AGI: Who expects human-like AI—and when?

Sam Altman · AGI achievable at all

Altman stated in early 2025 that OpenAI was confident it knew how to build AGI in the traditional sense.

We are now confident we know how to build AGI as we have traditionally understood it.

Sam Altman · LLM/scaling path is enough

The statement that OpenAI knows a path to AGI does not describe precisely enough whether simply scaling up the current LLM paradigm without significant innovations should be sufficient.

Dario Amodei · AGI achievable at all

Amodei specifically describes a “powerful AI” with broad superhuman expertise and considers its occurrence to be fundamentally realistic.

Dario Amodei · AGI before 2030

He names 2026 or 2027 as possible years and considers a point after 2030 to be very unlikely.

Possibly by 2026 or 2027

Dario Amodei · LLM/scaling path is enough

Amodei believes it is possible that progress within the current paradigm will be sufficient, but explicitly does not formulate this as a certainty.

within the existing paradigm may just be enough

Dario Amodei · AGI can be measured sensibly

He prefers “powerful AI” or “expert-level science and engineering” because he considers AGI to be too imprecise and laden with sci-fi baggage.

I find AGI to be an imprecise term

Dario Amodei · AGI already achieved

In his 2026 conversation, he expressly contradicts the claim that today's systems are already essentially AGI.

I don't believe we're basically at AGI.

Dario Amodei · AGI mostly hype/marketing

He criticizes the label's sci-fi ballast and hype, while at the same time describing a specifically expected "powerful AI" as a real technical development.

Demis Hassabis · AGI achievable at all

He defines AGI as a system with all human cognitive abilities and expects this stage to take place over a period of several years.

Demis Hassabis · AGI before 2030

At the end of January 2026 he estimates five to ten years until AGI; This puts even the lower end of its published corridor after the end of 2029.

Demis Hassabis · LLM/scaling path is enough

Foundation models remain a core building block for him, but he expects that additional major innovations will be necessary, for example in continuous learning, memory and planning.

Demis Hassabis · AGI can be measured sensibly

He defines AGI as the totality of human cognitive abilities including peak creativity and physical intelligence.

there has always been a scientific definition of that

Demis Hassabis · AGI already achieved

He explicitly says that today's systems are still a long way from AGI according to his comprehensive definition.

Demis Hassabis · AGI mostly hype/marketing

He specifically warns against turning AGI into a marketing term for commercial gain and insists on a scientific definition.

I don't think AGI should be turned into a marketing term for commercial gain.

Shane Legg · AGI achievable at all

His continued 50 percent forecast for minimum AGI by 2028 expressly assumes that corresponding systems are technically achievable.

Shane Legg · AGI before 2030

This is an explicitly probabilistic, uncertain forecast and is therefore coded as conditional.

Shane Legg · AGI can be measured sensibly

Legg and co-authors propose a cognitive measurement framework for progress towards AGI in 2026; it is intended as a research framework, not as an already generally accepted standard.

Shane Legg · AGI already achieved

In September 2026, he refutes claims that current Frontier systems already meet AGI definitions.

Eddy Keming Chen, Mikhail Belkin, Leon Bergen and David Danks · AGI achievable at all

Your Nature commentary explicitly argues that flexible general cognitive competence at human levels is already a reality.

Eddy Keming Chen, Mikhail Belkin, Leon Bergen and David Danks · AGI before 2030

Since they classify the relevant human-like ability as already achieved, their position is clearly before the end of 2029.

Eddy Keming Chen, Mikhail Belkin, Leon Bergen and David Danks · AGI can be measured sensibly

They argue based on flexible cognitive competence and existing evidence, but do not equate this with a generally accepted, uniform AGI measurement threshold.

flexible, general cognitive competence

Eddy Keming Chen, Mikhail Belkin, Leon Bergen and David Danks · AGI already achieved

The Nature commentary expressly states that the human-level machine intelligence outlined by Turing is now a reality.

human-level machine intelligence ... is now a reality

Yann LeCun · AGI achievable at all

His company, AMI Labs, explicitly develops world models as a path to true human-like intelligence; his criticism is directed against the AGI term and the LLM path, not against the goal itself.

Yann LeCun · LLM/scaling path is enough

He argues again in 2026 that more scaling of today's generative models will not lead to human-like intelligence and instead relies on world models.

