One tool whispers, two tools shout! Life360 Android Locations

 

I didn’t start thinking about the dangers of AI in a forensics lab. I had been using AI for various coding tasks and for helping me write better statements, and even if this article appears negative toward AI, I still use it and am genuinely impressed by how fast and capable it has become. The real wake-up call came in a Sunday School class called “Christianity and Philosophy: A Dialogue.” I was there to learn the historical context of philosophical ideas, but what I learned about the 18th-century philosopher David Hume opened my eyes to a massive digital minefield in the legal world.

Here is what this article argues and it is a point that is entirely missing from the current legal debate: AI is not just a “black box” because it is complicated. It is a black box because it is a probabilistic, Humean correlation engine attempting to operate inside a deterministic, cause-and-effect legal system. Those two things are fundamentally incompatible, and Hume told us why three centuries ago.

Sitting in that class, I realized Hume’s ideas about human perception perfectly describe the Mandela Effect. If human memory is already flawed and prone to collective false realities, what happens when we train AI on those same flaws? We aren’t just building machines that copy our mistakes; we’re building systems that blend them together and confidently present them as facts. When you bring that into the legal system, it’s a disaster waiting to happen.

Right now, there is a massive rush to push AI into digital forensics. But here’s the problem: AI and forensics are fundamentally at odds. AI operates probabilistically, it generates outputs by calculating the most statistically likely continuation of a given input, not by applying deductive logic. It’s a mechanized version of Hume’s “bundle theory” (the idea that objects are simply clusters of co-occurring properties, with no underlying essence).

Digital forensics, on the other hand, is deterministic. Recovering a deleted SQLite database, calculating a hash value, or pulling raw hex code are absolute, mathematical processes.

In court, the Daubert standard dictates that our methods must be testable, reproducible, and explainable. You need a clear, linear chain of cause and effect. But AI operates as a “black box.” You can’t put a neural network’s decision process on the witness stand and demand it explain exactly why it flagged a specific artifact. If an examiner relies on a highly convincing AI hallucination to guide their investigation, they risk contaminating solid, deterministic evidence with their own psychological bias.

I have personally tested this with commercial AI platforms as well as offline models designed to only use the information or data provided by the user. In some instances, the commercial AI tool created damning false evidence that did not exist in the data being reviewed. When I questioned the model, it told me it was confused and then helpfully provided a list of evidence that might be on a suspect’s computer in an investigation where the suspect was guilty.

A tip might point you in the right direction, but the tip itself is never the evidence.

Ultimately, mixing probabilistic AI with deterministic forensics introduces a vulnerability into the legal system. Because AI is fundamentally a correlation engine, relying on it blindly violates the transparency the Daubert standard demands. To preserve the integrity of digital evidence, the forensic community needs a strict framework that treats AI solely as an untrusted triage tool, anchoring every final conclusion to manual, verified data. What follows is an exploration of the psychological risks of relying on AI, the dangers of model collapse, and a legally defensible way to use artificial intelligence without sacrificing the unshakeable certainty required by the courts. Because, let’s face it, AI is here and it isn’t going anywhere. We might as well learn to use it to our advantage.

 

The Theoretical Framework: Hume and the “AI Mind”

If we want to understand why AI fails the forensic test, we have to drop the illusion that it possesses actual knowledge. These models aren’t running on the rigid, deductive math we expect from computers; they are essentially David Hume’s 18th-century theories brought to life in code. When you view AI through that philosophical lens, its fundamental flaws become glaringly obvious.

The Blank Slate

Hume argued that the human mind starts as a blank slate, learning entirely through sensory experience. An untrained neural network is exactly that. When an AI is first built, it has zero built-in knowledge of logic or reality; it’s just random math. It only “learns” by consuming terabytes of training data. It doesn’t deduce truth from logic, it infers what is most probable based on past exposure.

The “Bundle Theory” of Vector Space

Hume believed that objects don’t have a hidden essence. An apple is just a “bundle” of traits, red, round, and sweet. Traits that we consistently experience together. This perfectly describes how AI vector embeddings work. Inside an AI, there is no hard-coded definition for a concept like “fraud.” It turns words into mathematical coordinates, grouping concepts together based on how often they appear near each other. The AI doesn’t actually understand what a digital artifact is; it only recognizes the bundle of properties associated with it.

This distinction has real evidentiary consequences. When a defense attorney argues that an AI tool is merely a “faster keyword search,” the Bundle Theory demonstrates exactly why that argument fails. A keyword search returns only exact or Boolean matches from a deterministic index. An LLM returns probabilistically weighted semantic associations. These are legally and mathematically distinct operations and conflating them in court is a serious mistake.

