Lode Lauwaert

The through-line

Lauwaert's consistent move is to insist that the AI debate is aimed at the wrong targets. Public and policy attention, as he sees it, settles on "small benefits" like faster emails and "small problems" like bias and privacy, while a geopolitical lens shows what is actually at stake: superintelligence, loss of control, engineered pandemics, climate [1][2]. His argument is not that enthusiasm is misplaced. He describes himself as "een super groot technooptimist" and says "ik denk dat we geweldig moeten inzetten op artificiële intelligentie... het is niet omdat je geweldige voordelen ziet dat je ook niet bepaalde grote risico's ziet" [2]. The optimism and the alarm are held together, and the work of the argument is done by conceptual precision: separating AGI from superintelligence, separating technical brilliance from calibrated judgement, separating what is genuinely unforeseeable from what merely feels frightening.

On why AGI is a matter of convention, not discovery

There is no crisp definition of AGI, and Lauwaert does not think one is waiting to be found. He compares it to the boundary between day and night: whether the threshold sits at 50, 100 or 200 cognitive tasks performed at average human level is something that will have to be agreed by convention rather than discovered [1][2]. The comparative claim he does make is about level rather than count: "AGI zit op een gemiddeld niveau, superintelligentie ver voorbij het gemiddelde niveau, voorbij het expertniveau zelfs" [1][2].

On why superintelligence, not AGI, is where control is lost

The two concepts are routinely run together, and separating them changes the risk picture. AGI by itself carries no implication of losing control; it can remain in the hands and under the control of a person, a company or a state [1][2]. Control loss is tied to superintelligence, and the route there runs through automating AI research itself: a system that improves itself around the clock, without sickness, wages or weekends, sets off an intelligence explosion [1][2]. Lauwaert stresses that this is a gradual process rather than a single moment [1].

The specific danger he names is unforeseeable sub-goals. A superintelligent system pursuing a final goal will invent paths to it that we cannot anticipate, and it can do so precisely because it thinks further ahead than we do; Bostrom's paperclip thought experiment is the illustration [1][2]. The risk is structural rather than a matter of malice, and it follows directly from the capability gap.

On not trusting AI scientists with probabilities

Lauwaert treats extinction probability estimates coming from AI researchers with open scepticism, on two grounds. The first is a competence mismatch: "Het is niet omdat ze zeer intelligent zijn of zeer goed zijn in het ontwikkelen van AI systemen, dat ze ook goed zijn in het toekennen van waarschijnlijkheidsgraden" [1]. The second is motivational: researchers carry a positivity bias about what their own systems will shortly be able to do [1][2]. Both cut against taking headline numbers from labs at face value, whether those numbers are reassuring or alarming.

On treaties and why halting a technology is not naive

The proposal for an international treaty on superintelligence is usually dismissed as unrealistic, and Lauwaert rejects the dismissal on historical grounds. Nuclear arms control is the precedent: "Landen zijn wel degelijk in staat om te stoppen met een heel gevaarlijke technologie wanneer ze zien dat de risico's gelijk verdeeld zijn" [1][2]. The condition that makes agreement possible is symmetry of exposure, states recognising that the danger falls on themselves as well as on their rivals, and the necessary precondition for reaching that recognition is political awareness [1][2]. This is also why the misallocation of debate matters practically. As long as the conversation stays on bias and privacy, the awareness that would make a treaty conceivable never forms. He also notes that existing AI, not only future systems, already lowers the barrier for obscure sects to engineer pandemics [2].

On what people will do with much less work

If work shrinks, the drives that work currently absorbs do not vanish with it. Lauwaert points to deeply evolved motivations, status, prestige, community, competition and exploration, and argues they will migrate rather than disappear, into sport, into social media influence, into exploration [1][2]. His evidence that this is already happening is comparative: measured against life 150 years ago, we are living a foretaste of it [1].

On the Romantic bias in judging AI art

A finding he returns to concerns how people evaluate art they do not know the origin of. Rated blind, AI-made work comes out as the most interesting; once viewers are told a machine made it, they suddenly find it less interesting [1][2]. Lauwaert reads the reversal as cultural rather than aesthetic, a distinctly Western inheritance from the Romantic era that art must issue from human genius [1][2].

Takeaways

• Treat AGI as a term settled by agreement rather than by fact: the number of cognitive tasks at average human level that counts as the threshold, 50, 100 or 200, is a convention, like the line between day and night [1][2].
• Do not equate AGI with loss of control. AGI can stay under the control of a person, company or state; it is superintelligence, "voorbij het expertniveau zelfs", that breaks control [1][2].
• The mechanism to watch is automated AI research, a system improving itself 24/7 without sickness, wages or weekends, producing a gradual intelligence explosion rather than a single dramatic moment [1][2].
• The core danger is sub-goals we cannot foresee, since a smarter system finds paths to its goal that we are by definition unable to anticipate [1][2].
• Discount extinction probabilities offered by AI scientists: skill at building systems does not confer skill at assigning probabilities, and builders carry a positivity bias about their own technology [1][2].
• An international treaty on superintelligence is realistic once states see the risk as shared, exactly as with nuclear weapons; the missing ingredient is political awareness, not feasibility [1][2].
• Shift the frame from small benefits and small problems, faster emails, bias, privacy, to the geopolitical stakes: control loss, engineered pandemics, climate [1][2].
• Techno-optimism and risk-awareness are compatible positions: "het is niet omdat je geweldige voordelen ziet dat je ook niet bepaalde grote risico's ziet" [2].