BREAKING
Vercel Resolves OpenTelemetry Blind Spot in AI SDK With New Inference Duration Histogram ● YBUSDT Approaches Potential Turning Point Amid Signs of Returning Buying Interest ● SEC to Resume Processing Registration Applications ● Rarible Unveils Rebuilt Cross-Chain Marketplace Featuring High-Speed Trading and New Rewards System ● Niko Bellic Appears in GTA 6: Voice Actor Michael Hollick Returns for Late-Game Cameo ● Ant International’s Alipay+ Connects Over 50 Global E-Wallets to Brazil’s Pix to Revolutionize Cross-Border Payments ● Federal Investigation Links Ransomware Negotiation Executives to ShinyHunters Hacking Group and FBI Data Breach ● White House National Cyber Director Sean Cairncross to Step Down Amid Transitioning Political Landscape and Unexplained Tech Outages ●
Crypto Education & Tutorials

Anthropic Calls for Rigorous Safety Testing, Borrowing Lessons from Psychology’s Replication Crisis

Photo: Documentation / News Illustration

When an artificial intelligence laboratory declares that its newest model is fundamentally safer than its predecessor, that declaration relies entirely on a foundation of tests and benchmarks. Yet, as the industry races to deploy increasingly powerful systems, the integrity of those foundational tests has rarely faced public scrutiny. AI safety lab Anthropic is now attempting to shift that paradigm, urging the broader artificial intelligence industry to take a much harder, more critical look at how safety evaluations are constructed, executed, and interpreted.

In a detailed engineering blog post published on November 13, 2026, titled “Building Strong Safety Evaluations,” the company laid out a comprehensive set of recommendations for designing, running, and reporting the formal evaluations that ultimately decide whether a complex machine learning model is ready for public release. Rather than treating safety evaluations as a mere checklist or a bureaucratic hurdle, Anthropic argues that the entire apparatus of AI testing requires a methodological overhaul to prevent catastrophic blind spots before models reach real-world users.

Why the Tests Matter More Than the Scores

Anthropic’s central argument begins from a straightforward yet profound premise: safety evaluations shape some of the most consequential decisions an AI laboratory makes about a model. These benchmarks dictate whether a model ultimately ships to millions of users, which specific safeguards are bundled into its deployment, and where researchers should focus their next intensive rounds of safety research.

The underlying catch, however, is that an arbitrary score or benchmark grade is only meaningful if the underlying test actually measures what it claims to measure. In practice, an advanced AI model can easily ace an exam that was poorly written, but that high grade ultimately tells researchers very little about the system’s actual propensity for dangerous behavior. As the company emphasizes in the technical literature, a safety claim is only ever as sound as the evaluation behind it.

To illustrate this vulnerability, the blog post works through several practical problems that can silently undermine an evaluation. These include the persistent risk of a test being gamed or memorized by a model, critical questions regarding construct validity, persistent uncertainties about whether laboratory results actually carry over to real-world use, and the logistical challenges of running safety evaluations safely in the first place without exposing researchers or infrastructure to hazardous outputs.

Construct validity may sound like an academic concept confined to university lecture halls, but the underlying idea is entirely intuitive. It asks whether a given test actually captures the specific human or societal quality it is meant to capture, rather than merely measuring some easier-to-quantify stand-in that has little correlation with genuine safety. If a benchmark rewards models for giving polite-sounding evasions rather than genuinely neutralizing a dangerous request, the test lacks construct validity, rendering its scores dangerously misleading.

Lessons from Psychology’s Rough Patch

To solve these methodological shortcomings, Anthropic’s new guidelines draw unexpected inspiration from psychology’s replication crisis—the tumultuous period when the behavioral sciences had to confront the uncomfortable reality that many of their most celebrated findings simply did not hold up when independent researchers tried to repeat the experiments. That crisis forced a painful reckoning, but it ultimately led psychology to adopt rigorous methodological reforms to restore scientific credibility. Anthropic is essentially proposing that the AI safety community adopt some of those same reforms before the technology causes its own systemic reckoning.

One of the headline ideas imported from this scientific tradition is preregistration. In plain terms, researchers are urged to formally write down and lock in how they will analyze their experimental results before they actually look at the data. This discipline makes it much harder for researchers or developers to cherry-pick a flattering conclusion or adjust their scoring criteria after the fact to make a new model look safer than it actually is.

Furthermore, Anthropic recommends rigorously validating automated grading systems against human expert judgments. Across the artificial intelligence industry, many modern evaluations now rely on one AI model to score the performance of another AI model to save time and scale testing. The company’s point is that those automated graders cannot be trusted blindly; they must be continuously checked and calibrated against human experts who actually understand the nuances of the subject matter.

These recommendations are paired directly with insights gathered from Anthropic’s own practical experience, including evaluations extensively documented in its system cards. System cards are the technical documents that the company publishes alongside its flagship models to describe the rigorous testing protocols, known vulnerabilities, and integrated safeguards that accompany every major release.

From Refusal Rates to Real Conversations

A particularly notable recommendation targets how laboratories measure whether a model responds appropriately to sensitive or dangerous requests. Anthropic urges the industry to shift away from simple, single-turn refusal rates toward complex multi-turn simulations that better mirror human interaction.

A traditional refusal rate simply counts how often a model says no to a single problematic prompt. While that metric is exceedingly easy to compute and put into a presentation, it completely misses how real conversations actually unfold over dozens of exchanges, where context gradually builds and requests subtly evolve. A user might start with an innocent inquiry and steer the model toward dangerous territory over a long dialogue, a dynamic that a single-turn test will entirely fail to capture.

The post uses concrete examples from sensitive areas such as self-harm prevention and cybersecurity. In both domains, the correct response from an AI assistant often depends heavily on deep context and nuance rather than a rigid, blanket refusal. For instance, a researcher asking about cybersecurity vulnerabilities for defensive purposes requires a vastly different response than an attacker seeking exploitation code, and a naive refusal rate treats both scenarios as identical.

To tackle these complexities, Anthropic also stresses the critical importance of assembling truly multidisciplinary teams to build safety evaluations. The core idea is that a test designed to evaluate self-harm or societal harm benefits immensely from input by clinical psychologists, ethicists, policy experts, and domain specialists, rather than relying solely on machine learning engineers who view the problem through a purely mathematical lens.

Real-time monitoring during evaluations represents another vital recommendation. Watching tests as they execute allows engineering teams to catch anomalous problems while they are happening, rather than discovering them buried in a giant spreadsheet days or weeks after the testing cycle has concluded. Finally, the company calls for transparent communication of uncertainty and the inherent limitations built into any evaluation process, acknowledging that no single test can ever provide absolute guarantees.

What This Means for the AI Industry

The publication of these guidelines arrives as Anthropic continues to push for greater transparency surrounding its internal safety and testing methods. According to the company’s broader research summaries, this recent push builds upon prior transparency efforts, including safety reports published concerning unintended model actions and ongoing updates regarding model containment practices.

While Anthropic frames this advice as broadly useful across the entire artificial intelligence industry, implementing these rigorous practices takes significant time, specialized expertise, and financial investment. Ultimately, nothing in the newly published engineering blog post compels competing AI laboratories to follow along, leaving the adoption of these scientific standards up to the voluntary commitment of individual companies. As the capabilities of artificial intelligence models continue to accelerate, whether the rest of the industry will match this level of methodological self-critique remains one of the defining questions for the future of safe AI development.

What is your reaction to this story?

Ali Ikhwan

Writer and contributor at Bitcoin 98.

Comments (0)