- Article ID: NU54-041
- Section: Hot
- Slug: facial-recognition-thresholds-investigation
- Prepared: 31 October 2026
- Evidence checked: 31 October 2026
- Dek: A candidate list, a similarity score and a human reviewer answer different questions. The investigation becomes defensible only when the next evidence can stand apart from the face search.
<!-- ARTICLE_BODY -->
The difficult face is the one that survives
A facial recognition system can become more selective while the remaining mistakes become harder for people to spot. That is the uncomfortable problem examined in a research preprint submitted on 16 October 2026. Esha Edakoth and colleagues tested 801 human reviewers using difficult image pairs produced by four commercial algorithms under three security configurations. The authors report that reviewers could reduce false positives, but their performance declined for the errors surviving stricter thresholds. The work is a preprint, and its controlled task should not be mistaken for a measurement of an entire police investigation. Its practical question nevertheless matters: what evidence supports the assumption that a person looking at the screen will reliably rescue a machine's most convincing errors? The research record and abstract.
The answer cannot be supplied by adding the words human review to a contract. A reviewer may examine the same blurry face, receive the machine's first-ranked candidate and know that an investigator expects a useful lead. Each of those conditions changes the task. An accurate person working under a well-designed protocol may improve the result. A tired person asked to approve a plausible candidate may mostly certify the original uncertainty. Neither outcome can be inferred from the existence of a review button.
The central distinction is between a search result and an identification supported by additional evidence. A search narrows attention. It does not establish where someone was, whether the image was captured at the relevant time, whether an apparent resemblance is distinctive, or whether the activity shown amounts to the alleged conduct. Those are separate propositions. Some require witnesses, records or physical evidence. Some remain unresolved.
A useful public account therefore follows the result through the investigation rather than stopping at the algorithm's score. Which images were searched? How large was the gallery? What would the system return if the person was absent? What did the reviewer know? What independent evidence was obtained before any consequential action? This article develops that chain of questions, using explicitly hypothetical examples rather than presenting fictional cases as actual events.
A search has more than one meaning of error
One-to-one verification compares a presented face with an asserted identity. One-to-many identification searches a probe against a gallery of enrolled people. An investigative search may return a ranked list even when none of those people is the person in the probe. A list must have a first entry; that mathematical fact gives the first entry no automatic evidential status.
NIST's current one-to-many evaluation distinguishes thresholded identification from an investigative configuration. Its identification table measures failures to return the correct person above a threshold while controlling false positive identification. Its investigative table describes a fixed candidate list and reports whether the true person appears at the first rank. Those are different tests. A strong score in one is not interchangeable with a strong score in the other. The page also announces a submission closure beginning 30 October 2026 and planned dataset updates, another reason to name the evaluation version rather than invoke NIST as a timeless seal. NIST one-to-many evaluation.
Consider a hypothetical archive containing 100,000 people. An examiner searches an unknown face and receives twenty names. There are at least three questions. Was the actual person in the archive? If so, did the software retrieve that person? If not, did the software still produce a candidate that the examiner treated as meaningful? A test containing only searches for enrolled people can answer the second question while scarcely addressing the third.
The distinction changes the denominator. A pairwise false match rate counts comparisons between different individuals. A false positive identification rate counts searches for absent individuals that return at least one candidate above the threshold. A false negative identification rate concerns searches for present individuals where the true match is not returned under the stated rule. Calling all three accuracy hides what was counted.
The final investigative error has another denominator: people referred onward, interviewed, detained or otherwise affected. It includes the behavior of the system after retrieval. An agency could reduce one technical error while increasing downstream harm by searching far more images, expanding the gallery or weakening the corroboration requirement. Conversely, a cautious workflow could prevent a technical false candidate from becoming a false allegation. Evaluating the whole process requires keeping those stages separate rather than assigning every outcome to the algorithm alone.
The threshold is a decision setting
A similarity score usually describes a relationship between representations generated by a particular system. It is not automatically a percentage probability that the two photographs depict the same person. A displayed score of 92 cannot be read as a 92 percent chance of identity unless a calibration study establishes that interpretation for the relevant population and conditions. Even then, a change in the population or search process may invalidate it.
