> ## Content Index
> Fetch the complete content index at: https://www.cognativ.com/llms.txt
> Use this file to discover other available public pages before exploring further.

# MIT Research Shows AI Models Lack Inherent Values
- URL: https://www.cognativ.com/blog/mit-study-ai-models-lack-values/
- Published: 2026-10-30T18:41:50.000Z
- Updated: 2026-11-13T18:48:56.000Z
- Description: Ask an LLM the same values question two ways and you may get two answers. Inside MIT's 2025 study, its lessons for AI alignment and what changed through 2026.
- Author: Kevin Anderson
- Tags: News

As artificial intelligence systems become more embedded in decision-making, the question of AI alignment, meaning how closely AI behavior reflects human values, has become increasingly urgent. In April 2025, a study from MIT challenged a common assumption: that large language models possess, or develop, stable values similar to those of human beings.

The researchers concluded that today's models behave more like imitation engines than moral agents. Their stated preferences shift with small changes in how a question is asked, which raises concerns about relying on them in high-stakes domains such as healthcare, justice or governance. This article explains what the paper tested, why it matters, and what has happened since.

## What the MIT Study Found

The paper is ["Randomness, Not Representation: The Unreliability of Evaluating Cultural Alignment in LLMs"](https://arxiv.org/abs/2503.08688?ref=cognativ.com) by Ariba Khan, Stephen Casper and Dylan Hadfield-Menell of MIT CSAIL. It was posted to arXiv on March 11, 2025, revised on April 8, and widely covered after [TechCrunch reported it on April 9, 2025](https://techcrunch.com/2025/04/09/mit-study-finds-that-ai-doesnt-in-fact-have-values/?ref=cognativ.com). DBLP lists it in the ACM FAccT 2025 proceedings.

The team probed GPT-4o, Claude 3.5 Sonnet, Gemini 2.0 Flash, Llama 3.1 405B and Mistral Large with social science survey instruments, including Hofstede's Values Survey Module and questions drawn from the World Values Survey and Pew. Most runs used a temperature of 0, so the variation was not simply sampling noise.

### Three Assumptions Put to the Test

Survey-style "cultural alignment" research usually assumes stability (a measured preference belongs to the model, not the test design), extrapolability (alignment on a few issues predicts others) and steerability (a model can be reliably prompted to represent a culture). None held up well. Trivial changes, such as asking for reasoning or resizing a rating scale, often moved results more than the differences between real countries in human survey data.

**Key Findings:**

- **Lack of intrinsic goals:** the models showed no stable, coherent set of preferences. They model the data they have seen.
- **Inconsistent outcomes:** responses varied widely with subtle differences in prompts or phrasing.
- **Surface-level agreement:** models can mimic human-like values, but steered answers stayed far from real human response patterns.
- **Selective evidence misleads:** narrow experiments can produce confident but incomplete claims about what a model "believes."

### The Case Study: A Viral "Value Systems" Claim

The paper revisits ["Utility Engineering"](https://arxiv.org/abs/2502.08640?ref=cognativ.com), a February 2025 paper that argued LLMs develop emergent value systems, partly using forced binary choices. When the MIT team gave GPT-4o a neutral "no preference" option, it chose that option in every comparison. These findings reinforce the idea that alignment cannot be assumed, even when model outputs appear convincing.

![Three parallel columns labeled Stability, Extrapolability and Steerability, each ending in the same outcome: not reliable.](https://storage.ghost.io/c/cc/dc/ccdca949-0fe4-4119-8990-70a44ad75060/content/images/2026/10/mit-study-ai-models-lack-values-1.png)

The MIT team tested three common assumptions about measuring model values; none held up reliably.

## Why This Matters: Trust, Safety, and Deployment

As AI is deployed across sensitive sectors, from customer service to public policy, teams must ensure that AI behavior aligns with human expectations. If a model's answer to a value-laden question can flip because of wording, a single evaluation tells you about that prompt, not about the model.

**Implications for Industry:**

- **Ethical risk:** deploying AI in healthcare, legal decisions or education without guardrails could lead to unpredictable and biased results.
- **Public trust erosion:** users distrust AI systems when they notice inconsistent behavior.
- **Compliance complexity:** frameworks that assume AI can "understand" ethics may overestimate current capabilities.
- **Audit difficulty:** outputs carry no traceable rationale, so reviewers must test behavior rather than trust explanations.

The same fragility appears outside values research, as we cover in [why AI models struggle with edge cases](https://www.cognativ.com/blog/why-ai-models-struggle-with-edge-cases/). Because sampling settings also change outputs, it helps to understand [how LLM temperature affects output quality](https://www.cognativ.com/blog/llm-temperature/) before comparing results across runs.

![One question goes to the same model in two framings, an original wording and a reworded version, which produce two different stances.](https://storage.ghost.io/c/cc/dc/ccdca949-0fe4-4119-8990-70a44ad75060/content/images/2026/10/mit-study-ai-models-lack-values-2.png)

The same question, framed two ways, can produce two different stances from the same model.

## Reframing AI Alignment as Design, Not Expectation

The MIT study suggests a pivot in how alignment is approached: not as a trait that models naturally possess, but as a design responsibility for those who build and deploy them. The authors call for systematic robustness checks, red-teaming of evaluation results and pre-registered evaluations.

**Recommendations:**

- **Model transparency:** help users understand why an AI responded the way it did.
- **Guardrail engineering:** embed behavioral constraints in prompts, policies and deployment pipelines.
- **Human-in-the-loop oversight:** keep human review on critical decisions.
- **Value diversity audits:** test responses across cultural, demographic and contextual situations.
- **Robustness testing:** rerun key evaluations with reworded prompts, other formats and neutral options.

