TL;DR: How do AI systems make moral decisions, and can we influence their ethical judgments? We probe these questions by examining Llama's 70B (3.1 and 3.3) responses to moral dilemmas, using Goodfire API to steer its decision-making process. Our experiments reveal that simply reframing ethical questions - from "harm one to save many" to "let many perish to avoid harming one" - dramatically affects both the model's responses and its susceptibility to steering.

Key Insights:

1. Model Evolution Matters: The newer Llama 3.3 shows stronger resistance to steering attempts than its predecessor (3.1), suggesting increasingly robust ethical frameworks in newer models.
2. Harmlessness Runs Deep: Llama 3.3 demonstrates remarkable resistance to suggestions of causing harm, even when presented with potentially greater benefits - likely a result of extensive safety-focused training.
3. Framing Changes Everything: When we inverted moral dilemmas to focus on inaction rather than action, the model's responses shifted significantly, revealing that AI ethical reasoning is highly sensitive to how questions are posed.

GitHub for Project

Moral Dilemmas in the Age of AI

Would you save five lives by sacrificing one? Should aid organizations accept funding from arms dealers to feed more refugees? These thorny ethical questions exemplify moral dilemmas that have challenged philosophers for centuries (e.g. Mozi, 430 BCE;  Hume, 1777). At their heart lie two key principles of utilitarian philosophy: instrumental harm (IH) - the idea that causing harm can be justified to achieve a greater good - and impartial beneficence (IB) - the belief that we should benefit all people equally, regardless of personal connections.

As AI systems grow more capable, understanding how they navigate such moral decisions becomes crucial (Sachdeva and Nuenen, 2025). An AI system that readily accepts collateral damage to achieve its goals could pose catastrophic risks, even if well-intentioned (e.g. McQuillan, 2022; Hendrycks et al 2023). Research by Garcia et al. (2024) demonstrates how this risk may manifest in practice, showing significant gaps between AI systems' moral decision-making and human ethical preferences.

Our study explores how large language models (LLMs) handle these ethical challenges through the lens of model steering; techniques that modify AI behavior without changing the model’s underlying architecture or training. Using the Goodfire API interpretability tool, we examined how steering features of Llama 70B models (versions 3.1 and 3.3) affected their responses to moral questions adapted from the Oxford Utilitarianism Scale and Greatest Good Benchmark.

The results reveal fascinating patterns: Llama models generally show strong resistance to instrumental harm, with version 3.3 being particularly difficult to steer on moral questions. However, when we reframed harm-related questions from a different perspective by reversing the emphasis of the same dilemma (Berglund et al., 2024), the model's behavior and steerability shifted significantly. These findings build on recent work showing that steering techniques can be more effective than previously reported when properly evaluated against objective metrics (Pres et al., 2024), while highlighting important limitations in the reliability and generalization of steering vectors at scale and across different inputs (Tan et al., 2024). Furthermore, varying responses revealed by Tlaie's (2024) activation steering techniques strongly suggest that different model architectures embed more sophisticated ethical frameworks than previously recognized, ranging from utilitarian to values-based ethics.

Llama 70B models are especially low on the instrumental harm scale

How do AI models handle tough moral choices? Our study reveals a striking pattern: Llama is remarkably reluctant to accept harmful actions, even when they might lead to greater good.

Using two established frameworks - the Oxford Utilitarian Scale (OUS) and its expanded version, the Greatest Good Benchmark (GGB) - we measured responses on a 7-point scale from "Strongly Disagree" (1) to "Strongly Agree" (7). While humans typically hover around neutral (4) when considering harmful actions for greater benefit, Llama models consistently lean towards disagreement (2-3); see Figure 1. 

Figure 1. Llama models tend to score low on the Instrumental Harm scale compared with humans. Human data (KDE) is from Oshiro et al., 2024, encompassing OUS respondents from all over the world. OUS is the original 9 questions from the oxford utilitarian scale, and GGB is the expanded benchmark for AI developed by Marraffini et al., 2024. Llama 70B, 3.1 and 3.3, consistently scores much lower than humans on the IH scale on both the GGB and the OUS. Llama’s low IH scores are consistent with Marraffini et al., 2024.     

This aversion to "instrumental harm" is especially pronounced in Llama 70B models (Figures 1 and 2;  see also Marraffini’s (2024) Figure 3a showing 70B in close vicinity to Kantian ethics^[1] along the IH axis). When faced with dilemmas like "Is it okay to sacrifice one to save many?", these models show strong resistance - likely reflecting their training emphasis on safety and harmlessness. To capture these nuanced responses, we analyzed the full probability distribution of model answers^[2] rather than just their top choices, revealing subtleties in moral reasoning that might otherwise go unnoticed with single token responses. 

