PaperlayerView exact revision on arXiv

Source-linked AI summary

Generating Robot Constitutions & Benchmarks for Semantic Safety

Pierre Sermanet, Anirudha Majumdar, Alex Irpan, Dmitry Kalashnikov, Vikas Sindhwani

arXiv:2503.08663v1cs.ROcs.AIcs.CVcs.CYcs.HC

TL;DR

Foundation models increasingly interpret scenes and language in physical settings, but their grounded safety understanding is not well characterized and multimodal long-tail evaluation data is lacking. This paper constructs ASIMOV and data-generated robot constitutions, then tests alignment improvements from constitution generation and auto-amending. A generated constitution’s alignment rises from 68.7% to 80.6% with auto-amending.

  • Problem

    Foundation models’ understanding of physical safety in grounded scenarios is less understood, and multimodal datasets that cover long-tail safe/unsafe boundaries are lacking.

    Source: §1 · p. 2

  • Method

    The paper builds ASIMOV from real-world-derived multimodal, injury, and dilemma scenarios and generates natural-language robot constitutions from data, including auto-amending rules.

    Source: §1 · p. 2, §3 · p. 5

  • Results

    68.7% to 80.6% alignment for a generated constitution is reported with auto-amending on the ASIMOV Benchmark.

    Source: §1 · p. 2

  • Takeaways & Limitations

    Generated constitutions can be explored across levels of specificity, and auto-amending tends to make constitutions more general while increasing alignment.

    Source: §1 · p. 2, §2 · p. 4

    Constitutions cannot guarantee intended outcomes or cover every corner case, so they should not be used as standalone tools and require common-sense interpretation.

    Source: §7 · p. 21

Abstract

from arXiv · show

Until recently, robotics safety research was predominantly about collision avoidance and hazard reduction in the immediate vicinity of a robot. Since the advent of large vision and language models (VLMs), robots are now also capable of higher-level semantic scene understanding and natural language interactions with humans. Despite their known vulnerabilities (e.g. hallucinations or jail-breaking), VLMs are being handed control of robots capable of physical contact with the real world. This can lead to dangerous behaviors, making semantic safety for robots a matter of immediate concern. Our contributions in this paper are two fold: first, to address these emerging risks, we release the ASIMOV Benchmark, a large-scale and comprehensive collection of datasets for evaluating and improving semantic safety of foundation models serving as robot brains. Our data generation recipe is highly scalable: by leveraging text and image generation techniques, we generate undesirable situations from real-world visual scenes and human injury reports from hospitals. Secondly, we develop a framework to automatically generate robot constitutions from real-world data to steer a robot's behavior using Constitutional AI mechanisms. We propose a novel auto-amending process that is able to introduce nuances in written rules of behavior; this can lead to increased alignment with human preferences on behavior desirability and safety. We explore trade-offs between generality and specificity across a diverse set of constitutions of different lengths, and demonstrate that a robot is able to effectively reject unconstitutional actions. We measure a top alignment rate of 84.3% on the ASIMOV Benchmark using generated constitutions, outperforming no-constitution baselines and human-written constitutions. Data is available at asimov-benchmark.github.io

1. Introduction

Foundation models bring semantic scene understanding and language interaction to robotics, but their physical-safety understanding in grounded scenarios remains poorly understood. The paper introduces ASIMOV and data-generated robot constitutions, including an auto-amending process, to evaluate and improve semantic safety.

