The Algorithm as Witness: Errors in the Criminal Justice System Through the Lens of Technological Determinism

On 29 March 2026, CNN published a report on a case that might have been dismissed as yet another instance of bureaucratic negligence, were it not for one crucial detail: the basis for the arrest of a Tennessee resident was facial recognition carried out by Clearview AI — a company with a database containing billions of photographs collected from social media and open internet sources.

Angela Lipps, a 50-year-old grandmother and mother of three, spent over five months behind bars after Fargo (North Dakota) police linked her to bank frauds committed in a state where, by her own account, she had never been. It was only after her extradition to North Dakota that her lawyer discovered bank records confirming that Lipps had been in Tennessee at the time the crimes were committed. The charges were dropped on 23 December, and she was released on Christmas Eve.

This case calls for a much deeper analysis than simply noting yet another algorithmic failure. It demonstrates a structural pattern inherent in every era of technological transformation in criminal justice: a new identification tool is introduced with enthusiasm and uncritical trust, supplanting tried-and-tested methods of establishing facts, whilst specific individuals become the victims of systemic errors. To see this pattern for oneself, one need only look at a case that took place over 120 years ago in London.

THE CASE OF ADOLF BECK: ANATOMY OF AN IDENTIFICATION ERROR AT THE TURN OF THE TWENTIETH CENTURY

In 1896, the Norwegian national Adolf Beck was convicted by the Central Criminal Court of London (the Old Bailey) for a series of frauds against women — allegedly, posing as a lord, he would approach them on the street, promise expensive gifts, and steal their jewellery. The verdict was based almost exclusively on witness testimony: the victims identified Beck unanimously. Moreover, the police established that similar crimes had been committed in 1877 by a person using the same modus operandi — and this old case was attributed to Beck without sufficient grounds.

After serving five years of his sentence, Beck was released in 1901 — only to find himself back in custody in 1904 on identical charges. And once again: confident, unanimous eyewitness testimony. And once again: an innocent man. The real offender — William Thomas, operating under the alias John Smith — was apprehended only by chance: he committed yet another fraud after Beck’s second arrest and was caught in the act. Beck was released and pardoned. The total duration of his wrongful imprisonment amounted to approximately seven years.

What allowed this miscarriage of justice to occur not once but twice? The answer lies in the institutional attitude towards the new identification paradigm. The late nineteenth century was the era of what was then called ‘scientific policing’: Bertillon’s anthropometry, later fingerprinting, and systematised procedures for questioning witnesses. Eyewitnesses no longer simply ‘recognised’ a suspect — they participated in a procedure that carried the trappings of scientific rigour. The procedural authority of identification became self-sufficient evidence requiring no verification. Beck’s alibis were not seriously investigated precisely because ‘scientific’ identification appeared more reliable than any exculpatory counter-evidence.

The Beck case had significant legal consequences: in 1907, the Court of Criminal Appeal was established in England — directly as an institutional response to the demonstrated inability of the system to self-correct in cases involving identification errors.

STRUCTURAL PARALLELS: TECHNOLOGICAL DETERMINISM AND FAILURES OF PROOF

A comparison of the Beck and Lipps cases reveals a striking structural isomorphism, suggesting a consistent pattern rather than mere coincidence.

First, in both cases the identification technology was introduced into law enforcement practice without proper validation and without established standards for its admissibility. The Fargo police chief admitted at a press conference that the department had acquired its own facial recognition system ‘without the knowledge of senior management’ and without agreed protocols for its use. ‘We would not have allowed this to be used,’ he stated, noting that the system has since been prohibited. In the Beck case, systematic in-person identification was similarly introduced as a tool of ‘progress’ without adequate consideration of its inherent limitations.

Second, in both cases alibis or counter-evidence were available from the outset but were not properly investigated, due to excessive reliance on the identification outcome. Lipps’s lawyers emphasised that ‘the police officer used facial recognition as a substitute for basic investigation’ and that ‘no investigation was carried out to establish whether she was even in North Dakota at the relevant time’. The word ‘substitute’ is key: the technology did not supplement traditional investigative methods — it replaced them.

Third, in both cases the institutional machinery continued to move by inertia even after the first signs of error appeared. The police received information about potentially exculpatory evidence as early as 12 December, yet a substantive response — involving the investigator, the prosecutor, and the judge — did not follow until 23 December, more than five months after the arrest.

Fourth, the legal consequences for the victims proved catastrophic beyond the scope of the criminal proceedings themselves. For Beck, prolonged detention destroyed his reputation and social standing. For Lipps, ‘the trauma, loss of liberty and reputational damage cannot be easily remedied’, her lawyers state, noting that they are considering bringing civil rights claims.

This parallel points to a central theoretical problem: why do such errors recur with each technological shift in the identification paradigm? The answer lies in what might be called technological determinism in evidence law — the tendency, once a new tool acquires ‘scientific’ status, to treat it as a priori superior to traditional forms of evidence. This cognitive bias is reinforced by a structural asymmetry: the flaws of traditional methods (eyewitness testimony, documentary records) are well known and openly discussed, whereas the flaws of the new tool remain systematically invisible until the first high-profile scandal.

PROCEDURAL AND NORMATIVE DIMENSIONS

The Lipps case highlights several normative questions long debated in the doctrine of criminal procedure, which have taken on new urgency with the widespread introduction of AI into law enforcement.

The first is the admissibility standard for algorithmic evidence. The result of Clearview AI’s facial recognition served as the basis for issuing an arrest warrant with inter-state extradition — without any verification by an independent competent authority.

The second concerns procedural safeguards when using automated systems. The ‘human-in-the-loop’ principle has gained widespread formal recognition, yet the Lipps case demonstrates its practical erosion: a human is technically present, but their role is reduced to relaying the algorithmic output without critical scrutiny. ‘The detectives mistakenly assumed that they had also been sent photographs from the scene along with the identification conclusion,’ the police chief admitted — describing a process in which even the minimal verification step nominally built into the procedure was bypassed by those captivated by the authority of the algorithm.

The third concerns remedial justice. The police chief declined to apologise, citing an ongoing investigation into a ‘wide network of individuals’. This stance is symptomatic: institutional denial of responsibility is as much a structural component of the pattern as the original error.

CONCLUSIONS: THE PERPETUAL RACE BETWEEN TECHNOLOGY AND THE LAW OF EVIDENCE

The Lipps case is neither an anomaly nor the product of purely individual negligence. It is the predictable outcome of an incomplete institutional understanding of a new identification paradigm. One hundred and thirty years separate the Adolf Beck case from the Angela Lipps case, yet the nature of the error is unchanged. Technological innovation alters only the instrument; the first victims are always those whose alibi proves ‘less convincing’ than the authority of an identification system that has not yet stood the test of time.

For the doctrine of evidence law, the Lipps case reaffirms that proven reliability, reproducibility of results, and independent verification must remain the constant criteria for the admissibility of evidence, regardless of the technological nature of the tool. The foundational principle — that no evidence is self-sufficient and that all evidence must be tested in adversarial proceedings — is not superseded by the advent of artificial intelligence. On the contrary: the more powerful the tool, the stricter the procedural framework governing its use must be.

The genuine novelty of our era lies not in the fact that algorithms make mistakes — all identification systems do — but in the fact that the pace and scale of their deployment are outrunning the development of legal safeguards against their errors.