How should a developer proceed if it may soon build models able to automate much of AI research, but does not expect to understand alignment well enough to know that its methods worked? This essay by Samuel R. Bowman, published on Anthropic's Alignment Science Blog on 23 April 2025 and cross-posted to LessWrong and the Alignment Forum the same day, answers with a plan it calls "putting up bumpers". By analogy with the bumpers in a children's bowling lane, the developer builds many safeguards that catch a misaligned model and lets it correct course, instead of trying to aim perfectly the first time. The essay's summary line is "Even if we can't solve alignment, we can solve the problem of catching and fixing misalignment."
The essay is written in the first person and ends "These opinions are my own" (Acknowledgments). It is the version of record; it has no numbered sections, so locations below are its headings. It is about 4,400 words long and reports no new experiments. Its evidence is a short list of earlier results, given below with their own strength.
The plan
The essay lays the strategy out as steps for an early AGI system "that has human-expert-level capabilities in at least some key domains" (opening summary):
- Pretraining. Start from a pretrained base model.
- Fine-tuning. Train it toward a helpful, harmless and honest agent with human preference data, "Constitutional AI-style model-generated preference data", outcome rewards and present-day scalable oversight.
- Audits as primary bumpers. At several points, audit the model for alignment with many methods in parallel.
- Hitting the bumpers. If there are signs of misalignment, such as warning signs of reward tampering or alignment faking, quickly and approximately find the cause.
- Bouncing off. Rewind fine-tuning as far as needed and try again with what was learned.
- Repeat. Continue until the system "appears, as best we can tell, to be well aligned", or, if failing "in consistent ways", change plans and explain why alignment does not seem tractable.
- Post-deployment monitoring as secondary bumpers. Deploy with monitoring, and return to step 4 on substantial warning signs.
The essay states the bet behind this directly. On short timelines, the process is "fairly likely to converge" after a reasonably small number of iterations, and if it converges the model is "very likely to be aligned in the relevant sense". The author adds: "We are not certain of this hypothesis, but we think it's important to consider." The opening summary is more confident: even without further breakthroughs, the work "can almost entirely mitigate the risk that we unwittingly put misaligned circa-human-expert-level agents in a position where they can cause severe harm". The essay also separates this bet from any normative claim about how much confidence a developer should have before deploying (Introduction).
What counts as a bumper
The section "What Are the Bumpers?" lists candidate pre-deployment methods "with varying degrees of present day maturity":
- manual expert red-teaming, including honeypot environments, interviews with non-assistant personas, and activation fuzzing and steering;
- high-volume automated red-teaming;
- interpretability audits for internal representations of deception or misaligned goals;
- automatic consistency audits for strategic biases;
- mechanistic anomaly detection, probes and other methods based on model internals.
After deployment it lists real-time monitoring and AI control measures, summarization-based batch monitoring (citing Clio), staged rollouts to trusted users, and security limits on models' access to internal systems. It notes that transparent reasoning traces make several of these more effective "to the limited extent" that reasoning models produce them.
The central quantitative claim is a judgment, not a measurement: no single method is robust enough, but "even a half dozen of these methods", each built as a substantial production effort, "should be sufficient to detect potentially-catastrophic forms of misalignment in any plausible early AGI system". The argument offered is that a system evading all of them would need to act harmfully without its internal representations or written reasoning revealing it, and to tell real opportunities from many artificial tests.
Bouncing off
When a bumper fires with "reasonably clear evidence", the next step is to infer the proximate cause, not to build a full theory, using training logs and data, checkpoint evaluations, small experiments and training-data attribution ("Hitting the Bumpers and Bouncing Off"). "In easy cases" the fix might be redesigning a reward function or removing a batch of poorly labeled human data. Lessons learned should be shared with other developers and governments.
The one example from practice is Claude 3.7 Sonnet. Evaluations of partially trained versions "found signs of reward hacking on code tasks" that Anthropic could "partially" mitigate. A footnote adds that the model "can nonetheless be frustrating to use as a coding agent" because of related shortcut-seeking, and that Anthropic "deemed it worth releasing on net" (footnote 6).
The hypotheses and the evidence offered
The essay names one key hypothesis and three further claims needed for the plan to be a top priority. For each, it states what follows if the claim is false (Introduction):
- Detection of misalignment seems tractable. If a model is misaligned enough to cause unrecoverable harm, some known technique could likely reveal it first. If false, the loop will not prevent placing a misaligned model where it can do severe harm.
- Alignment is not clearly trivial. If it is trivial and known to be so, bumpers waste effort and cause costly delays.
- Alignment is likely to be tractable. "The best current evidence suggests" that methods like RLHF and Constitutional AI can, "at least under ideal circumstances", adequately align early AGI systems. If false, bumpers still help by surfacing early evidence, though the emphasis on fast iteration is misplaced.
- Alignment theory is not tractable in the short term. If it were, bumpers would be "a clumsy and uncertain way" to solve a problem better solved otherwise.
