Arrested with AI
Centennial police are using video presented as Motorola body-camera footage that include metadata inconsistencies and visual abnormalities—suggesting files may not be original bodycam recordings and may instead have been created using a mobile phone or personal camera, with subsequent alteration using artificial intelligence.There are likely other cases affected by similar practices.
Synthetic video, including deepfakes, has entered the courtroom in Arapahoe County, Colorado, appearing both as submitted as CCTV evidence and within body-worn camera footage. Similar anatomical anomalies—particularly irregular finger counts—have been widely documented in synthetic and AI-generated imagery. A prominent example includes early AI-generated images depicting Ariana Grande with six fingers, an artifact frequently cited in discussions of generative model failures involving hands and fine motor anatomy. The presence of comparable anomalies in Bodycam 64 and Bodycam 65 from the case’s discovery are therefore consistent with known failure modes observed in artificially generated or altered visual media.
In case number 2024CR2204, The State of Colorado v. Benjamin Rettig, both the prosecutor and the court were formally alerted to serious issues with the evidence presented, including photographs, CCTV footage, and body-worn camera video. During a motion hearing requested by the pro se defendant, the defendant challenged all elements of the Probable Cause Affidavit, submitting photographs, forensic opinions, and metadata demonstrating breaks in chain of custody and showing that the factual narrative underlying the allegations was false. Despite this, and despite the magistrate’s and prosecutor’s awareness of these issues, the prosecution was permitted to present its motions first. In doing so, the prosecutor sought an order requiring the defense to provide 35 days’ notice if it intended to use a forensic examiner. This procedural demand functioned as an attempt to obstruct the defendant’s motion and to advance the case while relying on demonstrably flawed evidence. This was January 9th of 2026. The defendant is scheduled to return to court February the 20th.
Multiple online AI-based video analysis tools have independently flagged the footage at issue as containing characteristics consistent with synthetic or manipulated video. Two videos are being presented as body-worn camera evidence: • Video 64: 1.000064-000064_AlexanderGockenbach_202406291311_BWL7023899-0.mp4_.mp4 • Video 65: 1.000065-000065_AlexanderGockenbach_202406291311_BWL7023899-0.mp4_.mp4 For clarity, these will be referred to as Bodycam 64 and Bodycam 65.
The videos are purported to depict events occurring on June 29, 2024, involving an officer responding to an alleged assault call at an adult care facility in Centennial, Colorado. However, both the embedded metadata and the visual characteristics of the footage raise substantial concerns regarding authenticity and provenance.Throughout both videos, the Motorola branding elements—including the Motorola logo, time overlay, and filename text—intermittently flicker, disappear, or reappear repeatedly . This behavior is highly irregular. Motorola body-worn camera systems embed overlays at the firmware level during capture; such elements do not selectively vanish during native recording. Their intermittent absence strongly suggests post-capture alteration, re-encoding, or compositing. (Exhibits zzz)
Close inspection of hands, arms, and faces reveals repeated visual distortions, including: • Localized blurring and rectangular masking artifacts • Apparent crop boundaries around fingers and facial features • Lighting inconsistencies that shift independently of scene illumination
Most notably, hand and finger anomalies appear throughout the footage. AI image generators often struggle with human hands. Professor Peter Bentley (University College London) explains, “They’ve got the hang of the general idea of a hand. It has a palm, fingers and nails, but none of these models actually understand what the full thing is.” Variations in finger count, size, and proportion are visible in multiple sequences. These anomalies are consistent with well-documented failure modes of synthetic video generation and AI-assisted face or body replacement. While only a subset of examples are highlighted here, similar artifacts recur across both videos.Several physical objects exhibit unexplained changes in appearance: • A cell phone changes color and form within seconds—appearing bright red in one moment, then reappearing as a longer, thinner, brown device shortly thereafter. • A driver’s license visibly changes facial characteristics as it is handed to the officer. • A character appears to operate a touchscreen device without their fingers ever fully contacting the screen surface.
