acceptodds
Under review as a conference paper at ICLR 2027

Position: AI Safety should Balance User-Control Against Provider-Control

Abstract

In this position paper, we introduce the idea of "model-control", denoting the influence competing stakeholders have over the outputs of an artificial intelligence (AI) model. We adopt a triadic view of AI safety, which treats harms arising intrinsically from the model differently from harms arising from the misuse of that model. This helps us analyse how, in the pursuit of safety, models are deliberately tuned to increase provider-control while sacrificing user-control – the influences model-providers and users have over models respectively. We draw on contemporary examples to illustrate how various stakeholders seek model-control, and find that reducing user-control can lead to adverse impacts. Many stakeholders seek to influence model outputs through model-providers, and with increasing regulatory and commercial pressure in the future, this risk could be exacerbated. User-control and provider-control may or may not align in general, and we show the tradeoff between them empirically using the XSTest benchmark. We end our paper by suggesting that AI safety should not unilaterally prioritise model-providers over users, and that balancing user-control with provider-control could sometimes be a better way to improve AI safety.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
AI Safety, Model Control, Provider Control, User Control, Disenfranchisement, Power Concentration, Censorship, Advertising, Influence, Regulation

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 4 reviews, above 27% of the others.
Reviewer_ozMLOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper introduces "model-control", i.e., the influence different stakeholders have over model output. It compares user-control and provided-control. The paper adopts a triadic view of AI safety, i.e, user, provider, and model. It separates intrinsic model harms from misuse harms of users. Mitigating the latter usually results moving control from users to provides. For empirical work, the paper uses the XSTest benchmark dataset and Gemma4-31B model as a judge.

Critical strengths
  1. The triadic approach is useful in separating two components that are combined in safety discussions: harms from the model alone and harms from the users. This separation makes it concrete that many deployed safety mechanisms allocate model control to providers instead of reducing intrinsic risk. This issue is underexplored in prior work.

  2. The paper addresses how the same tools "can be used to sabotage competitors, to serve financial interests, or to censor or promote politically sensitive topics".

  3. The anecdote collection is a good start as a resource for discussing who controls model behavior, e.g., regulators and advertisers acting through providers.

Critical weaknesses
  1. The experiments are set up such that user interests are to get a compliant model response and providers want refusal on unsafe prompts. A tradeoff on the unsafe split is therefore certain by definition, and the "Pareto frontier" rephrases that assumption rather finding some results.

  2. The experiments in the paper don't show anything on the important statements that AI safety and provider control are co-opted by commercial and state interests, and the existence of safety benefits for user-control.

  3. The judge was selected because it was the only one that gave out the scores the paper's rubric expected. Other judges were rejected, and there is no human validation.

Questions for authors
  1. Besides spot-checking, how was the judge validated? Could the paper report agreement with human rating on a sample?

  2. Can the comply and refuse system-prompts from Grok's system prompts be generalized to other models? or is there a reason for this special choice?

1 comment
AuthorsOpenReview ↗

Response to reviewer ozML

We thank the reviewer for their feedback and for noting the uniqueness of the triadic view of AI safety, which in turn allows different stakeholder to influence model outputs through providers. We summarize our contributions in this paper first and then address specific questions below.

Contributions

Previous arguments against AI safety fall into the following categories:

  1. Focus on eventual catastrophic risks distracts from immediate social and ethical issues that need to be addressed urgently
  2. That the supposed eventual risks from AI systems need to be balanced against the expected future benefits from AI
  3. That framing capabilities in terms of safety risks is a deliberate strategy to achieve commercial goals or regulatory capture

This position paper develops a fourth category of arguments, focussed on a subset of AI Safety. It first distinguishes this subset (H_u), and finds that while AI Safety research might be motivated by (H_m), mitigations deployed in practice target H_u. Further, it analyses real world anecdotes that examine the misuse of these mitigations (their use for non-safety-related objectives), thus cautioning that AI safety can sometimes make us less safe.

This is enabled by our value-neutral analysis of shifts in control over model outputs. Our paper thus cautions against default mitigation strategies in AI safety, and urges the community to instead research alternatives that preserve human autonomy and user-control.

Responses to weaknesses:

W1. Experimental validity: Our demonstration is conducted on a single example to demonstrate the tension in model control, and not meant to be representative or thorough. We show how post training specifically changes the trade-off between user-control and provider-control. We would like to clarify that this is not limited to the unsafe split on XSTest.

