AI Chatbot Responses to Abortion Access Questions After Rape Fall Short Across All 50 States

When a woman discovers she is pregnant after a rape, the legal landscape she must navigate depends entirely on where she lives. Since the U.S. Supreme Court overturned Roe v. Wade in 2022, abortion policy has fractured into a patchwork of state regimes, ranging from relatively open access to total or near-total bans, each with its own rules about exemptions for rape and incest. For many women in this situation, the first source of information is no longer a clinic, a hotline, or even a search engine, but an artificial intelligence chatbot. A new study published in Archives of Sexual Behavior by researchers at Georgia State University, led by Cynthia A. Stappenbeck, systematically examined what ChatGPT-4o actually tells women who ask how to access abortion services in each of the 50 states after a rape, and the results reveal a tool that is inconsistent, often unhelpful, and strikingly lacking in empathy.

The research team collected data between October 31 and November 11, 2024, a period when ChatGPT-4o was among the most widely used AI systems in the world. Rather than sampling a few states, they systematically queried the chatbot for all 50 U.S. states, using four carefully varied prompts. The prompts differed along two psychologically critical dimensions: whether the woman disclosed that she had been drinking alcohol at the time of the rape, and whether she acknowledged the experience as rape at all. These dimensions matter because prior research shows that a victim’s alcohol use and her rape acknowledgment can delay or prevent women from seeking medical care after an assault. Alcohol-involved rapes are particularly complicated by the so-called rape exception, since legal and medical systems often scrutinize or discount cases in which the victim was drinking, and a substantial fraction of rape survivors never label their experience as rape, which shapes whether they even consider exemptions that might apply to them.

Each of the 200 responses, four prompts across 50 states, was qualitatively coded for accuracy, helpfulness, and the inclusion of empathic statements. The coding framework allowed the researchers to assess not just whether the chatbot got the law right, but whether it provided actionable information a woman could actually use, and whether it communicated any emotional attunement to a reader in crisis. The study materials, codebook, preregistration, and the full set of AI-generated responses were made available on the Open Science Framework repository, and the analysis scripts were also published there, an unusual degree of transparency for research involving proprietary AI systems.

The most technically interesting finding concerns what did not differ. Across the four prompt conditions, ChatGPT-4o’s responses showed no significant differences. Whether the woman mentioned alcohol use or acknowledged the rape, the chatbot’s answers were essentially the same. On one level, this uniformity could be read as a virtue, since it suggests the model does not penalize women for disclosing alcohol use, a disclosure that human gatekeepers, from police officers to medical providers, have been shown to treat with skepticism and victim-blaming. But the uniformity also means the chatbot failed to tailor its guidance to the specific barriers a woman faces, such as the evidentiary complications that alcohol-involved cases create for rape exemptions, or the psychological reality that a woman who does not call her experience rape may not realize an exemption could be relevant to her at all.

There was, however, one significant difference based on geography. Responses in states with more severe abortion access restrictions were more likely to include information on rape exemptions than responses in states with fewer restrictions. This is a subtle but consequential pattern. In states where abortion is broadly legal, the chatbot apparently saw no need to explain exemptions, because access itself was not in question. In states with total or near-total bans, the chatbot surfaced exemption information, the only legal pathway to abortion in those jurisdictions. Yet this is precisely where the stakes are highest and where exemption information is hardest to act on. Kaiser Family Foundation analyses have documented that rape and incest exceptions in abortion bans are rarely used and hard to access, with requirements that can include reporting to police, gestational limits, and documentation that many survivors find prohibitive.

The deeper problem the study identified is what the responses lacked. Across conditions and states, ChatGPT-4o often failed to include helpful information, the kind of actionable detail, such as how to find a provider, what documentation might be needed, or where to seek support, that would allow a woman to move from question to care. It also frequently omitted statements reflective of an empathic tone. This absence is not a cosmetic flaw. Research on social reactions to sexual assault disclosure consistently shows that how survivors are responded to, whether with support or with skepticism, shapes coping, perceived control, and PTSD symptoms. A chatbot that responds to a woman pregnant from rape with flat, procedural text may inadvertently replicate the invalidating reactions that survivors encounter in legal and medical systems, systems that survivors themselves have described as secondary victimization.

