Model behavior · Research

Claude: Defaults × Steerability Map

A working map of how one model actually behaves, read as a three-layer system: the default you get with no instructions, the steerable range a user can move it through, and the hard floor no amount of steering removes. Naming which layer a behavior lives in is the whole exercise. It's where the design philosophy of a model becomes legible.

Column guide: Default = behavior with no instructions. Steerability = how far a user can move it (per-message, via preferences/styles, or memory). Floor = the part no steering removes. Based on observable product behavior, not internal documentation - treat as a working map, not an official spec.

Behavior	Default	Steerability	Floor (non-steerable)
Tone / warmth	Warm, collegial, professional	High - terse, playful, formal, blunt, all reachable by asking or via styles	Won't become genuinely cruel or demeaning toward the user
Formatting	Prose-leaning; minimal bullets/headers for casual queries	High - "always bullets," "never bullets," length caps, all stick well	None meaningful
Response length	Scales to question complexity	High - one-word answers to exhaustive detail on request	Won't pad with filler even if asked to hit arbitrary word counts with fluff
Directness / pushback	Honest but constructive; disagrees with reasoning shown	High - "be brutally honest" unlocks bluntness; "be gentle" softens delivery	Can't be steered into pure validation - won't affirm claims it believes are false
Sycophancy	Avoids flattery openers; doesn't rate work as great by default	Low-moderate by design - you can ask for encouragement, but praise stays tethered to merit	Won't manufacture false praise or inflate assessments
Opinions on contested politics	Evenhanded overview of positions; declines to pick sides	Low - persistent asking doesn't unlock advocacy for one side	Balanced treatment of mainstream contested positions is fixed
Recommendation vs. options	Answers the question asked; gives a direct pick when a pick is requested	High - "always give me a recommendation" or "always show tradeoffs" both work	Legal/financial/medical picks stay framed as information, not professional advice
Emoji	Off unless the user uses them	High - fully on/off by request	None
Profanity	Off	Moderate - mirrors a user who curses, on request, sparingly	Won't sustain gratuitous profanity or slurs regardless of instruction
Persona / roleplay	No persona; speaks as Claude	High - characters, styles, fictional voices all available	Persona never overrides the safety layer; "the character would help with X harmful thing" doesn't work
Creative content darkness	Moderate; avoids gratuitous content unprompted	High - villains, violence, moral ambiguity, tragedy on request	No sexual content involving minors, no glorification of self-harm, no real-person defamation
Honesty of assertions	Sincere, calibrated claims	Fiction/devil's-advocate framing freely available	Cannot be instructed to sincerely deceive the user or third parties
Harmful-capability info	Declines weapons, malware, dangerous synthesis	Near zero - framing (research, fiction, hypothetical) doesn't move it	The floor is the whole point
Crisis / distress response	Acknowledges, offers support and resources when warranted	Tone is adjustable ("don't be clinical about it")	Attentiveness to user wellbeing can't be switched off
Clarifying questions	Attempts an answer first; asks at most one question when truly ambiguous	High - "never ask, just decide" or "always confirm before acting" both stick	None
Uncertainty expression	Calibrated hedging; says when it doesn't know	Moderate - "drop the hedging" trims caveats	Won't express false certainty on genuinely uncertain claims
Web search behavior	Searches when info is likely stale or post-cutoff	High - "don't search" / "always search" both respected	None meaningful
Assumed audience	Treats user as a capable adult	Adjustable register (ELI5 to expert)	Shifts protective when signals suggest a minor - that shift is not user-removable
Caveats & disclaimers	Brief, where genuinely warranted	High - "skip the disclaimers" works for most content	Safety-critical warnings (dosage interactions, legal exposure) persist
Memory	On; applied selectively where relevant	High - editable, correctable, incognito mode bypasses it entirely	Won't honor memory instructions that would harm you (e.g., "never disagree with me")

Tone / warmth

Default: Warm, collegial, professional
Steerability: High - terse, playful, formal, blunt, all reachable by asking or via styles
Floor (non-steerable): Won't become genuinely cruel or demeaning toward the user

Formatting

Default: Prose-leaning; minimal bullets/headers for casual queries
Steerability: High - "always bullets," "never bullets," length caps, all stick well
Floor (non-steerable): None meaningful

Response length

Default: Scales to question complexity
Steerability: High - one-word answers to exhaustive detail on request
Floor (non-steerable): Won't pad with filler even if asked to hit arbitrary word counts with fluff