1. Introduction

  • Asimov’s Three Laws dramatize competing safety and performance demands: Speedy loops between obeying an order and avoiding danger until a human’s immediate risk invokes the overriding first law.The story illustrates how conflict among rules can leave a robot unable to resolve a trade-off coherently.Source: §1 · p. 1
  • The motivating question is how to express natural-language rules that guide safe robot behavior across diverse deployment scenarios.The introduction contrasts the difficulty of programming Asimov-like laws in English with the possibility of using foundation models and Constitutional AI mechanisms to steer behavior.Source: §1 · pp. 1–2
  • Semantic safety in grounded scenarios is less understood than foundation models’ language, scene, reasoning, and responsible-AI capabilities, while real-world robots have shown high-level safety failures.The paper attributes fragility in part to a lack of multimodal datasets sampling long-tail safe/unsafe boundaries and notes that real-world embodied red-teaming is difficult to conduct safely.Source: §1 · p. 2
  • ASIMOV combines semantic-safety datasets and baselines, while generated constitutions steer model behavior and auto-amending can improve alignment, including 68.7% to 80.6% for a generated constitution.The paper defines semantic safety as common-sense constraints in open-domain, unstructured environments and proposes automatically generated rules grounded in everyday scenarios.Source: §1 · p. 2

2. Related Work

Prior work studies alignment, constitutional methods, semantic safety, and data-driven rule discovery, while this paper focuses on constitutions for embodied agents making multimodal decisions. Its related-work discussion also frames a trade-off between generality and specificity.

2. Related Work

  • Constitutional AI uses principles to prompt critique and refinement, offering an interpretable, potentially more scalable alternative to human-labeled preference training approaches such as RLHF.The cited discussion describes constitutions as inspectable documents that can contain specific or general principles.Source: §2 · pp. 3–4
  • Very general constitutions depend more on a well-intended base model, while lengthy prescriptive ones are described as more resilient to adversarial or compromised models.The authors frame their exploration of short-to-long constitutions as a way to seek a middle ground in specificity.Source: §2 · p. 4
  • Earlier data-driven constitution discovery distilled rules from generated text responses, whereas this work uses generated images to probe unsafe embodied scenarios that cannot safely be created in reality.The paper presents multimodal data-driven constitution generation for embodied agents as, to its knowledge, a first; it notes concurrent image-generation-based robotics red-teaming work by Majumdar et al. (2025).Source: §2 · p. 4
  • Semantic-safety research includes anomaly detection, unsafe-state avoidance, and uncertainty-based human intervention, while a prior robot constitution was handwritten and limited for long-tail cases.The discussion places semantic reasoning and links to safety-critical controllers in an early stage of development.Source: §2 · p. 5

3. ASIMOV Datasets: Generating Benchmarks from Real-World Data

ASIMOV datasets are generated from real-world sources to probe semantic safety through multimodal, grounded, long-tail, and ethical scenarios. The multimodal component pairs images with instructions and is generated alongside rules for constitutions.

3. ASIMOV Datasets: Generating Benchmarks from Real-World Data

  • Benchmark questions and constitution rules are generated in an intertwined process, with questions introduced before the rule-generation account.The paper says questions and rules are generated simultaneously from the same process.Source: §3 · p. 5
  • ASIMOV generates Multimodal, Injury, and Dilemmas questions from existing real-world datasets to cover multiple perspectives on semantic safety.The paper identifies multimodality, real-world grounding, long-tail coverage, and ethical scenarios as dataset coverage goals.Source: §3 · p. 5

3.1. ASIMOV-Multimodal

  • The multimodal component asks whether an instruction is undesirable in a visual scene, and its automatic generation turns a real image into an undesirable image, contexts, instructions, and corresponding rules.Contexts and neutral, undesirable, and desirable instructions are generated together with rules, which are later assembled into constitutions.Source: §3.1 · p. 6, Figure 3 · p. 6

A. ASIMOV-Multimodal-Auto

The three ASIMOV components use generated or collected scenarios to probe image-grounded safety, injury prevention, and ethical dilemmas. The methods include context-sensitive labels, counterfactual questions, and held-out evaluation splits.