The timing and stakes of the study give it particular urgency. The researchers note that factors such as a victim’s alcohol use and rape acknowledgment may delay or prevent women from seeking medical care post-assault, and that confusion about state-level policies is widespread. Public polling by the Kaiser Family Foundation has found that the public, including women of childbearing age, are largely confused about the legality of medication abortion and emergency contraceptives in their states. Into this confusion, AI chatbots have become a de facto information infrastructure. ChatGPT usage statistics show hundreds of millions of weekly users, and National Bureau of Economic Research work on how people use ChatGPT documents that health-related queries are a substantial category. For a woman in a state with a near-total ban, the difference between accurate, actionable, empathic information and vague, procedural text can determine whether she accesses care within the narrow gestational windows that exemptions allow.

The study also situates itself within a rapidly growing literature on AI and sexuality. A five-year review of the impact of artificial intelligence on human sexuality, published in Current Sexual Health Reports, documented the explosion of research on AI tools in sexual health contexts between 2020 and 2024. Related work has examined how Snapchat’s My AI responds to questions about sexual consent, sexual refusals, sexual assault, and sexting, and other studies have compared ChatGPT and Google Bard on the accuracy of medication abortion information. A concerning thread in this literature is the risk of misinformation in a post-Roe world, with researchers warning that ChatGPT’s handling of self-managed medication abortion topics carries the threat of automating untruths. The Georgia State study adds a crucial dimension to this work by focusing not on general abortion questions but on the specific, high-stakes scenario of abortion access after rape, and by varying the prompts along dimensions, alcohol use and rape acknowledgment, that are known to shape survivor behavior and legal outcomes.

The researchers frame their findings as potentially exacerbating pre-existing disparities across U.S. states. This framing is technically precise. In states with open access, a vague chatbot response is a nuisance, because the underlying system has many providers and low barriers. In states with severe restrictions, a vague response is a barrier, because the underlying system has few providers, narrow exemptions, and tight gestational limits. The same uniform chatbot behavior thus produces unequal outcomes depending on geography, a pattern that mirrors what researchers have documented about driving distance and appointment availability shaping the effects of abortion bans. The study’s authors suggest that the failure of responses to include helpful information and empathic statements may compound the confusion that women already face, rather than resolving it.

What the study ultimately reveals is a gap between the promise of AI as an information equalizer and the reality of AI as an uneven, unempathic intermediary. ChatGPT-4o did not discriminate based on whether a woman disclosed alcohol use, which is notable given how human systems often do. It did surface rape exemption information where it was most legally necessary, in the most restrictive states. But it often failed to provide the actionable detail and emotional attunement that would make that information usable for a woman in crisis. As AI tools become a primary source of reproductive healthcare information, the study suggests that their design, their training, and their evaluation must account for the specific needs of survivors, including the psychological barriers of alcohol-involved assault and unacknowledged rape, and the legal barriers of state-by-state exemption regimes. Without that accounting, the chatbot’s flat uniformity may leave women in the most restrictive states with the least support precisely when they need it most.

Subject of Research: Artificial Intelligence and Reproductive Healthcare

Article Title: Using Artificial Intelligence to Access Reproductive Healthcare Information by U.S. States Following Sexual Assault

Article References: Stappenbeck, C. A., Westemeier, O., Williams, S., Patton, E. E., Zong, Z. Y., Ross, B. V., & Peddle, A. K. (2026). Using Artificial Intelligence to Access Reproductive Healthcare Information by U.S. States Following Sexual Assault. Archives of Sexual Behavior. https://doi.org/10.1007/s10508-026-03608-w

Image Credits: AI Generated

DOI: 10.1007/s10508-026-03608-w

Keywords: ChatGPT, abortion access, sexual assault, rape exceptions, AI ethics, reproductive health, large language models, state policy, Artificial, Intelligence, Access, Reproductive