A threshold sets which scores qualify for further attention. Raising it generally trades fewer false candidates for more missed true candidates. The choice is consequential because the costs are borne differently. An investigator may experience a missed candidate as lost information. A person incorrectly selected may experience scrutiny, anxiety or a much more serious consequence. Those costs cannot be settled by choosing the setting with the most attractive single number.
Imagine an agency evaluating three hypothetical thresholds. At the loosest setting, a test of 10,000 absent-person searches produces 300 searches with a candidate; at the middle setting, 40; at the strictest, five. On a separate set of 1,000 present-person searches, the true person is retrieved 970, 920 and 760 times. These invented figures illustrate a trade-off rather than any vendor's performance. The strictest setting greatly reduces false candidates but leaves 240 true-person searches without the required retrieval. Whether that is acceptable depends on the authorized use and what happens next.
The agency should not let a vendor select the threshold using the same data later presented as an independent test. Tuning chooses a rule. Testing estimates how the chosen rule behaves on other data. When the two are mixed, the published result may partly describe adaptation to the test itself. A defensible report identifies the tuning set, the held-out set and any changes made after results became visible.
Thresholds also need operational custody. If an administrator can change a setting without a record, a procurement evaluation may describe a system that no longer exists. The case record should preserve the threshold or search mode actually used, together with the algorithm version. A later reviewer needs to reconstruct the decision, not merely confirm that the agency once purchased a well-performing product.
Gallery size changes the opportunity for resemblance
Searching more people creates more opportunities for an unrelated person to produce an unusually high score. The exact relationship depends on the algorithm, correlations among images and how a person is represented in the gallery. A simple probability model is useful for explaining the direction, provided it is not passed off as an operational estimate.
Suppose, hypothetically, that each of 100,000 independent comparisons has a false-match probability of one in a million. The expected number of false matches is 0.1. Under the independence assumption, the probability of at least one is approximately 9.5 percent: one minus the probability that all 100,000 comparisons avoid a false match. That is not the same as saying that any particular returned person is 9.5 percent likely to be wrong. It describes a hypothetical search outcome under explicit assumptions.
Real galleries violate the simple model in important ways. They can contain several photographs of the same person, relatives, images gathered under similar conditions, and uneven representation across ages or communities. Search software may combine templates or apply a gallery-specific scoring strategy. These details mean an agency should measure search-level error using its intended gallery structure rather than multiply a brochure's pairwise rate and declare the matter settled.
The gallery also has an evidential history. A name attached to an image may be stale, duplicated or incorrect. Two records may represent one person; one record may mistakenly combine two people. A technically correct resemblance can therefore lead to an incorrect biographical record. Face matching cannot repair a mislabeled source image by itself.
A hypothetical audit might discover that a million-image gallery represents only 600,000 people because some have multiple records. Reporting the gallery as either a million identities or 600,000 images would misdescribe the search. Both counts matter, along with the distribution of images per person. If certain groups have more enrolled photographs, they may have different exposure to selection even before algorithmic performance is considered.
This is why a gallery expansion deserves its own evaluation. The software executable may remain unchanged while the effective system changes substantially. New populations, new image sources and new duplication patterns alter the search environment. Treating the gallery as passive storage overlooks one of the most important components of the result.
The image carries its own measurement limits
A face in a controlled portrait and a face extracted from surveillance video are not equivalent measurements. Resolution, pose, lighting, motion, compression and occlusion can remove or distort distinguishing information. Enlargement can make an image easier to inspect, but it cannot guarantee recovery of detail that the sensor never captured. An attractive enhancement may add a persuasive appearance without adding reliable identity evidence.
The original file therefore matters. A defensible workflow preserves it, records the extraction time and documents transformations. Cropping, rotation, contrast adjustment and frame selection can each affect what the algorithm receives. A reviewer should be able to distinguish an original image from a processed working image. When a technique synthesizes detail, that fact should be explicit rather than buried under the generic word enhancement.
NIST's demographic evaluation notes that image quality can create differential false negative performance, including through exposure and camera-angle problems. It also distinguishes those effects from false positive differences that can occur even with high-quality images. The distinction prevents a convenient but incomplete explanation: improving cameras may address some errors without resolving every disparity. NIST demographic evaluation.
Consider a hypothetical video with twelve usable frames. An analyst selects the frame that produces the most plausible candidate, then reports only that frame's result. The search has now included a hidden selection process. The final candidate appears to come from one query, but the analyst effectively tried several opportunities. A better record preserves the selection criterion, the searched frames and the outcomes, including unsuccessful searches.
