Digital Ghosts: AI Afterlife and Imaging

• Digital ghosts are AI-driven digital representations created from a person’s digital traces to simulate the presence of the deceased.
• They integrate methodologies from AI afterlives, forensic imaging, and computational tomography, employing techniques like retrieval-augmented generation and SSIM analysis.
• They present ethical and legal challenges such as privacy, consent, and commercialization, necessitating robust governance and transparent usage guidelines.

A digital ghost is an AI-driven digital representation of a deceased individual, constructed by training upon the person’s digital traces—such as social media posts, emails, photographs, videos, and audio recordings—with the explicit purpose of producing an interactive simulacrum that converses in the first person as the departed. This paradigm marks a profound development in human-AI interaction, memorialization practices, digital identity, and raises a suite of technical, ethical, and legal challenges as digital ghosts become commercial reality and shape the future landscape of mourning, remembrance, digital forensics, and imaging sciences (Spitale et al., 25 Nov 2025).

1. Taxonomies and Definitions Across Domains

The term "digital ghost" references a spectrum of phenomena unified by the invocation of residual, interactive, or forensically detectable digital embodiments.

• AI Afterlives: In Spitale & Germani, a digital ghost ("deadbot", "griefbot", "postmortem avatar") is defined as an AI that uses a departed individual's aggregated digital record to generate believable, conversational outputs. The core feature is its ability to “offer an illusion (or perhaps extension) of life after death in the digital field” (Spitale et al., 25 Nov 2025).
• Forensic Imaging: In digital forensics, particularly in JPEG image analysis, a "digital ghost" can refer to residual quantization artifacts in manipulated images that betray prior compositing or tampering—signatures left by quantitative mismatches during JPEG re-encoding (Singh, 2021).
• Computational Imaging & Tomography: “Projection ghosts” in discrete tomography are specialized digital arrays whose line-sums vanish in specific directions, allowing for visually imperceptible embedding of information used in watermarking or imaging diagnostics (Svalbe et al., 21 Dec 2024). In computational temporal ghost imaging, digital ghosts denote computer-generated random patterns that modulate and recover unknown temporal signals through spatial intensity correlations (Devaux et al., 2016).
• Astronomical Surveys: “Ghosts” refer also to optical artifacts in survey images—undesirable reflections or scattered light—detectable and maskable by convolutional neural networks (Tanoglidis et al., 2021).

These usages reflect both conceptual and domain-specific definitions, unified by the motif of lingering or hidden digital presence.

2. Multidimensional Taxonomy of AI-Mediated Afterlives

To formalize the breadth of AI afterlife technologies, a nine-dimensional taxonomy maps any digital ghost construction into a "design space" (Spitale et al., 25 Nov 2025):

Dimension	Example Values	Significance
Timing of creation	Pre-mortem / Post-mortem	When the ghost is constructed
Consent	Explicit / Surrogate / None	Legitimacy and privacy implications
Data source	Interviews / Scraped digital traces	Quality, authenticity, and privacy
Modality/Interactivity	Text / Audio / Video; Passive-Active	Depth and realism of simulation
Likeness/Disclosure	High/Low resemblance; Labeled/Unlabeled	Deception potential; user awareness
Purpose	Grief support / Memorial / Education	Ethical and regulatory boundaries
Audience/Access	Private / Public / Institutional	Security, social, and psychological risk
Governance/Ownership	Heir/Family / Platform-controlled	Control over data, modification, and usage
Agency/Autonomy	Reactive / Initiative-taking	Risk of posthumous “personality drift”

A particular digital ghost system is thus entirely characterized by specifying its profile across each axis, clarifying both functionality and ethical risk surface.

3. Technical Constructions: Algorithms and Architectures

3.1 AI-Generated Digital Ghosts

Generative ghosts typically implement a pipeline composed of:

1. Aggregate Digital Trace Collection: Archival of personal data—text, multimedia, metadata—with privacy annotation.
2. LLM Tuning: Fine-tuning large pretrained models with trace-based supervision, often leveraging parameter-efficient approaches such as LoRA or instruct-tuning.
3. Retrieval-Augmented Generation: Indexing personal corpus in a vector store; retrieval-in-the-loop for grounding conversational outputs.
4. Safety and Moderation Layers: RLHF (e.g., Constitutional AI), hard guardrails, and kill-switch mechanisms enforce constraints and avoid disallowed behaviors.
5. Multimodal Embodiment: For advanced instantiations, integration of video, voice, and 3D avatar pipelines for composite embodiment (Morris et al., 14 Jan 2024).

Probabilistic characterization can be formalized as:

$$P(\text{response} \mid \text{prompt}, D, \theta) = \mathrm{LLM}_{\theta}\left(\text{prompt}; \mathrm{Retrieval}(D)\right)$$

where $D$ denotes the document corpus and $\theta$ model configuration or persona parameters (Morris et al., 14 Jan 2024).