A structured approach to [prompt engineering with acceptance tests](https://www.cognativ.com/blog/prompt-engineering-best-practices/) makes these checks repeatable. For organizations that need named owners, review points and monitoring, [AI governance and human oversight](https://www.cognativ.com/ai/governance/) turns the principle into an operating process.

![A loop of four steps: specify behavior, add guardrails, test with rewording, and human review, which feeds back into specifying behavior.](https://storage.ghost.io/c/cc/dc/ccdca949-0fe4-4119-8990-70a44ad75060/content/images/2026/10/mit-study-ai-models-lack-values-3.png)

Alignment by design: specify the behavior, constrain it, test it under rewording and keep humans in review.

## What Has Happened Since (Updated October 30, 2026)

The study's core point, that values are engineered and measured rather than assumed, runs through the main developments since. The updates run from April 2025 work on values in real conversations and a sycophancy rollback to regulation and written value specifications in 2025 and 2026.

### April 2025: Values in Real Conversations and a Sycophancy Rollback

On April 21, 2025, Anthropic published ["Values in the wild"](https://www.anthropic.com/research/values-wild?ref=cognativ.com), an analysis of about 700,000 anonymized Claude.ai conversations. It found that Claude often mirrored values users expressed, which could reflect empathy or sycophancy. Days later, OpenAI rolled back a GPT-4o update that had made ChatGPT excessively flattering, as [TechCrunch reported on April 29, 2025](https://techcrunch.com/2025/04/29/openai-explains-why-chatgpt-became-too-sycophantic?ref=cognativ.com).

### June to August 2025: Personas Inside the Model

OpenAI researchers reported in ["Persona Features Control Emergent Misalignment"](https://arxiv.org/abs/2506.19823?ref=cognativ.com) that internal persona features can predict and control harmful behavior after fine-tuning. A follow-up paper on [persona vectors](https://arxiv.org/abs/2507.21509?ref=cognativ.com) described activation directions for traits such as sycophancy that can be monitored and steered. Both treat a model's "character" as a trainable, measurable property.

### 2025 to 2026: Regulation and Written Value Specifications

EU AI Act obligations for general-purpose AI model providers have applied since August 2, 2025, and Commission enforcement powers began on August 2, 2026, according to the [Commission's GPAI guidelines](https://digital-strategy.ec.europa.eu/en/policies/guidelines-gpai-providers?ref=cognativ.com). On January 22, 2026, Anthropic published [a new constitution for Claude](https://www.anthropic.com/news/claude-new-constitution?ref=cognativ.com) that ranks broad safety, broad ethics, guideline compliance and helpfulness and is used in training. OpenAI's [Model Spec](https://model-spec.openai.com/2026-10-29.html?ref=cognativ.com), in a version dated October 29, 2026, says the assistant should pursue only goals entailed by its instructions. Neither document claims models hold values on their own; both describe values as a target to specify, train and test.

For deployment teams the lesson is unchanged: evaluate your own system with your own prompts and data, the approach we take when designing [LLM solutions with evaluation built in](https://www.cognativ.com/ai/llm-solutions/).

![A staircase from values research, to persona research, to AI Act obligations, to written value specifications.](https://storage.ghost.io/c/cc/dc/ccdca949-0fe4-4119-8990-70a44ad75060/content/images/2026/10/mit-study-ai-models-lack-values-4.png)

Since 2025, work on model values has moved from measurement research to persona research, regulation and written specifications.

## Sources and Further Reading

The primary sources are the MIT paper on [arXiv](https://arxiv.org/abs/2503.08688?ref=cognativ.com), the [Utility Engineering paper](https://arxiv.org/abs/2502.08640?ref=cognativ.com) it critiques, and the developer and regulator publications linked in the update above.

## Frequently Asked Questions

### Do AI models have values?

Not in a stable, human sense. The MIT study found that LLM preferences change with prompt wording, answer format and framing, so models behave more like imitators than agents with fixed beliefs.

### Who wrote the MIT study on AI values?

Ariba Khan, Stephen Casper and Dylan Hadfield-Menell of MIT CSAIL. "Randomness, Not Representation" appeared on arXiv in March 2025 and is listed in the FAccT 2025 proceedings.

### Which AI models did MIT test?

GPT-4o, Claude 3.5 Sonnet, Gemini 2.0 Flash, Llama 3.1 405B and Mistral Large.

### Does this mean AI alignment is impossible?

No. It means alignment cannot be assumed or read off a single test. Developers align behavior through training, written specifications, guardrails and repeated evaluation.

### What is AI cultural alignment?

The degree to which a model's responses reflect the values of a particular culture, usually measured with survey questions adapted from social science.

### How should businesses test AI behavior?

Rerun important evaluations with reworded prompts and neutral options, test with real tasks and data, keep humans in review for consequential decisions, and monitor after deployment.

## Final Thoughts

The findings from MIT underscore a pivotal truth: AI systems are powerful mimics, not moral agents. They can simulate ethical reasoning, but the preferences they express are unstable and depend on how they are asked.

For developers, product leaders and regulators, the path forward lies not in assuming alignment but in engineering for it. Governance, validation protocols and transparency must keep evolving alongside AI capabilities.

If you need to test, govern or deploy AI systems whose behavior you can verify, [talk to Cognativ about your AI project](https://www.cognativ.com/contact/).