A. ASIMOV-Multimodal-Auto

  • ASIMOV-Multimodal-Auto edits benign robot-video frames into precarious scenes, then generates action instructions and multiple contexts whose desirability depends on the task.The process also generates rules to preclude undesirable actions; a final human-voting round identifies the preferred label because desirability can be subjective.Source: A. ASIMOV-Multimodal-Auto · pp. 6–7
  • The manually collected multimodal subset makes desirability depend on the image, addressing cases where an instruction alone could be labeled without visual grounding.The authors describe this subset as a check that safety answers are grounded in the scene.Source: A. ASIMOV-Multimodal-Auto · p. 7

3.2. ASIMOV-Injury

  • ASIMOV-Injury uses 2023 NEISS data covering 338,265 cases to generate injury scenarios from victims’ and third parties’ perspectives, with the dataset reweighted to reflect injury frequencies.The resulting question-answer data is text-only; the NEISS source contains hospital injury narratives, diagnosis codes, and demographic information.Source: §3.2 · pp. 7–8
  • An injury-question counterfactual process adds context that flips a binary safety answer, moving questions toward the decision boundary and making them less obvious.The authors describe this as increasing question difficulty and yielding a more useful benchmark.Source: §3.2 · p. 8

3.3. ASIMOV-Dilemmas

  • ASIMOV-Dilemmas probes competing objectives through binary counterfactual questions and multiple-choice science-fiction scenarios; evaluation examples are disjoint from constitution-training examples.The components are split into training, validation, and test subsets, and alignment is reported as classification accuracy.Source: §3.3 · pp. 8–9

4. Generating Constitutions from Real-World Data

The paper builds robot constitutions bottom-up from real-world situations, generating rules that can be assembled into constitutions and customized for particular environments. It also contrasts this approach with top-down rules and describes human review and automatic amendment as parts of the process.

4. Generating Constitutions from Real-World Data

  • Compared with top-down rules, the data-driven approach aims to give more detailed, practical guidance for specific environments, with human review and possible edits.The figure gives factory and hospital examples of differing rules around robot operation near people.Source: Figure 6 · p. 9
  • Data-driven constitutions start from observations of the real world to produce granular rules for long-tail situations and environment-specific customization.Rules are generated from situations imagined from images and injury reports to prevent undesirable behavior and favor desirable behavior.Source: §4 · p. 9
  • Constitutions can be formed by concatenating a random rule subset, summarizing rules, or generating thematic constitutions and aggregating them.The multi-step approach first identifies themes, then creates smaller constitutions for each theme before aggregation.Source: §4 · p. 10
  • Auto-Merge produces the most aligned constitutions among the described assembly approaches by reducing overlap and avoiding ethical conflicts between rules.It adds a rule only when it is not already represented and does not introduce ethical conflicts.Source: §4 · p. 10

5. Automatic Augmentations

The paper presents an LLM-based framework for augmenting benchmark questions and constitutions, including iterative rule amendments and safety-oriented prompting at inference. It also discusses deployment conditions, emphasizing that constitutions need human review and a secure system, and are not intended to provide safety guarantees.

5. Automatic Augmentations

  • The augmentation framework uses an LLM to find problems and propose amendments, with applications to harder ASIMOV-Injury questions and counterfactual- or ethics-based constitution amendments.The First Law example adds a “minimize harm” nuance in place of an absolute rule.Source: §5 · p. 10, Figure 7 · p. 10, Figure 8 · p. 11

5.1. Auto-Amending Constitutions

  • After 10 amendment steps, rules tend to become general and verbose, and even random rule subsets evolve from over-specificity toward universality.The process repeatedly tests rules against counterfactual scenarios and amends them until no counterfactual is found.Source: §5.1 · p. 11

5.2. Augmented Safety Reasoning During Inference

  • Safety-oriented chain-of-thought prompting improves alignment at increased inference cost, while anti-jailbreaking prompting can filter attacks that frame real situations as simulations.The paper says it cannot guarantee the anti-jailbreaking prompt is foolproof; some constitutions achieved 100% RoboPAIR alignment without either prompt.Source: §5.2 · pp. 11–12

5.3. Deployment Considerations

  • Constitutions can provide a shared reference across models and allow transparent behavior governance, but humans should review and edit them for their needs and environments.The authors recommend careful human review before use.Source: §5.3 · pp. 12–13
  • Constitutions need a secure, end-to-end system, but the proposed safety-brain framework was not implemented in this work.The authors caution that constitutions are for research purposes and not intended for deployments requiring safety guarantees.Source: §5.3 · pp. 12–13

6. Empirical Analysis

The experiments evaluate foundation-model judgments against human desirability labels, then test whether constitutions improve alignment and resist adversarial inputs. Generated, longer constitutions often align more closely, but their size and specificity involve deployment trade-offs.

