Memory manipulations in extended reality

Abstract

• Human memory has clear limitations such as ‘forgetting.’ We use external tools such as calendars to compensate.
• More VR applications exploit the limitations of perception: for example, redirected walking.
• Similarly, this study explores ways of exploiting memory limitations (forgetting, distortion, etc.) in XR: XRMMs (XR Memory Manipulations).
• A speculative design workshop with 12 participants produced 48 XRMM scenarios. Thematic analysis then extracted key elements in three categories.
• Discusses ethical issues and opportunities associated with XRMMs.

Introduction

Human memory has various limitations.

• Forgetting in long-term memory
• Distortion (changes from what was originally perceived)
• Persistence (some memories cannot be forgotten): “This is the unwanted recurrence of memories that you’d rather forget, like traumatic events or negative experiences. These memories can persist and cause distress even when you try to suppress them.” (The seven sins of memory - How the mind forgets and remembers)

We use various tools to compensate for these limitations.

• Recording anniversaries and appointments in calendars
• Keeping written records, photographs, and videos

XR devices are expected to become widespread tools for cognitive augmentation, particularly memory augmentation, in the near future. Studies suggest that recalling past experiences in VR can improve episodic memory.

Memory limitations can also be exploited to manipulate memories.

• Memory’s suggestibility: Studies show that manipulated photographs can implant false memories.
• As a richer medium than 2D photographs, XR enables not only the implantation of false memories but memory manipulation more broadly.
• XR also enables a wide range of perceptual manipulations (the authors’ previous research). But by influencing self-perception, social interactions, behavior, and so on, memory manipulation may have more invasive and longer-lasting effects than perceptual manipulation.
• Research into the possibilities and limitations of XR memory manipulation is therefore needed.

Summary of the research method:

• Three speculative design workshops to explore the scope of XRMMs broadly
• Each workshop involved two XR researchers and two cognitive psychology researchers with expertise in memory (4 people × 3 workshops = 12 participants).
• Experts in each workshop wrote near-future scenarios concerning possible memory manipulations using XR. A total of 48 scenarios were produced.
• Twenty-two scenarios judged technically and cognitively feasible were selected for deeper discussion (13 with positive intentions, 9 with negative intentions).
• Three of the paper’s authors coded the qualitative data and conducted thematic analysis over seven sessions (four hours each), extracting themes and mechanisms.

Summary of results:

• Definition of XRMMs: “perceptual manipulations grounded in memory limitations with the intention to impact memories, leading to long-term effects on users, through the self, social and directive functions of autobiographic memory”
• Five core elements of memory manipulation: context, memory mechanism, implementation, memory functions, and final goal
• Three categories extracted based on the technologies used and their outcomes:
  • 1. manipulations occurring at encoding (i.e., simultaneously as the original perception of the event), which aim at influencing the quality of memory encoding and predominantly rely on Augmented Reality (AR) technologies (e.g., steering user attention with AR to make the encoding of memories stronger);
  • 2. manipulations occurring at retrieval (i.e., when past events are digitally revisited), which aim at distorting the perception of a past event and predominantly rely on immersive reconstruction of scenes in Virtual Reality (VR) (e.g., altering a recording of prior event to overwrite the original memory);
  • 3. manipulations occurring pre-retrieval (i.e., just before the memory is remembered), which aim at inducing forgetting (e.g., hiding elements in AR to help someone to forget something).
• Most experts raised ethical risks during the discussions.

Contributions of this study:

• The definition of XRMMs and a framework presenting the core components that characterize these manipulations;
• The classification of three classes of memory manipulation in XR and insights about their unique technical characteristics and outcomes grounded in ongoing cognitive psychology and XR research.

Background

Human Memory

Memory takes various forms, but this study focuses on long-term memory.

The formation of long-term memory consists of three stages.

• Encoding: Acquiring new information through the senses
• Storage: Maintaining and managing information over time
• Retrieval: Recalling or recognizing stored memories

Within long-term memory, the focus is on autobiographical memory.

• What is autobiographical memory? A collection of personal experiences and general knowledge about oneself
• Functions of autobiographical memory:
  • 1. Self function: Maintaining self-coherence, including individuality and beliefs
  • 2. Social function: Forming and maintaining relationships with others
  • 3. Directive function: Guiding present and future thoughts and behavior
• Manipulating autobiographical memory may have long-term effects on users.

Extended Reality (XR)

This study considers the entire Milgram Continuum, rather than a particular form of VR or AR. These are collectively called XR.

What is the Milgram Continuum? The spectrum between a fully Real Environment and a fully Virtual Environment.

Related Work

Memory Aids

Existing HCI research on memory focuses on memory assistance.

• Technologies that support memory processes and technologies for offloading memory. For example, an XR memory palace.
• However, memory-offloading technologies can negatively affect memory and be vulnerable to manipulation. Calendars, for example.

Risks of Offloading Memory

Studies have examined risks associated with memory-offloading technologies.

• Recording tools can actually distract attention, affecting memory encoding.
• Excessive reliance on external tools may result in remembering less and less. Memories may also be distorted if records stored in those tools are manipulated.