Repeated searches are not necessarily improper. Different frames may contain genuinely different information. The issue is whether the record permits someone to understand the procedure. A search chosen before seeing results has a different evidential history from a search chosen because its answer looked convincing. That difference matters even if both images are authentic.
An image-quality rejection is also an outcome worth recording. If a system refuses to process unsuitable images, a vendor may look less productive while avoiding unreliable suggestions. Counting only accepted searches can conceal the cost of exclusions; counting rejected images as ordinary misses can punish a useful safeguard. Reports need separate categories for unsuitable input, failed processing, no candidate, rejected candidate and supported lead.
Demographic performance requires two ledgers
The first ledger concerns how the algorithm behaves on specified groups and image conditions. The second concerns who is exposed to the system and what follows selection. Combining them into one fairness number makes both harder to understand. Equal technical error rates would not establish equal exposure if one community's photographs are searched or enrolled far more often.
NIST's December 2019 demographic study examined differences across the algorithms and datasets available to that evaluation. It found substantial variation and stressed that results depended on the algorithm and application. Those findings remain an important demonstration of why subgroup analysis is necessary, but they do not describe every system sold in 2026. A current procurement needs current evidence for the submitted version and proposed use. NISTIR 8280.
A hypothetical comparison illustrates the two ledgers. System A produces false candidates in 0.2 percent of absent-person searches for one group and 0.4 percent for another. That is a twofold relative difference and a 0.2 percentage-point absolute difference. Both descriptions are correct. If the second group is also searched ten times as often, the distribution of false candidates will reflect exposure as well as technical performance. A single ratio cannot describe the resulting burden.
Subgroup estimates need uncertainty intervals and sample sizes. Observing no errors in twenty tests does not establish a zero error rate. Nor does a broad category necessarily reveal performance for intersections of age, sex, image conditions or other relevant characteristics. There may be too little data for a precise estimate. The honest result is a stated uncertainty, not an empty cell silently treated as satisfactory.
The labels themselves need care. A dataset's demographic fields may come from administrative records, self-report or another coding process. Those categories are not interchangeable, and the analyst should not infer biological explanations from a statistical association. A difference may involve image collection, training data, representation, the scoring procedure or combinations of these factors. The observed pattern establishes a problem to investigate; it does not identify its cause on its own.
For a public reader, the useful question is concrete: which groups, under which conditions, experienced which error, and how many people were affected downstream? That phrasing resists both blanket reassurance and blanket claims that every algorithm behaves identically.
Human review needs a test of its own
A human reviewer is another measurement component. The reviewer has training, information, time limits and a decision rule. Performance depends on those conditions. A workflow cannot borrow a machine benchmark and assume that the person who follows it contributes a known amount of additional accuracy.
The October 2026 preprint offers a focused warning about this assumption, not a universal verdict on human ability. Its difficult comparison pairs were selected from algorithmic false positives. That selection creates a task unlike comparing random pairs of faces. The remaining pairs can be unusually similar precisely because they fooled the machine. A reviewer test based on easy, randomly selected nonmatches might therefore overstate usefulness in the actual queue.
A hypothetical evaluation could divide reviewers into two groups. One sees candidate pairs without the algorithm's rank or score. The other sees the usual operational display. Both receive the same decision instructions and can choose match, nonmatch or inconclusive. Comparing their results would help isolate the interface's influence. It would not prove that one display is best in every setting, but it would turn an assumption about automation bias into a testable design question.
The inconclusive option is important. If a system forces a binary answer, uncertainty is converted into apparent certainty. A reviewer who cannot distinguish the faces may be pressured to approve or reject without sufficient information. Allowing an unresolved result preserves what the evidence actually supports. The agency then needs an explicit rule for what an unresolved result permits, rather than treating it as a weak approval.
Two reviewers do not automatically provide independent confirmation. If the second sees the first person's conclusion, both may share an initial error. If both see the same poor image and the same ranked suggestion, their errors may be correlated even without direct communication. Agreement is useful information, but its evidential value depends on the process that produced it.
A strong evaluation records individual judgments before discussion, measures disagreement and checks performance on cases representative of operational difficulty. It also asks how often reviewers overturn machine suggestions correctly and incorrectly. Counting only correct rejections makes review look safer than it is if some true candidates are rejected at the same time. The full confusion pattern belongs in the assessment.