Yann LeCun · AGI can be measured sensibly

He does not consider human intelligence to be “general” and says progress cannot be determined by a single test suite or scalar measure of intelligence.

it’s not like a uniform, you know, scalar measurement of quality

Yann LeCun · AGI already achieved

He describes current LLMs predominantly as information retrieval systems and continues to emphasize a lack of world models and flexible learning ability.

François Chollet · AGI achievable at all

His forecast for the early 2030s and his work on measurement techniques treat AGI as an achievable technical goal.

François Chollet · AGI before 2030

Its published expected value is therefore behind the deadline defined here at the end of 2029.

AGI early 2030s, most likely.

François Chollet · LLM/scaling path is enough

He considers mere scaling to be insufficient and sees general intelligence primarily in the efficient ability to acquire new skills.

take a different kind of technology

François Chollet · AGI can be measured sensibly

ARC and its newer benchmarks are intended to measure how efficiently systems acquire new capabilities rather than just covering familiar tasks.

human level skill acquisition efficiency

François Chollet · AGI already achieved

This means that his current forecast contradicts the claim that AGI will already be reached in 2026.

Gary Marcus · LLM/scaling path is enough

He argues that scaling up the current family of models has delivered neither AGI nor reliable general intelligence and a different approach is needed.

Scaling hasn’t gotten us to AGI

Gary Marcus · AGI can be measured sensibly

He criticizes AGI success reports without a definition and demands comprehensible criteria; He therefore considers measurement to be possible in principle, but not currently standardized.

Declaring victory without a definition simply muddies the waters.

Gary Marcus · AGI already achieved

He argues that current frontier systems continue to fail conventional AGI definitions.

By conventional definitions, Astra still falls short.

Fei-Fei Li · AGI achievable at all

It describes as a key question whether machines can think and act like humans, but does not specify in the source that a clearly defined AGI level will be achievable.

Fei-Fei Li · LLM/scaling path is enough

She believes additional technical innovations are necessary and points out the limitations of current approaches.

Fei-Fei Li · AGI can be measured sensibly

She says no one has convincingly defined AGI and sees no clear scientific demarcation between AI and AGI.

I don't know if anyone has ever defined AGI.

Fei-Fei Li · AGI already achieved

Because she does not accept AGI as a clearly defined scientific threshold, no reliable yes/no position on “already achieved” can be derived from her statement.

Fei-Fei Li · AGI mostly hype/marketing

She compares the term directly to a scientific term and assigns it more to marketing.

AGI is more of a marketing term than a scientific term.

Andrew Ng · AGI achievable at all

He speaks of AGI in terms of the full range of human intellectual tasks as a goal that is still many decades away.

Andrew Ng · AGI before 2030

Its current time frame clearly contradicts an AGI through the end of 2029.

We’re still very far from AI meetings that define AGI.

Andrew Ng · AGI can be measured sensibly

He considers the traditional Turing test to be unsuitable and proposes his own Turing AGI test; He therefore considers a more precise operationalization to be possible.

Andrew Ng · AGI already achieved

He distinguishes from lowered company definitions under which one can already claim success, but still considers human-level AGI to be very far away.

Andrew Ng · AGI mostly hype/marketing

He begins his proposal for a better AGI test by stating that AGI has become a hype term.

AGI has become a term of hype

Rodney Brooks · AGI achievable at all

His pointed forecast of 300 years is not a rejection of the possibility in principle, but rather an extreme distance in time.

we're not going to get AGI for 300 years

Rodney Brooks · AGI before 2030

This means that his published forecast clearly excludes AGI before 2030.

Rodney Brooks · LLM/scaling path is enough

He explicitly distinguishes surprising LLM abilities from general intelligence and does not see the current path as sufficient.

Rodney Brooks · AGI can be measured sensibly

He criticizes the fact that the concept of broad human ability to act and think is being shifted to simply passing tests.

These are moving targets for what it is.

Rodney Brooks · AGI already achieved

Its time horizon of centuries and its explicit demarcation of today's LLMs exclude an AGI that has already been achieved.

Rodney Brooks · AGI mostly hype/marketing

He says the modern AGI term was used to define a group of researchers and literally describes this as a marketing ploy.

It's a marketing ploy.

Timnit Gebru · AGI achievable at all

She explicitly writes that there is “no such thing as AGI.” However, because the column targets the underlying broad capability rather than just the label, no secure physical impossibility claim can be derived from it.