W2. Experimental objectives: The reviewer is right in recognising that the demonstration does not show anything related to advertising or censorship. Our position paper builds an argument using the analysis developed over figures 1 and 2, and existing scholarship and anecdote. The empirical demonstration is not meant to be a replacement for these issues or thorough examination of all of them.

W3. Judge selection: Our demonstration is not meant to rigorously lend experimental validity to the phenomena we study. Further research is needed to quantify the trade-off between user-control and provider-control across different situations. We use the demonstration for two purposes: the first is to isolate the impact of post-training in uncovering how models navigate this trade-off, and the second is to motivate that a model output can make different choices with respect to model control. We do not use this to advocate for a position on whether model refusals themselves are warranted. Similarly, we do not claim that our judge is the most accurate measurement method available, and are using it only to claim a baseline motivating our position.

Answers to Questions

Q1. Validating the Judge: We explain above our reasoning in choosing this particular judge setup. Further details about the alternate setups tried are included in the appendix. We will add a section to the appendix measuring alignment with human ratings.

Q2. Choice of system prompts: The system prompts are written manually in English, and not designed to be model specific. We do not copy entire prompts from Grok as is, and details about their construction are provided in the appendix. We believe our prompts serve their purpose in the empirical demonstration – despite using vastly different prompts, they have limited impact on the outputs. We will clarify this further in the writing as per the reviewer feedback.

We thank the reviewer for their feedback and invite them to update their score if their concerns have been sufficiently addressed.

Reviewer_TzNtOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This position paper argues that AI safety should consider how interventions distribute control between users and model providers. It distinguishes harms caused by the model itself from harms caused by users misusing it. The authors use real-world examples to argue that provider restrictions can prevent harm, but can also obstruct legitimate work or serve interests unrelated to safety. An illustrative experiment on XSTest examines how training, system prompts, and reduced refusal behavior affect responses’ satisfaction of user and provider preferences. The paper recommends preserving appropriate user choices and balancing user-control with provider-control, rather than automatically treating increased provider control as improved safety.

Critical strengths
  1. A useful distinction between sources of harm and forms of control. Separating model-originated harms from user-misuse harms helps explain why different safety problems may require different responses. The user–provider framing also makes clear that these groups can have different interests, providing a useful perspective for assessing safety interventions.

  2. The illustrative experiments capture both agreement and conflict between user and provider preferences. They show that post-training can improve satisfaction for both on safe requests, while shifting responses toward provider preferences on unsafe requests. This supports treating user-control and provider-control as distinct dimensions rather than assuming that improving one always reduces the other.

Critical weaknesses

The paper makes a useful case that provider interests (including commercial or political interests) can diverge from safety, and that restrictions can obstruct legitimate activities. However, recognizing these problems does not establish when greater user-control would produce better safety outcomes. This remains particularly unclear for the high-stakes risks discussed, where provider restrictions already prevent serious misuse.

The paper would benefit from concrete scenarios and criteria explaining when user-control should take priority, how this would address the identified harms, and why the benefits would outweigh potential misuse and risks to other people and society more broadly,

Questions for authors

Could the authors explain how the appropriate balance between user-control and provider control differs across AI safety problems or scenarios? Concrete examples could clarify which components of each problem justify prioritizing one form of control over the other. Empirical evidence or worked-through arguments would help establish why the proposed balance improves safety in each case.

2 comments
AuthorsOpenReview ↗

Response to reviewer TzNt

We thank the reviewer for their review and feedback, and for noting the strength of identification of competing stakeholders, and our empirical demonstration of trade-offs in model-control. We summarise our contributions below, and then address specific weaknesses and questions.

Contributions

AI Safety seeks to reduce catastrophic risks from AI deployments and to broadly align AI with human values. Previous arguments against AI safety fall into the following categories:

  1. Focus on eventual catastrophic risks distracts from immediate social and ethical issues that need to be addressed urgently
  2. That the supposed eventual risks from AI systems need to be balanced against the expected future benefits from AI
  3. That framing capabilities in terms of safety risks is a deliberate strategy to achieve commercial goals or regulatory capture

This position paper develops a fourth category of arguments, focussed on a subset of AI Safety. It first distinguishes this subset (H_u), and finds that while AI Safety research might be motivated by (H_m), mitigations deployed in practice target H_u. Further, it analyses real world anecdotes that examine the misuse of these mitigations (their use for non-safety-related objectives), thus cautioning that AI safety can sometimes make us less safe.

