Quick wins for a faster PC:
Fix the driver behind crashes, sound loss and screen glitchesFind Drivers →Clear out junk files and repair common Windows errorsFree Scan →Scan for outdated or missing drivers - takes under a minuteDriver Scan →iTechGuides is reader-supported. When you buy through links on our site, we may earn an affiliate commission. As an Amazon Associate I earn from qualifying purchases. Learn more
Anthropic and OpenAI both say they assess dangerous AI capabilities, apply safeguards, and use internal processes to guide development or deployment. Their published frameworks differ in how they set triggers, assign decision-making authority, and describe commitments—but those differences do not establish that one company is safer in practice. Anthropic’s 2026 policy revision makes the comparison more complicated: it retreats from some earlier conditional commitments while adding public roadmaps and recurring risk reports. Calling that change “safety theater” is an interpretation, not a conclusion demonstrated by the available evidence.
What “AI safety theater” means—and what the evidence can show
Calling a safety policy “theater” implies a gap between public assurances and meaningful action. A written framework can be assessed for what it promises, who makes decisions, and what outsiders can scrutinize. It cannot, by itself, show whether a company consistently follows the policy or whether its safeguards work in real deployments.
That distinction matters here. Anthropic’s and OpenAI’s public frameworks have a similar broad shape: they connect assessments of model capabilities or risks to safeguards and operational decisions. They differ in important design choices, but the sources available do not provide a comprehensive independent comparison of the companies’ incidents, deployment decisions, internal governance, or real-world safety outcomes. The defensible question is therefore how their stated frameworks compare—not which company is definitively safer.
What Anthropic changed in its 2026 policy
On February 24, 2026, Anthropic announced version 3.0 of its Responsible Scaling Policy (RSP). In its own explanation, it said the earlier approach had run into practical problems: capability thresholds can be ambiguous, evaluations do not always yield decisive answers, and government action has moved slowly. Anthropic said it had sometimes applied safeguards precautionarily despite uncertainty, while that uncertainty made it harder to argue for coordinated action across the industry. This is Anthropic’s account of its reasoning, not an independent finding about why the policy changed.
#1 Best Overall
A split between company actions and industry recommendations
The revised RSP separates mitigations Anthropic plans to pursue regardless of what other companies do from a broader set of measures it believes the AI industry should adopt. That is a change in emphasis: the policy no longer presents every proposed measure as a unilateral company commitment. Anthropic also calls for a Frontier Safety Roadmap spanning security, alignment, safeguards, and policy.
Public goals and recurring risk reports
Anthropic describes the roadmap’s goals as “nonbinding but publicly-declared” targets and says it will openly grade its progress. It also says Risk Reports will explain how model capabilities, threat models, and mitigations fit together, with publication planned every three to six months. Some reports are intended to receive external expert review. Anthropic says its current models do not yet require that review and that it is piloting the process.
These measures can make a company’s stated priorities and progress easier to follow. But a self-graded roadmap is not a binding safety threshold, and a report or review is not automatically an independent audit of every relevant decision. The policy’s value depends in part on what information the reports contain and how the stated goals are applied.
The Tool Desk
Outbyte Driver Updater FREEFix the driver behind crashes, sound loss and screen glitchesFind Drivers →Outbyte PC Repair FREERepair Windows errors before they cause bigger problemsFix Now →Why the change matters against Anthropic’s earlier promise
Anthropic’s September 2023 RSP version 1.0 said the company would “pause the scaling and/or delay the deployment of new models whenever our scaling ability outstrips our ability to comply with the safety procedures for the corresponding ASL.” The earlier document also described training pauses and deployment delays as parts of its conditional framework. That language made a more direct connection between a safety shortfall and a possible pause.
TIME reported that the new policy leaves Anthropic less constrained by its prior policy. Anthropic chief science officer Jared Kaplan told TIME: “We didn’t really feel, with the rapid advance of AI, that it made sense for us to make unilateral commitments … if competitors are blazing ahead.” TIME characterized the change as a retreat; Kaplan disputed that it was a U-turn and presented it as a pragmatic response to changed conditions. The contrast between the 2023 wording and the 2026 approach is real, but the company’s stated rationale and the reporter’s characterization should not be conflated.
How OpenAI’s published framework works
OpenAI’s April 15, 2025 Preparedness Framework update identifies three tracked capability categories: biological and chemical capabilities, cybersecurity capabilities, and AI self-improvement capabilities. It describes “High” capability as potentially amplifying existing severe-harm pathways, and “Critical” capability as introducing unprecedented new pathways.
OpenAI says systems at the High level need safeguards that sufficiently minimize associated severe-harm risks before deployment. At the Critical level, it says safeguards are needed during development as well. The framework describes a Safety Advisory Group that reviews reports and recommends action; “OpenAI Leadership” makes final decisions and can reassess them if new evidence appears. OpenAI also says it will publish capability and safeguards findings with frontier-model releases.
Those are the procedures OpenAI says it follows. The framework announcement is not an external audit showing how those procedures have been applied in each case.
Where the two frameworks align—and where they differ
Both companies present capability or risk assessment as a basis for safeguards and operational decisions. Both publish policies that can evolve. The differences lie in how those policies describe triggers, commitments, authority, and external scrutiny.
