Method
Parastance does not begin by assuming the critic is confused. It first identifies the strongest serious version of the claim, checks what the evidence supports, names the concession, and answers only the remaining disagreement.
What Parastance argues for
Governed, responsible uses of generative tools that expand human agency without treating artists, identities, workers, public trust, or the environment as unpriced inputs. A category-wide ban is usually overbroad; unrestricted deployment is indefensible.
What it refuses to defend
Opaque or unlawfully sourced training data, non-consensual identity simulation, intimate deepfakes, deceptive synthetic evidence, exploitative labour deployment, spam flooding, unsupported environmental claims, and corporate claims that cannot be verified.
How a response is assembled
- State the strongest credible version of the objection.
- Name what the critic is right about.
- Correct at most one central factual problem.
- Draw a narrow responsible-use distinction.
- Give one or two retrieved sources.
- End with a concrete safeguard, reform, or question.
What Parastance will not do
- Fabricate citations. Sources are stripped if they are not in the retrieved set.
- Answer on empty evidence. If nothing is retrieved, it says so.
- Debate a direct harm. The route changes to practical steps.
- Make first-person claims about an organization without a verified fact.
- Profile you, or steer you toward any ideological position.
How evidence retrieval works
Each claim is matched against a maintained packet of argument cards, FAQs, and a source ledger using hybrid keyword and similarity search with reciprocal-rank fusion. The agent may cite only source codes returned for that specific answer. Invented or unmatched citations are stripped and flagged.
Why it concedes real harms
Truth and source fidelity rank above persuasion. Losing an argument is preferable to making a false claim. Moral arguments are not treated as inferior to empirical ones.
Why direct personal harm changes the route
If you report a concrete harm — targeted imitation, a deepfake, lost work, harassment, a false AI-detection accusation — Parastance stops persuading. It acknowledges the harm and moves to practical documentation, reporting, and escalation steps. It is not a lawyer.
The truth gate
Parastance may only make first-person claims about a specific organization (“our model”, “our dataset”) when a verified fact supports it. Otherwise it says a responsible provider should publish verifiable evidence first.
What is stored, and limitations
Conversations are stored to your anonymous session unless you use private-session mode. The packet has a research cutoff of 2026-07-12; the agent will not present stale legal or product information as current. To report a correction, use “Report correction” on any answer.