Introduction
Artificial intelligence has entered the domain of visual evidence. Courts are now confronted with images and video that have been processed using systems far more sophisticated than the traditional tools investigators relied on for decades.
This shift has exposed an important distinction. Not all forms of AI enhancement operate under the same principles. Some systems generate new visual information. Others attempt to recover structure that already exists within the original data.
The difference between these two approaches is now beginning to shape legal outcomes.
A recent case demonstrated this clearly. In State v. Puluka, visual evidence enhanced using generative AI was rejected after forensic experts raised concerns about the reliability of the methodology. Around the same time, a different enhancement methodology based on deterministic processing was examined in a separate legal context and accepted as part of the analytical process surrounding the evidence.
The divergence between these two outcomes reveals an important truth about the future of artificial intelligence in forensic environments. The question is not whether AI can improve images. The question is whether the method preserves the integrity of the evidence.
Background and Context
Visual evidence has become one of the most common forms of digital evidence presented in modern courts. Surveillance cameras, body cameras, mobile phones, and traffic systems now generate enormous volumes of imagery that investigators must interpret.
Yet the tools traditionally used to examine these images have remained limited. Courtroom displays often enlarge footage using simple scaling algorithms such as bicubic interpolation. These algorithms were designed for convenience, not for forensic analysis. They enlarge images while introducing blur and distortion that can obscure important visual features.
Artificial intelligence has emerged as a potential solution to this problem. However, two fundamentally different technological approaches have developed within the field.
The first approach relies on generative models. These systems are trained on large datasets and attempt to reconstruct missing visual detail through statistical inference. When information in an image is degraded or absent, the model predicts what that region should look like based on patterns it learned during training.
The result can appear visually convincing. However the detail introduced by the system is not derived from the original evidence.
The second approach uses deterministic signal reconstruction. Instead of predicting what an image might contain, deterministic systems attempt to mathematically recover structure already present within the original signal but obscured by noise, compression, or blur.
This distinction may seem subtle from a technological perspective. In forensic environments it is profound.
Thesis
Artificial intelligence can assist courts in understanding visual evidence, but only when the enhancement process preserves the evidentiary integrity of the original data. Generative systems that introduce hallucinated detail undermine that integrity, while deterministic methods that reveal existing signal structure can operate within the established principles of forensic analysis.
What Happened
In State v. Puluka, enhancement produced through generative AI developed by Topaz Labs tools was introduced as part of the evidentiary process. The images appeared clearer than the original footage, but forensic specialists quickly identified a central problem.
The generative system was not recovering information from the source material. It was synthesizing plausible visual detail based on training data. This meant that the resulting pixels could not be traced directly to the original evidence. The model had effectively created new imagery that did not exist in the recorded footage.
The forensic community raised immediate objections. Courts have long required that scientific evidence meet standards of reproducibility, transparency, and methodological traceability. Generative AI systems are inherently probabilistic. The same input can produce different outputs depending on model conditions, and the internal reasoning that produced those outputs cannot be reconstructed.
Because of these factors the enhancement method was rejected.
Around the same period another case examined a different form of AI enhancement. The technology involved deterministic reconstruction methods developed by Predictive AI (Predictive Equations). The system applied structured transformations to the original footage in order to recover patterns that were already embedded within the signal but difficult to perceive in the degraded material.
The enhanced imagery itself was not introduced as a replacement for the original evidence. Instead it functioned as an analytical instrument used to interpret the source footage.
Experts were able to reproduce the transformation pipeline applied to the video. Independent analysts produced identical results when applying the same processing chain. Each stage of the enhancement process could be explained and documented.
This transparency allowed the court to evaluate the analytical conclusions drawn from the footage while maintaining the original evidence as the evidentiary anchor.
The reassessment of the visual evidence ultimately contributed to multiple defendant’s early release.
Why the Outcomes Were Different
The difference between the two cases was not the sophistication of the technology. It was the underlying philosophy of the methodology.
Generative enhancement prioritizes visual plausibility. The model attempts to construct imagery that looks realistic to a human observer, even when the original data is incomplete.
Deterministic enhancement prioritizes signal fidelity. The objective is not to imagine missing detail but to extract structure already present in the recorded data.
In forensic science this distinction determines whether the resulting imagery can be trusted.
