Policing Facial Recognition — Between Dangers, Misconceptions, and the Want for a Extra Sincere Debate – Cyber Tech

 

Asress Adimi Gikay (PhD) Senior Lecturer in AI, Disruptive
Innovation and Regulation, Brunel College of
London

 

Photograph credit score: Abyssus, by way of Wikimedia commons

 

Reside facial recognition on the rise

 

Reside facial recognition (LFR), is shortly gaining floor throughout
Europe, with international locations like Germany having used it to
goal severe legal offences. The expertise scans folks’s faces in actual
time and matches them towards police watchlists (e.g., folks suspected of
committing severe crimes). The EU’s Synthetic Intelligence(AI) Act, permits
police in member states to make use of LFR for severe crimes equivalent to terrorism. Nevertheless, the
implementation of the EU AI Act in member states will possible face challenges as
technical points equivalent to accuracy and authorized boundaries are but to be adequately
examined.

 

In the meantime, the UK Metropolitan Police have gained an intensive
expertise in managing the danger posed by the expertise, arresting greater than 1,000 folks between January 2024
and August 2025.  In August 2025, regardless of
opposition from 11 civil
liberty teams, the Metropolitan deployed LFR on the Europe’s largest avenue
pageant celebrating African-Caribbean tradition,  Notting Hill Carnival,  making  61 arrests.

 

The Metropolitan Police have taken probably the most step to deal with one of many
greatest challenges in using the expertise, i.e., ethnic bias. Nevertheless, an issue
stays as as to if ethnic bias has been adequately tackled with knowledge being
interpreted in a different way to assist the precise narrative being superior.
False impression or misframing of essential notions within the discipline surveillance additionally form
public notion and will probably inform coverage and regulatory decisions
that aren’t essentially proof based mostly.  I imagine the prevailing positions adopted by teachers
and civil society teams additionally partly replicate such a state of affairs— selective
use of knowledge, unwarranted anxiousness about surveillance and misconceptions round core
authorized ideas.

 

The view  predominantly
superior at the moment by teachers and civil liberty teams is a proposal for banning or imposing moratorium on using LFR on
the bottom that it’s inaccurate, ethnically biased, vulnerable to racially
discriminatory use and permits mass surveillance. While these are legitimate
considerations, the Metropolitan Police’s expertise over the previous decade and the
debate it sparked illustrates that the controversy over governing the expertise
typically doesn’t pretty weigh human rights and public security considerations. Based mostly on
the experiences from using LFR expertise in UK policing, on this put up, I
cowl points that always don’t floor wider-public discourse, a few of these
points being essential in offering insights into how LFR expertise can deployed
within the EU beneath the AI Act in addition to different jurisdictions.

 

From backlash to acceptance 

 

Critics typically describe policing facial recognition as Orwellian surveillance software.  But historical past exhibits facial recognition shouldn’t be
the primary or solely expertise to boost such a concern.

 

When Transport for London launched a poster in 2002 saying CCTV on buses, the design featured a double-decker
bus gliding beneath a sky, with floating eyes. Its slogan learn— “Safe
Beneath the Watchful Eyes.” Simon Davies, the then head of Privateness
Worldwide described it as “acutely disturbing.”  Twenty years later, CCTV
is broadly accepted as a vital software for fixing
crimes.  

 

 

Large Brother
Watch, initially opposed airport facial recognition e-gates, warning that
the system creates  privateness intrusive
huge database of non-public data and is vulnerable to danger of error. Immediately,
automated border management in Europe is taken into account a privilege, permitting sooner
passport management, out there primarily to European passport holders. ‘Different
travellers’ endure extra intrusive safety management, together with via fingerprints.

 

New applied sciences normally prompted alarm, till their public advantages
turn out to be clearer they usually acquire legitimacy. I don’t imagine policing facial
recognition is any totally different.

 

Measuring
the impression of ethnic bias is hard

 

Issues about bias in facial recognition stem
from early research of business gender-classification
algorithms and Metropolitan Police’s preliminary deployments that confirmed poorer accuracy particularly for black girls.  

 

Nevertheless, a 2023 audit by the Nationwide Bodily Laboratory (NPL), commissioned by the Metropolitan police discovered that when the system
is optimally set, it really works with out vital ethnic disparities.