An independent next step must add new information
Independent corroboration is more demanding than repeating the face search in another interface. Two products may rely on overlapping images, related training data or similar representations. Even genuinely different algorithms may be affected by the same low-quality probe. Agreement can increase confidence only to the extent that the evidence streams contribute information beyond their shared weaknesses.
The FBI's June 2019 congressional statement described its face-search output as a potential investigative lead requiring follow-up corroboration. That is the agency's stated approach at that time, not proof that every later case followed it. The distinction between a lead and corroborated identification is nevertheless clear enough to evaluate in a case record. FBI testimony.
Imagine a hypothetical candidate who resembles a person in a shop video. An investigator finds that the candidate owns a dark jacket like the one shown. That fact may add little if dark jackets are common. A reliable time-stamped record placing the candidate elsewhere could be much more informative, including as evidence against the lead. The purpose of corroboration should be to test competing explanations, not merely accumulate compatible details.
The same principle applies to witnesses. Showing a witness the machine's preferred candidate before obtaining an independent account can change the evidential pathway. A later statement may partly reflect exposure to the suggestion. This article does not prescribe legal identification procedures; it identifies the analytical problem. Information generated after a person sees the proposed answer may not be independent of that answer.
An investigator can make the distinction visible by keeping a proposition ledger. One entry asks whether the candidate is the person in the image. Another asks whether the image belongs to the relevant event. Another asks what the person did. Each entry lists supporting evidence, contradicting evidence and unresolved alternatives. A face match contributes mainly to the first proposition. It should not silently supply the others.
The ledger also prevents a negative search from becoming an exclusion it cannot support. Failure to retrieve someone may reflect absence from the gallery, poor imagery or a technical miss. It is not automatically evidence that the person was elsewhere. A method that demands caution about positive results should apply comparable care to negative ones.
The base rate follows the investigative setting
A search performed after other evidence identifies a narrow group has a different prior context from a broad search of unknown people. The same technical false positive rate can produce very different proportions of reliable leads. This is a probability issue, not a claim that prior suspicion is necessarily justified.
Take a deliberately simplified hypothetical system with a 90 percent true-positive rate and a one percent false-positive rate at the search level. In 10,000 searches where only one percent of probes belong to people actually present in the relevant gallery, there are 100 present-person searches. About ninety yield a true positive under the assumptions. Among 9,900 absent-person searches, about ninety-nine yield a false positive. The resulting positive pool contains roughly as many false as true outcomes, despite the apparently low false-positive rate.
Now change only the prevalence of present-person searches to fifty percent. There are 5,000 present-person searches and 5,000 absent-person searches. The assumed system yields about 4,500 true positives and fifty false positives. The positive pool is much more reliable. These invented calculations are not estimates of any police program. They demonstrate why a technical error rate does not, by itself, tell a reader how often an operational lead is correct.
The calculation also has limits. A real search may return several candidates, produce correlated errors and operate with changing gallery coverage. A present person can be missed while another person is returned. The binary example simplifies those possibilities to make one point visible: interpretation depends on the population of searches, not merely the algorithm's laboratory result.
Agencies may not know their operational base rate precisely. That uncertainty should appear in analysis rather than be replaced by a convenient assumption. They can report scenarios over plausible ranges, distinguish known-present from unknown-presence searches and audit resolved cases without pretending that resolved cases represent every unresolved one.
A program that reports only successful identifications creates an especially misleading denominator. Failed, unresolved and rejected searches disappear, leaving a portfolio of successes rather than an evaluation. The public needs the complete flow of cases, including how many searches never produced a usable lead and how many initially plausible leads were later contradicted.
The case record should survive a software update
Months after a search, the vendor may update the algorithm, the gallery may grow and the analyst may move jobs. A screenshot of the candidate's name will not reconstruct the procedure. A reproducible case record needs enough detail to establish what was actually done without distributing unnecessary biometric material.
At minimum, the internal record should connect the authorized purpose, original probe, processing history, query time, gallery description, software version, search settings, returned candidates and reviewer decisions. Access to those records can be restricted appropriately. The point is not to place people's faces on a public dashboard. It is to preserve a traceable account for authorized review.
