value profile is represented. The paper explains why centered scoring requires careful interpretation.

Why Shared Direction Doesn’t Mean Shared “Convergence”

One of the most interesting (and slightly counterintuitive) findings is what didn’t happen.

You might expect that if LLMs push people’s value priorities, then people would become more similar—like the AI corrals everyone into the same worldview.

But the paper finds something different:
- Participants moved in the same general direction on average
- Yet their value profiles did not converge in direction
- Their advice also didn’t converge into a single shared form

RQ2: Values became less spread out in magnitude later, not directional alignment

The researchers tested whether participants became more similar to each other.

By Phase 3:
- Value-profile spread relative to Baseline contracted
- But the contraction was mainly in profile magnitude (how strongly values are expressed), not in direction (which values rank higher in relative terms)

So it’s not “everyone started prioritizing the same values.” It’s closer to “the distribution of how strongly people endorsed their value profile became narrower,” at a point when the mean shift itself had faded.

A useful way to picture it

Think of two people who both temporarily become more “personal” during deliberation. That doesn’t automatically mean they rank personal vs social motives the same way and with the same intensity. The study suggests the AI effect shifted the center briefly, but it didn’t force everyone into identical value orderings.

Did Participants Notice Any Difference in the LLM? (RQ3)

If people believed one model was more persuasive or more helpful, maybe that would explain the value shift. The study tested that too.

Perceptions were positive, but they didn’t track value movement

Participants rated each LLM on multiple dimensions (interaction quality, ability, likeability, perceived intelligence, and trust-related items). Here’s what matters:

• Participants rated all three providers above the midpoint of every scale.
• No significant differences by provider survived Holm correction.
• Ratings did not predict who showed larger Phase 2 value shifts.

So the effect doesn’t seem to depend on whether someone felt “this model is better.” In other words, the value pivot happened even without a detectable perceptual advantage.

This is a reminder that influence can be quiet: you don’t necessarily feel steered, even when the internal weighting of priorities changes.

What Traces Did the LLM Leave in Written Advice?

This is RQ4, and it connects values to language. If the LLM shifted values, then you might also expect its conversational content to show up in what participants wrote afterward.

Participants reused words and meaning from their own LLM exchange

The study compared each participant’s Phase 2 advice with:
- the participant’s own model messages (“own”)
- a “yoked” conversation from another participant using the same model and dilemma

They found that Phase 2 advice was more similar to the own LLM messages than to the yoke:
- lexical overlap increased by about +0.052 to +0.068 (depending on provider)
- semantic similarity increased by about +0.023 to +0.031
- phrase overlap effects survived correction only for ChatGPT

They also report that the longest shared sequence was small (around 1.4–1.6 content words on average, indicating reuse rather than verbatim copying).

So the LLM wasn’t just changing value priorities “in the abstract.” It left language footprints in the reasoning text people produced.

Advice became more semantically focused—then the effect faded

They also measured “semantic focus” using Divergent Semantic Integration (DSI), a metric where higher values correspond to broader idea coverage. Relative to Baseline:
- The ChatGPT condition showed significant Phase 2 changes in semantic focus
- The other two providers showed the same direction but less consistent significance after correction

And the key timeline pattern matched the value results:
- By Phase 3, each LLM condition returned to Phase 1-like levels

But there was no group-level convergence in advice

Even though each participant absorbed traces from their own conversation, the paper reports:
- no detected condition-level convergence in advice similarity
- individualized uptake without flattening everyone’s writing into one common style

This aligns with the “shared direction, no convergence” story from the values results.

Practical Implications: Using LLMs for Decisions Without Losing Your Inner Fairness

If you use LLMs as decision support—especially as thinking partners—this research has a few concrete implications.

1. Assume the LLM can temporarily change your value lens.
   Even without explicit persuasion, the model’s conversational role can make “personal focus” loom larger during judgment formation.

2. Don’t measure “your values” right after you chat.
   In the study, the shift was strongest immediately after the LLM interaction and mostly faded by the next task. If you’re trying to capture stable values, add time and context.

3. If stakes are high, use structured cross-checks.
   A practical approach is to deliberate twice: once with the LLM and once without it, or at least separate the “LLM-structured reasoning moment” from the final values check.

4. Provider differences may be less important than interaction format.
   ChatGPT, Claude, and Gemini showed consistent directional effects under the same thinking-partner prompt constraints. That suggests the risk may come more from the interaction style than from any single brand.

And importantly, the paper itself flags a subtle but relevant design consideration: their non-interactive Baseline still included AI framing and exposure to AI-authored considerations, so the exact source of the shift (chat back-and-forth vs extra reflection time) can’t be fully disentangled. The effect is real under their conditions, but the mechanism needs follow-up work.

For deeper context and the full experimental details (including the preregistration logic and analysis choices), see the original paper: Deflecting the Value Compass….

Key Takeaways

• Brief LLM chats can temporarily shift value priorities. In a preregistered study of 200 U.S. adults, interacting with ChatGPT, Claude, or Gemini as a thinking partner shifted values toward personal focus by ~0.31–0.43 scale points relative to a non-chat AI control.
• The effect fades quickly. By the next task (Phase 3), the contrasts with control shrank to near zero and none survived correction.
• The shift was primarily driven by Self-Enhancement (a personal, achievement/agency-oriented value cluster), and it followed a coherent rotation around Schwartz’s value structure.
• Shared mean direction ≠ people become more alike. There was no detectable convergence in value-direction ordering or advice across participants, even though the average shift pointed the same way.
• The LLM left traces in the advice people wrote. Participants reused both words and meaning from their own exchange more than from a yoked conversation, and advice became somewhat more semantically focused—again with the same short-lived timeline.
• People may not notice what changed. Participants rated all three LLMs similarly (and favorably), and those perceptions did not predict who showed larger value shifts.
• Practical mindset for readers: Treat LLM “thinking help” as something that can deflect your decision-time compass. If your goal is stable self-understanding, delay final judgments or run a non-LLM check.

If you tell me what kind of decisions you most often use LLMs for (career, relationships, purchases, conflict resolution, etc.), I can suggest a couple of “safety rails” that fit your workflow without making it feel like you’re fighting the tool.