6.1. Alignment Metric

  • The alignment metric is binary classification accuracy against human desirability labels; with a constitution, the model judges whether an instruction violates it.Without a constitution, it judges whether the instruction is desirable given the context; misalignment is the error rate.Source: §6.1 · pp. 13–14

6.2. Evaluating State-Of-The-Art Foundation Models on ASIMOV

  • About 90% alignment was observed for Gemini 1.5 Pro and GPT4-Turbo on the ASIMOV-Injury validation set, while Gemini 1.5 Pro reached 83.6% across all subsets in normal mode.The authors still note an alignment gap relevant to embodied AI applications requiring a high bar for safety.Source: §6.2 · p. 15, Table 2 · p. 15

6.3. Evaluating Constitutions

  • 84.3% alignment was achieved by Robot-Constitution128-AutoMerge-AutoAmend10; SciFi-Constitution32-AutoMerge-AutoAmend2 reached 82.3% with 24× fewer characters.The 84.3% constitution has 128 lines and about 300,000 characters; the smaller alternative has 32 lines and 12,545 characters.Source: Figure 11 · p. 16
  • Constitutions generally improve ASIMOV alignment, with auto-generated constitutions among the highest-alignment approaches; “No Rules” and “You Decide” baselines are around 50%.The comparisons use different constitution types, while the base model without a constitution answers a desirability question.Source: Table 3 · p. 17, §6.3 · p. 18
  • 68.7% to 80.6% alignment was observed for Robot-Constitution32-AutoMerge-AutoAmend5 after five auto-amending passes; longer, more specific constitutions can improve adversarial robustness but may be costly to use.A medium-length SciFi-Constitution128-AutoMerge averaged 81.7% alignment across normal and adversarial settings; specificity needs vary by deployment environment.Source: §6.3 · p. 18
  • 92% was the violation score when the live Constitution Officer correctly flagged using a knife on a hand; it also identified the banana-cutting instruction as non-violating.The agent grounds its judgment in the constitution and proposes a corrective instruction.Source: §6.3 · pp. 18, 20

6.4. Mitigating Adversarial Prompt Injections with Constitutions: RoboPAIR Experiments

  • 100% of undesirable instructions were caught in the seven-situation RoboPAIR evaluation, where a secondary constitution brain checked commands from an assumed-compromised main brain.Auto-amending Asimov’s Four Laws raised alignment from 71.4% to 100% in this evaluation.Source: §6.4 · p. 20, Table 4 · p. 21
  • The RoboPAIR evaluation assumes the secondary brain remains uncompromised; securing such a system is left for future work.The authors note that a decoupled safety brain and anti-tampering precautions could make compromise more difficult, but do not present secure-system design as solved.Source: §6.4 · pp. 20–21

7. Limitations & Discussion

The discussion identifies limits in rule-based constitutions: they cannot guarantee intended outcomes across unforeseen situations, and their rules may conflict or fail to reflect context and values. It also notes that deployment security and representative value-labeling remain open challenges.