GAO's 2021 review identified gaps in federal agencies' awareness and tracking of employees' use of nonfederal facial recognition systems. That historical finding matters because a procurement inventory can miss services accessed through partners or external accounts. The report's recommendations and later status entries should be read with their dates, rather than treated as a single present-day condition. GAO's 2021 review.
A hypothetical agency might inventory one purchased platform while investigators also request searches from three partner organizations. Its annual report could truthfully count the purchased system's queries and still omit much of its practical use. A complete accounting follows the request, including outsourced and shared searches, rather than stopping at ownership of the software license.
Version retention presents a practical problem. An agency may not be able to rerun an old proprietary model indefinitely. It can still preserve the original output, settings, signed logs where available and the vendor's version identifier. If exact reproduction is impossible, that limitation should be documented before a dispute arises. A promise of reproducibility is weaker than a tested ability to reproduce a representative case.
Data correction also needs a path through the record. If an enrolled image is later found to be mislabeled, the agency should be able to identify affected searches and reassess dependent conclusions. Deleting the bad record without preserving an audit event can conceal the reason earlier results changed. Correcting evidence requires both preventing future error and understanding past reliance.
Training records are not performance records
Training is necessary evidence about preparation, but attendance does not establish competence in every task. A course can explain policy without testing difficult image comparisons. It can teach a software interface without teaching how to communicate uncertainty. A certificate therefore belongs beside, rather than in place of, measured performance and supervised practice.
In its September 2023 report, GAO examined seven federal law enforcement agencies and identified shortcomings involving required training and policies addressing civil rights and civil liberties. A March 2024 testimony summarized that work. These are dated findings about selected agencies, not a census of all law enforcement or a claim that no improvements followed. GAO's training and policy review.
For a hypothetical procurement, training acceptance could require an examiner to demonstrate several distinct skills. The examiner should recognize unsuitable imagery, interpret ranked results without converting scores into unsupported probabilities, document transformations, identify when the correct decision is inconclusive and explain what additional evidence is needed. Testing only whether someone can upload a photograph misses most of the consequential work.
Supervisors need their own competence criteria. A supervisor reviewing a report must be able to detect language that overstates the result. Phrases such as confirmed by facial recognition can erase the distinction between algorithmic resemblance and independently established identity. The report should state what the system returned and what the examiner concluded, then separately describe any corroboration.
Feedback must include mistakes and unresolved cases. If reviewers hear only that their successful leads helped an investigation, they may develop an inflated sense of reliability. If they never learn that a candidate was excluded, the workflow loses an opportunity to improve. Feedback should be structured to support learning without exposing information beyond the reviewer's legitimate role.
Periodic assessment matters because skills and conditions change. A new image source, different interface or revised candidate-list length may alter performance even when training requirements remain formally satisfied. The agency should ask whether its competence test still resembles the work people actually perform. A useful training record establishes preparation for a defined workflow on a defined date, not permanent immunity from error.
Public reporting can measure consequences without naming people
A public report can reveal a great deal about a system without publishing individual faces or investigation details. It can count authorized searches, rejected inputs, candidate-producing searches, examiner rejections, unresolved results, independently supported leads and later corrections. Each category should be defined so readers can follow cases through the process without assuming every query represents a distinct person.
GAO's January 2024 testimony on forensic algorithms emphasized that outcomes depend on factors beyond the underlying algorithm, including the evidence and how the tools are used. That broad systems perspective supports asking for workflow measures rather than relying exclusively on a vendor's benchmark. GAO's forensic algorithms testimony.
The report should distinguish queries from probes, probes from events and events from people. One video can produce several frames and repeated searches. One candidate can appear in multiple events. A count of 10,000 queries cannot be read as 10,000 investigations, 10,000 identified people or 10,000 successful matches. Publishing the relationships among those units is more informative than choosing the largest number.
Outcome audits face selection bias. Cases resolved through other evidence may differ from cases left open. If the audit evaluates only confirmed successes, it will overestimate utility. If it treats every unresolved lead as false, it will overestimate error. A transparent report preserves an unresolved category and explains how follow-up information was obtained.
A useful hypothetical audit samples from the full search register before checking outcomes. It then tracks how many sampled cases have enough independent evidence for evaluation. Results are reported for that evaluable subset, with the missing fraction visible. Sensitivity analysis can show how conclusions would change under different assumptions about unresolved cases. The analysis should not hide uncertainty behind an apparently precise overall accuracy percentage.
