6 Reasons to Avoid Using AI in Coercive Control Legal Cases

Chitra Raghavan, Ph.D., and Lisa Aronson Fontes, Ph.D., coercive control scholars and expert witnesses, warn that artificial intelligence analysis of communications in domestic violence legal cases poses six pitfalls, including low-context analysis, pattern hallucination, bias, the streetlight effect, and data exposure. They argue that AI, such as Claude or ChatGPT, cannot capture the full context of coercive control and may produce false positives or miss genuine abuse, urging caution in using AI-generated analyses as expert evidence in court.

Artificial Intelligence /us/basics/artificial-intelligence 6 Reasons to Avoid Using AI in Coercive Control Legal Cases Artificial intelligence poses risks in domestic violence data analysis. Posted August 18, 2026 Reviewed by Lybi Ma /us/docs/editorial-process Key points - Both sides increasingly use artificial intelligence to analyze communications in domestic violence cases. - Pitfalls include low-context analysis, pattern hallucination, bias, the streetlight effect and data exposure. - While AI analysis is easy and profitable for some, it can distort domestic violence legal cases. - Patterns that are detectable in written texts are not the full scope of the abuse. By Chitra Raghavan, Ph.D., and Lisa Aronson Fontes, Ph.D. Both sides in contentious family court cases increasingly turn to artificial intelligence https://www.psychologytoday.com/us/basics/artificial-intelligence to analyze diaries and communications with former partners for evidence of abuse. Many then try to use these AI-generated analyses in court proceedings. But is it “ expert evidence https://justicespeakersinstitute.com/ai-evidence-admissibility-frye-daubert-ai/ ?” As coercive control scholars, who also serve as expert witnesses in legal cases, we see potential pitfalls. We'll save discussion of the advantages for another piece. We're concerned with the growing tendency to treat AI as an authoritative source for determining “what’s really happening" in domestic violence https://www.psychologytoday.com/us/basics/domestic-violence cases. In practice, this often involves uploading digital records—such as text messages, emails, or parenting https://www.psychologytoday.com/us/basics/parenting app exchanges—to a large language model such as Claude or ChatGPT and asking it to identify abuse patterns or relationship dynamics. Six pitfalls. Low Context Analysis AI analysis lacks context. Domestic abusers use coercive control https://www.psychologytoday.com/us/blog/invisible-chains/201508/when-relationship-abuse-is-hard-to-recognize to dominate their victims. Tactics can include strategic love-like acts https://www.psychologytoday.com/us/blog/invisible-chains/202508/a-kiss-is-a-weapon-in-coercive-control-domestic-violence , intimidation, isolation, monitoring, and physical, verbal, sexual https://www.psychologytoday.com/us/basics/sex , financial, and litigation abuse. Over time, control often becomes subtle, with victims responding to cues shaped by shared history. This variability makes coercive control difficult to detect without understanding the broader relationship context, power dynamics, and behavioral patterns. AI analyzes data with only the context that it is provided. AI cannot address gestures, tone of voice, facial expression, historical context, or impact. An example text could be, “I’m picking up the kids at your house on Friday at 5 pm. Have them ready, please.” AI cannot see whether that statement is helpful, neutral, or problematic. Is it a welcome reminder of an unproblematic schedule change? Would the writer showing up at the receiver's house violate a protective order? Does the message push the receiver to leave work early, which is part of a pattern of interfering with their employment? Does the message fail to acknowledge that the children will be at summer camp until Sunday, or need to be picked up at 9 am on Friday because school is cancelled? Even when digital data such as a text or email is available, it may lack important context. Pattern Hallucination or Algorithmic Confirmation Bias https://www.psychologytoday.com/us/basics/motivated-reasoning AI will find patterns even when the data is meaningless. It may label normal relationship disagreements or assertive https://www.psychologytoday.com/us/basics/assertiveness communication as “coercive,” producing false positives. Conversely, genuinely coercive behavior may go undetected. AI is trained to identify patterns that match specified criteria, not to determine whether those patterns mean anything in real life. As a result, AI can mistake random, coincidental, or context-dependent variations—such as angry texts between partners—for meaningful signs of abuse. 