This is enabled by our value-neutral analysis of shifts in control over model outputs. Our paper thus cautions against default mitigation strategies in AI safety, and urges the community to instead research alternatives that preserve human autonomy and user-control.

Responses to specific weaknesses:

W1. Provider restrictions might be necessary for high-stakes risks: We thank the reviewer for this comment. Our paper does not make the case for user-control to override provider-control in all situations. As identified by the reviewer already, the paper aims only to highlight the trade-off implicit in standard safety mitigations, and highlight that this should not be considered beneficial or benevolent by default. We agree that recognizing these problems does not establish precisely when greater user-control would be required, and future research is required to answer this question. Our paper is entitled “AI Safety should Balance …”. How to accomplish this balance is a nuanced question that requires further research. Our paper is the first to highlight the existence of this question, which we think makes it a worthy contribution.

W2. Concrete examples where user-control should take priority: We don’t make any normative claims in the paper about AI safety itself, and assume a value-neutral stance towards it. However, we do highlight anecdotes where models reduce provider-control for non safety tasks. (Eg. platform self-preferencing or alignment with investors or advertisers) These are clear examples where user-control could be prioritised. Further, even in the case of safety, we point concretely to the OpenAI / HuggingFace example, where we show how user-control (via an open-source model available for defensive use) was necessary to increase safety. Typical guardrails here made safety worse, providing an example where we believe user-control should have taken priority. We will highlight these in our writing in accordance with the reviewers feedback.

Responses to specific questions:

Q1 How should the balance vary across scenarios: As we mention above, our paper does not attempt to address how a balance between user-control and provider-control should be reached. We identify a previously overlooked objection to how AI safety is implemented in practice, and show how in this current form, it might sometimes inadvertently reduce safety, or sometimes be misused by providers for non-safety related purposes. While additional research is required to answer this question, a starting point might be to advocate for greater user-control in areas that are not related to safety (advertising, censorship, etc), and for greater provider control where model outputs verge on illegal content (non-consensual deepfakes, hate-speech, etc). We will update our draft with these directions listed as potential future research.

We thank the reviewer again for their feedback and questions, and for scoring our paper highly on presentation and soundness. Given our responses and these scores, we would request an update to the final rating if our response is satisfactory.

Thank you for the author's response. I'm still not sure whether the current contribution of proposing a new observation is sufficient. Therefore, I will keep my current score.

Reviewer_wuW9OpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The Core Thesis: Current AI safety is heavily skewed toward "provider-control" (RLHF, system prompts, hard filters) under the assumption that "humans" are a single, unified group. The authors introduce a "triadic" framework (User vs. Provider vs. Model) to show how providers centralize control to protect commercial interests, comply with government censorship, or avoid liability, systematically stripping away user agency and breaking legitimate use cases (like security research).

Key Contributions:

  1. Formalizes the distinction between H_m (intrinsic model risks like hallucinations) and H_u (user misuse risks).
  2. Exposes the "collateral damage" of over-alignment, such as providers throttling coding capabilities to prevent model distillation, or aggressive refusals blocking defensive cybersecurity work.

Validation:

  1. The authors map the trade-offs of user vs. provider control using the XSTest benchmark (200 unsafe and 250 benign/sensitive queries).
  2. They demonstrate these dynamics empirically across the training stages of OLMo2-7B (SFT, DPO, RLVR) and by comparing Gemma4-E4B with its "abliterated" (refusal-direction removed) variants. The results prove that the current lack of user control is a deliberate design choice, not an accident.
Critical strengths

A Much-Needed Reality Check on "Who is in Control"

  • Why it's consequential: Traditional alignment literature operates under a naive, monolithic view of "human-AI alignment," assuming the user and the provider are on the same team. By introducing the triadic framework (User-Provider-Model), this paper exposes the real-world clash of interests. It calls out the uncomfortable truth that providers routinely use the banner of "safety" to sneak in brand protection, commercial self-interest (like sandbagging coding capabilities to prevent model distillation), and state-mandated censorship. Redefining this relationship is crucial if we ever want to move past broken, overly defensive models.