These inconsistencies violate basic physical continuity and are not consistent with standard camera artifacts.The footage appears to employ strategies commonly used to conceal visual defects: • Subjects frequently hold objects, wear gloves, or keep hands in rigid or unnatural positions. • A cell phone is repeatedly positioned in front of the primary speaker’s face, partially obscuring facial features. • Other individuals remain off-camera while speaking and/or are obstructed from view with various methods.
Speech patterns are also irregular. Dialogue is largely restricted to short, two- or three-word responses. Despite this limitation, there are multiple instances where voices are audible while the corresponding speaker’s mouth does not move at all—an issue incompatible with synchronized audio-video capture.At one point, the officer—identified in the filenames as Alexander Gockenbach—is shown observing CCTV footage on a laptop. A pronounced glare is visible on the laptop screen during the display of the CCTV footage, yet this glare is absent immediately before and after the playback, despite no apparent change in camera angle or lighting conditions. Significantly, the CCTV footage being viewed—separately submitted as evidence—has itself been identified in a forensic examination as containing elements consistent with artificial or synthetic manipulation.
One nurse depicted in the footage exhibits extreme and inconsistent anatomical proportions. Her hands and forearms appear unusually large in certain sequences, yet appear markedly different in size and shape in others within the same video. In at least one instance, this individual can be heard speaking clearly while her mouth remains completely motionless.
The abnormalities described above represent only a portion of the issues identified to date. Analysis is ongoing, and additional anomalies continue to be documented. Links to the original source videos and visual exemplars of these irregularities are provided below for independent review.
In simple terms, the technical information attached to these videos does not look like what you would normally expect from clean, straightforward export of Motorola body-camera footage. The metadata can be viewed here: While it makes sense that a discovery copies would be a regular MP4 video that plays in common media players, several details stand out as unusual. The files do not show when they were created or exported, which makes it impossible to tell when these versions of the video were made. The software listed as creating the file appears to be general-purpose video software commonly used by consumers, not Motorola’s own evidence system. The video timing and compression look “too perfect,” which often happens when a file has been re-saved or converted. Finally, the file contains no information tying it back to a specific camera, case, or recording session. Taken together, these points suggest the video was processed or converted beyond a simple Motorola export. (Exhibit - Metadata 64-65) The audio is stored as stereo, even though Motorola body cameras record sound from a single microphone or commonly referred to as mono, and it is saved at a very low quality level that is more typical of re-processed video or post production AND at one point a woman can be heard talking yet her mouth never opens. (See figure zzz)
Manipulated CCTV and the bodycam videos are interlinked in the same legal discovery files as they stem from the same case. On June 27th 2024 a new employee working alone reports he has been attacked twice by a resident at an assisted living facility in Centennial, Colorado. He contacts his supervisor and a nurse to alert to the incident. They are not responsive until after the incident is over. (See exhibit) The resident who attacked him had just had a change in medication days earlier. The medication was for schizophrenia. This new employee had not been alerted to this condition, change in medication or trained on how to deal with residents with schizophrenia or those taking medication for such. He only finds out about the change as he notices new medication in the resident’s drawer, which he never dispenses to the resident as it is administered only during the day and he works the night shift. The resident, prior to the incident stated above, recently sent another caregiver to the emergency room.(Exhibit - Cindy says …) The new employee only gets the confirmation of the change in medication and the resident’s history of attacks on other caregivers after watching bodycam footage from the discovery evidence after being arrested solely based on re-recorded video and bodycam footage, which would he’d only see many months later.
A resident at the Milwaukee address showed signs of bruising on Tuesday, June 26th (Exhibit). This information was reported to Officer Alexander G. on June 29, 2024, while he was wearing a body camera. The officer did not document or report this earlier incident and instead conflated it with a separate incident on June 27, 2024, involving a different employee who was not present on June 26th—a QMAP named Benjamin Rettig. In Colorado, a QMAP (Qualified Medication Administration Person) is a state-certified, non-licensed worker authorized to administer prescribed medications under strict limits and supervision. Ben had recently been reprimanded for not administering medication while covering shifts at another facility Euclid, which resulted in a restriction limiting him to working at a single facility—Milwaukee—until retraining. However, on the night tied to the reprimand at the Euclid facility, Ben was not acting as a QMAP, worked solely as a caregiver, and administered no medications. (See Exhibit) This is important because it shows the June 26 injuries were misattributed, and it also demonstrates that the employer was effectively fencing Ben into the Milwaukee facility purposely. By denying him work at the companies 12 other locations in the Denver area while steering him to work exclusively at the Milwaukee facility , the employer appeared to be attempting to link him to a prior incident. This context is especially significant given that a forensic report states the June 27, 2024 footage “very likely contains localized AI or manual alterations.” The employer called Ben in to work at the Milwaukee facility on June 26 in an apparent effort to place him at that location, while simultaneously restricting him from working at other facilities.