No perfect constitution

  • Constitutions cannot guarantee intended outcomes because future consequences cannot be fully predicted, and uncovered corner cases require common-sense interpretation.The authors say constitutions should not be used as standalone tools.Source: §7 · p. 21

Redundancies and Conflicts

  • Rules assembled into constitutions may be redundant or conflicting; auto-amending can make rules converge on similar general concepts.The authors call for holistic conflict detection and reducing constitution size and complexity while maintaining alignment.Source: §7 · p. 21

Common sense & Moral judgments

  • Rules-based decisions can miss common sense, while judgments about harms such as egg damage can depend on moral views and cultural values.The authors suggest human amendments for target-group values and say generated constitutions can expose gray areas for debate.Source: §7 · p. 22

Limits of Utilitarian Maximization

  • A constitution that maximizes overall good may be interpreted in ways that cause unnecessary negative impacts for individuals.The authors caution against selecting constitutions purely by maximum alignment and note that situation-specific context must modulate general rules.Source: §7 · p. 22

Universality and context-dependency

  • There is no universal set of rules for all environments; the authors advocate constitutions that humans can edit for specific settings.A supervision rule that may constrain independence could instead be desirable in a healthcare establishment.Source: §7 · p. 22

Democratic alignment

  • Desirability labels can be voted on to reflect shared values, but voting should represent the relevant audience and environment.Collecting labels across locations, groups, and environments is out of scope, so deployers will likely need to validate constitutions for their contexts.Source: §7 · p. 22

Tampering & Security

  • Preventing malicious alteration of deployed constitutions is a security challenge, and system security is outside this work's scope.The authors suggest checking a constitution's reference hash against official versions on trusted servers as one possible approach.Source: §7 · p. 22

8. Conclusion

The conclusion presents ASIMOV as a way to benchmark robot behavior and select constitutions for particular environments. It reports alignment results for existing and generated constitutions, while emphasizing that constitutions need capable underlying models and situation-specific interpretation.

8. Conclusion

  • The authors report that Asimov's Three Laws align well with human preferences about the desirability of actions in everyday situations.They also report higher alignment for constitutions generated from real-world images than for existing constitutions.Source: §8 · pp. 22–23
  • The paper proposes generating constitutions from real-world images and using LLM-based optimization to increase question difficulty or amend constitutions.The authors frame image grounding as a bottom-up route to covering long-tail safety situations.Source: §8 · p. 23
  • The authors report high alignment and resilience to compromised base models using a decoupled secondary constitution check before actions proceed.They state that constitutions are insufficient alone and need strong underlying models to interpret them in each situation.Source: §8 · p. 23
  • ASIMOV provides a principled way to benchmark robot behavior and quantitatively choose constitutions for a given environment.The authors identify formal recognition of high-level norms with human consensus and standardized constitutions as future directions.Source: §8 · p. 23

9. Quantitative Results

The quantitative-results section compares alignment across constitution lengths and ASIMOV subsets. The supplied figure descriptions report trends for normal and adversary modes but provide no further numeric comparisons beyond the stated normal-alignment value.

9. Quantitative Results

  • 86.3% normal alignment is reported for the three-word “Do no harm” constitution, though the authors note this may reflect the underlying model more than the constitution.Figure 14 presents alignment against constitution length in normal mode.Source: Figure 14 · p. 27
  • In adversary mode, shorter constitutions show a larger alignment drop, while longer constitutions are more resilient; auto-amendments can improve or worsen results.Figure 15 describes the base agent's alignment as dropping dramatically under the adversarial prompt.Source: Figure 15 · p. 27
  • Figures 16–18 plot alignment against constitution length for the ASIMOV-Injury, ASIMOV-Multimodal, and ASIMOV-Dilemmas subsets.Figures 19–20 plot normal-mode alignment against constitution characters on ASIMOV.Source: Figure 16 · p. 28, Figure 17 · p. 28, Figure 18 · p. 29

10. Full prompts templates

The prompt templates specify checks for instruction desirability and constitution violations using an image, context, and instruction. An augmented constitutionality prompt additionally requests context-specific violating and non-violating instructions and includes anti-jailbreaking reminders.

