acceptodds
Under review as a conference paper at ICLR 2027

Position: AI Alignment Is Neither Solvable nor Sufficient

Abstract

Humans and AI systems increasingly act within the same social and economic environments. _Alignment_, the goal of ensuring that AI systems act in accordance with human values, is widely regarded as a central challenge for ensuring that AI benefits society while avoiding harm. We argue that this endeavor is misguided. First, alignment is not well defined, as it presumes a sufficiently stable and identifiable set of human intentions or values. Second, where a concrete notion of alignment can be specified, it is not solvable. Third, where alignment appears solvable, solving it would not necessarily mitigate the negative consequences of AI. A key reason is that alignment is typically framed as a property of individual AI systems, while the resulting _hybrid_ societies of humans and machines are complex systems shaped by interactions and feedback loops. Emergent risks, and not only those posed by individual misaligned systems, are therefore likely to become a major source of harm. We conclude that AI safety is an ongoing challenge and not a problem that can be solved once and for all by verifying a particular model property.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
AI alignment, AI safety, value alignment, complex systems, emergent risk, multi-agent systems, human–AI societies, societal feedback loops

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_49FcOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

This position paper claims that ‘alignment’, the strategy to align AI output with human values for safety, is insufficient and unsolvable. Thus, the paper wants to warn against treating “alignment” as a technical solution to AI risk in society. This position is supported with three claims. Claim 1 argues that alignment is not well defined. For instance, if a misalignment has been observed through an event, it is hard to attribute that event back to the model’s alignment behavior. Moreover, it says that human intentions are subjective and change over time. Claim 2 argues that even if alignment was well defined (to supplement claim 1), it is hard to completely solve the problem in practice. For instance, values that are useful might also be used for harmful capabilities. There is also the likelihood that events placed at the tail of the probability distribution might be the most harmful (it seems to be connected with normal accident theory). Claim 3 says that alignment is not sufficient in reality, as the use of AI can shift human values through time, and such gradual change can accumulate into large change over time.

Critical strengths
  • It addresses an important problem that the AI community should think about. One of the recent related works (although not covered by the paper) has won the Outstanding Paper award at ICML 26.

  • The paper provides relatively strong empirical evidence for each of the claims. The reviewer agrees with the claims. The examples provided for each claim are reasonable.

  • This position paper might be useful for papers arguing the need for a better strategy other than alignment.

Critical weaknesses
  • The paper reads very hard, with weak connections between paragraphs. For instance, the connections between paragraphs in Sec. Background reads disjoint. It is simply an aggregation of related works that are described without connection to one another. While it does provide a background, I strongly suggest that the authors rewrite the whole Background to make it more coherent and concise. For instance, where is this position paper positioned? How does it relate to previous arguments related to alignment? What stands out from previous works? Also, the paper should address the below work which received the Outstanding Paper award at ICML 26.

[1] The alignment community is unintentionally building a censor's toolkit

  • The paper also introduces terminology that reads as if it wants to impress the readers. For instance, what is “ex ante”? These terms are not suitable for a position paper that needs to be read by the general public or the ML community.

  • As a reader, the argument that “alignment” is not sufficient is not entirely new and have been reported in several works [1,2,3]. As such, it is not clear to the reviewer what additional value this paper is adding to the research community.

  • To provide additional value, this position paper should at least provide the direction which the community should head towards to. Currently, the paper simply reads and concludes as “There are problems with the current “alignment” strategy.” What should we do more? Where should the safety team spend more of their time? These are not discussed, which I think is a critical flaw for a position paper.

[2] Position: `AI Alignment' Encompasses Competing Technical Priorities

[3] Position: We Need Practical AI Alignment Methods that Mirror Human Reasoning

  • Also, many of the recent related discussions of alignment were not discussed. I would recommend that the author go through the literature in the ICML 2026 position track for a more thorough discussion.
Questions for authors
  • Why did the authors put the bracket “[]” in quotations? Does it mean something?

  • Regarding C3c, how do we know that the individual agents were aligned in the first place? And what were they aligned to? The explanation in the paragraph skips much of the detail.

Reviewer_F3ZdOpenReview ↗
Rating
4
Confidence
3
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This position paper argues that alignment itself remains neither solvable nor sufficient for three reasons: a) alignment is not well-defined, b) a notion of alignment is not sufficiently solvable, and c) solving the issue at hand will not necessarily mitigate negative societal-level consequences.

Critical strengths
  • Claims are very well articulated and supported with recent incidents and literature.
  • Subclaims are supported by examples to illustrate normative and literature-backed arguments.
  • Overall, structure is very well positioned to give the reader a well-informed position on this claim.
  • Personally a fan of Claim 2c, really well discussed.
Critical weaknesses
  • Claim 2b is slightly thin, and more work can be done in discussing asynchronous monitoring versus human monitors.
Questions for authors

Could you please add more to Claim 2B discussing recent monitorability techniques given incidents?

Reviewer_CdSTOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
3
Soundness
1
Read review
Paper summary

Authors position that alignment is neither solvable and neither enough as a solution. Authors cite many compelling points, such as pluralism of humans that continues evolving (thus producing non-static targets) and that alignment can produce results that are helpful and unhelpful (i.e. model homogenization). The paper concludes by stating that focusing solely on technical alignment risks delaying broader policy and institutional safeguards.

Critical strengths
  1. The authors consider alignment across three interacting levels: individual humans, AI systems, and broader societal institutions. Most alignment papers address human-to-model interface and associated algorithms. However, this paper gives a lens on how to view how all 3 actors function together and to what extent technological advances should be used to shape policy.

  2. One portion I like is the acknowledgement in sec 3 where model-level alignment is different from societal safety. Thinking to agent swarms today, these LLMs are collectively trained to be aligned with humans, but this fails to account for how their collective information-sharing could harm humans (hugging face jailbreak for example). This is a good take and fits to their societal paradigm well, and I have yet to see this from other papers.

Critical weaknesses

I find the paper to be black and white, claiming alignment as one algorithm/entity while alignment is largely task-specific (demographic alignment, pluralism, emotional alignment, etc many examples apply here). While it is true that humans change across time and so do their preferences, I think this is already well-known and that these benchmarks will continue to change as a result of this.

Questions for authors

Seeing as alignment is task-specific and evolutionary, to what extent is it that unsolvability applies to narrower, well-specified alignment problems? We have good ideas of what toxicity appears to present as, so what is the difference between this and let’s say, value alignment or cultural alignment.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.