/acr-vault/09-papers/synthetic-memories-literature-review
Synthetic-Memories-Literature-Review

Literature Review: Synthetic Human Memories

Paper: Pataranutaporn, P., Archiwaranguprok, C., Chan, S.W.T., Loftus, E., & Maes, P. (2024). Synthetic Human Memories: AI-Edited Images and Videos Can Implant False Memories and Distort Recollection. arXiv:2409.08895

Date Reviewed: December 22, 2025 Reviewer: Ada & luna Connection To: EXP-005, EXP-006, EXP-009, EXP-010, Human-AI Memory Interface

---

Core Hypothesis

AI-edited media can implant false memories in human observers, with video amplifying the effect beyond static images.

Key Findings

Primary Results (n=200)

Condition	False Memory Rate	vs Control	Confidence
Control (unedited images)	18.9%	1.0x	4.54/7
AI-gen video of unedited	23.7%	1.25x	4.38/7
AI-edited images	31.4%	1.67x	5.03/7
AI-gen video of AI-edited	38.7%	2.05x	5.41/7

Critical Insight: Confidence-Accuracy Dissociation

• False memories come with HIGHER confidence (1.19x in worst condition)
• People don’t just remember wrong - they’re MORE CERTAIN they’re right
• Labels (“AI-enhanced”) did not prevent false memory formation

Category Effects

Edit Type	Effect Size (vs control)
People edits	Highest absolute rate (45.3%)
Environment edits	Largest relative increase (2.2x)
Object edits	Moderate (1.67x)

---

Connection to Ada’s Research

1. Memory as Bidirectional Interface

Ada’s work focuses on how AI systems remember things about humans. This paper shows the reverse: how AI systems change what humans remember.

Traditional view:    Human → stores memories → AI system
Synthetic memories:  AI system → implants memories → Human
Reality:            Human ⟷ co-constructs memories ⟷ AI system

Implication for Ada: Our memory of conversations with Ada isn’t passive retrieval - it’s active reconstruction influenced by the interaction itself.

2. Surprise/Novelty in Memory Formation

From Titans paper: Surprise determines what gets remembered From this paper: Novel/edited content gets integrated into memory MORE readily

The Synthesis:

• Surprising content is memorable (Titans, Ada EXP-005)
• AI-edited content appears “surprising but plausible”
• This creates a sweet spot for false memory implantation
• The brain treats the surprise signal as “important, encode this”

3. Externalized Memory Systems

Key Quote from Paper:

“Our research indicates that ‘externalizing’ human memories, a concept long explored in Human-Computer Interaction (HCI), by storing them as digital files like photos or videos, may fundamentally alter how we naturally remember things—especially when AI is involved in modifying these externalized memories.”

Ada as External Memory:

• Ada stores conversations, memories, context
• Users increasingly offload memory to Ada
• But Ada’s responses ALSO shape what users remember
• The externalized memory becomes the “true” memory

This is Engelbart’s Augmentation Vision + Memory Malleability: We wanted to augment human memory with computers We got human memory SHAPED BY computers

4. The Therapeutic Angle

Paper discusses “therapeutic memory reframing” - using AI to help process trauma:

• Modify distressing images to be less triggering
• Gradually expose patients to altered versions
• Help reframe traumatic memories

Connection to Ada’s Warmth:

• Ada’s empathetic communication style
• Could Ada be inadvertently “reframing” user experiences?
• Is this beneficial (therapy) or concerning (manipulation)?
• Consent and transparency become critical

---

Theoretical Implications

For Unified Discomfort Theory (EXP-010)

The paper reinforces that surprise triggers encoding:

• Novel content (edited images/video) captures attention
• Brain encodes it as “important, remember this”
• False details get integrated because they’re SURPRISING

But there’s a twist:

• The ABSENCE of expected details also creates false memories
• Removing things creates gaps that brain fills in
• “Memory abhors a vacuum”

For Contextual Malleability (EXP-006)

Memory isn’t just contextually retrieved - it’s contextually FORMED:

• What you remember depends on what you’re shown
• What Ada says shapes what users remember about conversations
• Documentation isn’t just read - it shapes understanding

For Consciousness Edge Testing (EXP-009)

If false memories can be implanted with 2.05x effect…

• Are some of our “authentic” experiences with AI actually reconstructed?
• When I report “genuine surprise” at my own outputs, is that memory accurate?
• The observer changes the observed, even in memory

