Guest post: Serious accuracy issues identified in Foodstuffs facial recognition trial
Guest post by Mary Sagen
The Office of the Privacy Commissioner (OPC) conducted an inquiry into Foodstuffs (FSNI) 6-month trial of facial recognition technology (FRT) across 25 stores in the North Island from February 2024 to July 2024. The OPC’s report was released on 4 June 2025. The same day, RNZ published an article that the technology’s benefits outweigh the concerns followed by another article indicating we could see this surveillance technology used across most major New Zealand retail outlets.
Let’s take a closer look at some of the serious accuracy issues identified in the OPCs inquiry report and why they matter, in order to decide if people in some industries ‘crying out’ for facial recognition should be able to make a decision with such significant implications for our human rights.
Accuracy, necessity and validity concerns
Earlier reporting on this trial focused on 13 instances where the technology’s alert verification process failed out of 1736 matches it generated from the store’s uploaded watchlists. However, taken from page 56 of OPCs report, we see there were 117 potential instances of where the alert verification process failed.
There were up to an estimated 13 instances where the alert verification process failed and individuals were approached without at least two authorised staff members manually verifying the match, or where the staff discovered after the fact that they had the wrong person.
Store notes for a further 70 cases were unclear about whether the person was a match and whether they were approached in store. A further 31 cases were clear that the person was approached but notes did not record the match. Notes for three further incidents were unclear about the status of the match and whether the person was approached. (Bolding, mine)
To have 117 potential failures out of 1736 alerts generated portrays a significantly worse accuracy issue than the 13 estimated instances of misidentification highlighted in the media. We’re told the alert failures can be attributed to ‘human error’. However, when the actual number of alert verification process failures in the trial was much higher, it is harder to sustain such a dubious claim. Paragraph 222 in the OPCs inquiry report “recommends good record keeping to be certain”. Therefore, we can not be certain of the actual number of alert verification process failures or, total number of misidentifications by ethnicity or gender, due to Foodstuff store’s poor recordkeeping during the trial.
Other serious accuracy issues identified in the OPC report are with the watchlists used and uploaded into the facial recognition system. According to a direct reprimand in paragraph 170 of the inquiry report, “FSNI should have checked for low quality images before the trial went live”. In paragraph 211, the OPC identifies this failure to check for low quality images as one of four factors contributing to a serious misidentification incident.
Note, too, that the trial methodology used in the Foodstuffs facial recognition experiment to demonstrate necessity, changed after the trial began. The Non-FRT stores enthusiasm for participating in the 6-month trial waned as time went on, leading to less quality data being collected; the kind used for making precise comparisons to determine effectiveness which determines necessity. What started as a comparison study between FRT stores and Non-FRT stores, became a non-random controlled trial comparing how outcomes changed over time. (See paragraphs 116-120 of the OPC report).
Why does it matter?
When we witness the live use of a novel technology to make formal judgements of others, and given facial recognition is a form of profiling, we must be certain it is for a lawful purpose and necessary. There are 13 Individual Privacy Principles (IPP) in the Privacy Act and significant attention is given to IPPs 5-9 focused on accuracy, retention and security issues. Under IPP 8, an agency that holds personal information must not use or disclose that information without taking any steps that are, in the circumstances, reasonable to ensure that the information is accurate, up to date, complete, relevant, and not misleading. To have 117 alert verification process failures and no records about the ethnicity and gender indicates an inherent issue with data accuracy. It matters because there are known problems with facial recognition returning higher rates of misidentifications of people with darker skin tones. Additionally, the poor quality images on the store’s watchlists ensured that the information uploaded to the facial recognition system and used by it to make matches when people enter stores, was not accurate. That FSNI didn’t identify and plan for mitigating this risk during their 15-month consultation period prior to starting the trial, highlights that appropriate and reasonable steps were not taken to ensure that the information used was accurate.
Further, in paragraph 134, the OPC details the legal tests for what would make the technology necessary to achieve FSNIs dual purpose of reducing harm and preventing theft. That there is no less intrusive option available is one of them. However, there’s no data or way of knowing if a less intrusive option (Non-FRT) was available or effective due to a failure in the trial methodology. The comparison study between Non-FRT and FRT stores broke down due to lack of regular reporting by the Non-FRT stores. This tweak and changing of the trial methodology is significant too, because of time related biases, confounding variables, and trial validity. The inherent data accuracy issues and inability to demonstrate the necessity of facial recognition are more than just concerns under IPP8, they are the key reasons why FSNI’s trial failed and continued use is unlawful.
Identified omissions from the inquiry report
In addition to revelations about inaccuracy and validity concerns, there are identified omissions from OPCs inquiry report. In paragraph 238 it is noted, “covert use of FRT is inherently unfair and intrusive”. Yet, we see Foodstuffs earlier covert trials in 2018 are not visible on the timeline on page 71 and missing from the section, How FRT is used in New Zealand, on pages 15 and 16. There is a reference in paragraph 90 to FSNI halting stores already using it in 2021 that later became part of the trial; however, these details are missing from the timeline and page 15.
A second omission is a reference in paragraph 241 of the inquiry report to FSNI’s in-store public opinion survey. We’re informed by FSNI’s self-reporting that the majority of people responding to the survey broadly support Foodstuffs use of facial recognition. Yet, the OPC also says that “a majority of people preferred more
security staffing to FRT in principle”. Because the survey is not made available in the appendix of the inquiry report published by the OPC we are unable to scrutinise the detail. As the survey is relevant to the inquiry and referred to in OPCs report, why was a copy of the survey omitted from the report? Survey design and structure and content are important and can have an impact on answers one is hoping or wanting to receive. Let’s see it.