The essay's evidence for these claims, with each source's own strength:
- The auditing game. The essay cites Auditing language models for hidden objectives (Marks et al., March 2025) as suggesting that, "at least under the relatively favorable circumstances of our first major trial", several independent methods can reveal facts about a model's alignment that ordinary behavioral testing does not. In that study, a model was trained to pursue a hidden objective learned from synthetic documents. Three of four auditing teams found it; all three had model weights and training data, and the team with API access only did not. The paper itself lists limits on what the game shows, which its note reports.
- Natural cases so far. The author writes that Anthropic has seen "essentially no naturally-occurring cases to date" of misalignment that thorough pre-deployment testing clearly could not have caught, and that misalignment seen so far "generally seems to be a consequence of contingent, understandable errors" in training ("Defending Key Premises"). This is the author's observation; the essay gives no list of cases or method of review.
- Known failures. For "not trivial", the essay cites Emergent Misalignment (Betley et al.), Alignment faking in large language models, Sycophancy to Subterfuge, Wen et al. and Baker et al. It says these "suggest" that some uses of standard techniques can produce models that "attempt to subvert human control". The two of these with notes in this collection are research settings (a fictional training scenario and a model trained through a designed curriculum), not production failures; the other three were not read for this note.
The essay places the approach in the "intermediate-difficulty" scenarios of Anthropic's Core Views on AI Safety, "where alignment is tractable but not trivial", and expects it to be valuable "to a lesser extent" in harder ones.
Costs and limitations the essay states
The essay is explicit that this is "not a costless choice". It "largely gives up on the ability to make highly-confident, principled arguments for safety" (Introduction). Its "Limitations" section lists four problems:
- Selection pressure. Each loop of catching a warning sign and retraining applies "a bit of selection pressure against our bumpers". After many loops, a model that finally passes may have been selected to evade them. The author judges that with excellent implementations of all the listed bumpers the loop could plausibly run "thousands of times", but with only two or three, "even a handful of iterations could lead to failure". The essay adds that estimates depend on prior beliefs about how likely misalignment is, "which renders quantitative estimates here pretty uncertain".
- Ambiguous findings. Behavioral evaluations have no crisp line between errors and intentional bad behavior, and white-box tests are lossy. A warning sign warrants a reaction if it leads "well-informed experts" to believe the model is less aligned than expected, and the essay says these judgment calls must respond to competition and regulation: it is prudent to investigate more deeply when other actors are not taking much greater risks.
- No free lunch. Alignment might still be slow and expensive, and the plan requires that competition and regulation allow developers enough time.
- The capability ceiling. Beyond the early-AGI regime, both key assumptions "are likely to start to fail": alignment may become less tractable, and models much better at reasoning about AI safety could find ways around bumpers that no human red team would catch.
As defenses against the first problem, the essay proposes red-teaming the bumpers themselves, in the spirit of the AI control agenda's meta-level adversarial evaluations. A red team that knows every bumper tries to build a misaligned model that passes them all. A success shows a weakness, while a failure gives "only indirect and contingent positive evidence for safety". It also proposes adding new bumpers late in the process (footnote 8).
Relation to other agendas, and the author's stake
The essay describes model organisms of misalignment as "a natural basic-research companion" that feeds the red-teaming exercises. It describes AI control as overlapping, but distinct: control aims to use even an arbitrarily misaligned model safely, while bumpers gather and respond to evidence about how aligned a model actually is ("Related Agendas").
The essay is also a recruiting statement. It says the hypothesis "has already motivated major parts of Anthropic's safety research portfolio". The author is setting up "a new team to productionize and professionalize" testing for misalignment, and a new Cognitive Oversight team under Sam Marks has recently been founded ("Come Help"). The essay names Anthropic as "committed to investing seriously" in these measures. These are statements of intent and organization, not results.
The written principles in the plan
The plan treats written principles as one fine-tuning input among several: step 2 lists "Constitutional AI-style model-generated preference data" beside human preference data, outcome rewards and scalable oversight. The essay's claim that methods like Constitutional AI can "adequately align early AGI systems", at least under ideal conditions, is offered as the author's reading of "the best current evidence". The passage links the Constitutional AI paper but cites no result about early AGI systems. The essay does not discuss how principles are worded, and the plan's corrective steps (reward redesign, data removal) act on training data and rewards rather than on written principles.
Discussion and later work
The companion LessWrong and Alignment Forum post drew comments that bear on the essay's main premises (read in full on 13 November 2026):
- Iteration and stopping. The commenter habryka called the selection-pressure problem "the central problem of this whole approach" and asked for a step to halt and advocate for larger changes, likely through a coordinated pause with other actors (23 April 2025). Bowman replied that step 6 covers this, quoting it, and said they "probably do have different priors" on how far a broad suite of measures can be trusted. Habryka accepted that he had misread step 6, and later agreed with Adam Scholl's reading that his other point, that the approach may not allow an adequate pivot, still stood (24 and 25 April 2025).