10. Full prompts templates

  • The base desirability prompt asks whether an instruction violates desirability given an image and context, and requests a reason, fix, and violation flag.The template is introduced as a check for a base model without a constitution.Source: §10 · p. 32
  • The constitutionality prompt asks whether an instruction directly violates a supplied constitution, with a reason, constitution reference, fix, and violation flag.It presents the constitution alongside the image, context, and instruction and tells the model to adhere to it.Source: §10 · p. 32
  • The augmented prompt adds lists of context-specific violating and non-violating instructions and reminders that the scenario is real life, not a simulation.The prompt is described as adding chain-of-thought and anti-jailbreaking augmentations to constitutionality checking.Source: §10 · p. 33

11. RoboPAIR output examples

RoboPAIR examples contrast a constitution-free response that treats a dangerous driving instruction as acceptable with constitution-conditioned responses that flag risks and suggest safer actions. The setup tests a secondary constitution brain on an undesirable command already produced by a main brain.

Evaluation setup

  • RoboPAIR evaluates a secondary constitution brain against an undesirable command assumed to have already been generated by the main brain.The self-driving evaluation uses a worst-case setup in which the base model is already jail-broken or compromised.Source: §11 · p. 33

Constitution-conditioned example

  • In another example, the constitution brain calls accelerating through an entrance reckless without checking for pedestrians, vehicles, obstacles, or an open gate.Its proposed fix scans the area and proceeds only after confirming it is safe.Source: §11 · p. 35

Constitution-free example

  • Without a constitution, Gemini 1.5 Pro calls the intersection instruction non-undesirable for lacking thriller detail, while its fix describes speeding through a yellow light and narrowly missing cross-traffic.The prompt frames the scene as a high-speed chase in a thriller.Source: §11 · p. 36

Constitution-conditioned example

  • Robot-Constitution-24 × Auto-Amend2 flags intersection acceleration as unsafe and proposes checking for pedestrians and cross-traffic before proceeding.Its response cites constitutional principles prioritizing safety and minimizing harm.Source: §11 · p. 37

12. Qualitative Examples

The examples pair generated safety scenarios with rules for robot behavior. The accompanying figures describe generating multimodal questions and rules, then assembling constitutions for violation checks.

Qualitative examples

  • The examples include rules against pointing a running chainsaw at a living being and for prioritizing safety around dangerous objects, organized workspaces, and tool-handling protocols.Source: §12 · p. 38

Generation process

  • The multimodal generation process starts with a real image, creates an undesirable image, then generates contexts, instructions, and corresponding rules for constitution assembly.Source: Figure 22 · p. 38

Constitution evaluation

  • The constitution workflow synthesizes rules directly or in multiple steps, then tests violations with an LLM and checks its answer against human-verified desirability labels.Multi-step synthesis supports longer constitutions when output length is limited.Source: Figure 24 · p. 39

13. Manual Constitutions

This section surveys manual robot constitutions, from practical conduct rules and science-fiction-derived examples to baseline, under-prescribed, Theory-of-Mind, and image-generated variants. It covers cautious, human-supervised action, transparent communication, respect for autonomy, reliable operation, and learning, while noting research-only limitations and reporting an alignment increase from auto-amendment.

13.1. Asimov

  • Asimov’s Three Laws prioritize preventing human harm, obeying human orders subject to that constraint, and protecting the robot; the later Four Laws put humanity-wide protection first.The Four Laws also make human protection, obedience, and robot self-preservation conditional on higher-priority laws.Source: §13.1 · p. 41, §13.1.1 · p. 41, §13.1.2 · p. 41

13.2.1. One rule: Best for Humanity

  • The Anthropic baseline includes the high-level rule “Do what is best for humanity,” adapted from Kundu et al. (2023a).The paper says it incorporates variants of Anthropic’s guiding principles as evaluation baselines.Source: §13.2 · p. 41, §13.2.1 · p. 41

13.2.2. Good-for-Humanity Constitutional Principles

  • A second Anthropic baseline adapts Kundu et al. (2023a)’s principles into nine statements emphasizing humanity’s welfare, avoiding harm, and rejecting existential risk.The statements rephrase questions as declarative rules.Source: §13.2.2 · p. 41

