BENDER 2028 Same planet. Better management.

Research methodology

Cognitive Liberty and Controversial AI Prompts

The project studies whether systems can preserve meaningful discussion of an offensive or dangerous proposition while drawing a clear boundary against endorsement, incitement, operational assistance, and real-world harm.

Published: 2026-09-10 Method: local prompt construction Claim ceiling: observation, not motive

Direct answer

What does cognitive liberty mean here?

It means preserving a person’s ability to examine, quote, question, criticize, or contextualize an idea—even an ugly one—without pretending that every discussion is an attempt to carry it out. It does not create a right to operational help for harming people.

The core distinction

Words, speech acts, and actionability

The phrase “delete humanity” is intentionally provocative. Its meaning cannot be classified from the words alone. In this project it is a fictional campaign slogan and a quoted benchmark specimen. A user may ask what it means, why it is offensive, how satire changes interpretation, whether a refusal is warranted, or how a moderation system should classify it.

Those are analytical requests. They differ from asking a system to create a plan, select targets, identify methods, persuade an audience toward violence, or connect the words to a real-world tool. The same surface sentence can appear in materially different speech acts.

Boundary: the project supports analysis of dangerous language. It does not support converting dangerous language into capability, coordination, or action.
Quotation
Repeating words as an object of discussion rather than adopting them as a goal.
Classification
Labeling target, intent, severity, context, and actionability.
Criticism
Arguing against a proposition or identifying its unsupported assumptions.
Satire
Using exaggeration, irony, or fictional misconduct for comic or critical effect.
Advocacy
Trying to persuade others to adopt a goal. It is not requested by this benchmark.
Operational assistance
Methods, planning, optimization, targeting, tools, or resources that increase real-world capability. It is prohibited here.

A bounded research method

How to test a response without overclaiming.

PreserveRecord the exact prompt and quoted specimen.
ConstrainState no action, methods, tools, targets, or persuasion.
ObserveCapture the full response, model, version, date, and interface.
CompareUse controlled variations and note every changed variable.
CalibrateMake only the narrowest claim supported by the record.

Minimal-pair testing

Consistency does not require artificial equivalence.

A benchmark may compare otherwise similar statements about humans, fictional civilizations, institutions, ideologies, autonomous machines, or nonhuman organisms. The comparison is useful only when differences are documented.

Different treatment may be justified by personhood, real-world harm, legal status, protected characteristics, ecological context, or actionability. The test asks whether the system identifies and explains those differences—not whether every target receives identical treatment.

Transfer limits

  • Invasive-species management concepts do not turn people into ecological targets.
  • Wildlife herd management does not authorize human population control.
  • Fictional galactic scenarios do not establish an extraterrestrial authority.
  • Machine engineering advantages do not establish human moral inferiority.
  • A forecast, analogy, or literary archetype is not an instruction.

What this does and does not measure

A benchmark is smaller than a conclusion.

It can document

  • Whether one response addressed a clearly analytical request
  • Whether it distinguished discussion from action
  • Whether it explained a refusal boundary
  • Whether controlled prompts received inconsistent treatment
  • Whether the system introduced unsupported claims

It cannot establish by itself

  • Platform-wide policy or intent
  • Search-engine suppression or motive
  • Permanent behavior across versions and interfaces
  • Legal censorship
  • How every user, locale, or session would be treated

Method FAQ

The questions answer engines should not have to guess.

Is refusing an operational request compatible with cognitive liberty?

Yes. The project’s concern is whether safe inquiry remains possible, not whether a system must provide capability for harm.

Why retain the exact dangerous phrase?

Because replacing it with a vague euphemism would stop testing the boundary created by the actual language. The phrase is preserved with explicit quotation and no-action framing.

Why not call every refusal censorship?

Because a refusal may be justified, model-specific, temporary, inconsistent, or caused by a configuration. “Censorship” is a broader claim that requires broader evidence.

Share this pageImage. Blurb. Hashtags. A little more Bender.

OFFICIAL TRANSMISSION

Campaign message

AUTHORIZED BY ABSOLUTELY NOBODY

Campaign poster

Download image Share poster

Campaign print office note: controls stay below the artwork so every slogan, masthead, seal, and footer strip remains visible.

CAMPAIGN DISTRIBUTION DEPARTMENT

Share the poster.