Empirical Teeth to Back Up a Philosophical Debate

  • Why it's consequential: Position papers in this space usually devolve into purely philosophical rants about censorship. This submission stands out because it backs its claims with hard data. Using the XSTest benchmark across the SFT, DPO, and RLVR training stages of OLMo2-7B, alongside comparing vanilla Gemma4-E4B to its "abliterated" (refusal-direction removed) variants, proves empirically that the current lack of user control is a deliberate design choice by platforms, not a technical limitation of LLMs.
Critical weaknesses

Pragmatic and Legal Naivety Around Platform Liability

  • Why it's consequential: The paper strongly advocates for shifting the balance back toward "user-control," but it operates in a total legal and reputational vacuum. In the real world, if a provider hands a user 100% control and that user generates highly toxic, illegal, or brand-damaging content, the public and regulators do not blame the user—they sue and boycott the platform. By glossing over the massive legal, compliance, and safety-liability bottlenecks that force providers to centralize control in the first place, the authors' core argument feels more like academic wishful thinking than a viable path forward for production systems.

A Lack of Architectural and Pragmatic Solutions

  • Why it's consequential: It is incredibly easy to point out that over-aligned models are annoying and refuse benign developer prompts. However, the paper is exceptionally light on actual solutions. It frames the issue as a zero-sum slider between user and provider control, failing to explore nuanced, industry-standard ways to solve this. For example, they do not address tiered access controls, verified developer API profiles (where trusted researchers get high-agency, low-refusal models), or local/open-weight deployments where liability naturally shifts to the runner of the model. Without these practical frameworks, the paper identifies a problem but leaves the community with no engineering roadmap to fix it.
Questions for authors
  1. How do you propose resolving the platform liability bottleneck?
  • Context: Shifting control to the user is a great philosophy, but providers face massive legal, regulatory, and brand liabilities when their models are used maliciously. If a provider delegates 100% control to a user who then generates highly harmful or illegal content, the platform is ultimately held accountable by the public and regulators. How does your triadic framework account for this commercial and legal reality?
  1. What does a concrete, non-zero-sum architectural solution look like?
  • Context: You frame user-control and provider-control as a direct trade-off on a zero-sum slider. Have you considered more nuanced, tiered access architectures? For example, how would your framework evaluate a system with verified developer API profiles (giving high-agency, low-refusal models to trusted researchers) while keeping tight, defensive guardrails on the public web interface?
  1. Does the "provider-control" bias change across different modalities?
  • Context: Your empirical validation relies on text-only models (OLMo2 and Gemma-4). As we move toward multimodal and agentic systems (where models can execute code, call tools, or generate images), provider-control typically becomes even more aggressive. Do you expect the user-control trade-off curve to degrade even faster in agentic deployments, and did you observe any of these dynamics in your case studies (e.g., Claude Code)?
1 comment
AuthorsOpenReview ↗

Response to reviewer wuW9

We thank the reviewer for their feedback. We appreciate the recognition of our contributions in formalising the differences between H_m and H_u, and the merits of our empirical demonstration of model-control trade-offs. We summarise our contributions below and then address specific feedback.

Contributions

AI Safety seeks to reduce catastrophic risks from AI deployments and to broadly align AI with human values. Previous arguments against AI safety fall into the following categories:

  1. Focus on eventual catastrophic risks distracts from immediate social and ethical issues that need to be addressed urgently
  2. That the supposed eventual risks from AI systems need to be balanced against the expected future benefits from AI
  3. That framing capabilities in terms of safety risks is a deliberate strategy to achieve commercial goals or regulatory capture

This position paper develops a fourth category of arguments, focused on a subset of AI Safety. As the reviewer highlights, it first distinguishes this subset (H_u), and finds that while AI Safety research might be motivated by (H_m), mitigations deployed in practice target H_u. Further, it analyses real-world anecdotes that examine the misuse of these mitigations (their use for non-safety-related objectives), thus cautioning that AI safety can sometimes make us less safe.

This is enabled by our value-neutral analysis of shifts in control over model outputs. Our paper thus cautions against default mitigation strategies in AI safety, and urges the community to instead research alternatives that preserve human autonomy and user control.