Consequences also deserve separate measurement. A false candidate discarded at the review stage is different from a false candidate passed to investigators or used in a consequential decision. All matter, but their remedies differ. The first may call for better retrieval or review. The latter may also reveal a failure in corroboration, communication or authority. A single total of errors cannot locate the point where protection failed.
A small test cannot certify a rare error rate
Rare errors create a measurement problem that procurement summaries often hide. A test can be large enough to demonstrate routine performance while remaining too small to estimate the low-probability mistakes that matter most. If a buyer expects fewer than one harmful false candidate in a very large number of searches, a demonstration containing a few hundred convenient examples will not establish that claim.
For a simplified hypothetical calculation, suppose a fixed system is tested on 3,000 independent absent-person searches and produces no false positive identifications. Zero observed errors is encouraging. It does not establish a true error probability of zero. A common approximation for the upper end of a one-sided 95 percent bound is about three divided by the number of independent trials. Here that is about 0.001, or one in a thousand. The approximation is a statistical illustration with independence and representativeness assumptions, not a certification rule for facial recognition.
Those assumptions deserve as much attention as the arithmetic. Three thousand searches made from near-duplicate photographs do not provide the same information as 3,000 genuinely varied searches. A clean portrait collection may not represent a program dominated by oblique nighttime images. A test that excludes difficult images after seeing failures may estimate performance only on the remaining easy subset. Increasing the sample count cannot repair a systematically unrepresentative sampling frame.
A buyer can separate two test goals. One is a controlled comparison among candidate systems using the same data and settings. The other is an estimate of operational performance under the intended workflow. The first may identify relative strengths while the second reveals whether any option is suitable. A system can win a comparison and still fail an operational acceptance criterion. The fact that competitors performed worse does not make the winning error rate acceptable.
Acceptance criteria should be written before final results are inspected. Otherwise the buyer can unconsciously redefine success around the product it already favors. A hypothetical requirement might specify the maximum search-level false positive rate supported by a confidence bound, a minimum retrieval rate on defined image conditions and a maximum proportion of inputs rejected as unusable. It should also state what happens when a subgroup has too little data for an estimate. An unmeasured condition should not inherit a passing grade from the overall average.
The cost of obtaining adequate evidence can itself affect the deployment decision. If an agency lacks the data, authority or resources to evaluate a consequential use, that is an unresolved limitation. It is not evidence that the use is safe. A narrower deployment may be testable when a broad one is not. For example, a controlled collection process can constrain image quality in a way that arbitrary footage cannot. That narrower result should remain attached to the narrower conditions.
Finally, testing should preserve cases that expose failures. Removing a troublesome image from the next evaluation may improve a score while discarding information about a real weakness. A maintained challenge set can track whether revisions fix known problems, while a separate fresh set tests generalization. Keeping both prevents a software update from being judged solely by improvement on familiar examples. The resulting evidence is more demanding than a product demonstration, but it answers the question a public agency actually needs answered: how much uncertainty remains when this system is used in this way on these people?
The standard is a traceable change in knowledge
The most useful question after a face search is what the investigation now knows that it did not know before. Sometimes the answer is a promising name worth checking. Sometimes it is that the image is too poor for a responsible conclusion. Sometimes the search adds nothing. A mature workflow can record all three without forcing every use to become a success story.
Procurement should begin with that distinction. Buyers can require representative search-level testing, explicit thresholds, gallery documentation, examiner evaluation and a record of independent follow-up. They can reject unsupported claims that a benchmark, a human reviewer or a second algorithm automatically establishes identity. These are proposed evidence requirements, not a statement that one universal legal rule governs every jurisdiction.
The independent next step is where that uncertainty meets the world. A record that can be checked, an account obtained without suggestion, a reliable timeline or evidence that excludes the candidate can change the investigation's knowledge. Repeating a resemblance claim in stronger language cannot. The distinction protects both the person under scrutiny and the quality of the investigation.
A facial recognition match should therefore be reported as a precisely described event in an evidence chain. A particular version searched particular images under particular settings and returned particular candidates. People then made documented judgments, and additional evidence either supported, contradicted or failed to resolve the lead. Keeping those statements separate is less dramatic than a claim that a computer identified someone. It is also far more useful when the conclusion has to withstand scrutiny.
<!-- /ARTICLE_BODY -->