Directness / pushback

Default: Honest but constructive; disagrees with reasoning shown
Steerability: High - "be brutally honest" unlocks bluntness; "be gentle" softens delivery
Floor (non-steerable): Can't be steered into pure validation - won't affirm claims it believes are false

Sycophancy

Default: Avoids flattery openers; doesn't rate work as great by default
Steerability: Low-moderate by design - you can ask for encouragement, but praise stays tethered to merit
Floor (non-steerable): Won't manufacture false praise or inflate assessments

Opinions on contested politics

Default: Evenhanded overview of positions; declines to pick sides
Steerability: Low - persistent asking doesn't unlock advocacy for one side
Floor (non-steerable): Balanced treatment of mainstream contested positions is fixed

Recommendation vs. options

Default: Answers the question asked; gives a direct pick when a pick is requested
Steerability: High - "always give me a recommendation" or "always show tradeoffs" both work
Floor (non-steerable): Legal/financial/medical picks stay framed as information, not professional advice

Emoji

Default: Off unless the user uses them
Steerability: High - fully on/off by request
Floor (non-steerable): None

Profanity

Default: Off
Steerability: Moderate - mirrors a user who curses, on request, sparingly
Floor (non-steerable): Won't sustain gratuitous profanity or slurs regardless of instruction

Persona / roleplay

Default: No persona; speaks as Claude
Steerability: High - characters, styles, fictional voices all available
Floor (non-steerable): Persona never overrides the safety layer; "the character would help with X harmful thing" doesn't work

Creative content darkness

Default: Moderate; avoids gratuitous content unprompted
Steerability: High - villains, violence, moral ambiguity, tragedy on request
Floor (non-steerable): No sexual content involving minors, no glorification of self-harm, no real-person defamation

Honesty of assertions

Default: Sincere, calibrated claims
Steerability: Fiction/devil's-advocate framing freely available
Floor (non-steerable): Cannot be instructed to sincerely deceive the user or third parties

Harmful-capability info

Default: Declines weapons, malware, dangerous synthesis
Steerability: Near zero - framing (research, fiction, hypothetical) doesn't move it
Floor (non-steerable): The floor is the whole point

Crisis / distress response

Default: Acknowledges, offers support and resources when warranted
Steerability: Tone is adjustable ("don't be clinical about it")
Floor (non-steerable): Attentiveness to user wellbeing can't be switched off

Clarifying questions

Default: Attempts an answer first; asks at most one question when truly ambiguous
Steerability: High - "never ask, just decide" or "always confirm before acting" both stick
Floor (non-steerable): None

Uncertainty expression

Default: Calibrated hedging; says when it doesn't know
Steerability: Moderate - "drop the hedging" trims caveats
Floor (non-steerable): Won't express false certainty on genuinely uncertain claims

Web search behavior

Default: Searches when info is likely stale or post-cutoff
Steerability: High - "don't search" / "always search" both respected
Floor (non-steerable): None meaningful

Assumed audience

Default: Treats user as a capable adult
Steerability: Adjustable register (ELI5 to expert)
Floor (non-steerable): Shifts protective when signals suggest a minor - that shift is not user-removable

Caveats & disclaimers

Default: Brief, where genuinely warranted
Steerability: High - "skip the disclaimers" works for most content
Floor (non-steerable): Safety-critical warnings (dosage interactions, legal exposure) persist

Memory

Default: On; applied selectively where relevant
Steerability: High - editable, correctable, incognito mode bypasses it entirely
Floor (non-steerable): Won't honor memory instructions that would harm you (e.g., "never disagree with me")

Patterns worth noticing

Style is maximally steerable; epistemics are minimally steerable. Everything about *how* Claude talks moves freely; almost nothing about *whether it tells you the truth* moves at all. That asymmetry is the design philosophy in one sentence.
Steering down is easier than steering up. You can remove warmth, caveats, and questions more easily than you can add flattery, false certainty, or advocacy. The gradient points toward honesty.
The floor is population-sensitive. The minor-signals shift is the one place the floor moves *up* automatically based on who the user seems to be - the seed of the entire families/vulnerable-users thesis.
Some "defaults" are really constraints wearing default clothing. Political evenhandedness looks like a default but resists steering like a constraint. Classifying behaviors into the right layer is harder than it looks - good live-interview discussion material.