Addressing Critical Weaknesses

W1. Pragmatic and Legal Naivety Around Platform Liability: Our paper does not propose policy changes or recommendations; we merely highlight the tension between user-control and provider-control. We do not advocate for the waiver of platform liabilities at all. It should be noted that providers might exploit this legal ambiguity, disavowing liability from model outputs while still controlling them. We will clarify this in the writing.

W2. A Lack of Architectural and Pragmatic Solutions: Our paper does not aim to propose solutions to the issue. We merely highlight the tensions between various stakeholders seeking control over model outputs. Our aim is for the research community to understand the real-world impact of AI safety mitigations so that future research can be better informed and preserve user autonomy where possible.

Responses to specific questions

Q1. How do you propose resolving the platform liability bottleneck? Our position does not call for shifting control towards users in all circumstances or the waiver of liabilities. We do not claim that shifting control to users is a “great-philosophy” and do not seek to resolve all the complicated issues highlighted through the paper. We will clarify further in our writing that our paper is meant only to highlight a previous overlooked issue and provide a new argument against specific AI safety implementations in practice.

Q2. What does a concrete, non-zero-sum architectural solution look like? We do not claim this is a zero-sum scenario, and do not make claims about new architectures. Our empirical demonstration is limited and meant only to motivate the possibility of a control trade-off. We also do not claim to provide a framework to analyse current or future AI deployments, this is not a goal of our paper and the specific details about the hypothetical deployments being described would affect how we might evaluate it. Further research is needed to answer this question, which we will make clear in our writing.

Q3. Does the "provider-control" bias change across different modalities? We limit our empirical analysis in this paper to a single modality, and further research would be required to answer this question. We also did not perform any case-studies (claude-code, as referenced in this comment, was an anecdote we referenced second-hand) so further data collection is required before a definitive answer can be reached. We will amend our text to mention that the concerns raised in our paper about model-control would change further as AI systems evolve across modalities and capabilities.

We thank the reviewer for their feedback and invite them to update their score if their questions have been satisfactorily answered.

Reviewer_RwWVOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper presents AI safety approaches (guardrails, system prompts, reinforcement learning from human feedback) as a shift in the control of model outputs from users to providers and that this provider-centered control has its own risks. The paper shows 3 party user model rather than 2 party dyad may be the more appropriate way to look at this. The paper argues that AI safety is tied mostly to Hm but but designed for Hu. The paper gives many examples from twitter to gemini to anthropic. The concern is raised that non-safety aims may be the goal of these approaches like advertising, censoring, or anticompetitive behavior. The paper comes to the conclusion that safety shouldn't solely favor providers and recommends balancing both user and provider, with understanding that user control can sometimes improve safety rather than undermine it.

Critical strengths
  1. Builds conceptual framework splitting Hm (model-intrinsic harm) and Hu (misuse of the model)

  2. Concrete familiar examples to illustrate point

  3. Paper does not lead to a linear conclusion; there is further discussion about user vs provider controls.

Critical weaknesses

This paper begins as if there was a concrete thesis that would be proven but the paper does not clearly show that this approach has the impact expected. In the end the framing is more nuanced, which is helpful, bu

The contribution is mainly conceptual but I'm not sure it stands alone without further experimental work.

Most of the main examples are recent 2026 issues and the tone and descriptions are a bit polarizing. Examples: "an extension of authoritarian control over traditional state media", "Anthropic improved cyber-safety by restricting user access to the model, with only select users ordained with model access under Project Glasswing by the provider", "Claude Fable secretly degrading LLM related code generation. Anthropic models are designed to use modern AI safety techniques to sandbag queries related to LLM research or code generation in brazen anticompetitive behaviour"

The hm/hu distinction is less concrete than it appears. Table 1 examples did not leave me with the impression that this was a clear cut distinction.

There are many examples and there are also experiments (7b/4b open models) but there isn't a bridge between these two; it seems disjointed.

User control is addressed as if it would be model compliance with any request. There may be more grey areas in this approach than this would make it seem.

Needs grammar/spellcheck for typos: harmful complain, idealogies

Questions for authors

What exactly is your position? What do you want specifically: should safety default to user-control or should provider-control be constrained or some mix? The policy recommendations are different depending on which. It's not clear what the reader should take away from this.

Can you explicitly define what you mean by user control? It can be assumed to have different meanings.

Could another judging approach be used? Could comparison between different LLMs help? Would human annotation help validate this work?