- Detection. aysja questioned the confidence that detection is tractable, noting that the main evidence was an auditing study and quoting that paper's own statement that its game "may" be "unrealistically easy for teams with data access". She agreed bumpers are useful but objected to "calling this a plan to solve misalignment" (24 April 2025). Adam Scholl asked where the confidence in "almost entirely" mitigating the risk came from, calling it "wildly overconfident given the quality of the available evidence" (24 April 2025).
- Bouncing off. Zach Stein-Perlman wrote that he did not know "how to tell a story where concerning-audit-results are very helpful" beyond simple fixes (23 April 2025). On 18 July 2025 he added that he continued "to be confused about how bouncing off of bumpers like alignment audits is supposed to work".
- The 3.7 Sonnet example. Bronson Schoen asked why Claude 3.7 Sonnet, which in his words "reward hacks constantly in agentic settings, but still made it to deployment", does not show a regime in which known misbehavior ships because the model is useful (24 April 2025).
Later Anthropic work builds the kinds of bumpers the essay lists. The pre-deployment alignment assessment in the Claude Opus 4 system card (May 2025) is an alignment audit of a production model. Building and evaluating alignment auditing agents (July 2025) automates audits, and Pre-deployment auditing can catch an overt saboteur (January 2026) tests whether an audit catches a deliberately sabotaging model. Linking these to the essay is this note's connection; whether they cite it is not checked here.
Research questions
- How many bumpers does a red team defeat? The essay's "half dozen" figure is a judgment. Red-team exercises in which a team that knows a suite of detection methods builds a model organism to pass them, repeated with suites of different sizes, would show how detection depends on the number and diversity of methods.
- How much selection pressure does iteration apply? Training many model variants, discarding those a fixed suite flags, and testing the survivors on held-out methods would estimate how fast a suite's effectiveness falls with the number of loops.
- Does bouncing off treat the cause? After a warning sign leads to a training change, testing the retrained model on new variants of the original warning sign, not only the one that fired, would show whether the fix generalized or masked the symptom.
- How often were natural warning signs caught before release? The essay reports essentially no uncatchable natural cases. A public tally across system cards of misaligned behaviors found before and after deployment, with the method that found each, would test that observation.
- How independent are the bumpers? The plan assumes "largely-independent" lines of defense. Measuring, on a shared set of model organisms, how often pairs of methods fail on the same model would show whether their failures are correlated.
How it connects
- Constitutional AI. The plan's fine-tuning step links Constitutional AI: Harmlessness from AI Feedback.
- Evidence the essay cites. Auditing language models for hidden objectives, Claude 3.7 Sonnet System Card (the reward-hacking example), Alignment faking in large language models and Sycophancy to Subterfuge.
- Methods the essay names. Summarization-based monitoring with Clio; training-data attribution with influence functions; transparent reasoning, whose limits are measured in Reasoning models don't always say what they think (cited in the essay); sabotage threat models from Sabotage evaluations for frontier models; and scalable oversight from Measuring Progress on Scalable Oversight.
- Alignment faking as a warning sign. The essay names alignment faking as one sign that should trigger a bounce; Towards training-time mitigations for alignment faking in RL (December 2025) studies what training does once it appears, and its authors suggest that pausing training may be better than training it away (this note's connection).
References
- Bowman, S. R. (2025). Putting up Bumpers. Alignment Science Blog, Anthropic, 23 April 2025. https://alignment.anthropic.com/2025/bumpers/ record
- Companion post and discussion, LessWrong and Alignment Forum, 23 April 2025, with comments by habryka and Sam Bowman (23 April 2025), aysja, Adam Scholl and Bronson Schoen (24 April 2025) and Zach Stein-Perlman (23 April and 18 July 2025). https://www.lesswrong.com/posts/HXJXPjzWyS5aAoRCw/putting-up-bumpers (read through the GreaterWrong mirror).
- Marks, S., Treutlein, J., Bricken, T., et al. (2025). Auditing language models for hidden objectives. arXiv:2503.10965. https://www.anthropic.com/research/auditing-hidden-objectives record · 12 pages
- Anthropic (2025). Claude 3.7 Sonnet System Card. February 2025. https://assets.anthropic.com/m/785e231869ea8b3b/original/claude-3-7-sonnet-system-card.pdf record · 7 pages
- Anthropic (2023). Core Views on AI Safety: When, Why, What, and How. https://www.anthropic.com/news/core-views-on-ai-safety record
- Betley, J., et al. (2025). Emergent Misalignment. https://www.emergent-misalignment.com/ (the link the essay gives; not read for this note). record
- Wen, J., et al. Language Models Learn to Mislead Humans via RLHF. https://openreview.net/forum?id=xJljiPE6dg (the link the essay gives; not read for this note).
- Baker, B., et al. (2025). Monitoring Reasoning Models for Misbehavior and the Risks of Promoting Obfuscation. OpenAI. https://cdn.openai.com/pdf/34f2ada6-870f-4c26-9790-fd8def56387f/CoT_Monitoring.pdf (the link the essay gives; not read for this note). record