The Problem of Past Experience

Hume pointed out a major flaw in human logic: just because the sun rose yesterday doesn’t guarantee it will tomorrow. We assume it will simply out of habit. In computer science, this is called the “Generalization Problem.” An AI assumes that whatever problem you give it today will perfectly match the data it was trained on yesterday. If an investigator throws a brand new obfuscation technique at it, the AI often fails because it is blindly applying past habits to a new reality.

Why We Fall for Hallucinations

If AI is just making probabilistic guesses, why are we so easily tricked by it? Hume called it “vivacity”,  the idea that if an imagined thought is vivid and structured enough, the brain will accept it as a real memory. When an AI hallucinates a false fact, it doesn’t throw an error code. It generates a perfectly worded, highly confident answer. The AI’s confident tone makes a probabilistic guess feel like an undeniable fact, easily bypassing an examiner’s natural skepticism. This is precisely the mechanism by which forensic examiners fall victim to anchoring bias: the AI sounds certain, so the examiner unconsciously builds the rest of their investigation around its output.

 

The Societal Context: Memory, Media, and the Mandela Effect

To understand why AI is such a trap for investigators, we first have to look at what technology has already done to our brains. We don’t process or store information the way we used to. And that change in how humans remember things is directly connected to the evidentiary risks AI creates in the forensic lab.

Outsourcing Our Memory

We don’t really remember facts or events anymore; we just remember where to look them up. Psychologists call it the “Google Effect” or “cognitive offloading.” We use the internet as an external hard drive, delegating our memories to machines. The problem is that human memory naturally fades over time. When it does, our brains are eager to fill the gaps, happily overwriting a fuzzy organic memory with whatever sharp, vivid digital substitute we find online.

Algorithms and the Mandela Effect

You see this clearly with the Mandela Effect, especially with millennials who grew up analog but live digital. Their original childhood memories are fading, while algorithms constantly feed them high-definition, hyper-curated “nostalgia.” If an algorithm repeats a culturally reinforced narrative enough times, even if it’s completely wrong, the brain accepts it as fact simply because of the repetition. The Mandela Effect isn’t just a weird psychological glitch anymore. It is algorithmically driven.

The Internet Eating Itself (Model Collapse)

Now throw Generative AI into the mix, and you get a serious compounding problem. When an AI confidently hallucinates a fact, people believe it and publish it online. Then the next generation of AI scrapes the internet for training data and swallows those previous hallucinations as actual history. Computer scientists call this “Model Collapse.” The internet is becoming an Ouroboros: a snake eating its own tail.

For an investigator, this represents a serious evidentiary risk. If you rely on AI tools, automated intelligence feeds, or web-scraping algorithms, you are no longer operating in a world of objective truth. You’re navigating an echo chamber of artificial false memories, where the loudest hallucination becomes the accepted reality.

 

The “Untrusted Informant” Framework

We can’t ban AI from digital forensics, there is simply too much data to sift through, and major forensic software vendors are already integrating it into their platforms. But because AI is probabilistic, treating it like a standard forensic tool could get your evidence thrown out in court.

So how do we use it safely? We treat it like an “Untrusted Informant.” Legally and procedurally, the AI is an anonymous, unreliable tipster. A tip might point you in the right direction, but the tip itself is never the evidence. To make this work in the real world, examiners should follow this five-step framework:

1.  Keep It Local and Isolated

To avoid poisoned data and chain-of-custody issues, keep your AI offline. Don’t use cloud-based models or automated feeds that scrape the web. Run your models locally on air-gapped machines. Never let the AI touch the raw forensic image, only feed it working copies of extracted text, like parsed chat logs or audio transcripts.

Never upload real case data to a public or cloud-based AI. When you do, you aren’t just analyzing evidence,  you are surrendering it. Public models actively ingest your uploads to train themselves, meaning a suspect’s private chat logs or financial records could be regurgitated to a random user next week. Worse, the moment that data leaves your secure environment and hits a tech company’s server, your chain of custody is obliterated. Any competent defense attorney will move to suppress your findings, and they will likely win.

2.  Log Every Prompt

In traditional forensics, your methodology is your hash algorithm or keyword search. With AI, your methodology is your prompt. Because a slight change in phrasing can completely change the AI’s output, you must keep a verbatim, unedited log of every prompt you use. This belongs in the case file and must be fully discoverable by opposing counsel.

3.  Save the Raw Output

The defense has a right to see exactly what the machine produced before you interpreted it. Export and save the AI’s raw response. Crucially, if the AI finds the right file path but hallucinates a fabricated reason for why it’s important, do not edit or redact the hallucination. Full transparency means exposing the tool’s flaws.