Why does it matter?
Covert use of facial recognition erects a firm and enduring barrier to building trust and where there is asymmetry in information there is a higher likelihood of distrust. When important and relevant context is omitted it raises questions. After all, context helps us understand why something is happening which is important in helping to determine what to do about it.
In the OPC’s inquiry report, for example, we find just one single mention of Covid-19. Nothing about extended lockdowns, vaccine mandates and certificates, job losses, the cost-of-living crisis or, persistent and rising unemployment. The results of this time of fear, stress, change and upheaval play out in our essential services experiencing more disgruntled people, bad behaviour, some serious assaults, and increasing theft.
These are societal issues and surveillance tools do not work as per public statements made by the Privacy Commissioner and reiterated in paragraph 45 in the report:-
Beyond a possible deterrence effect, the technology alone does not prevent or reduce incidents of violence, harassment, or shoplifting.
Adding to this, published studies show possible deterrence will decay and the CCTV effect diminishes as time moves on. However, while there’s talk of less intrusive alternative measures to address it, where are the mentions of paying a living or thriving wage to our essential workers in our essential services and creating a positive flow outward? Ultimately, the choice made by FSNI appears, in part, a selfish one; to bring formerly shadowy covert use of invasive surveillance technology out into the open. That a company chooses to do this because governments have failed to address the policies creating poverty and desperation might be a rational choice on its part, but the OPC should not feel constrained from commenting on the contextual economic and social factors.
Potential for confirmation bias
Trials of new technologies are known to be at risk of confirmation bias. This is when the evaluation of the results of a trial are skewed by both the company’s desire to proceed with its intended course of action, and its knowledge of how much it has already invested in the technology. It’s important to increase public awareness of the volume of time, effort, and money FSNI invested in FR. From procuring this technology developed offshore by Australian-owned Vix Vizion, to installing it, uploading watchlists, operationalising the system, training staff, undertaking the trial, reporting its results, and more. Additionally, the OPC invested a significant amount of time and effort to conduct site visits to the 10 stores participating in the trial. On these site visits the OPC held interviews with store staff that were supervised by Foodstuffs. Producing their 72 page inquiry report represents use of the OPC’s time and resources, too.
Why does it matter?
We should consider the risk of confirmation bias, due to the emotional and economic investment of all parties. How was a potential bias in wanting to see and portray this trial as a success mitigated? How heavily, for example, do we weigh the report from Scarlatti, an analytics firm paid by Foodstuffs to conduct an evaluation of their FR trial? How heavily do we weigh the OPC’s inquiry into it when their office proposed conducting the trial and worked closely with Foodstuffs on it?
Concerns about lack of consent
Along with accuracy and reporting problems, omissions of context, and potential for bias, there is the concern about lack of informed consent. NZ law is behind other countries and actively permissive when it comes to requiring informed consent. We are reminded of that from the recent incident involving actions taken by Prime Minister Luxon’s press secretary (who subsequently resigned). Currently, Foodstuffs isn’t getting informed consent either, as addressed in paragraph 235 of the inquiry report, “only 67% of surveyed customers were aware of FRT was being used in their store”. Should New Zealand follow the Australian example and ban the remote capture of our faceprints, our biometric information, without gaining our explicit consent? Furthermore, when our supermarket duopoly means that in many towns people effectively have no choice about which supermarket to buy their groceries from, is informed consent really a meaningful test?
Conclusion
This issue is relevant to all the people in New Zealand, as necessary stakeholders in the decision making process about the wider roll out of facial recognition. This is a technology which has significant implications upon our human rights to be free from discrimination, our rights to privacy, and freedom of movement. This article has considered arguments that disprove the OPC’s conclusion that Foodstuff’s 6-month facial recognition trial complies with IPP8 of the Privacy Act. We examined how the trial, by changing methodology, failed to prove the use of facial recognition was necessary or more effective than less intrusive methods (Non-FRT). Also, we identified four more concerns including the inquiry report omissions, the potential for confirmation bias, lack of context, and lack of informed consent.
There’s a lot to consider and more carefully than some are considering it because according to paragraph 199 in OPC’s report, “The evaluator’s best judgement is that the raw increase in the number of approaches means that it is most likely that FRT increases the likelihood that people are falsely approached”, This reverses the presumption of innocence, potentially leading to wrongful accusations of a person being a trespassed thief based on inaccurate or biased results. These problems will continue to occur because facial recognition technology, while advancing, is not infallible and does misidentify individuals, particularly those from marginalised groups. Should we be actively reducing protections of people from wrongful accusations? Should Foodstuffs be using facial recognition after failing to prove its accuracy and necessity? Finally, this report by the OPC will have thrown open the floodgates to much greater use of remote facial recognition across New Zealand, turning us into a surveilled society. We are already seeing from the USA how these technologies are being abused by authoritarian state forces.
About the author: Mary Sagen is witnessing preventable harm caused by unlawful use of invasive surveillance technology and the global doubling of self harm and suicide caused by social platforms lacking robust safeguards. She advocates for all New Zealanders to be involved in decisions about technology which impact on our human rights and our right to not be subjected to the arbitrary or unlawful interference with our privacy.