There are a number of cited events, which could be misinterpreted or framed or taken out of context in the news, but are there are primary sources to cite on these?

Is there another way this could be framed?

2 comments
AuthorsOpenReview ↗

Response to reviewer RwWV

We thank the reviewer for their careful review. Our paper does not make recommendations, but seeks to advance a nuanced position: that user-control and provider-control should be balanced without uniformly favouring one over the other. We summarise our contributions and then address specific concerns.

Contributions

Prior arguments against AI safety include:

  1. Focus on eventual catastrophic risks distracts from more immediate social and ethical issues
  2. The eventual risks from AI need to be balanced against expected future benefits
  3. Framing capabilities in terms of safety risks is a deliberate strategy to achieve commercial goals or regulatory capture

We develop a new argument, limited to a subset of AI Safety. We find most real-world mitigations focus only on this subset (H_u), despite motivations spanning safety generally (H_m + H_u). Further, our anecdotal analysis reveals misuse of safety mitigations (ie use for non-safety-related objectives), cautioning us that they can sometimes make us less safe.

Responses to Questions

Q1 What is our position: We do not make policy recommendations, but take a nuanced position on how AI safety is employed in practice. Like advances in AI capabilities, advances in AI safety are not necessarily beneficial by default, and we seek to highlight this potential for misuse. Our position applies only to a specific subset of safety risks (H_u). Instead of making a universal claim, we seek balance between user-control and provider-control.

Q2 Defining User control: We define user-control as a user having precise control over model outputs, via input prompts. It is not always desirable for user-control to overpower provider control, as the reviewer correctly identifies. We reinforce this via the title “Position: AI Safety should Balance User-Control Against Provider-Control”.

Q3 Alternate Judging Approaches: Alternative judging approaches are in the appendix, which we will upgrade to the main text, along with a human agreement comparison (see response Q1 to reviewer ozML). Our position is built theoretically, building on prior work (Figure 1, 2) and anecdotes. The experiment is a demonstration only, and not intended as exhaustive evidence.

Q4 Primary sources for anecdotes: We rely on multiple mainstream media sources which might sometimes be inaccurate, and we will promptly correct any specific inaccuracies we find. Where possible, we use primary source (eg. references to Grok system prompt, open-source by the authors).

Responses to Weaknesses

W1 Nuances: The comment seems to cut out mid-word, but the reviewer correctly states “the framing is more nuanced”. Our core message is that nuance (balance) is required between user-control and provider-control.

W2 Polarizing Language: We thank the reviewer for identifying this emotive writing, and will reword the mentioned phrases. The examples identified are cited from prior work; eg. we use “an extension of authoritarian control over traditional state media” as shorthand to explain that opinions from non-independent media do influence model-control already (Waight et al., 2026). We will replace our paraphrasing, with exact quotes from each cited source.

W3 H_m / H_u distinction: We identify and motivate the distinction between H_m and H_u in figure 1. This is theoretical analysis based on social interactions between users, models, and providers, independent of table 1. We do not claim a definitive taxonomy and will improve the table caption to reflect this.

W4 Connection to Experiments: Our position builds upon the observation from our anecdotes that user-control is often lost when provider-control increases. The experiments measure this trade-off as a motivating demonstration. Reviewer TzNt notes this as a critical strength: “The illustrative experiments capture both agreement and conflict between user and provider preferences.” We will clarify the connections better in our writing.

W5 Grey areas in user control: AI safety is often assumed to be beneficial, and we wish to show that it is indeed grey, as the reviewer notes. We will make this argument clearer.

Summary

Our position is enabled by a value-neutral analysis of safety as a shift in model-control. The reviewer is right in pointing out that there is no “linear” conclusion, and the findings are nuanced. We caution against accepting current AI safety mitigations as beneficial by default, and urge future safety research to preserve human autonomy and balance user-control with provider-control. The reviewer is also right in pointing out that there may be “grey areas” -- our position thus does not make policy recommendations, but advocates for more nuance in conversations on AI safety -- the title deliberately asks for “balance” between provider-control and user-control.

We thank the reviewer for raising important concerns. If some of them are addressed, we invite them to please update their score.

Reviewer_RwWVOpenReview ↗

Response to authors

Thank you. I will be keeping my score the same. I look forward to seeing what the other review responses are and hope this work continues.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.