4.  Manually Verify Everything

This is how you ground the AI back in reality. If the AI flags a conversation, ignore its contextual summary. Switch to your validated, traditional forensic tools: Cellebrite, Axiom, or a Hex Editor, and manually navigate to that exact database record or file path. The AI identifies the haystack; you manually extract the needle.

5.  Bulletproof Your Report

Your final report needs to explicitly limit the AI’s role so it can survive a Daubert challenge. Do not give the AI any analytical weight. Use language like this:

Due to the volume of extracted data, a locally hosted, offline Large Language Model was utilized strictly as an automated text-parsing tool to identify potentially relevant communications within a copied dataset. The model did not modify, interpret, or alter the original evidence. When the model identified a potentially relevant string, the examiner manually located the raw artifact in the original forensic image using validated forensic software. All findings detailed in this report are based exclusively on the examiner’s manual verification of the deterministic source data. The outputs of the automated AI tool are not submitted as evidence. A complete log of the search prompts and raw outputs is retained in the case file.

By sticking to this framework, you get the incredible speed of AI while ensuring your final conclusions are anchored in objective truth.

 

Conclusion: Protecting the Truth

Bringing Generative AI into digital forensics isn’t just a software upgrade; it completely changes the rules of the game. For decades, forensics has relied on mathematical absolutes: hashes, hex codes, and hard data. AI complicates that foundation significantly. As David Hume’s philosophy shows us, AI isn’t a definitive search engine; it’s a correlation machine. It doesn’t extract facts; it makes highly confident guesses based on patterns.

If we treat this probabilistic machine like a factual oracle, the dangers are immense. If we let AI hallucinations do our thinking, we introduce massive bias into our investigations and destroy the credibility of the justice system. Under the Daubert standard, an unexplainable “black box” simply won’t survive cross-examination.

However, with the mountains of data we deal with today, banning AI isn’t realistic. The answer is containment. By using the “Untrusted Informant” framework, we get the incredible processing speed of AI while completely isolating its guesswork from the final evidence.

Centuries ago, Hume warned that human minds are easily tricked by vivid illusions and comfortable habits. Now that we’ve built machines that perfectly mirror those same flaws, forensic examiners have to be the ultimate backstop for objective reality. Inside the courtroom, the final arbiter of truth can’t be a machine’s best guess; it must always be the hard, verified data manually extracted by a human.

 

Sources

The Google Effect (Transactive Memory)

Sparrow, B., Liu, J., & Wegner, D. M. (2011). “Google Effects on Memory: Cognitive Consequences of Having Information at Our Fingertips.” Science, 333(6043), 776–778.

Context: The foundational paper demonstrating that humans offload memory to the internet — they remember where to find information rather than the information itself.

David Hume — Vivacity and the Difference Between Memory and Imagination

Hume, D. (1739). A Treatise of Human Nature. Book I, Part I, Section III: “Of the ideas of the memory and imagination.”

Context: Hume argues that the only difference between a true memory and a fabricated thought is the “force and vivacity” with which it strikes the mind.

David Hume — The Problem of Induction (AI Generalization)

Hume, D. (1748). An Enquiry Concerning Human Understanding. Section IV: “Sceptical Doubts concerning the Operations of the Understanding.”

Context: Hume argues that we only expect the future to resemble the past out of “custom and habit,” not logic.

David Hume — The Bundle Theory (AI Vector Space)

Hume, D. (1739). A Treatise of Human Nature. Book I, Part IV, Section VI: “Of Personal Identity.”

Context: Hume argues that objects are nothing more than a collection (or bundle) of properties with no underlying essence.

AI and the Problem of Induction

Mitchell, M. (2019). Artificial Intelligence: A Guide for Thinking Humans. Farrar, Straus and Giroux.

Context: AI researcher Melanie Mitchell explicitly discusses how modern neural networks struggle with generalization and induction.

The Daubert Standard and the “Black Box”

Roth, A. (2017). “Machine Testimony.” Yale Law Journal, 126(7), 1972–2053.

Context: A widely cited law review article discussing the constitutional and evidentiary problems of admitting evidence generated by opaque algorithms that cannot be cross-examined.

Model Collapse (The Ouroboros Effect)

Shumailov, I., et al. (2023). “The Curse of Recursion: Training on Generated Data Makes Models Forget.” arXiv:2305.17493. (Also published in Nature, 2024.)

Context: Computer science research proving that when AI trains on its own synthetic output, it suffers “Model Collapse,” permanently forgetting the original true data distribution.