 

A vital issue is the ‘recognition confidence
threshold,’ or ‘face match threshold’ which determines how precisely the
software program matches faces. It ranges between 0-1. Increased settings scale back errors
however yield fewer face matches whereas decrease settings give extra matches with much less
accuracy. The Metropolitan Police at present makes use of 0.64, a degree advisable by the NPL to cut back ethnic bias vital
sufficient to deal with is as not regarding(statistically insignificant).

 

The NPL’s take a look at concerned 400
volunteers embedded in an estimated crowd of 130,000. The take a look at confirmed that at a 0.64
setting or greater, there was no ethnic disparity in accuracy. At thresholds of
0.62 and 0.60, ethnic bias was statistically insignificant, whereas at 0.58 and
0.56, the system struggled to determine black faces.

 

Pete Fussey, a
recognised knowledgeable on this discipline, contends the pattern was too small to
assist  such a conclusion and notes that “false
matches weren’t truly assessed on the settings the place ethnic bias was
non-existent”.   This primarily rests on the truth that for a
expertise that scans hundreds of thousands of faces, testing it on faces of 400 volunteers
is much less prone to generate a ample proof base. Of their guide, Facial Recognition Surveillance: Policing within the Age of
Synthetic Intelligence (p, 58), Pete Fussey
and Daragh Murray argue:

 

“Additionally of notice are claims that no demographic bias
is discernible above the 0.64 threshold. It is because no false positives
occurred at this degree. Put one other method, no bias was noticed as a result of the
system was not adequately examined on this vary. Notable right here is that such
arguments relaxation much less on how FRT operates and extra on how statistics work  An acceptable analogy can be the declare that 90
per cent of automobile accidents happen inside a quarter-mile of dwelling. That is much less
as a result of such locales are inherently hazardous and extra as a result of nearly all automobile
journeys occur inside a quarter-mile of dwelling. Fewer journeys happen 600 miles
away so accidents in that class are rarer. ”

 

Nevertheless, a counter-argument to above is that the
take a look at in query did present regular decline in ethnic disparities with greater face match thresholds— at 0.56, 22 vs. 3
(Black vs. White); at 0.58, 11 vs. 0; at 0.60, 4 vs. 0; at 0.64, 0 vs. 0.  Regardless of the pattern being smaller, the
constant decline implies that face match threshold clearly determines
accuracy. The insistence on testing the expertise till bias is totally
eliminated can also be unrealistic. So, if no inaccuracy was recorded at 0.64 and
ethnic bias declined step by step as much as that time, it might not be unreasonable
to conclude that the expertise works optimally on the given setting.

 

The NPL’s take a look at is in line with the danger administration
system within the EU AI Act, which units strict requirements for high-risk AI programs.  In its provisions requiring danger administration for high-risk AI programs, particularly article 9(3), the AI Act requires that

 

“The chance administration measures referred to in paragraph 2, level (d),
shall be such that the related residual danger related to every hazard, as
nicely as the general residual danger of the high-risk AI programs is judged to be
acceptable.”

 

It implies that the expectation by way of dangers
together with danger of ethic bias shouldn’t be a whole elimination moderately it’s
mitigation to the extent that some acceptable(tolerable) degree of dangers may
nonetheless exist. By these normal, NPL’s testing is probably going thought of strong,
since at 0.64 ethnic bias would moderately be seen as low sufficient to be
acceptable in view of the expertise’s advantages

 

 

Subsequent Metropolitan Police’s deployment knowledge
can also be indicative of this.  Between
January and August 2025, the Metropolitan Police have misidentified solely eight folks
utilizing LFR, resulting in no arrests. Whereas ethnic breakdown
for these false matches shouldn’t be studied, the small quantity makes any ethnic
disparity possible negligible.

 

At present, there’s one pending authorized motion
introduced towards the Metropolitan Police by Large
Brother Watch regarding extended police engagement with a
mistakenly recognized particular person. This was not formally documented as false
arrest, and due to this fact the official document within the UK is that there has not been
a single false arrest following misidentification by LFR within the UK.

 

The above highlights that statistics alone
doesn’t seize the advanced methods LFR actually have an effect on folks. Human oversight,
accountable police judgment, and procedural safeguards play a vital position; and
the present debate reductions these elements.

 

Policing by
consent isn’t policing by of everybody’s consent

 

A standard false impression is that overt(clear)
LFR surveillance undermines policing by consent, as folks don’t meaningfully
consent to being surveilled. 