Courts must be able to answer a simple question about every pixel presented as evidence. Where did it come from?
Generative systems cannot reliably answer that question. The pixels may represent patterns learned from millions of unrelated images rather than the evidence itself.
Deterministic systems can answer it. The output is mathematically derived from the source signal and can be reproduced independently.
Because of this difference deterministic enhancement behaves more like a scientific instrument than a creative process. The technology does not invent evidence. It allows investigators to observe the evidence with greater clarity.
Why Deterministic Enhancement Is More Appropriate for Courts
Forensic environments demand reliability above all else. Courts operate under evidentiary standards developed over centuries. Any analytical method introduced into that environment must satisfy those standards.
Deterministic enhancement aligns naturally with these evidentiary requirements. The process is reproducible, meaning independent analysts applying the same transformation chain will produce identical results. It is also transparent, as each stage of the reconstruction pipeline can be documented and examined. Most importantly, the relationship between the enhanced output and the original evidence remains intact, ensuring the analysis derives directly from the source material rather than introducing synthetic detail.
In practice this methodology has already begun appearing across multiple legal contexts. Predictive AI has participated in roughly fifteen public and private investigations and court matters where degraded visual evidence required technical interpretation. In those cases the courts did not reject the analytical conclusions or the methodology used to derive them. The important distinction is that the enhancement itself was not presented as replacement evidence. The original footage remained the evidentiary anchor, while the deterministic analysis provided a structured method for interpreting visual features already present within the signal. This is precisely why the approach aligns more closely with traditional forensic practice than with generative image reconstruction.
In practice this means the technology functions as an analytical aid rather than a replacement for the evidence itself. Investigators and courts can examine the enhancement to gain insight into the original footage while still evaluating the source material directly.
In practice this means the technology functions as an analytical aid rather than a replacement for the evidence itself. Investigators and courts can examine the enhancement to gain insight into the original footage while still evaluating the source material directly.
This approach mirrors how other scientific instruments are used in forensic contexts. A microscope does not replace the specimen. It reveals details within the specimen that might otherwise remain invisible.
Deterministic enhancement serves a similar role in digital evidence analysis.
Conclusion
Artificial intelligence is rapidly entering the forensic domain, but recent cases demonstrate that not every form of AI will meet the evidentiary standards required by courts.
The rejection of generative enhancement in State v. Puluka reflects a broader recognition that hallucinated visual detail cannot serve as reliable evidence. Courts require methods that preserve the integrity of the original data and allow independent verification of the analytical process.
Deterministic enhancement represents a path forward that aligns with these principles. By focusing on signal reconstruction rather than synthetic imagery, the technology enables investigators and courts to examine degraded visual evidence with greater clarity while maintaining methodological transparency.
Takeaway
The case of State v. Puloka (2024) stands as a definitive "line in the sand" for this era of digital evidence. In that matter, the defense attempted to introduce video of a 2021 shooting that had been enhanced using Topaz Labs' AI software—a tool the prosecution’s expert argued added approximately 16 times the number of original pixels through an opaque, "un-forensic" process. The King County Superior Court ultimately rejected the footage, ruling it failed the Frye standard of general acceptance and Rule 702 for reliability. The judge noted that the system didn't just clarify the scene; it synthesized "what the AI model thought should be shown," creating a high risk of a "trial within a trial" over non-reproducible software logic.
As we move through 2026, the legal framework is catching up to this technical reality. The newly proposed Federal Rule of Evidence 707 (which just closed its public comment period in February 2026) specifically targets "Machine-Generated Evidence." If formally adopted, it will require AI-derived outputs to meet the same rigorous Daubert standards as human expert testimony—demanding transparency in training data, known error rates, and peer-reviewed methodology. For investigators and legal counsel, the message is clear: if your AI tool acts like an artist, it’s a liability; if it acts like a scientific instrument—reproducible, traceable, and grounded in the source signal—it’s the new gold standard.
The future of AI in the courtroom will not be determined by which systems produce the most visually impressive images. It will be determined by which systems remain faithful to the evidence itself.
As artificial intelligence continues to evolve, the tools that succeed in forensic environments will likely resemble scientific instruments rather than generative engines. Their purpose will not be to imagine reality, but to reveal it.