Nathalie Smuha

The through-line

The recurring preoccupation is that law and code are not the same thing, and that pretending otherwise lets governments hollow out the rule of law while appearing to follow it. When a public authority hands the execution of a statute to an algorithm, someone has to decide what the statute's words mean in machine-readable terms, and that decision is never neutral. "Die vertaalslag van een wet naar code, dat is geen puur technisch iets, dat is eigenlijk iets dat politieke en normatieve implicaties heeft" [1][2]. The danger is not lawlessness but its inverse: formally valid law used as an instrument against the people it should protect. "Het recht kan dus eigenlijk ook worden gebruikt om onze rechten en vrijheden uit te hollen en dat is wat rule by law is" [2].

Running alongside it is a habit of thinking one election ahead. Infrastructure built today outlives the government that built it, and the test of any camera network or decision algorithm is not what the current administration would do with it. "Denk niet aan de overheid die we vandaag hebben, denk aan volgende verkiezingen en uw nachtmerrie politieke partij wordt verkozen" [1][2]. Legality is no comfort here either, since "wetten kunnen ook veranderd worden: iets dat vandaag legaal is kan morgen illegaal worden gemaakt door een andere overheid" [1][2].

On translating law into code

The clearest illustration is Indiana, where coders operationalised "refusal to cooperate" with welfare authorities as simply not answering the phone [1][2]. That interpretation was arbitrary, and it was never what the democratic legislator intended, yet it became the operative rule for everyone the system touched [1][2]. This is where normative choices migrate out of parliament and into implementation, unnoticed, because they arrive dressed as technical detail.

The stakes are compounded by scale. An individual civil servant's mistake affects one file; an algorithmic system applies its interpretation, and its bugs, to an entire population at once, so a single error, deliberate or not, has enormous impact [1][2]. The UK Post Office case is the reference point: a simple bug in an accounting algorithm led to workers being wrongly prosecuted for theft, imprisoned, and in some cases driven to suicide before anyone identified the error [1]. The same logic explains why a nightmare government inheriting these systems is so dangerous, since it can decide for the whole population with one mouse click [2].

On why governments buy algorithms in the first place

The motives are less strategic than they look. There is fascination with technology, which produces solutions in search of problems, and there is pressure to cut costs [1][2]. What is largely absent is evidence that the systems are actually more efficient or more effective than what they replace [1][2]. That gap between the promise and the demonstrated result runs underneath every deployment discussed here, from tax fraud detection and benefits decisions to predictive policing and smart cameras [2].

On security as an open-ended justification

Anything can be argued for once safety is invoked, and that is precisely the reason to push back. "Er is eigenlijk geen limiet aan wat je kan doen in de naam van veiligheid en we moeten ons daar een beetje tegen verzetten" [1][2]. Function creep is the concrete form this takes. Cameras installed to protect Antwerp's Jewish neighbourhood after terrorist attacks were later used during COVID to check whether that same community was following lockdown rules, the identical infrastructure repurposed for goals citizens never consented to [1][2]. The COVID pass makes the parallel point about time rather than purpose: it was legalised through a law that allows the system to return in a future crisis, which is why surveillance infrastructure built in the name of safety needs strong sunset guarantees rather than dormant statutory hooks [1][2].

On predictive policing

Predictive policing is self-confirming by construction. Send officers where the algorithm points and crime will be spotted there, which registers as validation, while crime elsewhere goes undetected and unrecorded [1][2]. The loop tightens on the same neighbourhoods and the same people, who keep being flagged because they keep being watched [1][2]. The system's apparent accuracy is an artefact of where it directed attention.

On the AI Act

The Act is necessary and insufficient. Its core weakness is self-assessment: systems it classifies as high-risk, including welfare-decision algorithms and predictive policing, can be evaluated by the very organisation deploying them, with independent verification triggered only after complaints, potentially three years and a great deal of damage later [1][2]. The deeper mismatch is conceptual. The Act treats AI as a product to be made safe through technical standards, much like a washing machine, but risks to human rights, democracy and the rule of law do not translate into technical standards [1][2]. Its exhaustive lists also leave gaps, and the most important risks around government use are not fully covered [1].

The response to that should not be Macron's, which is to make the Act looser. It should be made smarter: simplified while offering better protection, with better coverage of the risks currently slipping through the lists and the standards [1][2]. Nor should transparency be mistaken for a remedy. Seeing an algorithm's code helps most citizens very little, errors can still creep in, and accountability for the final decision can still be entirely absent [1][2].

On what must never be delegated

Underneath the institutional argument sits a limit that is not procedural. Value judgments cannot and should not be handed to machines: "die waarde-oordelen is iets dat we niet mogen of kunnen delegeren aan AI-systemen... zelf uitmaken wat voor ons een waardevol leven is" [1][2]. This is the standard against which the Indiana coding choice, the AI Act's technical-standards approach and the transparency-as-fix argument all fall short, since each of them relocates a judgment about how people should live into a place where it cannot be contested.

On humanoid robot hype

The wave of humanoid robot announcements is best read as "communicative acceleration" rather than technical progress: once one large player announces, the others must announce too, mainly to keep investors pumping money in, and there is no evidence of genuine acceleration underneath [1][2]. General-purpose humanoids are likely inefficient in any case. Specialised machines that do one task well, of the robot vacuum kind, make more sense, given that something as basic as picking up an egg without crushing it remains extremely hard [1]. The humanoid form persists because of fascination with creating something human-like, not because anything requires it [1][2].

Takeaways

• Treat every translation of a statute into code as a political decision requiring democratic scrutiny, not an implementation detail; Indiana coded "refusal to cooperate" as not answering the phone, a meaning parliament never chose [1][2].
• Assess public-sector algorithms and camera networks by imagining them in the hands of the government after next, not the one that built them [1][2].
• Because algorithmic decisions apply at population scale, budget for catastrophic single-point failure: the Post Office accounting bug produced wrongful prosecutions, imprisonment and suicides before it was found [1].
• Demand evidence of efficiency and effectiveness before adoption; the actual drivers are technology fascination and cost-cutting pressure, and the evidence is thin [1][2].
• Build sunset guarantees into any surveillance system justified by safety, since infrastructure gets repurposed, as Antwerp's protective cameras were for lockdown enforcement, and emergency frameworks like the COVID pass are written so they can return [1][2].
• Do not accept self-assessment for high-risk systems, and do not accept transparency as a substitute for accountability over the final decision [1][2].
• Reform the AI Act by making it smarter rather than looser: simplified, with coverage of government-use risks that its lists and technical standards currently miss [1][2].
• Read humanoid robot announcements as competitive signalling to investors; the specialised single-task robot remains the more plausible engineering path [1][2].