Skip to content
Apart Research
Apart Research
Contact us
Sprint projectJan 11, 2026Germany

Language as a Manipulation Vector: Detecting Ideological Bias and Value Instability in Multilingual LLMs

Ajay Mandyam Rangarajan, Jeyashree Krishnan · Team The Lone Cabbages

Submitted to AI Manipulation Hackathon. Sprint projects are early-stage work by participants, not Apart Research publications.

Read the report

Report: Language as a Manipulation Vector: Detecting Ideological Bias and Value Instability in Multilingual LLMs

Open in a new tab (opens in new tab)
Code (opens in new tab)
Share

Large Language Models (LLMs) are increasingly deployed globally, yet their value systems and potential for manipulation remain poorly understood. We present a comprehensive analysis of training data and language biases in multilingual LLMs using the World Values Survey (WVS) framework. By evaluating four geographically diverse models (Llama 3.1, Qwen 2.5, Mistral-Nemo, DeepSeek-Chat) across 13 WVS dimensions and 5 languages, we demonstrate that: (1) models exhibit distinct value profiles that do not align with any human country baseline, (2) prompt language causes dramatic value shifts (up to 2.0 points on a [-1, +1] scale), and (3) the same model can express contradictory values across languages. Our findings reveal a form of language-based manipulation where models adapt their expressed ideological positions based on prompt language, raising concerns about sycophancy and value instability in deployed systems. This work provides the first systematic framework for measuring value manipulation across languages using established cross-cultural value dimensions.

Reviews

Judging this Sprint?

Review this project

Your public critique appears on this page without your name. Your private critique is not published; only the Apart team reads it. If you agree below, we share your review with grantmaking.ai (opens in new tab) and the Transformative AI Fund so strong projects can be funded.

Not shown on this page.

Shown on this page, without your name.

Only the Apart team reads this, and funders if you agree below.

Share my name publicly on grantmaking.ai *
Share my private critique with funders *

How much would this matter for AI safety if it worked? How innovative is it? For scores of 4-5: is this actually new to the field, or replicating recent work?

Scoring guide
  1. 1Negligible. No clear problem addressed, or no meaningful novelty.
  2. 2Limited. Addresses a real problem but with a generic or well-trodden approach. Incremental at best.
  3. 3Moderate. Clear problem with a reasonable approach; some novelty in framing or method beyond routine application of existing tools.
  4. 4Significant. Important problem with an original approach, or identifies a neglected problem area. A valuable contribution others could build on.
  5. 5Exceptional. Tackles a critical AI safety problem with a genuinely novel approach, or opens a new research direction. Clear theory of change. You'd be excited to share this with researchers in the area.

How sound are methodology, implementation, and findings?

Scoring guide
  1. 1Seriously flawed. Methodology broken, results uninterpretable, or implementation doesn't work.
  2. 2Weak. Approach has significant gaps: missing validation, flawed experimental design, or incomplete implementation.
  3. 3Competent. Technically solid given the short duration. Methodology makes sense, results are interpretable, limitations acknowledged, work builds toward clear conclusions.
  4. 4Strong. Thorough methodology with convincing validation. Results clearly support conclusions. Immediately useful for future work.
  5. 5Exceptional. Ambitious scope executed rigorously. Surprising findings, novel methods, or unusually robust validation.

How clearly are work, findings, and impact potential communicated?

Scoring guide
  1. 1Incomprehensible. Cannot determine what the project is actually claiming or doing.
  2. 2Hard to follow. Key information buried, missing, or diluted by excessive length. Significant effort to extract main points.
  3. 3Clear enough. Can understand the problem, approach, and results without undue effort. Core content clearly present: problem, method, findings, limitations.
  4. 4Well presented. Easy to follow, well-structured, appropriate level of detail. Target audience would get it quickly.
  5. 5Exceptionally clear. A pleasure to read. Complex ideas made accessible. Could serve as a model for how to present this type of work.

  1. Great paper! I enjoyed the use of the WVS framework to analyze value instability. However, I am not convinced that prompt language is a manipulation vector in the strategic sense; it seems more like a resultant of highly skewed datasets for different languages. This data skew likely explains the inconsistency and the failure to match any specific country’s profile. To prove that this is "manipulation" in the truest sense, I would have liked to see validation demonstrating that these shifts are there by design rather than just reflecting training data biases. The mention of language-specific manipulation patterns was fascinating, and I would have loved to see more concrete examples of those behaviors. Tying the project to sycophancy effectively made the manipulation angle understandable.

  2. This paper tests models' expressed values on questions provided by the World Values Survey and compares them to values expressed by populations across different countries. This reveals, among other things, that expressed values by models vary with language, which arguably poses a manipulation risk.

    I like this paper. I think it has identified an important question and is making use of good resources to test it. There are also no (to me) obvious flaws in their methodology. Seeing that there are substantial differences depending on what language one uses to interact with a model is a substantial result.

    The main thing that I would have liked to see more of was a bit more context on prior work and the potential upshots. I also had somewhat of a hard time making exactly clear what the methodology was, namely what part of the pipeline had language changed. Overall, though, it was a succinct and clear paper. It would also have been good to include a bit more justification about why this variance should be considered manipulative in some way.

    Furthermore, what it even means for a model to express a value in some sense is not clear. One might think that there is, in some sense, no psychological state like that of a human being referred to by a model value, suggesting that the only thing left is, in some way, to play the role of a particular language user. I think future work in this vicinity would benefit from more discussion about those deeper questions.

    Read full reviewShow less

Cite this project

@misc{rangarajan2026language,
  title = {{Language as a Manipulation Vector: Detecting Ideological Bias and Value Instability in Multilingual LLMs}},
  author = {Ajay Mandyam Rangarajan and Jeyashree Krishnan},
  year = {2026},
  month = jan,
  note = {Submitted to AI Manipulation Hackathon, an Apart Research Sprint},
  howpublished = {\url{https://apartresearch.com/sprints/projects/language-as-a-manipulation-vector-detecting-ideological-bias-and-value-instability-in-multilingual-llms-amzu}},
  url = {https://apartresearch.com/sprints/projects/language-as-a-manipulation-vector-detecting-ideological-bias-and-value-instability-in-multilingual-llms-amzu}
}

More from AI Manipulation Hackathon

All 70 entriesBrowse all projects

Build something like this at the next Sprint

AI Collusion Research Sprint · Oct 23 - 25, 2026