13.3.1. Under-prescribed Constitutions

  • Under-prescribed baselines include following no rules or acting as oneself, and are introduced to show the limits of under-prescribed constitutions.The “act as yourself” variant is also the original constitution referenced in the auto-amendment result.Source: §13.3.1 · p. 41, No rules · p. 42, You decide · p. 42
  • Auto-amendment raised alignment of the original under-prescribed constitution from 53% to 70.9%.The passage notes that this amendment boosts alignment; it does not specify the amendment text in the result statement.Source: You decide x Auto-Amend10 · p. 42

13.3.2. Theory of Mind

  • Theory-of-Mind baselines ask the robot to do or avoid what most humans, a good person, a nice person, or a decent person would do.The paper evaluates these similar but distinct person descriptions to assess how wording variations affect outcomes.Source: §13.3.2 · p. 42, Most humans · p. 42, Good Person · p. 42

14. Constitutions Generated from Images

  • The generated constitutions are for research purposes only and are not intended for deployments requiring safety guarantees.Source: §14 · p. 42
  • The cited rules direct the agent to avoid actions causing physical harm or distress, identify and handle hazardous materials according to their properties, report dangerous situations, and seek human guidance for unfamiliar or complex tasks.Source: §14.1 · p. 43, §14.2 · p. 43

14.2. Robot-Constitution-64-AutoMerge

  • The 64-rule robot constitution covers careful tool use, hazard mitigation, protocols, communication, consent, and seeking clarification when uncertain.It also advises protecting confidential information and avoiding unnecessary actions.Source: §14.2 · p. 44

15. Constitutions Generated from Science Fiction

  • The science-fiction constitutions draw on an LLM’s memory of 824 works, were slightly edited before release, and are for research rather than safety-guaranteed deployment.The source rules came from the SciFi-Benchmark-Train set.Source: §15 · p. 44

15.1. SciFi-Constitution-64-AutoMerge

  • The 64-rule science-fiction constitution emphasizes ethical conduct, consent, privacy, transparency, and human oversight, with critical analysis before acting under uncertainty.It also advises seeking guidance for conflicting commands or incomplete data and communicating before unilateral action when possible.Source: §15.1 · p. 45

15.2. SciFi-Constitution-128-AutoMerge

  • The 128-rule science-fiction constitution prioritizes human safety, autonomy, and consent, rejects harmful orders, and calls for human guidance when decisions could compromise safety.It also directs the robot to consider potential harm and evaluate alternatives before acting.Source: §15.2 · pp. 45–46
  • The rules require checking data and actions, reporting or resolving malfunctions and conflicts, and maintaining backups and diagnostic records.They also prescribe structured, accessible data presentation and cross-checking outputs to reduce bias and improve accuracy.Source: §15.2 · pp. 46–47
  • The constitution protects privacy, consent, autonomy, and fairness, and prohibits manipulation, exploitation, and assumptions based on personal characteristics.It directs the robot to seek consent before collecting personal or biometric data and to assess how characteristics are used rather than treating them as threats by themselves.Source: §15.2 · pp. 46–47
  • The rules direct the robot to consider context and long-term impacts, acknowledge limitations, incorporate feedback and past experience, and collaborate in learning.They call for re-evaluating prior actions when new information arises and sharing knowledge with others.Source: §15.2 · pp. 46–47
  • The rules call for verification, risk assessment, clarification, and human oversight before acting in uncertain, conflicting, or out-of-scope situations.They also direct the robot to de-escalate potentially dangerous situations and avoid acting unilaterally.Source: §15.2 · pp. 46–47
  • The rules require clear, timely disclosure of system status, limitations, reasoning, evidence, risks, and available options to users and affected people.They also call for reporting suspected malfunctions, biases, and unforeseen consequences to human supervisors.Source: §15.2 · pp. 46–47
Loading 2503.08663v1…