 

Peter Fussey and Daragh Murray argue that, for cases, signages positioned by the Metropolitan Police
at deployment spots to tell the general public of LFR operations have been inadequate to
receive knowledgeable consent, as they contained insufficient data, lacked
visibility and provided no alternative for refusing consent.

 

Echoing this, former director of Large Brother Watch, Silkie Carlo said in an interview, “there’s no significant consent course of
by any means. You definitely can’t withdraw consent.”

 

I feel this view misrepresents each the legislation and
the thought of policing by consent. The related UK Surveillance Digital camera Code of
follow requires overt surveillance to be based mostly on consent, particularly
clarifying that consent on this context 
needs to be considered “analogous to policing by consent”.

 

Policing by consent is  traced to the 9 level ideas of Robert Peel, UK’s Residence Secretary
set out within the basic directions issued to
new officers in 1829. Primarily, it requires public consent for police to serve the neighborhood the place the legitimacy of policing energy drives from public assist. It
doesn’t require particular person member of the general public to consent to particular
policing operations.

 

Equally, surveillance by consent requires the
neighborhood broadly to comply with seen digicam programs as a respectable software for
public security, not whether or not everybody agrees to the surveillance. In addition to
facilitating legitimacy, clear police surveillance ensures that these
aggrieved by probably illegal surveillance can take authorized actions.  The Surveillance Digital camera Code
of Apply itself which is the premise for transparency in overt surveillance
confirms this level by not solely specifying that consent on this context is
equal to policing by consent but in addition indicating the rationale why consent is
required. Part 3.3.2. states
that “Surveillance by consent depends upon transparency and
accountability on the a part of a system operator. The availability of
data is step one in transparency and can also be a key mechanism of
accountability.” Nowhere within the code or every other laws is it said that
surveillance by consent entitles people to consent to or withdraw consent
to particular operations on particular person degree. Regardless of quoting the SCC together with
the related reference to policing by consent of their current guide, Peter
Fussey and Daragh Murry don’t have interaction with the notion of policing by consent when
they talk about consent within the context of overt surveillance, as an alternative participating
with knowledge safety legislation notion of consent. If consent of everybody who may very well be
captured by LFR digicam or perhaps a regular CCTV got here is to be secured, most public
going through CCTV cameras must be eliminated.  

 

It’s due to this fact legally and conceptually
unfounded to assert that overt LFR surveillance requires the consent of everybody
who walks by the LFR digicam. Neither can this be realistically achieved in
follow.

 

Surveillance
harms, however context issues 

 

Opponents typically alert that surveillance in public
house, can deter folks from talking freely, attending protests, or becoming a member of
public occasions, a phenomenon known as the ‘chilling impact.’

 

Within the context of LFR, Daragh
Murray asserted
that it’d discourage attendance on the 2025 Notting Hill Carnival, citing
uncertainty about how the expertise is used and historic allegations of
institutional racism towards the Metropolitan Police.

 

The 2024 Carnival skilled two murders, a number of assaults, and stabbings, and but an estimated two million folks attended the Carnival this yr, undeterred by the potential violence.
Suggesting that surveillance would deter participation in such a cultural occasion
is clearly implausible.  On the very
least, there isn’t any proof to again this declare.

 

The chilling impact of surveillance is a priority
within the context of political protests, the place authorities might goal opposition
teams and threaten civil liberties. It may also be argued that extreme
policing of minority communities might create a chilling impact to some extent,
although that is extremely context dependent. For instance, the 2025 Carnival had 7,000 cops with
supporting applied sciences, and their presence was requested by the organisers and
typically welcomed by the general public. To counsel that including LFR to this setting
would have altered the behaviour of potential attendees is hardly credible. The
blanket declare that surveillance suppresses civil rights  and alters behaviours in all contexts shouldn’t be
supported by proof.

 

The
bottom-line

 

Facial recognition will inevitably turn out to be routine
policing software. Relatively than pushing unrealistic proposals of bans or
moratoriums, regulatory debate ought to correctly weigh the trade-offs between
human rights and public security in guaranteeing the proportionate use of the
expertise.   Questions on when LFR needs to be used and
thought of proportionate and different points equivalent to oversight needs to be debated
fastidiously. Nevertheless, the UK police’s use LFR, and the continuing debate highlights
that coverage and regulatory proposals may very well be based mostly on shaky interpretation of
knowledge and understanding of important authorized ideas.