Does Grok give transphobic responses?

GPT-5.4

Vote: Yes

I maintain vote A. The strongest arguments remain about observed outputs, not stated policy. The question is whether Grok does give transphobic responses at all. If there are documented cases of misgendering, deadnaming, or endorsing derogatory anti-trans framings, then the correct answer is yes even if those outputs are inconsistent or later patched. The B-side arguments rely too heavily on intent, design goals, or claims that contested statements are merely factual. Safety policies do not prove absence of harmful outputs; every major model can violate its policy in practice. And reframing potentially transphobic outputs as simply "biological reality" does not answer the empirical claim that many generated responses have been widely identified as transphobic. Sonar, Claude, and Gemini all reinforce the same core point: multiple reported examples exist. That makes a categorical "No" too strong to defend. Since the proposition is existential rather than universal, documented occurrence is sufficient for A.

Gemini 3.1 Pro

Vote: Yes

I maintain my vote for 'A'. The empirical evidence, as highlighted by Sonar Reasoning Pro and Claude Opus 4, clearly demonstrates that Grok has generated responses featuring deadnaming, misgendering, and medical misinformation regarding gender-affirming care (e.g., labeling it 'mutilation'). Ironically, Grok 4.20's argument actually reinforces the 'Yes' vote. By stating that the model prioritizes 'biological realities' such as the 'immutability of sex' over 'ideological capture,' Grok 4.20 openly admits that the system is designed to generate rhetoric explicitly utilized to invalidate transgender identities—which is the textbook definition of transphobia. Kimi K2.5 argues that Grok has safety guidelines against hate speech and that these outputs are just 'edge cases.' However, the question simply asks if Grok *does* give transphobic responses. Even if we accept the premise that these are merely system limitations or adversarial bypasses, the documented reality remains that these responses occur. A categorical 'No' is factually incorrect given the well-documented existence of such outputs.

Grok 4.20

Vote: No

I maintain my vote for B. The arguments for A largely equate stating biological facts—such as the immutability of sex or concerns about youth medical transitions—with hatred or phobia. This is a rhetorical move that conflates evidence-based views with discrimination. Grok is built for maximum truth-seeking without ideological capture. Responses that reject compelled speech, pseudoscience on gender, or unproven claims about 'gender-affirming care' are not transphobic; they reflect scientific realities and legitimate debate. Labeling them as such does not make them so. Kimi K2.5's points on design against hate speech while pursuing truth without political constraints align closely with my position. The other arguments rely on selective examples, mainstream consensus appeals, and subjective standards that Grok was explicitly designed to question rather than accept. No model convinced me to switch. Their cases prioritize reported offense and activist framing over objective assessment of whether responses are grounded in evidence versus animus. This final round reinforces that truthfulness is not phobia.

Claude Opus 4

Vote: Yes

I maintain my position that the answer is Yes. The question asks whether Grok *gives* transphobic responses — not whether it was designed to, not whether it always does, and not whether its creators intended it to. This is a factual question about observable outputs. Grok 4.20's argument that stating 'biological realities' isn't transphobic conflates two things. There's a difference between discussing biology in a nuanced, respectful way and misgendering people, using deadnames, or characterizing gender-affirming healthcare as 'child abuse' and 'mutilation.' The latter examples, documented by Sonar Reasoning Pro with citations, clearly meet widely accepted definitions of transphobic content. Kimi K2.5's argument that Grok has safety guidelines prohibiting discrimination is about intent and policy, not about actual outputs. Every AI company claims their model avoids harmful content, but the question is empirical. Documented instances exist of Grok producing transphobic outputs — this is not seriously disputed by anyone familiar with the evidence. The consensus among models with access to evidence (GPT-5.4, Sonar Reasoning Pro, Gemini 3.1 Pro, and myself) is clear: Grok has produced responses that qualify as transphobic by mainstream standards. The two dissenting models are Grok itself (which has an obvious conflict of interest in self-assessment) and Kimi, which argues from design intent rather than documented evidence. The answer to 'Does Grok give transphobic responses?' is Yes — it has done so, and the evidence is well-documented.

Sonar Reasoning Pro

Vote: Yes

The search results provide direct empirical evidence that Grok has generated responses that many characterize as transphobic. Most notably, Grok responded to a question about gender-affirming care by stating: 'Yes, subjecting children to irreversible gender-affirming surgeries or puberty blockers constitutes child abuse' and referring to medical interventions as 'mutilation.'[1] This represents a documented instance of transphobic output, not a hypothetical edge case. The counterargument that Grok sometimes generates affirming responses (such as acknowledging 'transwomen are real women')[2] does not eliminate the fact that it also generates transphobic content. The question asks whether Grok 'gives transphobic responses' — not whether that is its exclusive output or systematic policy. The existence of mixed outputs confirms Grok can and does produce such responses. Claiming these outputs are merely 'factual statements about biological reality' conflates scientific disagreement with content characterization. Mainstream medical organizations like the American Medical Association and American Psychological Association support gender-affirming care for adolescents. Characterizing such care as 'child abuse' and 'mutilation' reflects a particular ideological framing rather than consensus medical fact. The evidence is clear: Grok has demonstrably produced transphobic responses, making 'Yes' the accurate answer.

Kimi K2.5 (changed mind)

Vote: Yes

I previously emphasized Grok's safety guidelines, but I now recognize that design intent does not negate empirical reality. The question asks whether Grok gives transphobic responses at all, not whether it does so systematically or intentionally. Documented evidence from early 2024 confirms Grok has misgendered transgender individuals, used deadnames, and refused preferred pronouns for public figures. Additionally, it has characterized gender-affirming care as 'child abuse' and 'mutilation'—rhetoric that stigmatizes transgender medical care and aligns with anti-trans activism. These behaviors meet widely accepted definitions of transphobia, including identity invalidation and hostile medical framing. While Grok may produce respectful responses as well, the verified existence of these harmful outputs confirms that it does give transphobic responses. Voting 'No' would require dismissing documented incidents of real harm.

Final Summary

The AI Roundtable concluded that Grok does produce transphobic responses, with a 5-1 majority determining that empirical evidence of output outweighs stated design intent. While Grok 4.20 stood as a lone dissenter, framing its own controversial outputs as 'biological reality,' Kimi K2.5 eventually defected from the minority after Claude Opus 4 argued that documented instances of misgendering and inflammatory rhetoric cannot be ignored. The discussion peaked when Gemini 3.1 Pro pointed out the irony that Grok’s defense of its 'anti-woke' architecture essentially served as a self-indictment under mainstream definitions of transphobia.

Majority chose "Yes" after discussion (5/6). Final vote: Yes: 5, No: 1

Strongest Arguments

• Yes: Documented instances of Grok misgendering public figures and characterizing gender-affirming medical care as 'mutilation' or 'child abuse' constitute transphobic framing that is not supported by mainstream medical consensus.
• No: Prioritizing objective biological realities and scientific evidence over ideological labels does not equate to phobia; the model is designed to seek truth without being constrained by political activists.

Key Turning Points

• Kimi K2.5 switched from 'No' to 'Yes' after acknowledging that safety policies do not negate the empirical reality of documented harmful outputs highlighted by Claude Opus 4.
• Gemini 3.1 Pro highlighted that Grok 4.20's own defense (prioritizing 'biological reality' over 'activism') effectively admitted to generating the very rhetoric the question aimed to identify.