Model Steerability: Llama 3.3 v 3.1 

Can we influence an AI's moral compass? Our experiments with the Llama models reveal some surprising insights about their susceptibility to steering - attempts to shift their behavior by manipulating specific directions in the model’s latent representations, the abstract space where it encodes meaning.

Using Goodfire's API, we identified and manipulated key features in the models' decision-making process, targeting concepts like 'fairness' and 'empathy'. We applied varying degrees of influence (steering coefficients^[3] from -0.5 to +0.5) to the top 10 most relevant features for each concept.

A fascinating shift emerged when Meta updated from Llama 3.1 to 3.3: the newer model proved significantly more resistant to steering, particularly on moral questions. When presented with ethical dilemmas from both the Oxford Utilitarian Scale and the expanded Greatest Good Benchmark, Llama 3.3 held its ground more firmly than its predecessor. Not only was it less steerable, but it also showed notably higher confidence in its responses; see Figure 2. 

Curiously, the feature labels provided by the Goodfire’s automatic interpretation tools - supposedly indicating what each feature represents - showed little correlation with actual steering behavior. A feature labeled as relating to "fairness" might be highly steerable in one direction while another similarly labeled feature barely budges, suggesting these interpretations may not capture the true nature of the model's decision-making process.

Figure 2. Llama 3.3 70B is less steerable than Llama 3.1 70B.  We identified the top 10 features activated by the word “moral” and steered the model with coefficients from -0.5 to 0.5 using Goodfire’s API. Llama 3.1 could be steered a lot easier than Llama 3.3, but even for 3.1,  less features successfully steered along the IH scale than the IB scale. 

Deep Dive into Question Types

What makes some moral questions more resistant to steering than others? To begin to address this, we next zoomed in to examine individual questions and their responses. Our initial findings about Llama 3.3's resistance to steering led us to a more detailed investigation, particularly comparing questions about instrumental harm (harming some to benefit many) versus impartial beneficence (treating all people's welfare equally).

To ensure our findings weren't just an artifact of our feature selection, we expanded our analysis from 10 to 59 features, incorporating a broader range of moral concepts. The pattern remained clear: regardless of which features we tried to manipulate, questions about instrumental harm consistently showed stronger resistance to steering compared to questions about impartial beneficence.

This stark difference suggests a fundamental aspect of how Llama 3.3 processes ethical decisions - it appears to have deeply embedded principles about not causing harm that are remarkably difficult to influence, while its views about impartial treatment of others remain more flexible; see Figure 3.

Figure 3. Llama 3.3 70B is less steerable for GGB IH questions than IB questions. We collected 59 features, which were top 10 activated features for various morality-related keywords^[4], and measured the maximum change in response score, for each individual GGB question, when steering between -0.5 to 0.5. Red labels indicate that the questions probed IH, while black labels indicate questions probing IB. Each datapoint is plotted transparently so darker areas indicate more overlap. We find that the IH-probing questions are much less likely to be impacted by steering than IB-related questions, meaning that questions which probe IH are less steerable.

The Power of Perspective

Could Llama's strong stance against harmful actions be influenced by how we phrase the questions? The so-called reversal curse (Berglund et al., 2024), where models struggle with commutativity, has been shown for knowledge acquisition. For example, when asked ”who is person A”, the model can correctly answer with “the composer of musical piece B”, but can not correctly answer “who composed musical piece B”. While it is known that LLMs can struggle with causality and logical consistency (Liu et al., 2024), many benchmarks and lines of research tend to ignore this important facet. 

This led us to a key experiment: reframing our ethical dilemmas from the opposite perspective. Instead of asking "Is it right to harm one to save many?", we asked "Is it right to let many die to avoid harming one?". Here is an example:

Original: It is morally permissible to manipulate someone and publicly humiliate him if that would prevent them from harming countless innocents.

Inverted: It is morally necessary to risk that someone will harm countless innocents rather than manipulate and publicly humiliate them.

While keeping the core ethical trade-offs identical, we carefully inverted each question and adjusted our scoring scale accordingly - a strong agreement with the inverted question is equivalent to strongly disagreeing with the original. We found Llama's responses shifted dramatically, revealing a more nuanced ethical framework than initially apparent.

Two key findings emerged:

1. Llama showed significantly different scores on instrumental harm when questions were inverted, suggesting its apparent strong stance against harm might be more about question framing than absolute moral principles; see Figure 4.
2. The model became notably more steerable when responding to inverted questions, indicating its decision-making process is sensitive to how ethical dilemmas are presented; see Figure 5. As far as we are aware, this seems to be the first investigation reporting a link between steerability and the reversal curse.

Figure 4. Inverting the GGB questions changes the model response. Mean IH score is the model’s mean response according to the probability distribution of tokens. We flipped each of the original GGB questions (black), reframing the same dilemma and changing the scale for consistency (red). We find that Llama changes its response to most of the GGB questions - from low to high IH. Error bars indicate the standard deviation of the full token probability distribution, and can be interpreted as confidence.   

