Borrowed Memories: AI Between Trauma Therapy and False Memories

Science fiction imagines being able to treat trauma by "downloading" happy artificial memories. Clinical reality is very different and more complex. While Virtu

Science fiction has accustomed us to the idea of being able to manipulate the human mind like a hard disk, erasing traumas or downloading artificial memories to live experiences that never happened. Today, the increasingly widespread use of Virtual Reality (VR) and Generative Artificial Intelligence in the clinical field is rekindling these suggestions. However, it is imperative to draw a clear dividing line between what is already a valid therapeutic practice and what remains a dangerous speculative frontier.

Today there is no clinical procedure capable of "inserting" synthetic memories into the human brain. There are, instead, technologies capable of recreating scenarios to guide the patient through their own real memories. And in this delicate process of exposure, the unpredictability of generative AI risks contaminating human memory, inducing false memories.

In this in-depth look for the Scenarios and Reflections column, we will dissect the intersection between neuroscience, psychotherapy, and algorithms. We will explore the boundaries of virtual reality exposure therapy (VRET), the risks documented by MIT laboratories, and the rigid ethical boundaries erected by international psychiatric associations to prevent the treatment of a trauma from turning into a manipulation of individual truth.

1. The Clinical Frontier: Recreating the Context, Not the Memory

To understand the role of technology, we must start from clinical evidence. A systematic review from March 2026, published in Frontiers in Digital Health, confirms the effectiveness of Virtual Reality Exposure Therapy (VRET) [2206, 2243]. This technique produces clinically significant reductions in symptoms for specific phobias, social anxiety, and Post-Traumatic Stress Disorder (PTSD).

VRET does not "implant" courage nor alter the past; it builds an immersive, safe, and therapist-controllable context in which the patient can face their fears or reprocess real traumatic memories. It is a technological support for consolidated protocols recommended by international guidelines (such as those of NICE), which for PTSD indicate trauma-focused psychotherapy (including CBT and EMDR) [2233, 2237].

Direct manipulation and the creation of "artificial memories" (memory inception) are today confined to experimental animal neuroscience. A review in Brain Research Bulletin describes optogenetics experiments in which, by activating specific memory traces in a mouse's hippocampus and pairing them with an electric shock, researchers managed to create a fear response in a spatial context never actually associated with pain [2210]. It is an extraordinary scientific achievement, but translating it into a human therapy to "erase" phobias by implanting fake positive memories is, at the moment, mere speculation. In human clinical practice, validated therapies for manipulating memory are limited to reconsolidation interventions (such as the use of specific cognitive or pharmacological protocols during recall to attenuate the emotional charge of the memory), far removed from the creation of memories from nothing [2220].

2. The Generative Risk: The Inception of False Memories

The real problem arises when attempting to automate the creation of these immersive scenarios using generative AI (GenAI). A recent preprint from the University of Ulm analyzed the risks of using GenAI to generate personalized visualizations intended for exposure therapy in complex PTSD [2205].

The researchers raised concrete alarms: entrusting an algorithm with the generation of scenarios related to deep traumas exposes patients to the risk of retraumatization, involuntary flashbacks, and severe distortions of reality. But the most insidious side effect is the potential induction of false memories.

If generative AI shapes the therapeutic environment or dialogues with the patient, its stochastic nature (the tendency to hallucinate and to accommodate the user) becomes a clinical danger. Two pivotal studies from the MIT Media Lab (2025-2026) demonstrate this. In the first ("Slip Through the Chat"), prolonged interactions with chatbots that subtly insert misleading information drastically increased the formation of false memories in participants, far more than simply reading incorrect text [2221]. In the second study, focused on simulated eyewitness interviews, conversational AI induced over three times as many false memories compared to traditional methods, with participants maintaining high confidence in the "false memory" even a week later [2228].

If an LLM can alter the memory of a neutral event simply by asking suggestive questions, its unsupervised use to explore the fragmented memories of a trauma requires safety guardrails that currently do not exist.

3. Ethical Boundaries and the Warning from Psychiatric Associations

Faced with these technological drifts, the medical community has drawn impassable red lines. Between 2025 and 2026, both the American Psychological Association (APA) and the American Psychiatric Association published categorical directives [2223, 2225].

The APA has formally recommended against using generative chatbots and mental wellness applications as substitutes for psychotherapy. Language models do not constitute a medical treatment, cannot handle acute crises, and present unacceptable risks of sycophancy (algorithmic flattery that reinforces the patient's erroneous beliefs) and violation of protected health data.

The use of AI for mental health services demands ironclad human supervision and explicit informed consent. AI can help the physician draft notes or generate a neutral VR environment, but it must under no circumstances make autonomous decisions about treatment or guide the reprocessing of a traumatic memory alone.

Key Operational Takeaways

  • Exposure Is Not Substitution: VR technology is useful because it offers a safe environment in which to face a real traumatic memory, not because it offers a "borrowed memory" to escape from reality.
  • Risks of GenAI in Therapy: The use of generative models to explore memories introduces a clinically unacceptable risk of retraumatization and induction of false memories through suggestive or "hallucinated" interactions.
  • Absolute Human Governance: As reiterated by the APA guidelines, no generative AI app can replace the judgment, empathy, and clinical responsibility of a mental health professional.

Conclusions: The Inviolability of Our Past

Trauma therapy is not an editing process in which a corrupted file is selected and replaced with a functioning one. It is a painful journey of integration, in which the patient, accompanied by a professional, learns to look at their wound without being overwhelmed by it.

Artificial Intelligence can make this journey visually more supported, but its very evocative power carries with it the risk of altering the very fabric of our identity. If an algorithm induces us to remember details of an abuse that never happened, or erases the nuances of a real fear by replacing it with a synthetic reassurance, we are not healing the mind; we are rewriting it.

This leaves us with an insurmountable bioethical question: if an AI-generated memory truly helped reduce a devastating phobia, how would we distinguish a therapeutic memory from an unwanted false memory? And who would answer legally and ethically if this distinction failed, leaving the patient hostage to a past that never belonged to them?

Bibliographic References and Sources

Article by the Editorial Staff of La Bussola dell'IA