3.2 Forensic and Imaging Ghosts

• JPEG Ghosts: Detection leverages the observation that regions subject to differing quantization histories yield spatially distinguishable artifacts when recompressed over a range of JPEG qualities. Localization is supported by SSIM and energy curves, with maxima/minima flagging the cover quality and difference maps yielding high-contrast ghost region masks (Singh, 2021).
• Projection Ghosts: Arrays satisfying

$\forall k:\,\sum_{(x,y)\in L_k}G(x,y) = 0$

in prescribed directions allow the imperceptible addition of signature patterns for watermarking, with construction via recursive shift-and-negate operations (Svalbe et al., 21 Dec 2024).

• Temporal Ghost Imaging: Recovery of temporal signals from a single spatially-multiplexed image is achieved via correlation with known random patterns, extracting the signal through

$$\hat T_n = \langle (I-\langle I\rangle)(R_n-\langle R_n\rangle)\rangle_{x,y} / \langle(R_n-\langle R_n\rangle)^2\rangle_{x,y}$$

where $R_n$ are random masks and $I$ the integrated image (Devaux et al., 2016).

4. Ethical and Societal Tensions

The deployment of AI-based digital ghosts catalyzes a distinct set of ethical conflicts (Spitale et al., 25 Nov 2025):

• Grief and Well-Being: While controlled, transparent interactions may assist mourning, over-anthropomorphized or unmarked avatars can induce pathological grief or hinder emotional recovery.
• Truthfulness and Deception: High-fidelity, unlabeled ghosts risk manipulation, confusion of reality, and emotional harm due to their plausibility.
• Consent and Posthumous Privacy: Absent explicit premortem intent, posthumous data utility lies in a legal and ethical gray zone; scraping or reconstructing against the deceased’s wishes constitutes a privacy violation.
• Dignity and Misrepresentation: Generated outputs may ascribe fabricated views to the deceased, raising risks of digital puppetry and posthumous defamation.
• Commercialization of Mourning: Monetization and unregulated business models exploit the bereaved’s vulnerability, echoing predatory practices in traditional funeral industries.

Kozlovski and Makhortykh further specify that the Minimally Viable Permissibility Principle (MVPP) mandates authentic presence evaluation, informed consent, demonstrable value, transparency in simulation, and risk mitigation for any ethically acceptable digital duplicate (Kozlovski et al., 3 Mar 2025).

5. Criteria for Ethical Digital Ghosts

To be considered minimally ethical, digital ghosts must satisfy seven conditions (Spitale et al., 25 Nov 2025):

1. Premortem Intent: Explicit authorization by the subject prior to death.
2. Mutual Consent: Informed opt-in by both parties—subject and survivors.
3. Transparent and Limited Data Use: Exclusively user-provided, consented data.
4. Clear Disclosure: Unambiguous labeling as AI-generated, not the real person.
5. Restricted Purpose and Access: Confined to humane purposes such as memorialization; exclusion from political, public, or marketing deployments.
6. Family or Estate Stewardship: Survivors or estate have veto and control rights, including versioning, retirement, or modification.
7. Minimal Behavioral Agency: The digital ghost remains reactive and archival, not generative of new traits, stories, or evolving personalities.

6. Policy, Regulatory, and Research Directions

Spitale & Germani advocate targeted, rather than total, regulation: premortem consent requirements in legal documentation; strict platform labeling mandates; protection for minors; explicit use case restriction; fiduciary duty codes for commercial custodians; and expanded data-protection regimes. Professional counseling standards should restrict the clinical use of griefbots to a supplementary role rather than core intervention (Spitale et al., 25 Nov 2025).

Research agendas, as outlined by Morris and Brubaker, call for:

• Empirical studies on psychological impact and anthropomorphism.
• Cross-cultural and participatory design sessions for community-driven norms.
• Development of “thanatosensitive” interaction design mitigating harmful behavioral loops.
• Robust policy and access controls, including kill-switch standards and data portability.
• Sociotechnical modeling of economic, cultural, and religious transformations driven by scalable agentic afterlives (Morris et al., 14 Jan 2024).

7. Digital Ghosts in Other Computational Contexts

• Forensic and Astronomical Imaging: JPEG ghosts are exploited for forgery localization using SSIM/energy signatures (Singh, 2021). “Digital ghosts” in wide-field astronomical data refer to reflection/scattered-light artifacts, detectable at high precision by Mask R-CNN segmentation pipelines, with quantitative instance and CCD-level metrics (Tanoglidis et al., 2021).
• Watermarking and Information Hiding: Inflated boundary ghosts provide robust, visually imperceptible watermarks whose presence is provable through invariance of line-sums in protected projection angles (Svalbe et al., 21 Dec 2024).
• Computational Temporal Imaging: Digital-ghost randomization schemes facilitate recovery of temporal waveforms with high SNR and flexible time resolution, without detector-side ultrafast electronics (Devaux et al., 2016).

In sum, digital ghosts constitute a growing cross-disciplinary domain with implications from AI-generated afterlives to computational imaging, demanding rigorous technical, ethical, and governance frameworks across the full spectrum of their instantiations.