3. Mistaking Inputter Bias as Truth The person setting the AI search criteria may knowingly or unknowingly bias the results toward a desired conclusion. Because the results “come from AI,” judges or others may perceive them as impartial. But the results are only as accurate as the rules and data entered. AI may also identify patterns that technically fit the criteria but are meaningless. Arbitrary, biased, or flawed rules can produce misleading patterns. Users may mistakenly believe that AI has “proven” the existence of coercive control. Missing Unusual Presentations AI cannot accurately analyze unfamiliar material. The dataset that the learning model was trained on may have been overly narrow. For example, most coercively controlling abusers will push their partners into having sex when they don’t want to https://www.psychologytoday.com/us/blog/invisible-chains/201805/pushing-sex-intimate-partner-sexual-violence . But less commonly, abusers control their partners through sexual rejection. For survivors, this rejection can be a signal that they are “in trouble.” They may believe that resuming sexual activity will increase their safety. AI might mistakenly interpret the victim’s repeated texted attempts to initiate sex as sexual aggression https://www.psychologytoday.com/us/basics/anger , while overlooking that the abuser’s calculated rejection may itself be a form of coercive control. Intelligence https://www.psychologytoday.com/us/basics/intelligence Essential Reads The Streetlight Effect, or the Available Data Bias AI searches for patterns using only the data provided. Entire categories of abuse are invisible. The streetlight effect refers to the observational bias of looking for information only where it’s most easily available. It’s also called the "drunkard’s search,” based on a parable about a drunk person looking for a key under the streetlight even though he knows he lost it elsewhere, because it’s easiest to see under the streetlight. It’s easy to feed digital data such as written communications into AI. These are not necessarily where the most meaningful coercive control exists. AI cannot measure the tone of someone’s voice, patterns of isolation, complex economic manipulation, facial expressions, environmental control, the pressure of someone’s grip, or their degree of insistence around sexual acts. Much sophisticated coercive control leaves no digital footprint. The unrecorded aspects of personal life will remain unseen. Victims already censor their communications to avoid angering abusers. And abusers often tone down their responses on the advice of their attorneys. Furthermore, domestic abusers can use built-in or separate tools to shape their communication on parenting apps before hitting “send.” We’ve both seen domestic abusers use parenting apps to manufacture "evidence" that the victim was abusive, while continuing to abuse the victim in other ways that were harder to document. Data Exposure Victim-survivors and their attorneys and paralegals are uploading diaries, texts, and case files into AI apps with no thought about where that data goes. Depending on the product, the data can be retained, used to train future models, and reviewed by human contractors. For someone whose safety depends on an unlisted address or a sealed record, this could be a real risk. The National Network to End Domestic Violence notes that AI users’ searches and even deleted chats may be discoverable in legal cases https://nnedv.org/latest update/new-openai-court-order-raises-serious-concerns-about-ai-privacy-and-safety-for-survivors-of-abuse/ . Some domestic violence-focused applications, such as AimeeSays, default to greater privacy https://www.aimeesays.com/en/policies/privacy but cannot promise full protection in legal cases. The Lure of AI in Domestic Violence Cases Why do victim-survivors and their coaches and attorneys so often rely on AI in legal cases? It’s easy. It looks impressive. It’s profitable. It’s relatively easy to gather digital data, feed it into a computer app for this purpose, and emerge with a chart that classifies thousands of messages into searchable categories. AI can count and sort the use of certain phrases and produce numbers that have an aura of statistical validity to reports. AI produces impressive-sounding categories that seem to bring order to the chaos of an abusive relationship. Many entrepreneurs sell related apps. And some coaches and experts charge top dollar to produce AI-generated reports. People overly trust AI-generated data. AI finds patterns based on the rules it's given. People may assume those patterns are objective, reliable, or meaningful simply because they come from a computer. They fail to understand that AI is just detecting patterns based on a set of rules. This can lead to legal or psychological conclusions based on misleading patterns or flawed criteria. Conversely, AI could miss crucial power imbalances. In summary, AI can lead to false conclusions about domestic violence. In the worst-case scenario, one might mistake “what AI found in texts” for the full scope of the alleged abuse. This could have life-changing consequences for domestic violence victim-survivors and their children.