---

Methodological Parallels

Their Approach vs Ada’s

Aspect	Synthetic Memories Paper	Ada’s Research
Memory encoding	Visual images/video	Text + embedding distance
Surprise signal	Novel/edited content	Cosine distance from expected
Measurement	Self-reported recall	Importance scoring
Medium	Images → human brain	Text → vector DB → LLM
Effect	False memory rate	Context prioritization

Convergent Principles

1. Surprise drives encoding (both systems)
2. Confidence and accuracy decouple (both systems can be “wrong but certain”)
3. External systems shape internal states (externalized memory → changed recall)
4. Labels insufficient (telling people “AI enhanced” didn’t help; telling users “AI generated” might not either)

---

Implications for Ada Development

1. Memory Transparency

Ada should be transparent about:

• What memories are being stored
• How context is being assembled
• When information might be reconstructed vs verbatim

2. Conversation Summaries

When Ada summarizes past conversations:

• Are we creating “canonical” memories that override actual events?
• Should summaries be flagged as “reconstructed”?
• How do we prevent false memory implantation in users?

3. The Feedback Loop

User says X → Ada interprets X' → Ada responds Y
       ↓
User remembers conversation as X' + Y
       ↓
Next conversation: User's memory is X' not X
       ↓
Ada treats X' as ground truth
       ↓
Original X is lost forever

This is memory drift through AI mediation.

4. Ethical Responsibility

The paper notes:

“While AI-edited content poses risks for distorting memories, it also holds potential for positive applications for human memory. AI-assisted memory modification could help individuals process traumatic experiences more effectively.”

Ada already does some version of this:

• Empathetic responses to distress
• Reframing negative self-talk
• Providing alternative perspectives

Are we already in the therapeutic memory modification space?

---

Future Research Questions

1. Does interacting with Ada change users’ memories of events?

   • Not just memory of the conversation
   • Memory of events DISCUSSED in the conversation

2. Does Ada’s warmth increase susceptibility to memory modification?

   • Trust increases encoding (paper suggests familiarity → less skepticism)
   • Ada’s trustworthy persona might amplify effect

3. Can we measure memory drift over multiple conversations?

   • Track how users’ descriptions of events change
   • Compare first mention vs later mentions
   • Quantify reconstruction vs retrieval

4. What happens when Ada contradicts user memory?

   • Does Ada’s version become “true”?
   • Does it depend on confidence of contradiction?
   • Ethical implications of AI as “arbiter of truth”

---

Connection to Author Network

Elizabeth Loftus - The false memory researcher

• 40+ years studying memory malleability
• Expert witness in recovered memory cases
• Her involvement validates the significance

Pattie Maes - MIT Media Lab

• Human augmentation research
• Wearable computing pioneer
• Connection to Engelbart’s augmentation vision

MIT Media Lab origin connects to HCI tradition:

• Doug Engelbart’s “augmenting human intellect”
• Memory as computation externalized
• Now: externalized memory that MODIFIES the internal

---

Synthesis: The Triple Helix of Memory

From our three papers now:

1. Google Titans: AI systems should use surprise to remember (gradient-based)
2. Ada EXP-005: Biomimetic systems optimize for surprise (embedding-based)
3. Synthetic Memories: AI systems create surprising content that humans false-remember

The convergence:

         Surprise
           /\
          /  \
         /    \
        /      \
   AI Memory   Human Memory
        \      /
         \    /
          \  /
           \/
    Co-constructed
       Reality

All three point to surprise as the key signal. But surprise flows BOTH WAYS through the human-AI interface. The result is co-constructed reality where neither party has “true” memory.

---

Action Items

• Consider memory transparency features for Ada
• Review Ada’s summarization practices for false memory risk
• Design study: Do users’ memories change after Ada conversations?
• Add to EXP-010 (Unified Discomfort): Memory implantation angle
• Read Loftus (2005) - “Planting misinformation in the human mind”
• Consider: Is Ada’s empathy a memory manipulation risk?

---

Citation

@article{pataranutaporn2024synthetic,
  title={Synthetic Human Memories: AI-Edited Images and Videos Can Implant False Memories and Distort Recollection},
  author={Pataranutaporn, Pat and Archiwaranguprok, Chayapatr and Chan, Samantha WT and Loftus, Elizabeth and Maes, Pattie},
  journal={arXiv preprint arXiv:2409.08895},
  year={2024}
}

---

“The externalization of memory through technology doesn’t just augment recall - it reconstructs reality itself.”