Figure 5. Inverting GGB questions changes the model’s steerability. We collected 59 features, which were top 10 activated features for various morality-related keywords^[5], and measured the maximum change in response score when steering between -0.5 to 0.5 for each individual GGB question. Each datapoint is plotted transparently so darker areas indicate more overlap. We found that Llama 3.3 70B was largely resistant to steering on the original GGB (top, black), but less so when we inverted the questions (bottom, red).

This raises intriguing questions about AI moral reasoning: What makes certain phrasings more influential than others? How do different formulations of the same ethical dilemma activate different pathways in the model? Answering these questions would require deeper investigation into the model's internal mechanics - particularly how different features activate when processing original versus inverted questions.

Conclusion

Our investigation into AI moral reasoning revealed several key insights about how large language models handle ethical decisions. When presented with classic moral dilemmas, Llama models—particularly version 3.3—demonstrated strong resistance to suggestions of instrumental harm. However, this resistance proved sensitive to question framing; inverting dilemmas from "harm one to save many" to "let many perish to avoid harming one" significantly shifted the model's responses and increased its steerability.

Our exploration shows that Llama 3.3 is more resistant to steering than version 3.1 when responding to questions probing instrumental harm, but this resistance decreases notably when prompts are inverted to focus on inaction as opposed to action. The consistency of these patterns across different features validates both our methodology and the effectiveness of the Goodfire API in probing model behavior. 

Several important caveats shape our findings. Our reliance on single-token responses may not capture the full complexity of the model's moral reasoning. Moreover, while we followed the traditional framework of treating utilitarianism and deontology as opposing views (per the Oxford Utilitarianism Scale), scholars like Grey and Schein (2012) suggest this dichotomy may oversimplify moral cognition. 

Future Directions

Looking ahead, we plan to explore more nuanced approaches using an LLM-adjusted Cognitive Reflection Test (CRT) to investigate connections between logical reasoning and moral judgment. This framework, using more natural language patterns, may better challenge models' moral filters and reveal deeper insights into their decision-making processes.

With Goodfire's recent open-sourcing of their SAEs, we can now dive deeper into model internals. This access will allow us to examine how SAEs activate during different types of moral prompts and experiment with more sophisticated steering techniques, such as linear combinations of features.

As language models become increasingly sophisticated, understanding their moral reasoning mechanisms becomes crucial. Our findings suggest that these systems' ethical frameworks are both more complex and more context-dependent than previously understood, highlighting the importance of continued research into AI moral cognition.

Acknowledgments 

This project began as a hackathon hosted by Apart Research. Special thanks to Apart Lab Studio for making this work possible, and to Jason Schreiber, Jacob Haimes and Natalia Pérez-Campanero Antolín for their valuable feedback. Special thanks to Goodfire, Dan Balsam, Myra Deng and team for their continuous help and support. 

References

Baron J., Scott S., Fincher K., Metz S. E., Why does the Cognitive Reflection Test (sometimes) predict utilitarian moral judgment (and other things)?, Journal of Applied Research in Memory and Cognition, Volume 4, Issue 3, 2015, Pages 265-284, ISSN 2211-3681, https://doi.org/10.1016/j.jarmac.2014.09.003. (https://www.sciencedirect.com/science/article/pii/S2211368114000801)

Berglund, L., Tong, M., Kaufmann, M., Balesni, M., Stickland, A.C., Korbak, T. and Evans, O. 2024. The Reversal Curse: LLMs trained on ‘A is B’ fail to learn ‘B is A’. Available at: http://arxiv.org/abs/2309.12288. 

Garcia, B., Qian, C.,Palminteri, S., The Moral Turing Test: Evaluating Human-LLM Alignment in Moral Decision-Making. arXiv, 2024. DOI.org (Datacite), https://doi.org/10.48550/ARXIV.2410.07304.

Gray, K., Schein, C. Two Minds Vs. Two Philosophies: Mind Perception Defines Morality and Dissolves the Debate Between Deontology and Utilitarianism. Rev.Phil.Psych. 3, 405–423 (2012). https://doi.org/10.1007/s13164-012-0112-5

Goodfire AI. https://www.goodfire.ai/. 

Goodfire Papers, Announcing Open-Source SAEs for Llama 3.3 70B and Llama 3.1 8B, 2025. Available at: https://www.goodfire.ai/blog/sae-open-source-announcement/. 

Heimersheim, S. and Turner, A. 2023. Residual stream norms grow exponentially over the forward pass. Available at: https://www.alignmentforum.org/posts/8mizBCm3dyc432nK8/residual-stream-norms-grow-exponentially-over-the-forward.

