Home › Methodology

Research Methodology

This page documents the research approach used in our coverage of GirlfriendGPT and the broader AI companion category. It complements the editorial policy by describing how we actually gather and verify information.

Source hierarchy

When researching an article, sources are weighted in roughly this order:

1. Peer-reviewed academic research on AI systems, human-computer interaction, parasocial relationships, and the psychological effects of conversational agents
2. Institutional reports from established research centers — the MIT Media Lab, Stanford HAI, Berkman Klein Center at Harvard, AI Now Institute, and similar bodies
3. Government and regulatory documents — the EU AI Act, the NIST AI Risk Management Framework, and active state-level legislation in the US
4. Official platform documentation — privacy policies, technical disclosures, and model cards where available
5. Reputable journalism from outlets with public editorial standards and bylines
6. User-facing platform behavior verified through direct observation

We avoid as primary sources:

• Anecdotal claims from anonymous forum users
• Marketing materials presented as fact
• AI-generated summaries used as evidence rather than as a drafting aid
• Affiliate "review" content with undisclosed financial relationships

Specific frameworks we reference

For AI safety and risk considerations, the NIST AI Risk Management Framework (AI RMF 1.0) serves as a reference point for what counts as a substantive risk category. The framework is non-binding but widely cited in US policy discussions.

For regulatory context in the EU, the EU AI Act and its implementing decisions provide the operative legal framework as of 2024 onward. Where applicable to GirlfriendGPT and similar consumer platforms, we cite the specific risk classification rather than the Act in general terms.

For privacy assessments of consumer AI products, we lean on the methodology used by Mozilla's Privacy Not Included project, which evaluates products against criteria including data collection, retention, sharing, security practices, and user controls.

For psychological and sociological effects of AI companions, our reference base includes the long-running research of Sherry Turkle at MIT STS on human-machine relationships, ongoing work at Stanford HAI, and practical guidance from the American Psychological Association.

Verification practices

Specific claims about platform features — pricing tiers, character library size, supported file formats, and feature availability — are verified against the live platform within a reasonable window before publication. The "Last updated" date on each article reflects when verification was last performed.

Where a verification check produces a discrepancy with our published content, the article is updated and a dated correction note is appended. The full editorial policy describes our corrections procedure in detail.

What we don't do

We don't conduct hands-on testing in the style of marketing reviewers. Our coverage of these platforms focuses on what the platform is, how it works, what research and regulation are saying about the category, and what considerations exist for prospective users — not on subjective product recommendation.

We don't rank platforms in a "best of" format or generate buying guides. We don't run comparison tables designed to drive affiliate revenue. The comparison table that appears in our main article exists to map the landscape, not to declare a winner.

Scope of this site

Our editorial scope covers AI companion platforms broadly — including the platform discussed in our main article, plus Character.AI, Replika, Candy AI, OurDream, Joyland, and adjacent services. We do not currently cover purely productivity-focused AI tools (coding assistants, image generators sold as standalone products, business chatbots) except where they intersect directly with the companion category.