A Guardian Assessment of Age Assurance Technology’s Impact on Marginalized Groups
An analysis by The Guardian of trial data concerning age assurance technology reveals a concerning trend in the government’s impending ban on social media access for teenagers. The research illustrates that the implementation of age verification measures is poised to disproportionately affect already marginalized demographics.
According to the published trial data, the age estimation software displayed a marked lack of accuracy when applied to individuals with Indigenous or Southeast Asian backgrounds. As a result, youth belonging to these communities may be erroneously classified as over the age limit, while older individuals could be misidentified as minors.
Furthermore, individuals from these backgrounds experienced notably longer wait times for results when utilizing the age estimation technology, which relies on facial recognition to estimate age.
The report also highlights that the age verification software, which depends on scanning documents such as driver’s licenses or passports, proved to be similarly unreliable for Indigenous people, although the sample size was insufficient for definitive statistical analysis.
Minimized Bias Concerns
The summary of the report conveys a tempered view of these accuracy discrepancies, asserting that “systems performed broadly consistently across demographic groups, including Indigenous populations.” It acknowledges a shortfall in training the age analysis systems with data representative of Indigenous groups, yet claims “no substantial difference” in outcomes for First Nations peoples compared to other multicultural communities.
Subsequent sections, however, exhibit a more nuanced perspective, admitting to “certain known challenges in accuracy for underrepresented skin tones or facial features.”
The £6.5 million technology trial, orchestrated by the UK-based Age Check Certification Scheme (ACCS), evaluated various technological modalities intended to restrict access for individuals under the age of 16 or 18 when the Australian social media ban takes effect in December.
- Age inference
- Age estimation
- Age verification
The ACCS tested age estimation and verification through real-world scenarios involving school students across Australia, as well as “mystery shopper” testing. Automated testing was also carried out using an image dataset.
While the ACCS has made some of the data public, the Guardian’s analysis is confined to this accessible information. The findings indicate that Indigenous participants exhibited an accuracy rate seven percentage points lower than their counterparts categorized under “Oceania and Antarctica,” which comprises individuals who identified as “Australian” during the trial.
For those with Southeast Asian backgrounds, the accuracy rate was five percentage points lower. The ACCS report focused solely on evaluating the age estimation software’s performance by skin tone through automated testing, refraining from publishing accuracy rates based on demographic backgrounds.
The trial’s assessment outlined that the systems underperformed on average for individuals with darker skin tones, yet the report concluded a lack of significant evidence for systemic disparity, stating: “no significant evidence of adverse impact across skin tone groups.”
Recent examinations, including research from the US government and a leading age estimation firm, have revealed heightened error rates in age estimation for women and individuals with darker skin tones.
No detailed analysis of accuracy based on gender was provided in the age assurance technology trial. Nonetheless, the report made references indicating that “fairness testing showed acceptable parity across age and gender,” while recognizing “variability in model output by gender presentation.”
The ACCS clarified that the report did not include such an analysis due to recommendations from an independent ethics committee, which expressed concerns about potential misinterpretation.
Methodological Scrutiny
The Guardian raised several queries concerning the methodology employed in the trial with the Age Check Certification Scheme (ACCS). The public dataset for age verification tests encompasses merely 328 participants. Due to small sample sizes, the report did not investigate accuracy based on skin tone or demographic background.
However, the data indicated that six of these participants had Indigenous backgrounds, and the age verification system failed for three, suggesting a 50% error rate among Indigenous individuals, although this number is statistically insignificant.
The methodology for testing age inference systems was more generalized and lacked thorough scrutiny, acknowledging that participation could be inconsistent in specific groups such as new migrants and remote First Nations communities.
The report also identified persistently existing digital disparities in remote regions, citing a lack of foundational credentials as a barrier to access. When queried about the report’s claims of consistent performance across demographic groups, the ACCS maintained that these assertions derived from a composite of evidence, including various testing methodologies.
The Guardian scrutinized the claimed overall accuracy of age verification software, contrasting the ACCS’s assertion of 97% accuracy with its own analysis, which suggested a figure closer to 92%.
The ACCS clarified that the 97% statistic was based on additional unpublished lab tests beyond those conducted with the mystery shopper trials.
When questioned about the exclusion of specific results from the public dataset, the ACCS cited these lab results as falling outside the parameters of the anonymised dataset release.
In the report’s analysis, it was noted that two underperforming providers were omitted due to their consistent poor performance, skewing the overall results. The ACCS did not elucidate the rationale behind this exclusion.
Minimal Impact on Most Adults
Despite these concerns raised during the analysis, findings suggest that most adults are unlikely to experience substantial inconvenience from the age assurance technology.
Professor Toby Walsh from the University of New South Wales, who provided independent advice on the report’s methodology, commented that adults would face minimal disruption due to a “cascade” of technological solutions for age determination.
“Many adults will not encounter difficulties, as age inference will often suffice for established social media platforms,” Professor Walsh asserted, emphasizing that for many adults, age estimation technology is remarkably accurate.
Challenges for Younger Users
The report indicates a significant degree of uncertainty within the age estimation tools. Experts highlight that the predictive differences between actual and estimated ages can span several years. Navigating age estimation is particularly complex for adolescents undergoing rapid physiological changes.
Professor Tama Leaver from Curtin University elaborated that the tools may effectively differentiate between ages such as 16 and 30, but struggle to accurately distinguish narrower age ranges, including individuals just shy of or exceeding 16.
Data from the trial reveals alarming false positive rates—ranging from 25% to 73%—for individuals under 16, diminishing significantly only when ages reach 18 or above.
Given these challenges, the report suggests implementing a buffer, necessitating supplementary verification for anyone predicted to be two to three years younger or older than 16.
This development likely implies that more than 1.3 million Australians aged 16 to 19 may need to provide additional information confirming their age when seeking to create social media accounts.
The Guardian reached out to Communications Minister Anika Wells regarding these findings. She deferred questions to the eSafety Commissioner, whose spokesperson indicated that the trial produced valuable independent evaluations of various technologies.
“Enhancing age assurance tools necessitates ongoing training to improve their precision—particularly concerning the identification of diverse ethnic backgrounds,” the spokesperson affirmed.
Thanks to Prof Falk Scholer at RMIT for providing insight during our analysis.
Nick Evershed serves as the data and interactives editor, while Josh Nicholas is a data journalist for Guardian Australia.
Source link: Theguardian.com.