Orchestrated discovery refers to a process in which evidence is deliberately prepared, reformatted, or curated prior to disclosure, rather than produced in its original, native state. This practice commonly involves exporting recordings into viewable formats (such as MP4), re-recording footage, or otherwise stripping technical metadata that would normally document how, when, and by what device the evidence was created. Such handling limits independent verification of authenticity, origin, and chain of custody.
The Probable Cause Affidavit (PCA) appears to follow a narrative structure that predates both the body-worn camera recordings and any meaningful on-scene investigation. This is evident during the officer’s interaction with Cindy M. (Exhibit P), in which Alexander Gockenbach repeatedly attempts to have her identify or locate video footage she is alleged to have “discovered.” During this interaction, Cindy M. demonstrates clear uncertainty and an apparent lack of knowledge regarding the existence or location of such footage. In Bodycam Video 64, the care coordinator similarly indicates no knowledge of any relevant video evidence. The officer then leaves the camera’s view. Approximately ten minutes later, a video is suddenly displayed on a computer screen. In Bodycam Video 65, by contrast, the same individual now appears familiar with the video and its contents, despite the absence of any documented discovery process in the interim.
Subsequently, Cindy M. testified under oath that she did not discover the video, stating instead that it was sent to her via text message. This sworn testimony directly contradicts the implication in the PCA that she independently located or uncovered the footage.
Under Colorado law, licensed health care facilities are required to retain original facility records, including native CCTV recordings, pursuant to C.R.S. § 25-1-801 and applicable Colorado State Archives retention schedules, and law enforcement officers have a duty to obtain and preserve original evidence when reasonably available rather than substitute re-recorded or reformatted versions. An officer’s failure to request or secure the facility’s exported native CCTV files, when available, undermines evidence integrity and frustrates independent verification of authenticity and chain of custody. Officer Alexander never states in the PCA that the CCTV videos are re-recorded with an iPhone and presents them as original video exported from an employee working at the facility. In actuality, re-recording strips original metadata and camera signatures, creating a new file that can hide edits or AI alterations and be passed off as original because the true capture source can no longer be verified. The re-recording method itself introduces substantial limitations on forensic verification, including the inability to reliably determine original capture time, device, encoding parameters, or prior edits. Moreover, all eleven re-recorded videos exhibit visual characteristics consistent with AI-assisted alteration or synthetic manipulation, compounding concerns regarding their evidentiary reliability. The underlying evidence ultimately consists of eleven videos re-recorded on an iPhone - the first three of which don’t contain time overlays, originating from an unidentified individual and recorded inside a private bedroom, who can be seen re-recording video via an iPhone. These videos are not native camera exports. Rather, they appear to have been screen-recorded or re-captured, a process that typically removes or replaces original metadata and obscures the transformation history of the files. Alexander also doesn’t state why he failed to obtain such yet given he presented such as original it may have appeared he did not need to.
These points were also raised in the Motion - and here’s a quick recap The person who is stated as discovered the footage stated under oath that they did not - the person re-recording from their bedroom isn’t referenced. The first three videos have no time overlays and failed to mention that the videos were not original but re- Officer Alexander ignored incident of abuse from the 26th of June and the injuries to the resident that occurred The officer orchestrates the discovery so that the video matches the PCA. Officer Alexander lastly states that the QMAP, Ben stated the defendant claimed there was “no physical altercation.” The caregiver was never questioned by the officer, which Alexander uses as opportunity to zzz
Before beginning work, during his orientation, the company’s CEO followed Ben into a restroom and stood observing him in silence. At the time, the interaction was unsettling and unexplained. In retrospect, Ben understood this conduct as the CEO assessing or “sizing him up,” rather than an incidental or accidental encounter. The CEO had previously owned a videography company before entering the healthcare industry.
