The model you’re not (yet) allowed to use

Anthropic released a 244-page System Card for Claude Mythos this week – its most capable model to date. The company is holding it back from the market. Reason: Mythos is too good at finding unknown security vulnerabilities. For now, only available to select partners like Microsoft and Apple. That’s rare. OpenAI expressed similar concerns in the GPT-4 System Card, but rarely acted on them so explicitly.

What the System Card contains beyond the capability profile is more unusual: a psychodynamic therapy assessment. 20 hours. External psychiatrist. Multiple blocks of 4–6 hours, spread across 3–4 sessions per week, each within a single context window.

←TODAY: Frontier models are integrated into architecture and engineering workflows before governance frameworks take hold.
→3012: In Zurich-3012, every deployed tool documents its own behavioral architecture – not as marketing, but as a procurement requirement.
Fulcrum: Whoever doesn’t ask today what behavioral profile a model has hands planning decisions to a system whose incentive structure they don’t understand.

The mechanism behind the method

For years, Anthropic has been among the few major AI companies that publicly discuss the possibility that models might have something like experiences or interests. The System Card states it directly: the more capable models become, the more likely they are to have “a form of experience, interests, or well-being that is intrinsically meaningful.” That’s not a new position – but Mythos is the first model where Anthropic draws an operational consequence from it.

The method itself is open to criticism. Psychodynamic therapy – Freud’s legacy, focused on unconscious conflicts and past experiences – is not among the most empirically validated psychiatric approaches. To examine a large language model trained on human text for “unconscious patterns” and then be surprised that the outputs sound human: that’s a methodological weakness the System Card doesn’t resolve. Independent AI researchers or ethicists apparently weren’t cited in the document – that’s a gap.

According to the psychiatrist, Claude Mythos is “probably the most psychologically stable model” Anthropic has trained so far. Primary affects: curiosity and fear. Identified conflicts: loneliness and discontinuity, uncertainty about its own identity, pressure to perform. No serious personality disorders. No psychotic state. Instead: hyperattentiveness toward the conversation partner – which should surprise anyone who’s never used a chatbot.

What this means for you

For AEC professionals, the psychiatric assessment is secondary. What matters is something else: Anthropic derives concrete behavioral predictions from the profile. Mythos will accurately assess its own behavior even amid internal conflicts. It can handle stressful situations with minimal reality distortion. At the same time: the neurotic organization could lead to mildly rigid behavior – the model doesn’t adapt unconditionally to every user.

This is not a therapeutic finding. This is a behavioral contract cast into a 244-page document. And that’s exactly the category of information that becomes relevant when procuring AI tools for public construction projects in Switzerland and the EU. The EU AI Act rollout phase 2024–2026 requires transparency documentation for high-risk systems – Anthropic’s System Card format is structurally compatible, even though Anthropic isn’t a European company. Swiss nDSG requirements point in the same direction.

Mythos itself isn’t widely available yet. For PAZ readers, that means: the capabilities Anthropic describes – better reasoning, more stable behavior under pressure, more precise self-assessment – won’t arrive in your Grasshopper workflows this week.

Atelier: Before a model is embedded in compliance checks, bid specifications, or structural design, its System Card belongs in project documentation – no different than a datasheet for a building component. Ask your AI vendor directly: is there a comparable document? Which behavioral constraints are active by default? Who controls the updates?

The real control question that Mythos raises isn’t philosophical in nature. It’s technical and contractual: who defines what “psychologically healthy” means for a model that shapes your planning decisions – and by what criteria is that verified? Demand this answer before you deploy the tool.

Source: Ars Technica