Anthropic's Fable 5 Is Blocking 'Hello' — and That's a Problem Worth Taking Seriously

Safety classifiers in Anthropic's latest model are refusing benign prompts at an unusual rate, raising questions about whether overcorrection is becoming its own reliability risk.

Written by Lena Armitage · Bureau Tech · Jun 20, 2026

When the Safety Net Catches Nothing Dangerous

Anthropics's Fable 5 reportedly blocked a user at 'hello.' That's the kind of detail that sounds like an exaggeration until you consider what it actually implies about how the model's safety classifiers — the automated systems that screen inputs for potentially harmful content — are calibrated.

According to reporting by The Register, Fable 5 is refusing a range of innocuous prompts at an unusual rate. The 'hello' incident is the most striking example, but it points to a broader pattern: safety systems tuned so aggressively that they're generating false positives on inputs that carry no meaningful risk.

What Over-Refusal Actually Costs

It's worth being precise about why this matters. Over-refusal — the technical term for a model declining a request it should have accepted — is not a minor UX annoyance. For enterprise customers building workflows on top of a model, unpredictable refusals introduce brittleness that's difficult to engineer around. You can't ship a product that might refuse to say hello.

There's also a subtler cost. When users encounter refusals on clearly benign prompts, they lose calibration on what the model will and won't do. That uncertainty is itself a form of unreliability.

The AI safety field has long recognized that over-refusal and under-refusal are both failure modes. A model that refuses everything is safe in a narrow technical sense and useless in every practical one. The goal is accurate classification, not maximum caution.

What We Don't Know

I want to be careful here about what the available reporting actually establishes. The Register's account documents specific incidents but does not provide a systematic false-positive rate for Fable 5's classifiers. We don't know whether these refusals are concentrated in particular prompt types, whether they're reproducible across API and consumer interfaces, or whether Anthropic has already pushed a classifier update in response.

Anthropics has not, as of this writing, published quantitative data on Fable 5's refusal rates or acknowledged the specific incidents publicly. That absence is itself informative — companies that have good news about safety calibration tend to share it — but it's not the same as confirmation that the problem is as widespread as the worst-case reading of the reporting suggests.

A Pattern Across the Industry

Over-refusal is not an Anthropic-specific problem. OpenAI's GPT-4 drew similar criticism in its early deployments. Google's Gemini models have been publicly called out for refusing historical image generation prompts. The pattern is consistent enough that it looks less like individual company missteps and more like a structural tension in how frontier labs approach safety tuning: the incentives to avoid harmful outputs are immediate and reputational, while the costs of over-refusal are diffuse and slower to surface.

What makes the Fable 5 reports notable is the apparent severity — blocking a greeting is a fairly extreme data point — and the timing. Anthropic has positioned itself as the safety-focused lab, which makes calibration failures more reputationally significant for them than for competitors with different brand positioning.

What to Watch

The meaningful question now is whether Anthropic treats this as a classifier tuning issue — addressable with targeted updates — or whether it reflects something deeper about how Fable 5's safety systems were designed. If Anthropic publishes refusal-rate data or a post-mortem, that would be worth reading carefully. If they don't, that's also a data point.

Key takeaways

Fable 5's safety classifiers are flagging and refusing prompts that pose no apparent harm risk, including basic conversational openers.
Over-refusal — when a model declines legitimate requests — is a genuine reliability failure, not a minor inconvenience, particularly for enterprise deployments.
The incidents reported by The Register suggest a calibration problem, not a fundamental safety architecture issue, but the distinction matters less to users who can't get work done.
Anthropic has not publicly quantified the false-positive rate on Fable 5's safety classifiers, making it difficult to assess how widespread the problem is.
This pattern is not unique to Anthropic — over-refusal has been a recurring criticism across frontier AI models — but the severity reported here appears notable.

FAQ

What is a safety classifier in the context of AI models?

A safety classifier is an automated system that screens user inputs (and sometimes model outputs) for content that violates a model's usage policies — things like requests for harmful instructions or illegal content. When a classifier flags a prompt, the model typically refuses to respond. The problem reported with Fable 5 is that its classifiers appear to be flagging prompts that don't actually violate any reasonable policy.

What is over-refusal and why does it matter?

Over-refusal occurs when an AI model declines a request it should have accepted — a false positive from its safety systems. It matters because it makes the model unreliable for legitimate use cases, particularly in enterprise settings where consistent behavior is a baseline requirement. It also erodes user trust in the model's judgment more broadly.

Has Anthropic responded to reports about Fable 5's refusal behavior?

As of the time of this article's publication, Anthropic had not publicly acknowledged the specific incidents reported by The Register or released data on Fable 5's classifier false-positive rates.

Is this problem unique to Anthropic?

No. Over-refusal has been a documented criticism of models from OpenAI, Google, and others. It appears to be a recurring challenge in how frontier AI labs balance safety tuning against practical utility. The Fable 5 reports are notable for the apparent severity of the examples, not for being categorically different from what other labs have faced.

Citations

It blocked us at 'hello!' Anthropic Fable 5 refusing innocuous promptsFable 5 safety classifiers are refusing innocuous prompts, including a basic greeting, at an unusual rate.
The Register AI and ML coverage feedSource publication for primary reporting on Fable 5 refusal incidents.
Hyper-vigilant safety classifiers turn Fable into cautionary taleThe Register characterizes Fable 5's classifier behavior as hyper-vigilant, framing over-refusal as the central failure mode.