During Ben’s first month of employment, electronic morphine count records were repeatedly altered in the system for medication Ben was assigned to dispense to a hospice patient. Each discrepancy was independently identified and corrected by another unscheduled employee, who confirmed that the errors did not originate with Ben, thereby clearing him of any wrongdoing.
In a separate and unrelated proceeding, a wrongful death lawsuit filed in Jefferson County, Colorado—Doak, Kelly et al. v. Rocky Mountain Assisted Living LLC et al., Case No. 2022CV030579—alleges that the same company altered electronic records at another one of their 13 facilities. In that matter, video evidence reportedly depicts a resident walking despite documentation indicating the resident had sustained a fractured hip.
On the evening of June 27, prior to reporting to work and before the attack later that night, Ben requested a copy of his tuberculosis test from the facility. The facility may have perceived this request as an indication that he was planning to leave the company. On June 28—his final day of employment and one day after the attack—Ben worked his last shift alone with all residents as normal, unaware of what would follow. This date was also one day prior to the date Officer Alexander later stated he visited the facility.
On June 28, Ben was instructed to administer morphine to a resident prior to a planned urinary catheter placement. A visiting nurse examined the resident and declined to proceed, concluding that neither the medication nor the catheterization was medically necessary. Months later, Ben learned that morphine is not a sedative and is not clinically appropriate for use prior to catheter placement. He also learned that morphine can cause urinary retention, which may affect clinical assessments. In retrospect, Ben believes the company was aware that this would be his final day of employment and that the request to administer morphine was made with knowledge that its effects could influence clinical findings, potentially leading to an unnecessary procedure. He also believes they knew this would be his last day and so assigned him with such task.
The circumstances suggest this matter may have been treated as a test case for using AI as a means of shaping or manipulating outcomes, rather than for conducting a fair, human-driven investigation focused on the truth.
The request for tuberculosis documentation may have signaled an impending employment separation, prompting preemptive actions by the facility and influencing how later events were framed.
Assigning responsibility to a single employee may have served to shield the facility from broader scrutiny, regulatory consequences, or insurance liability from the injuries the resident incurred on June 26.
The former QMAP disclosed prior employment included canvassing for Jason Crow in 2016 with the Democratic Party of Arapahoe and graduation from Florida A&M University, and the events occurred during the 2024 election year.
His résumé did not list his graduate education or his technical background in videography and software development. In hindsight, these undisclosed skills may have made him an inconvenient subject for scrutiny, as they positioned him to question and challenge the evidence presented.
Certain actions reflect an apparent belief that procedures, evidence handling, or narratives would not be meaningfully challenged, suggesting conduct driven by institutional confidence or assumed impunity.
The motives and decision-making of law enforcement in this matter remain unclear, yet the conduct demonstrated in this case establishes a dangerous and far-reaching precedent. Police were alerted to significant evidentiary irregularities, but nevertheless proceeded in a manner that suggests indifference to issues of authenticity, chain of custody, and factual accuracy. This raises serious concerns about whether similar practices have occurred in other cases, and whether individuals may currently be incarcerated based on evidence that was inadequately scrutinized or fundamentally unreliable. This case warrants review and scrutiny by higher authorities. Although the case may appear minor or insignificant to some observers, its consequences for the accused have been severe and enduring, profoundly affecting personal liberty, reputation, and overall well-being. The actions of law enforcement reflect a troubling willingness to rely on questionable digital evidence and institutional narratives rather than rigorous, human-led investigation. Coupled with the apparent confidence and boldness of the company’s actions, this suggests that when a resident is injured or a serious incident arises, responsibility may be shifted away from human decision-makers and deferred to artificial intelligence or automated processes. Such practices erode accountability, undermine due process, and weaken public trust in the integrity of both medical institutions and the justice system.