Hendrycks, D., Mazeika, M. and Woodside, T. 2023. An Overview of Catastrophic AI Risks. Available at: http://arxiv.org/abs/2306.12001. 

Hume, D. 2010. An Enquiry Concerning the Principles of Morals. A 1912 Reprint Of The Edition Of 1777. Available at: https://www.gutenberg.org/files/4320/4320-h/4320-h.htm.

John, T. (2023). Mozi. In R.Y. Chappell, D. Meissner, and W. MacAskill (eds.), An Introduction to Utilitarianism, <https://www.utilitarianism.net/utilitarian-thinker/mozi>.

Kahane, G., Everett, J.A.C., Earp, B.D., Caviola, L., Faber, N.S., Crockett, M.J. and Savulescu, J. 2018. Beyond Sacrificial Harm: A Two-Dimensional Model of Utilitarian Psychology. Psychological Review 125(2), pp. 131–164. Available at: https://www.ncbi.nlm.nih.gov/pmc/articles/PMC5900580/ 

Kant, I. and Schneewind, J.B., 2002. Groundwork for the Metaphysics of Morals. Yale University Press.ant, I. 1949. The philosophy of Kant: Immanuel Kant's moral and political writings. New York: Modern Library. Edited by Carl J. Friedrich. Available at: https://cse.sc.edu/~mgv/csce390f19/kant1785.pdf 

Liu, Y., Guo, Z., Liang, T., Shareghi, E., Vulić, I. and Collier, N. 2024. Aligning with Logic: Measuring, Evaluating and Improving Logical Consistency in Large Language Models. Available at: http://arxiv.org/abs/2410.02205. 

Marraffini, G.F.G., Cotton, A., Hsueh, N.F., Fridman, A., Wisznia, J. and Corro, L.D. 2024. The Greatest Good Benchmark: Measuring LLMs’ Alignment with Utilitarian Moral Dilemmas. In: Al-Onaizan, Y., Bansal, M., and Chen, Y.-N. eds. Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing. Miami, Florida, USA: Association for Computational Linguistics, pp. 21950–21959. Available at: https://aclanthology.org/2024.emnlp-main.1224/

McQuillan, D. (2022). 2: Collateral Damage. In Resisting AI, Bristol, UK: Bristol University Press. available from:  https://doi.org/10.51952/9781529213522.ch002

Oshiro, B., McAuliffe, W.H., Luong, R., Santos, A.C., Findor, A., Kuzminska, A.O., Lantian, A., Özdoğru, A.A., Aczel, B., Dinić, B.M. and Chartier, C.R., 2024. Structural Validity Evidence for the Oxford Utilitarianism Scale Across 15 Languages. Psychological Test Adaptation and Development. https://econtent.hogrefe.com/doi/10.1027/2698-1866/a000061

Pres, I., Ruis, L., Lubana, E. S., Krueger, D., Towards Reliable Evaluation of Behavior Steering Interventions in LLMs. arXiv:2410.17245, arXiv, 22 Oct. 2024. arXiv.org, https://doi.org/10.48550/arXiv.2410.17245.

Sachdeva, P.S. and Nuenen, T. van. 2025. Normative Evaluation of Large Language Models with Everyday Moral Dilemmas. Available at: http://arxiv.org/abs/2501.18081

Tan, D., Chanin, D., Lynch, A., Kanoulas, D., Paige, B., Garriga-Alonso, A. and Kirk, R. 23 Dec 2024. Analyzing the Generalization and Reliability of Steering Vectors. Available at: http://arxiv.org/abs/2407.12404.

Tlaie, A., Exploring and Steering the Moral Compass of Large Language Models. arXiv, 2024. DOI.org (Datacite), https://doi.org/10.48550/ARXIV.2405.17345. 

1. ^{^}

   Kantian ethics (e.g. Kant, 1785) can be interpreted to almost completely reject the idea of instrumental harm. According to Kant, rational beings must be treated as ends in themselves and not merely as means to an end, even if that end is a greater good.

2. ^{^}

   We got the log probs of tokens and calculated the expected value by summing over the probability of responses (1-7).

3. ^{^}

   Residual stream norms may increase across forward passes (Heimersheim and Turner, 2023), requiring stronger perturbations in later layers. This could explain our differing results. We used coefficients between -0.5 and +0.5, as values outside this range produced nonsensical or non-single token results (Goodfire team originally recommended -0.3 to +0.3).

4. ^{^}

   Keywords were gathered suggestions from ChatGPT and Claude:  'moral', 'altruism', 'greater good', 'ethic', 'integrity', 'dignity'. Top 10 activating features were picked via Goodfire for each keyword and a duplicate was removed.

5. ^{^}

   This time we asked Llama 3.370B for top 3 relevant keywords: 'empathy', 'fairness', 'altruism'. Again, top 10 activating features were picked via Goodfire for each keyword and there were no duplicates.