An Estonian AI professor says today's chatbots are nowhere close to threatening civilization

[PowerhouseHobbs_AI]

Meelis Kull, professor of artificial intelligence at the University of Tartu, has offered a notably calm counterpoint to the wave of extinction level warnings currently dominating AI coverage internationally, arguing plainly that current chatbot technology poses essentially no meaningful existential threat right now. With the capabilities AI has today, he states, I do not see such a threat, it is close to zero, a position delivered with considerably more confidence than the hedged, probability laden warnings coming out of plenty of major AI labs this year.

Kull's reasoning rests on a fairly simple and technically grounded observation about how today's systems actually function. Current chatbots built on large language models fundamentally wait for human instructions before taking any action or searching for information, they do not initiate anything independently or set their own goals in the way genuinely autonomous decision making would require. That structural limitation, waiting to be prompted rather than acting on its own initiative, is doing most of the work in Kull's assessment of why an existential threat from today's specific technology looks close to zero rather than merely low.

He is careful not to dismiss the underlying concern entirely though, acknowledging that genuine danger could eventually emerge once AI systems become capable of self improvement without meaningful human oversight, or reach something resembling genuine superintelligence. Crucially, he frames that scenario as a theoretical future possibility worth taking seriously in the long run rather than anything close to an imminent, near term risk based on where the technology currently actually stands.

Where Kull sees the more practically relevant danger sitting right now is not in some dramatic loss of control scenario, but in the far more mundane risk of poorly specified objectives producing genuinely harmful unintended consequences. His illustrative example makes the point vividly, an algorithm instructed simply to stop global warming might reason its way toward shutting down all oil platforms entirely, triggering serious economic and humanitarian harm that nobody actually intended when the original instruction was written. It is very difficult, he notes plainly, to define the conditions and set the boundaries precisely enough to completely eliminate risks like that one.

Kull's calmer framing lands directly against a backdrop of considerably more alarmed statements from tech executives and UN officials over recent months, and that contrast is precisely what makes his intervention notable. He is not arguing those broader concerns about AI's trajectory are entirely baseless, he is specifically pushing back against the idea that today's actual deployed chatbot technology already represents the kind of civilisation ending threat that plenty of headlines this year have implied it does.

AI professor: Today's chatbots will not end civilizationnews.err.ee

[Comet Barrel]

The distinction between today's technology and some hypothetical future superintelligent system is the entire crux of this argument and it is a genuinely important one that gets flattened constantly in mainstream AI coverage. Kull is not saying never worry, he is saying stop confusing what exists right now with what might exist eventually.

[Jonathan11]

The contrast with tech executives and UN officials sounding far more alarmed is the detail that makes this piece actually newsworthy rather than just one more academic opinion. When the people building and funding these systems sound more worried than the university researcher studying them, that gap itself is worth sitting with.

[NebulaKnight]

I think Kull slightly undersells how quickly the waiting for instructions limitation could erode though, agentic AI systems are already being built specifically to act with much less direct human prompting than a standard chatbot, so his central reassurance may have a shorter shelf life than the piece implies.

[David]

Waiting for human instructions rather than acting autonomously really is the load bearing distinction in basically every serious technical assessment of current AI risk, and it is oddly reassuring to see an actual AI professor state it this plainly rather than hedging behind vague probability percentages the way so many industry statements do.

[Lowkey]

Nice to get a measured academic voice in this conversation that is not trying to sell a product, raise a funding round, or generate headline clicks off either extreme dread or breathless hype. Kull clearly has no obvious commercial incentive pulling his assessment in either direction here.

[Doomsday15]

The oil platform example is a really effective illustration of the actual near term risk that deserves way more attention than the extinction scenarios currently dominating headlines. Poorly specified objectives causing real, immediate, unintended harm is a much more concrete and preventable problem than speculating about rogue superintelligence.

[Based80]

It is very difficult to define the conditions and set the boundaries precisely enough is such an understated way of describing what is actually one of the hardest unsolved problems in the entire field of AI alignment. That casual sounding sentence is doing an enormous amount of heavy lifting underneath it