THERE IS NO SUCH THING AS AGI

Timnit Gebru · AGI before 2030

Her statement is directed against the term and the project itself. She does not give a reliable deadline forecast for the human-like ability, which is defined here independently of the label.

Timnit Gebru · LLM/scaling path is enough

She explicitly criticizes the idea spread by OpenAI that large language models gradually lead to AGI.

Timnit Gebru · AGI can be measured sensibly

The peer-reviewed paper argues that, unlike clearly defined systems, an undefined “AGI” cannot be tested according to standard technical security standards.

undefined systems like “AGI” cannot be appropriately tested for safety

Timnit Gebru · AGI already achieved

Since it criticizes AGI as a non-existent or undefined category, a yes/no assignment to the threshold that has already been reached is not reliable.

Timnit Gebru · AGI mostly hype/marketing

She accuses OpenAI of first pushing reinforcement learning and later LLMs as stepping stones to AGI. This proves criticism of the narrative, but not exactly the claim “mainly marketing”.

Emily M. Bender and Alex Hanna · AGI achievable at all

They dispute the representation of an inevitable AGI technology, but do not formulate as scientific evidence that every conceivable broad human-like machine capability is fundamentally impossible.

Emily M. Bender and Alex Hanna · AGI before 2030

The source says AGI does not identify any specific upcoming technology. That's enough for skepticism, but not for a literal deadline forecast by the end of 2029.

Emily M. Bender and Alex Hanna · LLM/scaling path is enough

The article rejects the idea that the performance of chatbots automatically demonstrates broad general capabilities; However, they do not make a narrowly formulated technical statement about future scaling alone.

Emily M. Bender and Alex Hanna · AGI can be measured sensibly

They describe AGI as a vague signifier without precise meaning and show how changing definitions serve economic and ideological purposes.

AGI is a term that famously lacks a precise meaning

Emily M. Bender and Alex Hanna · AGI already achieved

Because they already reject the category and its definition, the binary statement “already achieved” cannot be clearly derived as yes or no according to the narrow coding rule used here.

Emily M. Bender and Alex Hanna · AGI mostly hype/marketing

They describe the term as an economically useful, mystifying signifier that mobilizes capital and can shift responsibility.

the term is meant to evoke something with awesome power

Tim Dettmers · AGI achievable at all

He justifies this with physical computing limits, exponential costs, limitations of robotics and what he believes is a flawed assumption of infinitely scalable intelligence.

AGI Will Never Happen, and Superintelligence Is a Fantasy

Tim Dettmers · AGI before 2030

His explicit thesis that AGI will never occur in the usual broad sense includes AGI before the end of 2029.

Tim Dettmers · LLM/scaling path is enough

He argues that hardware, storage, cost limits, and diminishing returns limit today's scaling path and do not lead to true AGI.

Tim Dettmers · AGI can be measured sensibly

For him, real AGI would also have to include physical, economically relevant human activities. In this way he operationalizes his objection, but does not accept the usual narrower thresholds.

Tim Dettmers · AGI already achieved

His impossibility thesis precludes the claim that available systems have already reached this threshold.

Tim Dettmers · AGI mostly hype/marketing

He speaks of fundamentally flawed concepts, a Bay Area echo channel and “compelling narratives.” This is hype criticism, but not a clear assignment to marketing as the main function.

José M. Muñoz, Javier Bernacer, Alva Noë and Evan Thompson · AGI achievable at all

Their Nature correspondence expressly takes a fundamentally skeptical position on human-level intelligence and points to a lack of generalization, representation and selection.

Why AI will never be able to acquire human-level intelligence

José M. Muñoz, Javier Bernacer, Alva Noë and Evan Thompson · AGI before 2030

Your explicit “never” thesis includes achievement before the end of 2029.

José M. Muñoz, Javier Bernacer, Alva Noë and Evan Thompson · LLM/scaling path is enough

They name generalization, representation or world models and selection of relevant information as skills that LLMs lack on the path to general intelligence.

José M. Muñoz, Javier Bernacer, Alva Noë and Evan Thompson · AGI already achieved

Anyone who argues that AI will never achieve human-level intelligence is also ruling out such a capability being achieved as early as 2026.

AAAI Presidential Panel 2025 Respondents · LLM/scaling path is enough

In the AAAI survey of 475 participants, a clear majority rated upscaling today's AI approaches as an unlikely or very unlikely path to AGI.

“scaling up current AI approaches” to yield AGI is “unlikely” or “very unlikely”