Perceptual Illusions and Manipulations

Redirected walking research shows that manipulating vision and hearing can influence walking.

Perceptual manipulation (the authors’ previous research) can harm users if misused.

XR Memory Manipulations

There is much research on compensating for memory limitations, but little on exploiting them in other ways. This study broadly explores the possibilities of XR memory manipulation to deepen our understanding of its opportunities and risks.

Method

Participants

To explore possible XRMM scenarios, the authors recruited six HCI researchers studying XR and six cognitive psychology researchers studying human memory who were familiar with XR (age: M=28.7, SD=4 years).

Three workshops were held, each with two XR experts and two memory experts. This allowed both discussions within a discipline and discussions across disciplines in each session.

Workshop Procedure

Because XR is not yet as commonplace as smartphones, speculative design workshops were used to explore possibilities broadly.

Each three-hour session took place online using Zoom and Miro, in the following order:

• Introduction
• Individual brainstorming: Each participant generated as many scenarios as possible.
• Discussion: Participants briefly introduced their scenarios and discussed them together. They were encouraged to evaluate technical/cognitive feasibility and classify benefits or risks. All discussions were recorded and transcribed.

Scenario template: “In this scenario, USER NAME AND DESCRIPTION uses XR to PURPOSE. This app manipulates user’s memory by MANIPULATION IMPLEMENTATION DESCRIPTION. This app leverages MEMORY FLAW LEVERAGED in order to IMPACT / AIM OF MANIPULATION which is facilitated by XR FEATURE.”

Data Collection

A total of 74 scenarios were generated, of which 25 were discussed (Table 5).

Thematic Analysis

The discussion transcripts for the 25 scenarios were qualitatively analyzed (coded).

Analysis procedure:

1. Three coders created a codebook (deductive coding).
2. They first read the text individually, then discussed each scenario.
3. They began coding according to the codebook.
4. If a code absent from the codebook seemed necessary, it was added only with all three coders’ agreement (inductive coding).
5. Codes were grouped into themes and organized into broader structures and concepts.

Scenarios Exclusion

After coding, the authors defined XRMMs and removed scenarios that did not fit the definition (S23, S24, S25 in Table 3). Adding 26 scenarios that had not been discussed yielded 48 XRMM scenarios in total (25 - 3 + 26 = 48).

However, subsequent mentions of scenarios in the paper refer to the 22 discussed scenarios (S1–S22 in Table 3).

Results

Definition XRMMs

Perceptual Manipulations: “Perceptual manipulations are mechanisms grounded in limitations of users’ cognition and perception with a clear intention to influence users towards a specific outcome”

XR Memory Manipulations: “XRMMs are perceptual manipulations grounded in memory limitations with the intention to impact memories, leading to long-term effects on users, through the self, social and directive functions of autobiographic memory.”

Virtual-Physical Perceptual Manipulations: “VPPMs are perceptual manipulations that are grounded in visual-haptic limitations with the intention to nudge the user’s physical movements.”

XRMMs Framework

Five core elements:

• Context: The manipulator, the person being manipulated, and the application in which manipulation occurs
  • Ten of the 22 scenarios involved helping or harming people experiencing trauma or memory-related problems.
  • In six of the 22 scenarios, the manipulator and the person manipulated were the same person.
  • Applications varied, including training programs and social media.
• Memory mechanism
  • Fourteen scenarios exploited limitations in memory mechanisms, positively or negatively.
  • Eight scenarios compensated for those limitations.
• Implementation: Adding information absent from the real world, removing existing information, or modifying it
• Outcome: Effects on memory encoding or retrieval, or perceiving past memories differently
• Final goal: The outcomes above contribute to goals such as boosting confidence or treating trauma.

Threat Model

Manipulators may use these techniques maliciously, for example to interfere with others’ decisions. Possible pathways:

• Non-consensual manipulation: Hacking, etc.
• Exploiting a position of power: Malicious use by caregivers, professionals using XR devices for medical purposes, etc.
• Consensual, incentivized manipulation: Deceptive patterns, etc.

Classes of XRMMs

At Encoding, Pre-retrieval, At Retrieval. (Figure 4)

XRMMs at Encoding

Intervening at the very moment perception occurs. This matters because encoding determines what enters long-term memory.

Influencing Encoding by Inducing Emotions

• Basis: Emotion is involved throughout cognition and is particularly important to memory. Studies suggest that positive emotions encourage learning. Negative emotions such as stress or fear can also affect memory quality or alter memories, depending on their intensity.
• Scenarios:
  • S2: “You’ll Remember the Unlikely Risky Event”. A driving practice app creates moderately stressful situations to support training.
  • S1: “Peak-End Rule”. Inducing positive emotions during an XR experience leaves generally positive, happy memories, improving self-perception.

Influencing Encoding by Steering Attention

• Basis: Humans cannot remember everything they perceive. Attention strongly affects memory.
• Scenarios:
  • S6: “This You Should Remember”. AR that highlights key points.
  • S4: “Attention Detector”. Using eye trackers and the like to monitor and guide the attention of people with attention deficit disorders, supporting learning.
  • S5: “Encoding Enhancer”. Guiding attention to support learning, for example by blurring distracting elements in the surroundings.
  • S3: “Make Forget”. An attacker manipulates a target’s AR display to steal their wallet.