Rank #4
| Comparison | Anthropic: RSP version 3.0, announced February 24, 2026 | OpenAI: Preparedness Framework update, April 15, 2025 |
|---|---|---|
| Risk or capability structure | Risk Reports are intended to connect model capabilities, threat models, and mitigations. The announcement does not describe a matching set of three tracked categories in the terms used by OpenAI. | Names three tracked categories: biological and chemical capabilities, cybersecurity capabilities, and AI self-improvement capabilities. |
| Triggers and thresholds | The 2026 announcement describes difficulties with ambiguous capability thresholds and evaluations that may not give dispositive answers. It does not establish a directly comparable High/Critical trigger scheme. | Defines High and Critical capability levels, with different stated safeguards associated with each. (OpenAI, Preparedness Framework update, April 15, 2025.) |
| Stated response | Separates measures Anthropic says it will pursue regardless of others from broader industry recommendations; adds roadmap goals and recurring Risk Reports. The 2026 announcement does not state a directly comparable, universal pause rule. | States that High-capability systems need safeguards before deployment and Critical-capability systems need safeguards during development as well. (OpenAI, Preparedness Framework update, April 15, 2025.) |
| Decision process | Risk Reports and external expert review in specified circumstances are described; the announcement does not specify a decision body and final-authority structure comparable to OpenAI’s in the same terms. | The Safety Advisory Group reviews reports and recommends action; OpenAI Leadership makes final decisions and can reassess them if new evidence appears. |
| Public disclosure and outside review | Risk Reports are planned every three to six months, and some are intended to receive external expert review. Anthropic says current models do not yet require such review and that the process is being piloted. | OpenAI says it will publish capability and safeguards findings with frontier-model releases. The cited framework announcement does not specify external expert review in the same terms. |
| Force of stated commitments | Roadmap targets are explicitly described as “nonbinding but publicly-declared.” | The update describes safeguards and decision procedures; the cited announcement does not characterize them as nonbinding roadmap goals. |
The table compares what the companies said in the cited policy announcements; it does not establish how completely either company has implemented those statements. Nor does the absence of a matching provision in one announcement prove that a company has no related practice elsewhere.
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.What external reviews say about both companies
The Future of Life Institute’s Summer 2025 AI Safety Index gave Anthropic a C and OpenAI a C in its published safety-framework domain. The report credited strengths in Anthropic’s risk identification and mitigation approach and OpenAI’s commitment to publishing evaluation results. Its panel also identified limited risk scope as a common weakness, alongside concerns about external oversight, insufficiently specified conditional pauses, and reliable enforcement mechanisms.
What’s actually slowing this PC down?
Pick the symptom - the matching free tool is one click away.
Those are grades and criticisms from a particular report and assessment panel—not universal or timeless safety scores. They offer a reason to scrutinize both frameworks, but a shared grade does not mean the policies are identical or that the companies have equal safety performance.
Best Value
Other critiques address narrower questions. A 2025 arXiv paper by Sam Coggins and coauthors argues that OpenAI’s April 2025 framework does not guarantee meaningful mitigation across the risks the authors examined. That is an academic critique of OpenAI’s policy, not an official finding or a comparative audit of Anthropic. MIRI’s 2025 analysis argues that frontier-lab frameworks leave gaps in risk assessment, accounting for residual risk, and responses to unknown risks; MIRI explicitly advances a critical position. An earlier Federation of American Scientists analysis treats company frameworks as potentially useful but unilateral, and notes the case for broader public policy.
What cross-lab model testing can—and cannot—tell us
OpenAI published results from a pilot cross-lab evaluation that included Claude Opus 4 and Sonnet 4 alongside several OpenAI models. The findings varied by test category: Claude models generally performed well on instruction-hierarchy tests, while Claude performed less well than OpenAI o3 and o4-mini on some jailbreak evaluations. OpenAI also reported low rates in some scheming evaluations for o3 and Sonnet 4.
OpenAI cautioned that the comparison was not exact or fully apples-to-apples, that evaluation setups had small inconsistencies, and that the tests measured model propensities in deliberately difficult environments rather than real-world likelihoods. It explicitly warned against sweeping conclusions. These results are a limited example of cross-lab testing, not a ranking of the companies’ safety programs.
Recommended Free Tools
How to judge the “smaller difference” claim
The claim is plausible if it means the frameworks share a broad design: assess risk or capability, connect it to safeguards, and publish some account of the process. It is not a sound basis for saying their policies are interchangeable. OpenAI’s update specifies tracked categories, High and Critical levels, an advisory group, and leadership authority. Anthropic’s 2026 revision foregrounds the distinction between unilateral action and industry recommendations, a public but nonbinding roadmap, and recurring reports with external review in specified circumstances.
The clearest policy-level difference is also the one most relevant to the “theater” label: Anthropic’s earlier policy expressly described pausing scaling or delaying deployment when safety procedures could not keep pace, whereas its 2026 presentation emphasizes a different mix of company actions, industry goals, and reporting. That shift warrants scrutiny. It does not, by itself, show that Anthropic’s safety work is performative—or establish that OpenAI’s framework is stronger in practice.
To determine whether either policy is more than a public statement would require evidence about implementation: what assessments found, what safeguards were actually applied, how decision-makers responded to uncertainty, and what independent reviewers could inspect. The policy texts and limited model tests described here do not settle those questions.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