Discussion

• Technical challenges:
  • Various studies show that XR can evoke emotions.
  • Various studies show that XR can guide attention.
  • Guiding attention requires AI that can detect what matters in real time. Object recognition and related technologies have advanced rapidly recently. (For example, Segment Anything.)
• XR’s role:
  • Conventional apps could do this, but XR (especially AR) is more powerful because it can directly manipulate the visual field.
• Purposes:
  • Mostly positive scenarios for strengthening memory or improving attention. Potentially useful for people with attention deficit disorders as well as nondisabled people.
  • S3 misuses AR through techniques such as distraction (MI - Dead Reckoning?).

XRMMs at Retrieval

Memories can change during retrieval, and retrieval can also form new memories.

Influencing Perception through Reflection

• Basis: Reflection is revisiting and reframing past experiences. Memories are actively reconsidered in this process.
• Scenarios:
  • S8: “Traumatic Memory”. A person with PTSD is encouraged to reexperience an event during therapy, revisiting it with information they now know but did not know then. Considerable research already exists under the name “Virtual Reality Exposure Therapy.”
  • S9: “Couple Game”. Helping partners understand each other better by experiencing each other’s memories.
  • S12: “Shame Reminder”. Inducing shame by making people recall embarrassing memories.
  • S13: “Naked Man VR”: Triggering users’ trauma, then monetizing their escape by requiring payment.

Distorting Memories through Suggestions

• Basis:
  • Misattribution: A memory exists but is linked to the wrong source. For example, remembering directly experiencing an event when one actually saw a photograph of it. This may arise through erroneous inference during retrieval.
  • Suggestibility: New information added between encoding and retrieval can distort memory.
• Scenarios:
  • S18: “Buy a Friend”. “Inserting” yourself into someone else’s AR video recording so that replaying it distorts their memories.

Discussion

• Technical challenges:
  • Recording events in 3D. Volumetric capture has already been commercialized to some extent.
  • Inserting 3D objects: Advances in generative AI make this considerably more feasible.
• XR’s role:
  • Earlier media could do this too, but XR can amplify the effects.
• Purposes:
  • Understanding the past better and framing it more positively to make it less painful
  • Modifying one’s own past more positively to increase confidence
  • Tormenting people by triggering trauma
  • Becoming closer to others by distorting their memories

XRMMs Pre-retrieval

Intervention is possible between an event occurring and its being stored in memory.

Inducing Forgetting by Hiding Memory Cues

• Basis: Transience is the weakening or disappearance of memories over time. Frequent retrieval can prevent it.
• Scenarios:
  • S21: “AR Eraser”. A program that removes objects associated with a particular person to help the user forget them.

Inducing Forgetting by Inducing Stress

• Basis: Blocking is when a memory exists but cannot be retrieved at the appropriate moment, as when one cannot remember a famous actor’s name.
• Scenarios:
  • S22: “False Training App”. Placing users under stress so that they perform poorly on quizzes, inducing them to buy a learning program.

Discussion

• Technical challenges:
  • Implementing AR Eraser requires automatically classifying objects associated with a particular person, a technical challenge.
• XR’s role:
  • Even without XR, objects associated with someone can be put away in a box. But AR can also apply to objects that cannot be put away (a tree in a park, for example).
  • False Training App depends on inducing stress. XR environments can induce stress more actively than conventional methods.
• Purposes:
  • With AR Eraser, it is questionable whether simply erasing bad memories is always good.
  • False Training App is a scam.

General Discussion

Thematic analysis yielded a way of describing XRMMs through five core components (context, memory mechanism, implementation, outcome, final goal).

XRMMs could be classified into three categories (encoding, retrieval, pre-retrieval).

Combining XR researchers with cognitive science researchers enabled scientifically grounded scenario generation in the speculative design workshop.

The Good: One Step Closer to Augmenting the Human Intellect

Douglas Carl Engelbart explored augmenting human cognitive abilities in Augmenting human intellect: A conceptual framework. XRMMs differ from D.E.’s framework in acting on memory itself, rather than as ‘external’ tools. (Not a particularly meaningful distinction from an extended cognition perspective. —ak)

The Bad: One Step Closer to Losing Perceptual Integrity

Influencing human perception and memory may pose risks to perceptual integrity.

It may affect autonomy, agency, independence, and so on.

The Ugly: Who is Responsible for the Trajectory and what can the HCI Community do?

HCI, XR, security, and cognitive science need to collaborate to examine the positive and negative aspects of XR perceptual manipulation broadly.

Limitations

It is unfortunate that current XR devices are not advanced enough, necessitating speculative design workshops.

Collaborative coding meant that inter-coder reliability could not be measured in the thematic analysis.

Cultural diversity was limited: 11 of the 12 participants were from Western Europe.

Conclusion

Defined XRMMs, extracted five components, and classified them into three types. Showed that many feasible XRMM techniques exist, which could be used positively or misused.